Social Security Programs and Retirement around the World: Micro-Estimation 9780226309989

Citation preview

Social Security Programs and Retirement around the World

A National Bureau of Economic Research Conference Report

Social Security Programs and Retirement around the World Micro-Estimation

Edited by

Jonathan Gruber and David A. Wise

The University of Chicago Press Chicago and London

J G is professor of economics at the Massachusetts Institute of Technology, director of the research program on children at the National Bureau of Economic Research, and research associate of the National Bureau of Economic Research. D A. W is the John F. Stambaugh Professor of Political Economy at the John F. Kennedy School of Government, Harvard University, and the director for Health and Retirement Programs at the National Bureau of Economic Research.

The University of Chicago Press, Chicago 60637 The University of Chicago Press, Ltd., London © 2004 by the National Bureau of Economic Research All rights reserved. Published 2004 Printed in the United States of America 13 12 11 10 09 08 07 06 05 04 1 2 3 4 5 ISBN: 0-226-31018-3 (cloth)

Library of Congress Cataloging-in-Publication Data Social security programs and retirement around the world : microestimation / edited by Jonathan Gruber and David A. Wise. p. cm. “A National Bureau of Economic Research conference report”— Half t.p. An analysis and country-by-country comparison of the effects of social security incentives on retirement behavior in Belgium, Canada, Denmark, France, Germany, Italy, Japan, The Netherlands, Spain, Sweden, the UK, and the United States. Includes bibliographical references (p. ) and index. ISBN 0-226-31018-3 (cloth : alk. paper) 1. Social security. 2. Retirement—Economic aspects. 3. Retirement income. 4. Aged—Employment. I. Gruber, Jonathan. II. Wise, David A. HD7091.S6244 2004 368.4′3—dc21

2003050721

o The paper used in this publication meets the minimum requirements of the American National Standard for Information Sciences—Permanence of Paper for Printed Library Materials, ANSI Z39.48-1992.

National Bureau of Economic Research Officers Michael H. Moskow, chairman Elizabeth E. Bailey, vice-chairman Martin Feldstein, president and chief executive officer Susan Colligan, vice president for administration and budget and corporate secretary

Robert Mednick, treasurer Kelly Horak, controller and assistant corporate secretary Gerardine Johnson, assistant corporate secretary

Directors at Large Peter C. Aldrich Elizabeth E. Bailey John H. Biggs Andrew Brimmer John S. Clarkeson Don R. Conlan George C. Eads Martin Feldstein

Jacob A. Frenkel Judith M. Gueron Robert S. Hamada George Hatsopoulos Karen N. Horn Judy C. Lewent John Lipsky Laurence H. Meyer

Michael H. Moskow Alicia H. Munnell Rudolph A. Oswald Robert T. Parry Richard N. Rosett Marina v. N. Whitman Martin B. Zimmerman

Directors by University Appointment George Akerlof, California, Berkeley Jagdish Bhagwati, Columbia William C. Brainard, Yale Michael J. Brennan, California, Los Angeles Glen G. Cain, Wisconsin Franklin Fisher, Massachusetts Institute of Technology Saul H. Hymans, Michigan

Marjorie B. McElroy, Duke Joel Mokyr, Northwestern Andrew Postlewaite, Pennsylvania Uwe E. Reinhardt, Princeton Nathan Rosenberg, Stanford Craig Swan, Minnesota David B. Yoffie, Harvard Arnold Zellner, Chicago

Directors by Appointment of Other Organizations Mark Drabenstott, American Agricultural Economics Association Gail D. Fosler, The Conference Board A. Ronald Gallant, American Statistical Association Richard C. Green, American Finance Association Robert Mednick, American Institute of Certified Public Accountants Angelo Melino, Canadian Economics Association

Richard D. Rippe, National Association for Business Economics John J. Siegfried, American Economic Association David A. Smith, American Federation of Labor and Congress of Industrial Organizations Josh S. Weston, Committee for Economic Development Gavin Wright, Economic History Association

Directors Emeriti Carl F. Christ Thomas D. Flynn Lawrence R. Klein

Franklin A. Lindsay Paul W. McCracken Peter G. Peterson

Bert Seidman Eli Shapiro

Relation of the Directors to the Work and Publications of the NBER 1. The object of the NBER is to ascertain and present to the economics profession, and to the public more generally, important economic facts and their interpretation in a scientific manner without policy recommendations. The Board of Directors is charged with the responsibility of ensuring that the work of the NBER is carried on in strict conformity with this object. 2. The President shall establish an internal review process to ensure that book manuscripts proposed for publication DO NOT contain policy recommendations. This shall apply both to the proceedings of conferences and to manuscripts by a single author or by one or more coauthors but shall not apply to authors of comments at NBER conferences who are not NBER affiliates. 3. No book manuscript reporting research shall be published by the NBER until the President has sent to each member of the Board a notice that a manuscript is recommended for publication and that in the President’s opinion it is suitable for publication in accordance with the above principles of the NBER. Such notification will include a table of contents and an abstract or summary of the manuscript’s content, a list of contributors if applicable, and a response form for use by Directors who desire a copy of the manuscript for review. Each manuscript shall contain a summary drawing attention to the nature and treatment of the problem studied and the main conclusions reached. 4. No volume shall be published until forty-five days have elapsed from the above notification of intention to publish it. During this period a copy shall be sent to any Director requesting it, and if any Director objects to publication on the grounds that the manuscript contains policy recommendations, the objection will be presented to the author(s) or editor(s). In case of dispute, all members of the Board shall be notified, and the President shall appoint an ad hoc committee of the Board to decide the matter; thirty days additional shall be granted for this purpose. 5. The President shall present annually to the Board a report describing the internal manuscript review process, any objections made by Directors before publication or by anyone after publication, any disputes about such matters, and how they were handled. 6. Publications of the NBER issued for informational purposes concerning the work of the Bureau, or issued to inform the public of the activities at the Bureau, including but not limited to the NBER Digest and Reporter, shall be consistent with the object stated in paragraph 1. They shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The Executive Committee of the Board is charged with the review of all such publications from time to time. 7. NBER working papers and manuscripts distributed on the Bureau’s web site are not deemed to be publications for the purpose of this resolution, but they shall be consistent with the object stated in paragraph 1. Working papers shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The NBER’s web site shall contain a similar disclaimer. The President shall establish an internal review process to ensure that the working papers and the web site do not contain policy recommendations, and shall report annually to the Board on this process and any concerns raised in connection with it. 8. Unless otherwise determined by the Board or exempted by the terms of paragraphs 6 and 7, a copy of this resolution shall be printed in each NBER publication as described in paragraph 2 above.

Contents

Preface

ix

Introduction and Summary Jonathan Gruber and David A. Wise

1

1.

Micro-Modeling of Retirement in Belgium Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

41

2.

Income Security Programs and Retirement in Canada Michael Baker, Jonathan Gruber, and Kevin Milligan

99

3.

The Impact of Incentives on Retirement in Denmark Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

4.

Estimating Models of Retirement Behavior on French Data Ronan Mahieu and Didier Blanchet

5.

6.

Micro-Modeling of Retirement Decisions in Germany Axel Börsch-Supan, Reinhold Schnabel, Simone Kohnz, and Giovanni Mastrobuoni Micro-Modeling of Retirement Behavior in Italy Agar Brugiavini and Franco Peracchi

153

235

285

345

vii

viii

7.

8.

9.

Contents

Social Security and Retirement in Japan: An Evaluation Using Micro-Data Takashi Oshio and Akiko Sato Oishi

399

Incentives and Exit Routes to Retirement in the Netherlands Klaas de Vos and Arie Kapteyn

461

Micro-Modeling of Retirement Behavior in Spain Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

499

10.

Income Security Programs and Retirement in Sweden 579 Mårten Palme and Ingemar Svensson

11.

Pension Incentives and the Pattern of Retirement in the United Kingdom Richard Blundell, Costas Meghir, and Sarah Smith

643

The Effect of Social Security on Retirement in the United States Courtney Coile and Jonathan Gruber

691

Contributors Author Index Subject Index

731 735 737

12.

Preface

This is the second volume presenting results of an ongoing project on social security and labor supply organized through the Program on the Economics of Aging at the National Bureau of Economic Research. Funding for the project was provided by the National Institute on Aging grant numbers P01-AG05842 and P30-AG12810 to the National Bureau of Economic Research. Funding for individual papers is noted in specific paper acknowledgments. Any opinions expressed in this volume are those of the respective authors and do not necessarily reflect the views of the National Bureau of Economic Research or the sponsoring organization.

ix

Introduction and Summary Jonathan Gruber and David A. Wise

If they aren’t paid, people don’t work. This fundamental economic principle is dramatically demonstrated by social security provisions and retirement. The social security programs in most developed countries are financed on a pay-as-you-go basis. Under this arrangement, most countries have accumulated large unfunded liabilities and, in many countries, face looming financial burdens. The aging of the populations in almost all countries is often cited as the reason for the financial burdens faced by the social security programs. Many of the programs are very generous and thus are increasingly costly as the population ages because there is a greater proportion of the population retired and collecting benefits, relative to the fraction of the population that is in the labor force and paying for the benefits. Perhaps just as important, although not as widely appreciated, is that the provisions of the programs themselves typically encourage retirement by reducing pay for work. This penalty on work induces older employees to leave the labor force early and thus magnifies the financial burden caused by population aging. This volume represents the second stage of a research project to study the relationship between social security provisions and retirement. The first stage of the project is reported in Gruber and Wise (1999). In that volume, we documented the enormous disincentives for continued work at older ages in many countries. The analysis also revealed a strong correJonathan Gruber is professor of economics at the Massachusetts Institute of Technology, director of the research program on children at the National Bureau of Economic Research, as well as research associate of the National Bureau of Economic Research. David A. Wise is the John F. Stambaugh Professor of Political Economy at the John F. Kennedy School of Government, Harvard University, and the director for Health and Retirement Programs at the National Bureau of Economic Research.

1

2

Jonathan Gruber and David A. Wise

spondence across countries between social security program incentives to retire early and the proportion of older persons that has left the labor force. The weight of the evidence suggested that this relationship was largely causal. The results of the second phase strongly affirm the causal relationship between retirement and social security incentives to quit work. In this second stage, we turn to country-by-country analysis of retirement behavior based on micro-data. The research teams in each of twelve countries compiled comprehensive, large databases of individuals. The data in each country match information on retirement decisions to the retirement incentives inherent in the social security provisions of each country. Retirement models were estimated based on the micro-data. The results show the enormous effect of social security incentives on retirement, and the uniformity of findings is striking. In every country, the quantitative magnitude of the incentive-program effects on retirement is very large. The key advantage of this micro-estimation approach is that in each country the effects on retirement of changes in social security provisions can be predicted. To demonstrate the effect of such changes, the country papers in this volume include simulations of the effects of two illustrative reforms: One illustrative reform delays the benefit eligibility ages in each country. A second illustrative reform assumes common provisions in each of the countries—reducing retirement incentives in some countries and increasing incentives in other countries. Under the first reform, the simulations show a large reduction in retirement in each country, and a corresponding increase in the labor force participation of older workers in each country. Under the second reform, the simulations show an increase in retirement in some countries and a decrease in other countries, in accordance with the relationship between the current country provisions and the common reform provisions. In short, the findings confirm the conclusions based on the first stage of this ongoing project and, in addition, illustrate the enormous magnitude of the effect that changes in social security provisions would have on retirement and thus the labor force participation of older people. Like the first stage, this second stage of the project relies on the analyses of a large group of economists who conduct the analysis for each of their countries. The authors of the individual country papers in this volume are Belgium: Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman; Canada: Michael Baker, Jonathan Gruber, and Kevin Milligan; Denmark: Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen; France: Didier Blanchet and Ronan Mahieu; Germany: Axel Börsch-Supan, Simone Kohnz, Giovanni Mastrobuoni, and Reinhold Schnabel; Italy: Agar Brugiavini and Franco Peracchi; Japan: Takashi Oshio and Akiko Oishi;

Introduction and Summary

3

The Netherlands: Arie Kapteyn and Klaas de Vos; Spain: Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi; Sweden: Mårten Palme and Ingemar Svensson; The United Kingdom: Richard Blundell, Costas Meghir, and Sara Smith; and The United States: Courtney Coile and Jonathan Gruber. The central feature of the project is the presentation of comparable analysis in each of the countries. Each of the country studies follows essentially the same format, although country-specific issues are often discussed as part of the analysis done for that particular country.

Background: The First Stage The goal of the first stage of the project was to describe the incentives inherent in the social security provisions in the project countries and to relate the incentives to the labor force participation of older workers. Each of the studies in the first volume begins with a description of the historical evolution of labor force participation and then presents data on the current age-specific activities and income sources of men and women. Each of the papers presents data for men and women in nine areas: 1. Labor force participation rates by age interval between 1960 and the present; 2. The proportion of employees covered by the public pension system and the proportion of persons over fifty-five receiving public pensions from 1960 to the present; 3. Replacement rates under the public pension system from 1960 to the present; 4. Current labor force participation rates by age; 5. Labor force status—employed, unemployed, disability, or retired; 6. Proportion receiving various public pensions—such as old age, disability, or survivor—by age; 7. Proportions receiving employer-provided pensions by age; 8. Source of household income by age; 9. Retirement and public pension hazard rates by age. Each paper then describes the institutional features of the country’s social security system, highlighting any interactions with other public and private programs that might also influence retirement behavior. The core of each paper is a detailed analysis of the retirement incentives inherent in the provisions of that country’s retirement income system. By making the same analytic calculations and by presenting the same simulations in each of the countries, the individual studies provide a means of comparing the retirement incentives among the countries. Because it provides the background and the motivation for the continu-

4

Jonathan Gruber and David A. Wise

ing project, we summarize here the key results from the first stage of the project. The decline in the labor force participation of older persons is perhaps the most dramatic feature of labor force change over the past several decades. The decline has been striking in all but one of the countries studied here. The labor force participation rates of men aged sixty to sixty-four for the years 1960 to 1996 are shown for each of the eleven countries in figure 1, which for ease of exposition is presented in two panels. (Denmark was added to the project after the first stage was completed.) The decline was substantial in each of the countries, but was much greater in some countries than in others. In the early 1960s, the participation rates were A

B

Fig. 1 LFP trends for men ages 60 to 64: A, Japan, Sweden, US, UK, Germany, Belgium; B, Japan, Spain, Netherlands, France, Italy

Introduction and Summary

5

above 70 percent in each of the countries and above 80 percent in several countries. By the mid-1990s, the rate had fallen to below 20 percent in Belgium, Italy, France, and the Netherlands. It had fallen to about 35 percent in Germany and 40 percent in Spain. Although U.S. analysts have often emphasized the “dramatic” fall in that country, the U.S. decline from 82 percent to 53 percent was modest in comparison to the much more precipitous decline in these European countries. The decline to 57 percent in Sweden was also large, but modest when compared to the fall in other countries. Japan stands out with the smallest decline of all the countries, from about 83 percent to 75 percent. Each of the country papers presents completely parallel labor force and other data for men and women, including current labor force participation and labor force departure rates by age, which are key components of the analysis in this volume. By considering the labor force participation rates by age in each country it is possible to calculate the proportion of persons in a given age interval who were out of the labor force. These unused-productive-capacity measures for all of the countries are shown in figure 2, for men in the age range from fifty-five to sixty-five. For the entire age range from fifty-five to sixtyfive, unused capacity ranges from 67 percent in Belgium to 22 percent in Japan. The goal was then to consider how this measure of labor force participation was related to the provisions of the social security programs in the countries. The incentive measure calculated in this first volume was the implicit social security tax on work. To understand that measure, it is useful to think of wage compensation for working an additional year in two components.

Fig. 2

Unused productive capacity

6

Fig. 3

Jonathan Gruber and David A. Wise

Sum of tax rates on work from early retirement age to 69

The first is wage earnings. The second component is the increase in the expected present discounted value of promised future social security benefits, known as the accrual in social security wealth (SSW), which is equal to SSWt
1 – SSWt . It is natural to think of this difference as positive, or at least not negative—that is, if a person works for an additional year and thus forgoes one year of benefits, it might be expected that benefits begun one year later would be increased enough to offset the fact that they are received for one fewer years. This is true, for example, for the typical worker in the United States: If a worker forgoes claiming benefits at the earliest possible age (sixty-two) and works another year, subsequent benefits are increased by 6.67 percent to account for the fact that benefits will be received for one fewer year. In most of the countries studied in this project, however, the accrual is significantly negative. This is a consequence in large part of not increasing benefits enough if the age of benefit receipt is delayed so that benefits are not “actuarially fair.” Thus, what the worker gains in wage earnings is partially, or even largely, offset by a loss in future social security benefits. We call the ratio of this loss to wage earnings (after tax) the social security implicit tax on earnings. In many countries this tax can be 80 percent or more at certain ages. Suppose that the tax rates for each of the years from the early retirement to age sixty-nine are summed: We call this the tax force to retire. It is shown for all the countries in figure 3. The relationship between the (logarithm of the) tax force to retire and the proportion of men age fifty-five to sixty-five that is out of the labor force (unused capacity) is shown in figure 4. There is a striking correspon-

Introduction and Summary

Fig. 4

7

Unused capacity versus tax force

dence between the two series, showing a clear relationship between the social security tax on work and departure from the labor force. A number of examples in the previous volume, some of which are replicated below, suggest that the relationship is largely causal. We concluded in the introduction to that volume that: It is clear that there is a strong correspondence between the age at which benefits are available and departure from the labor force. Social security programs often provide generous retirement benefits at young ages. In addition, the provisions of these programs often imply large financial penalties on labor earnings beyond the social security early retirement age. We conclude that social security program provisions have indeed contributed to the decline in the labor force participation of older persons, substantially reducing the potential productive capacity of the labor force. It seems evident that if the trend to early retirement is to be reversed, as will almost surely be dictated by demographic trends, changing the provisions of social security programs that induce early retirement will play a key role. (Gruber and Wise 1999, 35)

The Second Stage Analysis and Issues The first stage of this project established two key findings: (a) The social security systems in many countries provide enormous incentives to leave the labor force at older ages, and (b) A strong correspondence exists between social security incentives to retire and the withdrawal of older work-

8

Jonathan Gruber and David A. Wise

ers from the labor force. This implies that social security incentives to retire are likely an important cause of the low labor force participation of older workers in many countries. The relationships in the first volume, however, do not provide a means of estimating the magnitude of the effect on labor force participation of changes in plan provisions. Thus, in this second stage of the project, we undertake analysis to estimate how much the retirement age would change if social security provisions were changed, based on within-country analysis of the determinants of retirement. The analysis is based on the micro-data for each country that considers the relationship between retirement and the incentives faced by individual employees. That is, rather than considering system-wide incentives for representative persons (such as those with median earning histories) and comparing these incentives with aggregate labor force participation across countries, we now turn to micro-econometric analyses within countries. The results of these analyses are based on differences in individual circumstances within a given country. This approach has two key advantages. First, the analyses in this volume show that social security retirement incentives have very similar effects on labor force participation in all countries. In particular, the results strongly confirm that the relationship between labor force participation and retirement across countries is not the result of cultural differences among countries that could yield different norms for work at older ages. That is, the within-country analyses show similar responses to retirement incentive effects, even though the countries differ with respect to cultural histories and institutions. Second, the analysis of micro-data also allows consideration of several features of social security systems, as well as individual attributes, that may simultaneously affect retirement decisions. In particular, we can consider jointly the age at which benefits are first available and the incentive to retire once benefits are available. The first stage of the project showed that both of these features were important determinants of retirement. The importance of benefit eligibility ages presents a particular challenge for the analyses in this volume. We believe that much of the effect of social security provisions is likely to be through the choice of benefit eligibility ages, which in some instances may tend to establish social norms for retirement. For example, a common finding of many of the analyses in this volume is that even very detailed models of retirement incentives cannot explain the large jumps in retirement rates at normal and early entitlement ages. As a result, the retirement effects of major system reforms (like raising the early entitlement age) may be greatly understated by comparisons among individuals within a given retirement income system who all face the same eligibility ages. In addition, small private saving may limit the option of most persons to retire before pension benefits are available (many people are liquidity constrained), and such effects may not be cap-

Introduction and Summary

9

tured by retirement models. We discuss later how we deal with this critical issue. In addition, as emphasized in the first volume, unemployment and disability benefits often provide early retirement income before the nominal social security normal retirement age. Thus in many countries, to estimate the effect of the plan provisions on retirement, it is necessary to consider all three programs jointly, which the micro-econometric analyses in this volume accomplish. Considered jointly, we believe that the analyses in this volume provide overwhelming confirmation that the provisions of social security programs play a key role in the determination of retirement decisions. This result complements the conclusion from the first stage of the analysis. In addition, the estimates in this volume allow prediction of the effects on retirement of changes in program provisions as well as the effect of changes on program costs, which will be taken up in the next stage of the project. As in the first volume, the analysis in each country follows a template so that results can be compared across countries. The micro-analysis in each country is based on a sample of individuals. In some cases, the data come largely from administrative records. In other cases, the data were obtained from special surveys. The coverage is not precisely the same in each country. For example, the data for Italy pertain to private sector workers only, excluding public sector employees. Nonetheless, it has been possible to estimate the same models in each country, even though the population covered by the country data sets may differ in some respects. In this section, we first describe the incentive measures used in the analysis, as well as other features of the model specifications. The alternative incentive measures are constructed economic variables that describe the financial gain or loss from continuing to work. Then we discuss the method used to obtain estimates when there are multiple routes to retirement. Next we briefly summarize the parameter estimates obtained in the analyses across the twelve countries. In the next section, we discuss a key empirical regularity that strongly influences the analyses undertaken in this volume: the correspondence between benefit eligibility ages and retirement decisions. In the concluding section we describe the simulations undertaken to summarize the implications of the estimated models, and discuss the simulation results. The simulations describe the effect of illustrative policy changes. The goal is to provide an understanding of the nature of the findings, focusing on selected portions of the analyses described in detail in the country papers. The simulations demonstrate the implications of the retirement model estimates. As part of this discussion, we take some care to explain the different methods of simulation that are used in the analysis and why certain features of the simulation methods are of key importance. A central aspect of

10

Jonathan Gruber and David A. Wise

the analysis is experimentation with various approaches to estimation and to simulations based on the estimates. The aim is to determine the most reliable methods to use in the subsequent stages of the project. Some of the discussion is necessarily somewhat technical. We believe, however, that it is important to present an overview of the critical features of the analysis so that readers can approach the individual country papers with a broad understanding of the issues and rationale behind the approach taken in the country analyses. Thus, in the text of this introduction, we have explained the main features of the analysis and have included some additional, more technical detail in an appendix. The Estimation Models: Incentive Measures and Control Variables The goal of the analysis in the country papers is to estimate the probability of retirement based on the provisions of the country’s social security system—which provides differential incentives depending on individual employee circumstances—and on other individual attributes of employees. The focus of the analysis is the plan provisions. In particular, the way the incentives to retire, inherent in plan provisions, are in fact related to the retirement choices that individuals make. There are several ways that the incentives might be measured. All of the measures describe the financial gain or loss from continuing work. The specifications used in the country papers are summarized in the following table. The most important variable in each specification is the incentive measure, which is noted across the top of the table. In addition to these measures, the models control for various other variables (covariates). The specifications are summarized in table 1. Each of the specifications includes SSW. The expectation is that, all else equal, persons with greater SSW are more likely to retire. In principle, total wealth should be controlled for, but in most countries the data do not provide measures of other forms of wealth. The focus of the analysis is on forward-looking measures of the incentive for retirement, or for continued work. A natural starting point is a Table 1

Estimation Method (Incentive Measure) and Variables Estimation Method (Incentive Measure)

Variables SSW Linear age Individual age indicators Earnings Sector Demographics

Single Year Accrual X X X X X

X X X X X

Peak Value X X X X X

X X X X X

Option Value X X X X X

X X X X X

Introduction and Summary

11

measure that looks ahead only one year, the single-year accrual measure. This measure captures the effect of another year of work on future benefits. Thus, as a basis for comparison, the country analyses present the single-year accrual incentive measure. However, it has been shown in other contexts,1 as well as in the first volume of this project, that the financial gain from continuing to work may vary from year to year. That is, the gain from working one more year may be large, for example, but once that single additional year is worked, the gain from working one more year may be small or even negative. Likewise, the gain from one more year may be small, but might then be followed by a year of large gain. In this case, a person who decides to retire based on looking forward just one year, would forego the gain in pension wealth that would be gained by continuing to work for two years. Thus a key principle of the approach followed here is that the estimation should account for the pension accrual not just in the next year, but many years into the future. The benchmark approach for considering the entire future path of accruals is the option value model.2 To summarize, this model evaluates the expected present discounted value of incomes for all possible future retirement ages and then measures the value of retirement today versus the value at the optimal date (perhaps today, but more likely in the future). If looking ahead suggests gains from work at some time in the future, there is an incentive for the person to remain in the labor force to take advantage of these gains. The relationship between the measures can be explained briefly. As previously described, the social security accrual from one year to the next is given by: SSWt
1  SSWt

(1)

That is, this measure describes the change in promised future social security benefits from working one additional year. A simplified version of the option value measure at age t can be described by Simplified OVt (r∗) (2)





 


discounted  future wages
through age r∗




discounted discounted benefits if  benefits if retire at r∗ retire at t



discounted  future wages
[Peak Value]. through age r∗ 1. For example, see Lazear (1983), Kotlikoff and Wise (1985, 1989), and Coile and Gruber (2000a). 2. See Stock and Wise (1990a,b).

12

Jonathan Gruber and David A. Wise

In this formulation, a person considering whether to retire at age t considers the present value of benefits if he retires now (at age t) with the benefits if he retires at some later time. If the person retires at some later age he will gain from future wage earnings and from any gain in future pension benefits. The gain in wage earnings is represented by the first bracket and the gain in pension benefits by the difference between the terms in the second bracket. The age at which the total of the two components is the greatest is denoted by r∗. The option value prescription is that the person will continue to work if this option value is positive. Notice that the option value approach as set out above combines both of the components of compensation from working: one component is wage earnings, the other is the change in promised future social security benefits. We label this second component the peak value.3 It includes only social security benefits and not wage earnings. The peak value occurs at the age that gives the greatest discounted value of social security benefit. That age need not be at r∗, although for simplicity the two are assumed to be the same in this description. A more precise discussion of the differences in the measures is presented in the appendix. As previously emphasized, a crucial issue in the analyses in this volume is identification—that is, determination of the separate effect of each variable on retirement, as distinct from each of the other variables. Determining the effect of plan incentives on retirement is a key goal, but other individual attributes also influence the decision to retire. For example, persons are more likely to prefer retirement to work as they age. A linear age variable will potentially capture this effect, but only if preferences for leisure evolve linearly with age. Individual wage differences may also proxy for differences in the preference for work versus retirement. A wage-earning covariate may help to control for this form of heterogeneity among individuals. But both age and wage earnings also determine in part the value of the incentive measures. Thus, including age and wage covariates may make it more difficult to accurately determine the effect of the incentive measures or to accurately isolate the program incentive effects from the effect of worker heterogeneity. To put it another way, the importance of controlling for differences in taste regarding work may suggest the inclusion of the wage and age variables, separate from their incorporation in the option value. But there is a countervailing consideration: Much of the estimated effect of these variables is likely to reflect the influence of financial incentives and not individual heterogeneity. Thus, the full effect of the program incentives may be understated when the separate age and earnings controls are included. The issue of identification also arises in considering the option value as compared with the peak value incentive measures. If individual heterogeneity were not a concern, the option value measure would be the most 3. As proposed by Coile and Gruber (2000a,b).

Introduction and Summary

13

parsimonious incentive measure to use, as it captures the full financial incentive on retirement of both future wage earnings and retirement benefits combined. But, to the extent that wages proxy for the taste for work, the option value variation across individuals may reflect in part this wage proxy for heterogeneity, rather than the financial retirement incentive. The peak value measure recognizes this possibility by measuring the retirement incentive by the future stream of retirement benefits only, without including the future stream of wage earnings. But to the extent that future wage earnings have an important incentive influence on retirement, the peak value approach understates the full effect of financial incentives on retirement. Perhaps the most important identification issue arises when age-specific variables are included to estimate the retirement effect of program eligibility ages. It is common to find that the retirement rate at certain ages is larger (or smaller) than would be predicted on the basis of an incentive measure alone. For example, in the United States, the retirement rate at age sixty-five is noticeably higher than is predicted based on financial incentives alone. Perhaps this is due to a customary retirement age effect: Since age sixty-five is the normal retirement age, many employees may think that age sixty-five is the age to retire. In addition, in virtually no instance in any country do employees typically retire before some form of retirement benefits are available. The retirement rate at the early retirement age—the age at which a person is first eligible for benefits—is typically substantially greater than would be predicted on the basis of financial measures alone. This empirical regularity likely reflects a liquidity constraint; most employees have not saved enough to retire without receiving public social security or employer-provided pension benefits. To capture this effect, some of the specifications allow an indicator variable for each age. These indicator variables allow retirement to jump or decline at each age, after controlling for the financial incentive measures. But the inclusion of these age indicator variables in particular raises the identification question: Here, the issue is whether the effect of the financial incentive measures can be distinguished from the effect on retirement of plan eligibility ages. This is a critical issue and is discussed in more detail later. There is no right answer to these identification dilemmas. The country analyses in this volume follow what is perhaps a conservative as well as a flexible approach. All of the specifications control for background variables, including sex, education, industry of employment, and both current and average lifetime earnings. This is the conservative part. The flexibility is reflected in the different incentive measure specifications, each estimated using linear age and then again using indicator variables for each age. In this way the sensitivity of the findings to different incentive measures and to the controls that are included in the analysis can be assessed. One additional note on estimation: The option value model as set out by Stock and Wise, and in several subsequent applications, was estimated by

14

Jonathan Gruber and David A. Wise

maximum likelihood methods to obtain the relevant behavioral parameters. In this cross-country context, however, we concluded that that approach would likely pose numerical complexities that would best be avoided for this comparative analysis over a large number of countries. Thus the estimation undertaken in the papers in this volume is typically based on a regression counterpart to the option value model as well as two other approaches, as previously explained. In some countries, however, the option value parameters in the Stock and Wise specification have been estimated by a grid search, and in other instances the Stock and Wise option value parameter estimates have been used to calculate the option value that is used in the regression. In at least one country, the option value model was estimated by maximum likelihood. Multiple Retirement Options In some countries, like the United States, social security is the single public program that provides retirement benefits to the vast majority of retirees. The only retirement decision is then at what age to choose to retire under this program. (A small fraction of persons retire under the public disability program and many employees are covered by employer-provided pension plans.) In other countries, however, there are two or more programs under which a person can retire. Germany is a good example. Figure 5 shows paths to retirement for men in Germany between 1960 and 1995. The figure shows clearly the changes in the pathways to retirement after the 1972 reform, which is discussed later. Here, we draw attention to the multiple paths to retirement. All persons are eligible to retire at age sixty-five, the social security program normal retirement age, but only a small proportion of employees work until that age. A large fraction of

Fig. 5

Germany: Pathways to retirement for men, 1960 to 1995

Source: Data provided by Axel Börsch-Supan.

Introduction and Summary

15

employees retires under the disability program before age sixty. Others retire under the social security disability program available after age sixty. Some can retire under the social security system unemployment program after age sixty. Still others are eligible to retire at age sixty-three under the flexible retirement program that allows persons with long service to retire at that age—essentially the early retirement age. In addition to these programs, liberal interpretation of unemployment plan provisions allows persons to retire with unemployment benefits before age sixty. Whether a given person is eligible for a program depends on specific plan provisions, like eligibility for flexible retirement at age sixty-three. Eligibility for other programs, like disability or unemployment, is uncertain. In Germany, the approach is to assign (or predict) eligibility probabilities for each of the programs at each age, depending on the empirical probabilities of retirement under each of the programs at that age. Then the incentive measures are weighted averages, with the weights given by the probabilities. This instrumental variable method is described in more detail in the appendix, using the situation in Denmark as an example. Parameter Estimates We do not attempt in this introduction to provide a detailed discussion of the estimates. Rather, we rely on the simulations based on the parameter estimates to indicate the implications of the estimated models. The simulations are later discussed. Here we highlight the strikingly common finding in virtually all the country papers: The retirement incentives inherent in most social security programs are strongly related to early retirement. The estimation results are summarized in table 2. For each incentive specification, the table shows the sign and the statistical significance (at the 5 percent level) of the estimated effect of the incentive measure. The table also shows the sign and the statistical significance of SSW. For each incentive measure, the sign and significance level are shown when linear age is used and when the age-specific indicator variables are used. The results in table 2 are striking. In ten of the twelve countries, almost all of the estimated incentive measure effects are negatively related to retirement and significantly different from zero. (With respect to the following discussion, it is also notable that the sign and significance of the incentive measures rarely depends on whether age indicator variables are used in the specification.) In two countries—Italy and Spain—the peak value and option value effects are typically not significant and sometimes of the wrong sign.4 Also in these two countries, the single year accrual effect is negative and significantly related to retirement in four of the six cases. 4. In the United Kingdom, the option value incentive measures are significant when a “bootstrap” method, which accounts for repeated observations on the same person, is used to calculate standard errors. Also in the United Kingdom, both the peak value and the option value incentive measures are very significant—under conventional standard error estimates—when cohort indicator, instead of age indicator, variables are used.

16

Jonathan Gruber and David A. Wise

Table 2

Summary of Parameter Estimates, by Country and Specification Estimation Method (Incentive Measure) Single Year Accrual

Country Belgium Canada Denmark France Germany Italy Japan The Netherlands Spain Sweden United Kingdom United States

Peak Value

Option Value

Linear Age

Age Indicators

Linear Age

Age Indicators

Linear Age

Age Indicators

ACC:–* SSW:–* ACC:–* SSW:+* ACC:+* SSW:+* ACC:–* SSW:–* ACC:–* SSW:– ACC:–* SSW:–* ACC:–* SSW:+ ACC:+* SSW:+* ACC:–* SSW:+* ACC:–* SSW:+* ACC:–* SSW:+* ACC:+* SSW:+*

ACC:–* SSW:–* ACC:–* SSW:+* ACC:–* SSW:+ ACC:–* SSW:–* ACC:–* SSW:– ACC:–* SSW:– ACC:–* SSW:– ACC:+* SSW:+ ACC:+ SSW:+* ACC:– SSW:+* ACC:– SSW:+* ACC:+ SSW:+

PV:–* SSW:–* PV:–* SSW:+* PV:–* SSW:+* PV:–* SSW:–* PV:–* SSW:– PV:+ SSW:+ PV:–* SSW:– PV:–* SSW:+* PV:–* SSW:– PV:–* SSW:+* PV:– SSW:+* PV:–* SSW:+*

PV:–* SSW:–* PV:–* SSW:+* PV:–* SSW:+* PV:–* SSW:– PV:–* SSW:– PV:– SSW:– PV:–* SSW:– PV:–* SSW:+* PV:+ SSW:+ PV:–* SSW:+* PV:– SSW:+* PV:–* SSW:+

OV:–* SSW:– OV:–* SSW:+* OV:–* SSW:+* OV:–* SSW:–* OV:–* SSW:– OV:+ SSW:+ OV:–* SSW:+ OV:–* SSW:+* OV:– SSW:+ OV:–* SSW:+* OV:–* SSW:+* OV:–* SSW:+

OV:–* SSW:– OV:–* SSW:+* OV:–* SSW:+* OV:–* SSW:–* OV:–* SSW:–* OV:– SSW:+ OV:+ SSW:– OV:–* SSW:+* OV:+ SSW:+ OV:–* SSW:+ OV:–* SSW:+* OV:–* SSW:–*

Notes: See text for explanation of abbreviations. Regarding the United Kingdom, the option valued estimates are significant when standard errors accounting for repeated observations for the same person are used and when cohort indicator variables, instead of age indicators, are used. Both the option value and the peak value incentive measures are very significant. Spain: The indications in this table pertain to the Regimen Especial Trabajores Autonomos (RETA) employee group. *Statistically significant at the 5 percent level.

The estimated effect of SSW, however, is often not statistically different from zero and in many cases is of the wrong sign. In many countries, it is likely easier to identify the effect of the incentive measures than the effect of wealth levels. Because of program provisions, there is much more variation in the incentive measures than in SSW. Thus, overall, the results from these twelve separate analyses seem amazingly consistent to us. The incentives inherent in retirement income programs are clear determinants of individual retirement behavior. The estimates themselves strongly suggest a causal interpretation of the crosscountry results presented in our first volume. The results point to an im-

Introduction and Summary

17

portant relationship between incentive effects and labor force participation, independent of cultural differences among countries. The magnitude of the implied effects are also vary comparably across countries, as shown by the simulations discussed later. Eligibility Ages and Retirement: A Key Empirical Regularity The effect on retirement of the changes in benefit eligibility ages perhaps presents the most difficult prediction challenge of the project. Thus, we give special attention to a consistent empirical regularity in retirement that highlights this challenge. In each country, retirement rates are strongly related to particular eligibility ages prescribed in country-specific plan provisions. Perhaps most importantly, retirement rates increase sharply at ages of first eligibility for benefits. The age of first eligibility may differ from person to person and varies by program (e.g., social security, disability, or unemployment) in many countries. In the absence of eligibility for benefits, retirement is rare in each of the countries. One way to see this relationship is to consider retirement hazard rates. The hazard rate shows the proportion of persons employed at a given age who retire over the subsequent year. The empirical regularity between hazard rates and eligibility ages across countries is shown in some detail in the first volume of the project. Here, we show additional country examples that help to motivate, in particular, the simulation and estimation methods used in this volume. United States Labor force departure rates for men in the United States are shown in figure 6. The hazard rates are close to zero before age fifty-four and then in-

Fig. 6

Hazard rates for men in the United States

18

Jonathan Gruber and David A. Wise

Fig. 7 U.S. retirement hazards for men, social security only and employer defined benefit

crease gradually through age sixty-one. All employees are eligible for social security early retirement benefits at age sixty-two, and there is a sharp jump in the hazard rate at that age. The rise in departure rates at age fifty-four can be attributed not to the social security program, but to eligibility for early retirement under employer-provided pension plans. The importance of eligibility can be seen clearly in figure 7. This figure shows hazard rates by pension plan coverage. The hazard rates for men who have no private pension coverage and are covered only by the social security program are indicated by the line with round markers. The hazard rates for these persons are very close to zero until the social security early retirement age, when they are first eligible for retirement benefits. At that age, there is a sharp increase in the departure rate. The important feature of the pattern is that there is essentially no retirement before that age.5 The other line in the figure shows the hazard rates for men who are covered by an employer-provided defined benefit pension plan. The early retirement age under these plans is often age fifty-five and is rarely over age sixty. The hazard rates are very low before age fifty-five. But for these employees, there is a sharp jump in the hazard rate at age fifty-five, when many in this group are first eligible for benefits. And then another jump at sixtytwo, when social security benefits are first available. If the early retirement age for the U.S. social security program were raised from sixty-two to sixtyfive, for example, these data strongly suggest that the jump in the hazard rate at age sixty-two would no longer occur at that age, but would shift to age sixty-five instead. The critical question is whether the hazard rates would remain close to zero until age sixty-two. 5. These rates are based on labor force participation rates of Health and Retirement Study respondents. The precise age of departure from the labor force is obscured by the two-year interval between survey waves, and thus the jump in the hazard rate does not match the early retirement age of sixty-two exactly.

Introduction and Summary

Fig. 8

19

Hazard rates for men in Germany

Germany The German social security system provides additional and perhaps better examples. Figure 8 is analogous to figure 6 for the United States and shows hazard rates, with respect to departure from the labor force, for men in Germany. The detailed provisions of the 1972 social security legislation (effective until 1998) are mirrored in the retirement rates by age. In particular, there is a jump in the hazard rate at each important eligibility age. The ages of key plan provisions are also noted on the figure so that the correspondence between provisions and retirement is easily seen. Men who are disabled or unemployed at age sixty and have a certain number of years of employment under the social security system are eligible for early retirement at that age. There is a corresponding large jump in the retirement at that age. Men who have been employed for thirty-five years are eligible for early retirement at age sixty-three, and there is a corresponding jump in the retirement rate at that age. The normal retirement age is sixty-five and all men are eligible for benefits at that age. Again, there is a corresponding spike at that age as well. (By age sixty-five, however, fewer than 29 percent of men are still in the labor force in Germany.) In addition, even before age sixty, liberal interpretation of disability and unemployment plan provisions effectively serves to provide early retirement benefits, so that many men are eligible for effective early retirement before age sixty. There is a corresponding jump in the hazard rates between fifty-five and fifty-six. Like the U.S. data, the German data also suggest that if the eligibility

20

Fig. 9

Jonathan Gruber and David A. Wise

Germany: Retirement ages, ages before and after 1973 reform

Source: Data provided by Axel Börsch-Supan.

ages were increased, for example, the observed jumps in the hazard rates would also shift upward. This sort of pattern is exhibited in all countries and is shown in detail in the first volume of the project. The German system provides additional examples that reinforce the importance of plan-specific eligibility ages. First, the provisions of the German social security program were changed in 1972, as mentioned above. Second, the provisions for men and women are different. The difference in the pre- and post-reform hazard rates, as well as the difference in the hazard rates for men and women, highlight the point. Figure 9 shows the distribution of retirement ages for men under the social security system provisions in 1970 and in 1976. Before the 1972 reform, retirement under the social security program was essentially only possible at the normal retirement age of sixty-five, and there is a correspondingly large spike in the distribution at that age. The 1972 reform provided for early (flexible) retirement at age sixty-three, and in 1976 there is a large concentration of retirement at that age. Notice that retirement at ages before sixty and after sixty-five was essentially unaffected by the change in plan provisions. The change in hazard rates was essentially confined to the ages affected by the legislation. (These data do not cover retirement under the disability and

Introduction and Summary

Fig. 10

21

Germany: Retirement ages 1973, 1976, and 1980

Source: Data provided by Axel Börsch-Supan.

unemployment programs before age sixty. Retirement under these programs, however, is evident in figure 8 and in figure 5.) Figure 10 is like figure 9, but adds the distribution of retirement ages in 1980. After the 1972 reform, men quickly took up retirement at the early retirement age of sixty-three—as seen in 1976. Over the next few years, retirement through lenient disability and unemployment rules was increasingly taken advantage of, and by 1980 a large retirement concentration at age sixty (through these programs) is evident. Apparently, the ease of retirement at the earlier age was not at first recognized. In addition, the eligibility ages for men and women differ. The effect of the differences can be seen in figure 11. This figure shows the distribution of retirement ages, under the social security system, for men and women in 1995. For both men and women there is a concentration of retirement at age sixty-five, the normal retirement age. For both men and women, there is also a concentration at age sixty, but for different reasons. Age sixty is the early retirement age for women, but for men, the concentration at age sixty is due to eligibility for disability and unemployment benefits at that age. Women are also eligible for these benefits at age sixty, if they have enough years of employment. The key feature of this figure is the retirement behavior at age sixty-three. For men there is a concentration at age sixtythree, the early (flexible) retirement age. But this option is not available for

22

Fig. 11

Jonathan Gruber and David A. Wise

Germany: Retirement ages for men and women in 1995

Source: Data provided by Axel Börsch-Supan.

women, and there is no concentration at age sixty-three for women (or, too few women have enough employment years to retire at that age). The United Kingdom The U.K. program also has different provisions for men and women. Men can begin to receive benefits under the public social security program at age sixty-five; women can begin to receive benefit at age sixty. These differences are clearly reflected in the retirement patterns of men and women, as shown in figure 12. This figure shows the labor force “survival probabilities” for men and women who do not have an occupational pension. The proportion of women employees still in the labor force drops by about 20 percentage points (from 60 to 40 percent) at age sixty, but there is essentially no decline for men at this age. On the other hand, there is a 20 percentage point drop (from about 40 to about 20 percent) for men at age sixty-five, when they can receive benefits. Thus within-country differences in labor force departure rates by gender, by pension plan coverage, and over time show clearly that retirement is strongly influenced by eligibility ages. It seems clear that differences in labor force departure rates among countries are also strongly influenced by differences in eligibility ages. We would like the estimation and simulation methods used in the analysis to capture the essence of the empirical regularity between benefit eligibility ages and retirement and, at the same time,

Introduction and Summary

Fig. 12

23

U.K. survival rates for men and women

capture the effect of the other plan incentive measures as previously defined. We give particular attention to this issue later. Simulations: Method and Results Simulation Method Perhaps the best way to judge the implications of the country estimates is to consider the simulations based on the estimates. Thus, the main focus of this introduction is on the two illustrative simulations that are performed for each of the countries. The first simulation predicts the effect of delaying all program benefit eligibility ages by three years. In countries in which disability, unemployment, or other retirement pathways are important, the eligibility age for each of the programs is delayed by three years. The second simulation is intended to predict the effect of the same reform (the “common reform”) in each country. Under the common reform, the early retirement age is set at age sixty and the normal retirement age at sixty-five. Benefits taken before age sixty-five are reduced actuarially by 6 percent for each year before age sixty-five. Benefits taken after age sixtyfive are increased by 6 percent for each year the receipt of benefits is delayed. In addition, the replacement rate at age sixty-five is set at 60 percent of (projected) age-sixty earnings. The simulations are summarized in the table 3. The country papers show simulations done in nine different ways. For each reform, simulations are

24

Jonathan Gruber and David A. Wise

Table 3

Simulations Simulation and Estimation Method Common Reform—ER at 60, NR at 65, replacement rate of 60% (of age 59 earnings) at 65, 6% reduction before 65, 6% increase after 65

Three-Year Delay in Eligibility Ages

Simulation Method S1: Without age indicators S2: With age indicators but without increment age effects S3: With age indicators and increment age effects

Accrual

Peak Value

Option Value

Accrual

Peak Value

Option Value

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

Notes: ER = early retirement age; NR = normal retirement age.

based on the accrual, the peak value, and the option value methods. For each of these estimation methods, three simulation approaches are used. The three simulation methods are based on different uses of the age indicators. In every case the simulations are done by recalculating the relevant incentive (accrual, peak value, or option value) to correspond to the program change. In addition to the incentive measure per se, the SSW measure and wage earnings are also recalculated to correspond to the change in the program. Then the retirement rates are reestimated using the new measures. The key feature that distinguishes the three methods is the use of age indicators. The first simulation method (S1) does not use age indicators at all, either in the estimation or in the simulation. Only a linear measure of age is used as a control variable. In this case, only the incentive measures (and the related variables) are recalculated to simulate the effect of the reforms. The second simulation method (S2) uses age indicators in the estimation, but does not use the age indicators in the simulation. The third method (S3) uses age indicators in the estimation and, in addition, uses adjusted age indicators to simulate retirement under the program changes. Method S2 likely minimizes the simulated effect of the program changes. The effect of the incentive measures is estimated conditional on the individual age indicators, but in predicting the effect of program changes, the simulations account only for the change in the incentive measures and do not account for the age effects. This will clearly understate the true effect of the program changes, assuming that there are important program eligibility age effects. Simulation method S3 may typically yield the largest simulated effect of the program changes. The estimated age indicator effects, as well as the

Introduction and Summary

25

program incentive effects, are used to predict the effect of the program changes. For example, for the three-year eligibility delay, the age indicator for a given age is taken to be the estimated age indicator three years prior to the given age. The age-sixty indicator, for example, is taken to be the estimated age-fifty-seven indicator. The result is that under the three-year eligibility delay, the projected retirement rate at age sixty is approximately the same as the current program age-fifty-seven retirement rate. The spike at the early retirement age under the current program, for example, shows up there years later under the reform. This approach assumes that all of the estimated age effects can be attributed to the eligibility-age program provisions. (The ages include the age at which persons are eligible for one or more programs, as well as the normal retirement age.) Method S1 perhaps provides a middle ground. In this case, the estimation method does not explicitly allow for increases in retirement at given eligibility ages, and thus these effects are not allowed to influence the simulated effect of program changes. These effects will only be reflected in the simulations if they are captured by the estimated incentive effects. Recall that in our terminology, the incentive measures are the option value, the peak value, and the single-year-accrual financial measures; they do not include the eligibility age effects per se, which may also reflect additional incentives. Changing the early entitlement age by three years, for example, will change the incentive measures and SSW at every age, and the effect of this change is captured by the S1 simulations. But any other eligibility-age effects, such as social norms, liquidity constraints, or other reasons to retire at given eligibility ages, are not captured by the S1 simulations. As the simulation results reported afterwards show, method S3 most often yields the largest estimated effect of program changes—but this is not always the case. In several instances, method S1 yields larger effects than method S3. In the next section of this introduction, we discuss simulation results based on the option value (OV) estimation results and using the S1 and the S3 simulation methods, marked by an X in the table. Without undue complexity, this allows us to describe the general features of the results and to direct attention to the most important issue in estimation and simulation—the use of age indicators. We also focus solely on the results for men for expositional convenience, but results for women as well as men are presented in each country’s chapter. Simulation Results We begin by considering the results for the three-year delay in program eligibility. We first show results based on S3. Then, for the three-year delay simulation, we compare results based on S1 and S3. Next, we consider the predicted effects of the common reform, and then we compare results under the common reform and under the three-year eligibility delay. Before

26

Jonathan Gruber and David A. Wise

proceeding to cross-country comparisons, however, we briefly explain why country-to-country differences in simulated results should be expected. Differences Across Countries Although the overall simulated effects of the illustrative reforms are large in all countries, the magnitude of the effect differs from country to country. There are several reasons for the differences: The first and the single most important reason is that the current programs differ substantially among countries, and thus the effect of given reforms should differ as well. A second reason is that the data files upon which the estimates are based differ from country to country; in a few countries, the data pertain to only a portion of the workforce. A third reason is that there may be differences across countries in individual responses to a given incentive. A fourth reason, related to the second, is that the precise calculation of the incentive measures may differ somewhat from country to country. A fifth reason is that the precise implementation of the simulations may differ among the countries. It is, of course, not possible to apportion the quantitative effect of each of these reasons. The models described above are used to predict the effect of program incentives (and other variables) on hazard rates, or the likelihood that a person in the labor force at a given age will leave the labor force at that age. The simulations begin with base hazard rates, that is, the predicted hazard rates under the current program. Then new hazard rates are predicted based on the provisions of the illustrative reform. These hazard rates are then used to predict the proportion of persons out of the labor force at given ages, and these proportions are used to determine the proportion of persons out of the labor force in given age ranges, such as fifty-six to sixtyfive. Thus any of the reasons for differences that affect the base hazard rates (under the current program) or the predicted hazard rates (under the illustrative reforms) will lead to different results among the countries. Three-Year Delay in Eligibility Turning to cross-country comparisons, we begin with simulations of the effect of a three-year delay in eligibility. Results for men aged fifty-six to sixty-five are shown in figure 13. The figure shows the out-of-the-labor (OLF) percentage for the base case and under the eligibility delay. The eligibility delay estimates are based on S3. In all countries there is a noticeable reduction in the proportion of men out of the labor force when the eligibility ages are increased by three years. The comparison among countries may be confounded, however, by the wide variation across countries in the age at which retirement begins. Thus, the change in the proportion out of the labor force may vary more among countries at younger ages than over the entire fifty-six-to-sixty-five age range. To help to standardize for this effect, we define the first age at which

Introduction and Summary

Fig. 13

27

OLF ages 56–65: Base versus three-year delay, OV-S3

at least 25 percent of men are out of the labor force, which we call the “25 percent age.” Then we consider the five years beginning with the “25 percent age”—“25 percent age plus four years.” The results for the “twenty-five percent age plus four years age” range are shown in figure 14, in which the “25 percent age” is shown at the top of the bars for the base case. The “25 percent age” ranges from age fifty-three in Italy to age sixty-two in Spain and the United States. Within the “twenty-five plus four years” age range, the OLF proportion is currently (under the base case) between 40 and 50 percent in ten of the twelve countries. Within this more standardized age range, there is typically a greater reduction in the OLF percent—when eligibility is delayed—than for the fifty-six-to-sixty-five age range. The percent reduction in the OLF proportion for the “twenty-five percent age plus four years” age range is shown in figure 15. The average reduction is 47 percent, with a range from 14 percent in Canada to 77 percent in Germany. Figure 16 is a comparison of results under S1and S3. The figure shows reductions in the OLF proportion for the “twenty-five percent age plus four years” age range. Recall that under S1, predictions are based on changes in the incentive measures alone, while under S3, age-specific indicators are used as well.

Fig. 14 OV-S3

OLF 25 percent age plus four years: Base versus three-year delay,

Fig. 15 OV-S3

OLF change 25 percent age plus four years: Base versus three-year delay,

Introduction and Summary

Fig. 16

29

OLF 25 percent age plus four years: Three-year delay, OV-S1 and OV-S3

There are two notable features of this figure. The first is that the overall reductions are large under either method, and in many of the countries, the two methods yield quite similar results. The average reduction is 47 percent under S3 and 28 percent under S1. Under S3, the reduction is at least 34 percent in eleven of the twelve countries. Even using S1, the reduction is greater than 23 percent in six of the twelve countries (and in two of these countries the reduction is greater under S1 than under S3). The reduction is between 12 and 18 percent in four countries. Only in Italy and Japan is the estimated effect under S1 quite small. For the first six countries—Spain, the Netherlands, France, Canada, Germany, and Denmark—the two methods yield rather similar results. In the first eight countries, the reduction under S1 is at least 47 percent of the reduction under S3. Thus on the whole, the reduction in the OLF proportion is large under either approach. The second notable feature of the figure is the similarity across countries in the reduction under S3. The reduction is between 34 and 55 percent in nine of the twelve countries (in Germany and Sweden the reductions are 77 and 68 percent, respectively). This similarity reflects the similarity in the estimated age effects at program eligibility ages. In all countries, there are spikes in the hazard rates at these program eligibility ages similar to those for the United States and Germany (shown in figures 6 and 8).

30

Jonathan Gruber and David A. Wise

In summary, for almost every country, the estimates under either method show very large reductions in the OLF proportion when program eligibility ages are raised. The reductions, however, are typically larger under S3, which allows age-specific variables to capture the effect of benefit eligibility on retirement. Based on the strong empirical regularity between retirement rates and program eligibility ages, as discussed previously and in the first volume, we believe that S3 likely provides the more tenable predictions of the long-run reductions in the OLF proportions (although responses to increases in eligibility ages may not parallel responses to reduction in eligibility ages). However, even under the more restrictive S1—which does not directly allow for eligibility age effects—the predicted effect of the delay in eligibility ages is large in almost all countries. Common Reform We turn now to simulation results for the common reform. Recall that the common reform has four key features: (a) It sets the normal retirement age at sixty-five; (b) it sets the early retirement age at sixty; (c) benefits are reduced actuarially if they are taken before age sixty-five; and (d) the replacement rate at the normal retirement age is set (approximately) at 60 percent of age-fifty-nine earnings. Figure 17 shows the OLF proportion for

Fig. 17

OLF ages 56–65: Base versus common reform, OV-S3

Introduction and Summary

31

the fifty-six-to-sixty-five age range under the base case and under the common reform, based on S3. Only in Germany and in the Netherlands is the OLF proportion reduced substantially. In five of the twelve countries, the common reform increases the OLF proportion. But there appears to be no clear pattern between the base-proportion OLF and the reduction in the OLF proportion under the common reform. The variation in the OLF proportion, relative to the base, is reduced under the common reform, but substantial variation across countries remains. For the “twenty-five percent plus four years” age range, however, there is a strong pattern regarding the change in the OLF percent. The OLF proportion under the base and under the common reform for this age range is shown in figure 18. In this figure, the countries are ordered by the “25 percent age.” For the “twenty-five percent plus four years” age range it is clear that the greatest reductions in the OLF proportion under the common reform are realized in the countries with the lowest “25 percent age” rates. The change in the OLF proportion in the “twenty-five percent” age plus four years age range is shown in figure 19. For the six countries with a “25 percent age” less than sixty, the average reduction in the OLF proportion is 44 percent. For the six countries in which the “25 percent age” is sixty or more, there is, on average, a 4 percent increase in the OLF proportion.

Fig. 18

OLF 25 percent age plus four years: Base versus common reform, OV-S3

32

Fig. 19 OV-S3

Jonathan Gruber and David A. Wise

OLF change 25 percent age plus four years: Base versus common reform,

The systematic pattern of these results shows a strong correspondence with intuition. For the countries with the lowest “25 percent age,” the common reform represents a substantial increase in the youngest eligibility age, and the actuarial reduction means that benefits at this age are much lower than under the base country plans. Thus, for these countries, the OLF proportion should decline under the reform, which is the case for every country but Canada. But for the countries with a “25 percent age” of sixty or greater, the common reform may reduce the earliest eligibility age—as in the United States—and may provide a greater incentive to leave the labor force. In addition, the 60 percent replacement rate at the normal retirement age represents an increase for some countries, like the United States, and a reduction in the replacement rate for other countries. Consequently, in three of these six countries, there is an increase in the OLF proportion under the common-reform simulation, and on average there is an increase in the OLF proportion. The seemingly anomalous result for Canada is explained by the fact that Canada is the only country in which the “25 percent age” is below the nominal social security entitlement age; the “25 percent age” is fifty-eight, while the social security entitlement age is 60. In

Introduction and Summary

33

Fig. 20 OLF change 25 percent age plus four years: Base versus common reform, OV-S1 and OV-S3

addition, Canada has relatively low benefits at the early retirement age (age 60). Thus the common reform significantly increases benefit levels, providing an additional inducement to retirement. The simulated changes under the common reform based on S1 and S3 are compared in figure 20. In each of the countries, both methods either predict a reduction or an increase in the OLF proportion. Overall, the magnitude of the simulated changes based on the two methods is rather similar as well. The most apparent exceptions are Italy, Canada, and Germany. Both methods, on average, show reductions in the OLF proportion in the six countries with the lowest “25 percent age” rates and small changes in the OLF proportion for the six countries with the highest “25 percent age” rates. The differences between the groups are more muted, however, under S1. Based on S3, the average OLF change is –44 percent for the first six countries and 4 percent for the last six, as noted previously. Based on S1, the OLF change is –21 percent for the bottom six countries and –4 percent for the top six. Most of the difference between the methods is accounted for by the differences for Italy, Belgium, Germany, and Canada. (The anomalous result for Canada is already explained.) Again, on the whole, the two methods suggest similar results. Like the simulated

34

Jonathan Gruber and David A. Wise

effects of the three-year eligibility delay, we believe that S3 is likely to give the most reliable long run predictions. Comparing the Three-Year Eligibility Delay and the Common Reform Finally, figure 21 compares the results for the three-year eligibility delay with the common-reform results. The figure is based on S3. It shows the percent change from the base under the two reforms for the “twenty-five percent plus four years” age range, which is shown as the label on the common-reform bars. Recall that the three-year delay reduces the proportion out of the labor force in all countries. The average reduction in the OLF proportion is 47 percent, and there is little difference in the “25 percent age” (–49 percent for the six countries with the lowest “25 percent age” rates and –45 percent for the six countries with the highest “25 percent age” rates). The results under the common reform, however, should depend on the base program provisions. As shown in figure 19, for the first six countries the average reduction in the OLF proportion is –44 percent and for the last six countries the average increase is 4 percent. In particular, under the common reform, benefits are not available in any country until age sixty. For

Fig. 21 OLF change 25 percent age plus four years: Three-year versus common reform, OV-S3

Introduction and Summary

35

many countries, current benefits are available well before age sixty. (The “25 percent age” helps to identify the early eligibility countries.) In most of these countries—Italy, the Netherlands, Belgium, France, Canada, and Germany—the delay in eligibility should tend to reduce the OLF percent, and this is the result in all of these countries, with the exception of Canada. In addition, the actuarial reduction in benefits if they are taken before the age-sixty-five normal retirement age should tend to further reduce the OLF percent in most of these countries. Moreover, the common reform represents a reduction in the replacement rate in some of the countries. In other countries, such as Spain and the United States, current benefits are not available until age sixty or later. In these countries the common reform could increase the OLF proportion. Overall, the relative effects of the two reforms are plausible, lending credence to the estimation approach. Under the three-year eligibility delay, which should reduce retirement ages in all countries, there is a reduction in the OLF proportion in each countries, and in many countries, the reduction is very large. But for the common reform, the effects should depend on country-specific program provisions, as the simulations show. Conclusions Our introduction to the first volume of the project concluded with a striking graph showing a strong relationship across countries between social security program incentives to retire and the proportion of older persons out of the labor force (figure 4 of this introduction). From the weight of the evidence, we judged that the relationship was largely causal. The strong response of retirement decisions to within-country changes in program provisions over time, and to different provisions for different groups at a point in time (shown in figures 6–12), also point to a casual relationship between program provisions and retirement. The results of the country analyses reported in this volume confirm the strong causal effect of social security program retirement incentives on labor force participation. But perhaps more importantly, the results in this volume show the large magnitude of these effects. Across twelve countries with very different social security programs and labor market institutions, the results consistently show that program incentives accord strongly with retirement decisions. The magnitude of the estimated effects varies from country to country, but in all countries, the effects are large. The magnitude is illustrated most clearly by the simulations reported in each country’s paper, and we have emphasized the simulations in this introduction. Considering the average across all countries, a reform that delays benefit eligibility by three years would likely reduce the proportion of men aged fifty-six to sixty-five out of the labor force between 23 and 36 percent, perhaps closer to 36 percent in the long run. For the “twenty-five per-

36

Jonathan Gruber and David A. Wise

cent plus four years” age range, the average reduction would likely be between 28 to 47 percent, and perhaps closer to 47 percent in the long run. The effects are much larger than this in some countries, and in virtually every country, the effects are large regardless of the estimation method. On the other hand, an illustrative common reform—with early retirement at age sixty, normal retirement age sixty-five, and actuarial reduction in benefits between sixty-five and sixty—has very disparate effects across the countries, depending on the provisions of the current program in each country. For the countries in which the current modal retirement age is younger than sixty, this reform typically implies a reduction in retirement incentives, and under this reform, the simulated proportion of older persons out of the labor force declines substantially in most countries. But for countries in which the modal retirement age is sixty or older, this reform may represent an increase in retirement incentives, and the proportion of persons out of the labor force may increase, on average, in these countries. The strong correspondence between the simulation results and a priori expectations lends credence to the estimation procedures used in the country papers. In short, the results in this volume provide an important complement to the first volume. The results leave no doubt that social security incentives have a strong effect on retirement decisions, and the estimates show that the effect is similar in countries with very different cultural histories, labor market institutions, and other social characteristics. While countries may differ in many respects, the employees in all countries react similarly to social security retirement incentives. The simulated effects of illustrative reforms reported in the individual chapters make it clear that changes in the provisions of social security programs would have very large effects on the labor force participation of older employees. In the next stage of the project, we will use the estimation results and simulation methods developed in this stage to estimate the financial implications of changes in program provisions.

Appendix Incentive Measures In this appendix, we review the relationship between the two forwardlooking incentive measures—the option value and the peak value. Under the option value formulation, the value at age t of retirement r is given by r1

S

St

Sr

Vt (r)  ∑ StEt(Y S )
∑ StEt (kBS (r)) 

Introduction and Summary

37

using the Stock-Wise specification. Here Y is future wage income and B is social security benefit income, which depends on the retirement age r. For simplicity, the probabilities of being alive to collect the income or the benefits have been suppressed. The gain from postponing retirement to r, versus retiring at age t, is given by r1

S

S

St

Sr

St

Vt (r)  ∑  StEt (Y S )
∑  StEt(kBS (r))   ∑ StEt (kBS (r)) . If r∗ is the retirement year that gives the maximum expected gain, the option value is given by r1

OVt (r∗)  ∑ StEt (Y S )
St








discounted utility
of future wage



S

∑

Sr*




S



StEt (kBS (r∗))   ∑ StEt (kBS (t)) 

discounted utility of benefits if retire at r∗

St




discounted utility  of benefits if retiree at t



.

Considering this equation, we can see that there are two ways to calculate the option value used in the analyses in this volume: One way is to use prior estimated values for the utility parameters , , and k. The second is to set assume a value for  and to set   k  1. Multiple Pathways to Retirement and Combining Programs In the United States, the social security program is the principle public program route to retirement. Only a small fraction of older persons enter retirement through the disability program. In some European countries, however, there are several public programs that provide routes to retirement. The case in Germany is discussed in the text and illustrated in figure 5. Thus, in considering the incentive to retire, it is important to recognize that retirement incentives under several programs may matter. The key question is which program, or programs, a person could choose to enter, out of those available to a given person. For example, who could retire under the disability program? In some instances, administrative provisions limit the universe of persons who might be eligible. In other instances, a large fraction of persons could be eligible, but which persons are eligible is unknown. Thus, the incentives facing a given individual must be estimated probabilistically. We would like to have the probability that each person is eligible for each program. Suppose that the incentive measure under each possible program is calculated for each person for each age. Then for each age, these probabilities could be used to obtain a weighted incentive measure, in which the weights are the probabilities that the person is eligible for each program. This is essentially an instrumental variable approach. In principle, eligibility probabilities should be estimated for each person

38

Jonathan Gruber and David A. Wise

for each age, depending on administrative rules as well as individual attributes. During the course of this part of the project, several different approaches were tried in various countries. Eligibility for disability is a good example. Based on administrative rules, it might be assumed that every person is eligible for disability beginning at some age, or it might be assumed that eligibility probabilities correspond to actual empirical take-up rates by age and other variables (where the take-up rate is estimated based on personal attributes). Here, it is implicitly assumed that the take-up rate for a person with a given set of attributes represents, on average, the eligibility rate for persons with that set of attributes. There is no correct way to do this without knowing eligibility for a sample of persons and then being about to predict eligibility. In few, if any, countries was this an option. Thus, for the purposes of the estimates in this volume, we have elected to assign weights based on empirical take-up rates. In this case the disabilityeligibility probability will typically increase with age, for example. (Deviations from this method are noted in the individual chapters.) To explain the procedure we use the situation in Denmark, which is likely the most complicated of the country situations. Here are the programs in Denmark, together with the eligibility age and information to determine eligibility, are shown in table A.1. An important calculation is the probability that a person is eligible for social and disability pension (SDP). To obtain the probability of SDP eligibility, the approach is to use actual takeup rates by age, year, and sex cells, by disability level. Suppose the calculation pertains to pension-based SSW. The goal is to obtain a weighted measure based on the probability a person is eligible for a specific program at a given age. At sixty-seven and beyond, a person is only eligible for the old age pensions (OAP) and possibly the public employees pension (PEP). Persons who retire under other programs convert Table A.1 Program

Eligibility Age

Determine Eligibility Based On

PEW (post-employment wage)

60 to 66

Age and insurance fund information

TBP (transitional benefits program)

55 to 59; unemployed at ages 50 to 59 during 1994 to 1996

Age and insurance fund and unemployment information in two years prior to age

PEP (public employees pension)

60 to 69

Employer pension contributions over required period

SDP (social and disability pension)

18 to 66

Probability: Based on observed participation rates by age-year-gender in each of the three levels of the disability program

OAP (old age pensions)

67

Age = 67

Introduction and Summary

39

to OAP and start receiving benefits under that program at age sixty-seven. Before age sixty-seven, a person could be eligible for the post employment wage (PEW) or the transitional benefits program (TBP), but not both. Assuming that we know for sure whether a person is eligible for these programs, these are the potential sources of wealth: SSWOAP , which is available to all persons; SSWPEW , which could be available between ages sixty and sixty-six; SSWTBP , which could be available between ages fifty-five and fifty-nine or fifty and fifty-nine; SSWPEP , which is available to public employees who meet certain criteria; and SSWSDP , which could be available even before age fifty. Now SSWOAP can be thought of as a base that is available to everyone. The question is then what else is or might be available. Assuming that a person is not eligible for either PEW or TBP (or PEP), but that the person is eligible for SDP with probability p, then the weighted average SSW would be SSW  SSWOAP
p∗ max [0, (SSWSDP  SSWOAP )]. That is, with probability p the person has more than SSWOAP (or, with probability p the person would have SSWSDP and with probability 1 – p the person would have SSWOAP only). The formula as set out above accounts for the (unlikely) possibility that SSWSDP is lower than SSWOAP , in which case the disability option would be ignored. (SSWSDP can be collected beginning at an age much younger than sixty-seven, so it will almost surely be greater than SSWOAP , which can only be received beginning at age sixty-seven.) After age sixty, a person could be eligible for PEW, for example (if you retire before you are eligible for this program you never get these benefits). In that case, SSW is SSW  max(SSWOAP , SSWPEW )
p∗ max [0, (SSWSDP  max (SSWOAP , SSWPEW ))]. That is, the certain amount in this case is max(SSWOAP , SSWPEW ). Once again, the maximum will almost surely be SSWPEW , since the person can take benefits SSWPEW , which incorporates benefits SSWOAP beginning at age sixty-seven. With probability p, the person could be eligible for more, assuming that SDP would provide more. A similar procedure is used to estimate SSW in two consecutive years, and from SSW in those two years, the social security accrual from one year to the next can be calculated. The peak value and option value measures are obtained in a similar fashion, but in these cases, wealth measures must be calculated for all ages into the future.

40

Jonathan Gruber and David A. Wise

References Coile, Courtney, and Jonathan Gruber. 2000a. Social security incentives for retirement. In Themes in the economics of aging, ed. David Wise, 311–341. Chicago: University of Chicago Press. Coile, Courtney, and Jonathan Gruber. 2000b. Social security and retirement. NBER Working Paper no. 7830. Cambridge, Mass.: National Bureau of Economic Research, August. Gruber, Jonathan, and David A. Wise, eds. 1999. Social security and retirement around the world. Chicago: University of Chicago Press. Kotlikoff, Laurence J., and David A. Wise. 1985. Labor compensation and the structure of private pension plans: Evidence for contractual versus spot labor markets. In Pensions, labor, and individual choice, ed. David A. Wise, 55–87. Chicago: University of Chicago Press. ———. 1989. The wage carrot and the pension stick. Kalamazoo, Mich.: W. E. Upjohn Institute for Employment Research. Lazear, Edward. 1983. Pensions as severance pay. In Financial aspects of the United States pension system, ed. Zvi Bodie and John Shoven, 57–85. Chicago: University of Chicago Press. Stock, James H., and David A. Wise. 1990a. The pension inducement to retire: An option value analysis. In Issues in the economics of aging, ed. David A. Wise, 205– 24. Chicago: University of Chicago Press. ———. 1990b. Pensions, the option value of work, and retirement. Econometrica 58 (5): 1151–80.

1 Micro-Modeling of Retirement in Belgium Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

1.1 Introduction The Belgian social security systems face an uncertain future. One major reason is the financial burden imposed by the aging of the population. For the systems to survive this demographic process, higher contribution levels, lower benefits, or both will have to be introduced, given the pay-as-yougo (PAYG) nature of these systems. Indeed, a straight increase in the public debt to finance this demographic transition is not an option as it would mean pushing the already high ratio of public debt to gross domestic product (GDP) to even further astronomical heights. Most recently however, the successive Belgian governments have successfully brought down the debt-GDP ratio to close to 115 percent from a level of 130 percent by means of strict budgetary policy. The government has even achieved a small budget surplus in 2001 and has reached the same goal in 2002 even in the presence of an economic slowdown. This fiscal rigor will no doubt increase the margin of maneuverability of the federal government in its attempts to cope with the demographic aging process. Another factor of uncertainty pertains to the consequences of increased labor mobility on the way the social security systems are organized. First Arnaud Dellis is a Ph.D. candidate at Cornell University. Raphaël Desmet is a researcher at the University of Liège. Alain Jousten is professor of economics at the University of Liège. Sergio Perelman is professor of labor economics at the University of Liège. We are indebted to the Belgian Institut National de Statistiques (INS) and in particular to Frans Desmedt, Paul Everaerts, and Jean-Marc Sobrie for their outstanding contributions in building the database used in this study. Also, we want to thank Jan Lesthaeghe, from the Free University of Brussels, for agreeing to perform the estimation of life tables by educational group. We thank Pierre Pestieau for helpful discussions. All remaining errors are entirely our own. Financial support from the Belgian National Science Foundation is acknowledged (FRFC Project no. 2.4.544.01).

41

42

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

of all, increased mobility between jobs in the public sector, in the private sector, and in self-employment may induce large changes in the way the three corresponding social security systems work. Recent reform proposals by the Belgian federal government to improve the way the public sector works also has to be seen in this light. Mobility between sectors will most likely increase once people’s behavior is determined by similar factors measuring achievement and productivity in both the public and the private sectors. In those circumstances, a review and harmonization of the corresponding public retirement-income systems seems warranted. Second, the question of international job mobility is becoming more and more important, particularly for a small open economy in the heart of Europe like Belgium. Jousten and Pestieau (2002) study the implications of an expected increase in labor mobility from a European perspective, and the authors pay particular attention to the degree of redistribution inherent to the different systems. They argue that both levels of intra- and intergenerational redistribution will be widely affected, even if we replace the assumption of perfect labor mobility between member countries with a more restrictive and plausible one of mobility limited to individuals belonging to particular income groups. However, even leaving these two challenges aside, the Belgian social security system needs reform. The widespread use of a variety of early retirement programs makes Belgium the country with the lowest average retirement age in the Western world, which is approximately fifty-seven years old for men.1 This chapter studies the incentives pushing people toward retiring early. We explicitly model the incentive structure built into the various public retirement and early retirement systems. First, we compute indicators of benefit entitlement, such as social security wealth. Then, we define several different incentive measures based on the notion of social security wealth. In a third step, we perform an empirical estimation of microeconometric probit and option value models. From our exceptionally rich and broad database, we are able to compute a rather accurate measure of all individuals’ pension wealth, as well as of the implicit tax rates the elderly workers face in case of delayed retirement. The structure of the paper is as follows: Section 1.2 describes the essential features of the various public retirement and early retirement systems. In section 1.3, we explain the different, mostly administrative components of our large data set. The following section touches on the problem of the earnings process used in the simulations and estimations. Section 1.5 explains the process and logic underlying the construction of the different incentive measures used, while section 1.6 contains regression results ob1. The average retirement age of 57.6 years for men was estimated by Blöndal and Scarpetta (1998) on the basis of the Labor Force Surveys. In this study we estimated an average retirement age of 58.4 years for men and 57.4 years for women.

Micro-Modeling of Retirement in Belgium

43

tained using these latter incentive measures. Section 1.7 delivers two policy simulations using the previously estimated coefficients. The first simulation consists of an increase by three years of the eligibility age in the various retirement systems, and the second consists of a policy in which early retirement would be possible at age sixty at the earliest, while normal retirement age would be sixty-five. Section 1.8 is devoted to the conclusions. 1.2 Social Security Schemes The Belgian retirement income system rests upon three very unequal pillars. First of all, there are the public social security programs that represent the largest part of pension income for a wide majority in the population. The second pillar consists of company pension schemes, which only play a minor role as a source of income for the average Belgian worker. Essentially, they are currently confined to the higher-income individuals in the private sector and to the self-employed, a finding that is at least in part due to their tax treatment. A third type of retirement income comes from individual retirement savings. These take multiple forms: there are tax-favored individual-pension savings accounts with a maximum annual contribution of €580 per person (approximately U.S.$615),2 or there are more traditional savings vehicles, such as the tax-favored savings accounts, investments in trust funds, life insurance, and so forth. The dominance of the first pillar can also be represented in numbers: Whereas the first pillar represents pension entitlements of more than 250 percent of GDP, assets in private-pension funds only amount to 10 percent of GDP.3 The first pillar, public retirement programs, essentially consists of four components. There are three large sectoral social security programs; one for the public sector, one for the private-sector wage earners, and one for the self-employed. Some special categories of workers, such as coalmine workers and military personnel, have special retirement systems that we will not analyze in the present paper. A fourth large category of public retirement income consists of the guaranteed-minimum-pension system that operates on a means-tested basis. Aside from these pure retirement programs, the Belgian government has introduced early retirement provisions that either operate under the name of early retirement scheme, or alternatively as a form of old age unemployment. Table 1.1 gives a brief outline of the importance of the different categories of social security programs for the year 1995. 2. In this paper, we apply an exchange rate of €0.942 per $1, which approximately corresponds to the exchange rate in place on 31 December 1999. 3. See the Organization for Economic Cooperation and Development (OECD; 1994), Bouillot and Perelman (1995), and the European Economic Community (EEC; 1994) on the subject.

44

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

Table 1.1

Categories of Social Security Schemes in 1995 Benefits as % of GDP

Number of Pensionersa

Average Amount of Benefits in Relative Terms (%)

5.72 0.71 3.38 0.64 0.14 10.59

1,347 246 405 128 50 2,175

87.3 59.1 170.7 103.8 56.7 100.0

Wage earners Self-employed Public-sector workers Mandatory early retirement Minimum old age pensioners All schemes

Source: Bouillot and Perelman (1995). a The number of pensioners is reported in thousands and also includes surviving spouses. A possibility of double counting exists.

1.2.1 Wage Earners’ Scheme The wage earners’ scheme is the largest program, according to the number of people affiliated with the program. The program allows for retirement between the ages of sixty and sixty-five, with the choice of the retirement age not inducing any actuarial adjustment.4 The system worked according to a different logic till the early 1990s. Until 1992, the wage earners’ scheme had an actuarial adjustment factor of 5 percent per year of early retirement before the age of sixty-five. However, in the case of most workers, the choice of the retirement age is not completely neutral with respect to the benefit amount, as most men still accrue extra pension rights by working additional years between the ages of sixty and sixty-five. This is so because a full earnings history consists of forty-five years of work for men, a condition that many people do not satisfy at the age of sixty. For those having more than forty-five working years, a dropout-year provision replaces low-income years with higher ones. The situation so far has been slightly different for women. Indeed, until very recently, women only needed forty years of earnings to have a complete earnings history. In reaction to successive court rulings on the illegality of this sex discrimination, the Belgian government introduced a reform a few years ago that aims at progressively increasing the complete career condition to forty-five working years for women over the time period 1997–2009. However, for most women included in our data set, a full career still consists of forty years of work. Benefits are computed based on earnings during periods of affiliation. The benefit formula can be represented as follows: 4. Notice that, from a legal point of view, age sixty-five does not represent a compulsory retirement age, but rather an age at which a worker loses the social protection associated with his job.

Micro-Modeling of Retirement in Belgium

45

n Benefit
  average wage  k, N where n represents the number of years of affiliation with the wage earner’s scheme, N the number of years required for a full career (in our case either forty or forty-five), and k is a replacement rate, which takes on the value of 0.60 or 0.75 depending on whether the social security recipient claims benefits as a single person or as a household. The variable average wage corresponds to indexed average wages over the period of affiliation, with indexation on the price index combined with additional discretionary adjustments for the evolution of growth. A peculiar feature of the Belgian wage earners’ scheme is that periods of the life spent on replacement income (e.g., unemployment benefits, disability benefits, workers compensation, and the like) fully count as years worked in the computation of the average wage, and hence of the social security benefit. For any such periods, fictive wages are inserted into the average wage computation. In line with the general philosophy of the Belgian social insurance system that any such spell on a replacement-income system is purely involuntary, imputed wages are set equal in real terms to those that the workers earned before entering these replacement-income programs. An additional category of linked benefits is payable to surviving spouses or, more generally, to surviving dependents of deceased wage earners. All the different types of benefits provided for under the wage earners’ social security system are covered against erosion by the means of inflation through an automatic consumer price index (CPI) adjustment. The system works both with floors and ceilings, which are either indexed to the evolution of prices or to average wages. The minimum household pension represents a floor for workers that have contributed during their entire working life to the systems. It is approximately equal to 56 percent of average net wages. At the opposite extreme, a ceiling operates on pensionable but not on taxable earnings. The earnings entering the above pension formula are strictly limited to a maximum of 120 percent of average gross wages. Wage earners’ pensions are also subject to an earnings test. Currently, the earnings limit is approximately €7,450, or $7,900, per year. For earnings above this limit, pension entitlement is suspended. The wage earners’ system is essentially based on the PAYG principle and financed through payroll taxes that are levied both on the side of the employers and of the employees with a combined contribution rate of 16.36 percent. The system also receives a subsidy from the Belgian federal government that is approximately equal to 11 percent of overall benefits. Next to the official wage earners’ scheme, several forms of early retirement programs have developed over the last few years, some officially carrying that name and others that do not (e.g., unemployment, preretirement, and so forth). Those schemes can be broadly subdivided into two

46

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

groups: collective mandatory retirement and individual early retirement. During the 1980s and the 1990s, an arsenal of mandatory early retirement schemes was put in place. All of these arrangements were and are based on collective agreements, which are negotiated with the active involvement of employees and employers, sometimes at the sectoral level or sometimes at the level of an individual company or production site. For some companies in a difficult economic position, mandatory retirement ages as low as fifty were introduced. The federal government did not necessarily object to such arrangements, as it considered early retirement as a good tool in the fight against youth unemployment. Indeed, some of these early retirement schemes required the employers to rehire young workers. Lately however, these early retirement schemes have undergone some scrutiny. Not surprisingly, the beneficial labor market effects have been rather modest, if not completely absent. Recent discussions and decisions at the government level clearly move in the direction of lifting the effective early retirement age and hence also the sector-specific mandatory retirement ages. This has to be seen in contrast to the massive costs these programs induce for the federal budget, as well as for the society as a whole. First of all, the effect of these waves of early retirement on the federal budget operates both through missing contributions during the period spanning from early retirement to the normal retirement age as well as through additional costs. Thus, the federal government pays a large fraction of early retirement compensation.5 On the other hand, an individual’s pension rights in the wage earners’ pension schemes are essentially unaffected by the decision to retire early or not. This is due to the previously discussed feature of the Belgian social security system that days spent on replacement income count as working periods in the computation of average pensionable earnings and of periods of activity. Hence, retiring early does not induce any loss of income during retirement. Individual early retirement differentiates itself from its collective counterpart by the fact that it is based on an individual’s decision to retire from work. The most prevalent way is to pass through the unemployment system in which the unemployed aged fifty or older are considered “aged unemployed,” and thus neither subject to controls on availability to work, nor to benefit cuts due to long-term unemployment. Therefore, people unwilling to continue to work can ask their employer to lay them off. The latter has no incentive not to lay the worker off, unless the employer considers the employee to be a crucial wheel in the working of the company. Laying the worker off allows the employer to replace an expensive old worker by a cheaper young one. Furthermore, the employer’s behavior does not add any costs to his unemployment contributions, as the system is not experience rated. 5. Depending on the early retirement scheme, the employer pays part of the income to the worker from the early to the normal retirement age.

Micro-Modeling of Retirement in Belgium

47

Next to the unemployment path, some people also attempt to proceed to retirement through the disability-insurance scheme. However, in the Belgian context, we think that disability is not a very prominent means of departure, at least not for private-sector employees. Incentives to claim disability benefits are rather limited: Medical screening is relatively severe, and benefits are not significantly more interesting than early retirement provisions. 1.2.2 Public-Sector Employees The social security scheme for public servants is the oldest one of the three, dating back to as early as 1844. Public pensions are paid out of the general federal budget. Officially, the public sector pensions are considered as deferred income rather than old age insurance. The only official insurance aspects are the 7.5 percent payroll taxes that the public-sector employees have to pay to finance survivor benefits. Benefits are essentially individualized, that is, there are no additional spousal benefits available for no- or low-income spouses. Civil servants’ pensions are compulsory as of age sixty-five for both men and women. However, as for the private sector, there is a multitude of ways of retiring earlier than this normal age of sixty-five. It is possible to opt for an incomplete career and retire at sixty. For some particular categories of workers, the normal retirement age is lower than sixty-five, and early retirement provisions are sometimes extremely generous. This is particularly the case for military personnel and for teachers, who have always enjoyed a much more favorable treatment. For example, secondary-school teachers in the French-speaking community can either retire at age fifty-five if they have sufficient years of service, or alternatively take a less demanding route in terms of career requirements and retire at the age of about fifty-eight. Public-sector pensions are based on the income earned by an individual during the last five years before retirement. Benefits are computed according to a rather complicated formula, but can never exceed 75 percent of the average wages over the last five years. The benefit formula can be represented as follows: Benefit
average wage over last five years  min (fract; 0.75), where fract is a fraction with a numerator consisting of the number of years the person worked in the public service, and the denominator is a benefit accrual factor. This latter benefit accrual factor, also called tantième, depends on the rank the person occupied in the hierarchy. This denominator ranges from 30 to 60, taking the value of 30 for the highest-ranking civil servants (university professors and so forth) and 60 for the lowest ranks. In addition to the aforementioned limit on pensions of 75 percent of the average gross wage, there is also an absolute limit to the amount of a public-sector pension, which corresponds to about three times average

48

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

gross wages in the economy. Furthermore, minimum pensions are also available to aged civil servants, which corresponds to 56 percent of average wages for a single individual and 70 percent for a one-earner couple.6 Notice the rather marked difference of these floors and ceilings with those applicable to the private-sector employees. A major conclusion we can already draw at this stage is that higher-income civil servants get a much better deal than their private-sector counterparts. This finding is even reinforced once we introduce another aspect, namely the indexation rules. As opposed to the private-sector social security system that only indexes benefits to the CPI, public pensions are indexed to average wages ( péréquation). Civil servants therefore enjoy the benefits of productivity increases in the economy even beyond the moment when they actively contribute to them as workers. Aside from this official route to retirement, civil servants have another alternative to quit work early through disability protection. This route seems a much more plausible route to early retirement for public-sector employees than for private-sector wage earners as the screening seems to be much less severe. 1.2.3 Self-Employed The self-employed retirement scheme is the latest one to have been introduced as it has only existed since 1956. It is also the least generous of the three big social security systems. For a very long time, pensions have been independent of earnings levels. However, since 1984, the system has been progressively transformed to allow for a stronger link between contributions and benefits. Additional earnings past 1984 enter the pension computation formula at their correct value, instead of some fictive amount. Given the period of analysis we will be looking at in our econometric analysis, pensions of the self-employed are still essentially independent of their earnings histories and contributions. Full benefits are available at age sixty-five for men with a complete earnings history of forty-five years. However, early retirement is possible as early as age sixty with a reduction of 5 percent per year of anticipation. Women are currently in a transitory system that progressively increases their normal retirement age from sixty to sixty-five and the complete career requirement from forty years of work to forty-five in line with the reform of the wage earners’ scheme. Again, most women in our data set are still subject to a normal retirement age of sixty and a full career condition of forty years of work. The social security system of the self-employed is financed through two broad categories of income. First, there are direct social insurance contri6. The latter feature is the only instance in which the family structure matters for the amount of civil service pension.

Micro-Modeling of Retirement in Belgium

49

butions levied under the form of a 16.7 percent tax on the first €46,035 ($47,368) of income and 12.27 percent on the income in the bracket between €46,035 and €67,352 ($89,302). Income above the latter threshold is not subject to social-insurance taxation. More than 75 percent of the contributions raised using this social-insurance taxation are used for the pension system of the self-employed, the remainder serves to cover health care and other social-insurance benefits for the self-employed. Second, the federal government pays a large subsidy to the system that corresponds to approximately 37 percent of benefits. Self-employed do not have access to the unemployment-insurance system, and no other special regime has been put in place to allow them to retire early. A public disability system exists, but in our opinion, it cannot be seen as an early retirement vehicle, as it is based on criteria even more stringent than those of the wage earners’ scheme. Hence, if the self-employed desire to retire early, they generally have to transit through some private retirement-income arrangement, be it a formal pension plan or simple savings. 1.2.4 Guaranteed Minimum Income The guaranteed-minimum-income pensions are fully paid for by general government revenue, and are means tested. This type of pension is only available after the legal retirement age. 1.2.5 Labor Market and Benefit-Program Participation Labor market participation rates start to decline at a rather early age. Table 1.2 illustrates the picture for men between the ages of fifty and sixtynine. Notice the very rapid decline in the labor force participation for men in their late fifties. Part-time work plays a totally marginal role in the Belgian retirement landscape. However, the Belgian government currently plans to introduce bigger incentives for people to retire progressively through a period of part-time work and part-time retirement. The corresponding panel for women shows that early retirement schemes are much less important for female workers than for males. The reasons for this finding are at least two. First, women sometimes do not fulfill some career length requirements for access to these early retirement schemes. Second, women tend to work in somewhat different jobs. Many mandatory early retirement schemes were set up in male-dominated vocations involving structural difficulties, such as heavy industry and mining. 1.2.6 Pathways to Retirement Pathways to retirement vary somewhat according to the social security system the worker is subject to. Table 1.3 summarizes the differences for the case of the wage earners and the self-employed. Unfortunately, we have been unable to separate out the different pathways for the civil servants.

50

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

Table 1.2

Labor Market and Benefit Program Participation (%) 50–54

55–59

60–64

65–69

Total

Working full time Working part time No benefit receipt Unemployment benefits Disability benefits Early retirement benefits Public retirement benefits Total

24.71 1.53 0.73 0.63 0.82 0.45 0.55 29.42

Men 15.39 0.93 1.10 0.62 1.23 3.94 1.80 25.01

6.64 0.45 0.74 0.52 1.57 6.63 7.93 24.48

1.59 0.19 0.03 0.01 0.02 1.41 17.84 21.09

48.33 3.10 2.60 1.78 3.64 12.43 28.12 100.00

Working full time Working part time No benefit receipt Unemployment benefits Disability benefits Early retirement benefits Public retirement benefits Total

9.58 3.65 11.22 1.55 0.93 0.22 1.50 28.65

Women 5.02 1.70 12.12 1.15 1.01 1.11 2.84 24.95

1.66 0.42 6.83 0.04 0.04 0.28 16.34 25.60

0.29 0.09 5.46 0.00 0.00 0.00 14.95 20.80

16.55 5.86 35.62 2.74 1.99 1.60 35.64 100.00

Source: Authors’ calculations. Table 1.3

Pathways to Retirement (%) Wage Earners

Public-Sector Employees

Self-Employed

Men Directly to SS Early retirement then SS Disability then SS Unemployment then SS Total

34.85 46.97 8.21 9.97 100.00

Directly to SS Early retirement then SS Disability then SS Unemployment then SS Total

54.85 20.02 5.25 19.88 100.00

94.97 5.03 0.00 0.00 100.00

90.02 0.00 9.98 0.00 100.00

93.82 6.18 0.00 0.00 100.00

98.18 0.00 1.82 0.00 100.00

Women

Source: Authors’ calculations. Note: Impossible to separate the disability path for public-sector employees because of missing data.

The reason for this rather disappointing fact is that our data source for the corresponding information is the income tax returns, which do not separate the type of pension income for public-sector employees. Focusing on the wage earners, we see the importance of the early retirement provisions. Furthermore, we also see the important role of the unemployment system

Micro-Modeling of Retirement in Belgium

51

that absorbs some of the mandatory early retirees, but also some of the voluntary early retirees. 1.3 Data Overview Our data set stems from five sources, which are mostly administrative databases. We were able to match all of these using an individual national identification number. Our merged data set has the big advantage of being extremely rich, as it includes data from multiple sources for a very large fraction of the Belgian population. In the following, we briefly present the various data sources, as well as their major advantages and disadvantages. 1.3.1 Statistiques Fiscales des Revenus (SFR) The data used for the SFR are originally collected by Finance Ministry. The INS (National Statistical Office) then processes the raw information to produce the SFR. Starting in 1989, the SFR data include the national identification number, thereby making 1989 the first year for which we can merge the different data sets needed. In our analysis, we focus on the years ranging from 1989 to 1996. The INS records all information relevant for the computation of individual’s tax liabilities. Variables available include wage income and income from other professional activities, household size and type, number of dependents in the household, age and income of spouse and any other dependent, social-insurance transfers and privatepension receipt, house ownership status (owner or renter), taxable real estate income, contributions to second and third pillar pensions, and so forth. 1.3.2 Comptes Individuels de Pension (CIP) The CIP is collected by the wage earners’ pension administration (ONP) since the mid-1950s. It includes all career information relevant for the wage earners’ pension computation: gross wages, days of work, days on social insurance programs, and the like. 1.3.3 Institut National d’Assurance Sociale des Travailleurs Indépendants (INASTI) The INASTI dataset includes good information on periods of affiliation as a self-employed worker. There is however no information on earnings levels, other than those available for last years from the SFR dataset. As a result, we decided to apply the fictive earnings amounts that are used by the social security administration in the benefit computation for a large part of the self-employed. 1.3.4 Finance Ministry, Department of Pensions This segment of our database contains the information on the periods of affiliation as a civil servant, as well as some information on wages during

52

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

the last year of work. Again, information on periods of affiliation allows us to complete the wage data using the SFR file for the period 1989–1996. 1.3.5 The 1991 Census We only use a very limited amount of information available in the census. Essentially, we focus on education levels that are classified according to nine categories. Doing so, we are able to match education-specific life tables to all the individuals in the database. Furthermore, we use industry indicators from the census that give information on the activity sector to which each individual belongs. The major disadvantages of our merged data set are the slightly incomplete earnings information for the self-employed and public servants, as well as the sparse information on occupational and individual privatepension arrangements. In our analysis, we analyze men and women separately. Our sample selection procedure operates in four steps departing from the original SFR file. First, we select households with at least one member in the fifty to sixty-four age range. Then, we draw a 2.5 percent random sample out of this group, which gives us a total household number of 21,818. In a third step, we match all the information from the other data sources. Finally, we eliminate all inactive people using the following definition of retirement: The retired are all those who have either pension or early retirement income and have income from work smaller than a threshold of €7,450 ($7,895), or who have unemployment or disability income and no income from work. Table 1.4 summarizes all the relevant information for our sample. The total number of observations, 23,238 and 9,707 for men and women, respectively, corresponds to three successive years—1993, 1994, and 1995—on which we focus our analysis. 1.4 Earnings Histories and Projections Our data set contains very different earnings information for periods of affiliation with the different systems. The CIP data allows us to reconstruct the complete earnings histories for wage earners. For public-sector workers, however, we only have wage information for the years from 1989 until 1996. Given the requirement that we need five years of earnings to compute public-sector pension entitlements, we do a backward projection in case there is a missing observation for one of the last five years of earnings. For the self-employed, we have very insufficient earnings information to compute wage-dependent pension entitlements. Fortunately, the social security system of the self-employed has heavily relied on fictive income figures for past years of earnings, which we use for pension computation. As for forward projections, we decided to apply wage increases so as to keep real wages constant (price-inflation adjustment) after some experi-

Micro-Modeling of Retirement in Belgium Table 1.4

53

Summary Statistics Total

No. of observations Retired (%) Age mean Married (%) Inactive spouse (%) Age difference mean Earnings mean Spouse’s earnings mean Life earnings (wage earners)

SD

Men

Women

23,238 8.6 54.9 80.6 66.4 2.7 24,017 6,163 33,207

9,707 9.9 54.2 66.1 30.7 –1.9 15,252 19,865 18,610

Women

3.7

3.5

4.0 19,758 9,990 12,540

3.9 11,901 14,004 9,238

Retiring Within the Next Year (%)

Number

Age structure 50–54 55–59 60–65 Social security program Wage earner Self-employed Civil servant Region Brussels Flanders Wallonia

Men

Men

Women

Men

Women

11,938 8,200 3,100

5,664 3,149 894

3.1 9.8 26.3

5.3 9.0 41.3

13,135 3,984 6,119

5,242 1,080 3,385

9.6 5.0 8.6

11.0 7.5 8.8

1,850 14,715 6,673

1,330 5,197 3,180

9.1 8.6 8.4

8.7 10.4 9.4

Note: Observations correspond to person-year cells. SD = standard deviation. Blank cells indicate that data is not applicable (i.e., SD does not apply).

mentation with other wage regression models. In particular, it has the advantage of allowing for reasonable projections for people belonging to multiple systems. 1.5 Construction of Incentive Measures To measure the impact of the social security systems’ incentives we use several different indicators. A first one is the concept of social security wealth (SSW), which is the present discounted value of all future benefit flows from a given social security system. Discounting is done allowing both for time preference and mortality adjustments. Mortality adjustments are based on education-specific life tables, as computed by Deboosere and Gadeyne (2000) based on the 1991 Census and population registers. Depending on the household situation and the system, SSW also

54

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

includes an element that is a function of dependent or survivor benefits. Furthermore, SSW also has to allow for the possibility of people being subject to different retirement-income systems. In the Belgian context, private pensions are not integrated as they only play a minor role in the pension landscape. However, the case of people having pension entitlements in two or even three of the public social security programs is not rare. We apply the official rules that exist for cumulating benefits from the three main public systems. We compute this SSW measure for every pathway toward retirement that is accessible to the individual. After the construction of these SSW –figures, we then proceed in a second step to the computation of weighted SSW indicator, which sums the previously derived path-specific components. We attach age-specific weights to the different retirement paths, as described in table 1.5. The weights on the early retirement and the unemployment and disability routes correspond to the sum of observed frequencies of these routes among all people of any given age up to age sixty-five, and the public retirement system takes the residual weight. For wage earners, we add the unemployment-insurance and disabilityinsurance paths as the two systems produce very similar benefit structures. Doing so, we give an upper bound on incentives for people to retire, as we render all of disability voluntary. Given the lack of information for the public sector, we consider as early retirees all people retiring before the age of sixty. As for the self-employed, we only allow for one path in our computation as the disability system provides quite low levels of benefits with very strict conditions. Furthermore, it would be very difficult for us to compute any reasonable amount of disability benefits due to the lack of good earnings information that we mentioned earlier. A last important remark relates to retirement benefits of two-worker couples. It sometimes happens that the dependent benefit of a spouse is larger than the benefits based on the individual’s own work history. In that case, we apply the official rule of supplementing the pension based on the personal earnings history by the difference between the potential dependent benefits of the spouse and the pension for work. Furthermore, the SSW measure that will be used hereafter includes both the worker’s and the spouse’s potential SSW, independent of whether or not the individual continues to work.7 Based on this weighted SSW, we then compute different secondary incentive measures. A first one is the concept of accrual, which simply represents the variation in SSW that is obtained by retiring next year rather than 7. We assume that the spouse retires on a standard retirement path (using the previously defined weights) as soon as the spouse is entitled to access any one of the three main public retirement programs.

Micro-Modeling of Retirement in Belgium Table 1.5

Weight of the Different Pathways to Retirement (by age) Wage Earners Public Retirement

Early Retirement

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

0.335 0.351 0.362 0.382 0.401 0.417 0.456 0.502 0.605 0.757 0.807 0.844 0.903 0.970 1.000 1.000

0.479 0.483 0.485 0.479 0.473 0.472 0.442 0.415 0.331 0.185 0.145 0.112 0.065 0.018 0.000 0.000

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

0.526 0.570 0.586 0.623 0.658 0.700 0.756 0.835 0.924 0.995 1.000 1.000 1.000 1.000 1.000 1.000

0.210 0.214 0.219 0.201 0.176 0.169 0.143 0.122 0.048 0.000 0.000 0.000 0.000 0.000 0.000 0.000

Age

55

Civil Servants Unemployment/ Disability

Public Retirement

Early Retirement

0.186 0.166 0.153 0.138 0.126 0.111 0.102 0.083 0.064 0.058 0.048 0.044 0.032 0.012 0.000 0.000

0.948 0.948 0.949 0.953 0.952 0.956 0.963 0.968 0.977 0.982 0.990 1.000 1.000 1.000 1.000 1.000

0.052 0.052 0.051 0.047 0.048 0.044 0.037 0.032 0.023 0.018 0.010 0.000 0.000 0.000 0.000 0.000

Women 0.264 0.216 0.195 0.176 0.166 0.131 0.101 0.043 0.028 0.005 0.000 0.000 0.000 0.000 0.000 0.000

0.937 0.947 0.950 0.961 0.965 0.973 0.984 0.991 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000

0.063 0.053 0.050 0.039 0.035 0.027 0.016 0.009 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

Men

the present. Table 1.6 gives the summary statistics we obtain by applying this first incentive measure to the entire sample of individuals belonging to the different social security systems. At all ages from fifty to sixty, we find that the value of the SSW median is always higher than $200,000. For accruals, notice the large spread between the values we obtain for people at the tenth and the ninetieth percentile of the distribution. Another feature that our tabulation reveals is that once men and women attain age sixty, more than 90 percent of them

56

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

Table 1.6

Social Security Wealth and Accrual Accrual

Age

Obs.

SSW Median

10th %

Median

90th %

SD

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

2,821 2,586 2,305 2,150 2,076 2,045 1,894 1,701 1,409 1,151 1,027 694 473 375 302 229

248,352 250,816 250,300 250,413 251,578 255,997 255,076 258,560 255,560 258,025 259,078 240,421 207,874 199,347 199,683 203,135

Men –10,441 –9,579 –10,465 –9,944 –9,285 –12,164 –10,868 –14,785 –13,758 –6,919 –18,113 –18,405 –18,895 –18,148 –17,999 –20,918

–5,722 –4,878 –5,232 –5,005 –1,052 –4,990 –4,163 –3,917 –4,228 –658 –9,673 –8,995 –8,382 –7,660 –8,051 –12,376

4,905 5,021 4,925 5,404 6,722 5,168 4,745 4,954 5,155 5,864 –167 –187 –205 –301 –374 –6,596

8,517 8,006 7,746 7,600 8,213 8,481 7,665 8,363 7,726 6,050 6,868 7,083 7,577 7,101 7,421 6,318

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

1,464 1,273 1,118 940 869 824 747 621 514 443 391 227 97 78 57 44

227,610 231,872 232,794 233,461 229,748 229,591 230,956 231,328 234,003 232,094 232,243 217,335 207,054 197,000 172,528 143,350

Women –9,380 –5,528 –7,348 –6,706 –6,223 –7,276 –8,408 –9,494 –8,755 –9,027 –18,179 –17,306 –17,430 –17,753 –14,555 –15,120

–2,775 –1,028 –1,824 –1,354 –1,268 –1,707 –1,855 –1,789 –989 –514 –11,053 –10,448 –10,319 –9,778 –8,455 –8,899

4,418 5,231 4,768 5,403 5,228 5,000 4,846 5,235 5,512 6,109 –6,473 –6,213 –5,928 –5,361 –4,992 –3,434

5,889 5,533 5,560 5,455 5,112 5,261 5,461 6,201 5,760 6,089 5,616 4,896 6,694 5,242 4,553 4,259

Note: Obs. = observations; SD = standard deviation.

face negative accruals. These results are essentially due to the fact that, under the rules outlined in section 1.2, workers are hardly penalized for retirement at age sixty rather than sixty-five. Also, notice the effect of the early retirement systems that are made generally available to workers at key ages, such as fifty-two, fifty-five, and fifty-eight. For men, we observe quite substantial negative accruals at very early ages because of the availability of very generous early retirement provisions. For women, the situation is slightly different. First of all, the value of the accrual for women is much

Micro-Modeling of Retirement in Belgium Table 1.7

57

SSW, Accrual, and Tax or Subsidy for Male Wage Earners SSW

Accrual

Tax or Subsidy

Age

Obs.

Median

10th %

Median

90th %

SD

Median

Pestieau-Stijns

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

1,708 1,552 1,397 1,318 1,253 1,212 1,083 953 750 562 464 336 193 143 116 95

269,845 269,298 267,181 266,273 265,461 268,248 266,265 268,312 265,781 266,845 271,955 269,417 257,141 247,500 232,136 224,505

–15,009 –13,339 –13,443 –13,969 –12,251 –17,706 –15,479 –17,663 –15,762 –8,237 –17,832 –18,631 –19,053 –18,069 –18,236 –19,964

–8,064 –7,262 –7,665 –7,248 –6,182 –8,559 –8,061 –9,486 –10,026 –2,660 –10,776 –11,061 –10,660 –11,256 –9,750 –12,425

–4,204 –3,310 –2,852 –2,678 8,467 –3,449 –2,659 –1,939 –3,985 241 –6,018 –6,225 –4,070 –5,596 –5,057 –7,979

7,967 7,642 7,070 6,754 9,206 7,823 6,976 7,382 5,209 5,737 5,402 5,812 6,073 5,372 5,738 5,134

0.362 0.321 0.354 0.337 0.279 0.410 0.396 0.474 0.489 0.126 0.498 0.556 0.523 0.526 0.499 0.585

n.a. n.a. n.a. n.a. n.a. 0.821 0.809 0.789 0.771 0.811 0.496 0.497 0.491 0.489 0.473 0.529

Source: Pestieau and Stijns (1999, table 1.9) and authors’ calculations. Note: Obs. = observations; n.a. = not available.

less negative for ages up to fifty-nine than the values observed for men. Women usually have shorter periods of affiliation with the social security systems and thus have more limited access to these early retirement schemes. Furthermore, the change in accruals is much more pronounced at the ninetieth percentile for women than for men. We see two broad reasons for this finding. First of all, for single women with a rather complete earnings history, the same logic applies as the one we already saw for men. For married women however, the husband’s earnings history plays an important role in the determination of the value of the wife’s accrual. To illustrate this point, it is easiest to use a simple example. Consider a couple in which the husband is still working when the wife turns sixty. Suppose, rather plausibly, that the wife has an incomplete and low-income earnings history. This woman will face a large negative accrual, as she knows that when her husband retires, she will give up her own pension entitlement to claim spousal benefits based on her husband’s earnings history. Therefore, an additional year of work implies a net loss both in terms of pension income (based on her earnings record) and in terms of further social security contributions that will in the end not affect her pension entitlement. The findings are very similar once we restrict our attention to the wage earners’ scheme, as table 1.7 indicates. Approximately 90 percent of the population face negative accruals starting at age fifty. Under the rules of the unemployment or early retirement systems for wage earners, fictive

58

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

income is imputed to the earnings histories of workers on these types of replacement income. This way, the pensions payable to a large number of low- and medium-income workers under the wage earners’ scheme are almost immune to a decision to retire early. The same finding can also be represented using a different incentive measure, namely the implicit tax or subsidy rate imposed by the social security system. This tax or subsidy rate is defined as being the negative of the accrual divided by the potential income during the next year.8 To allow a comparison with the results of the previous study by Pestieau and Stijns (1999), we restrict our attention to the subsample of male wage earners. The simulated numbers of these authors match up pretty nicely with the ones we derive for our sample from age sixty upward. At earlier ages, the difference is rather substantial. This result is due to the different weighting of the pathways into retirement as we use a weight on the early retirement route that is smaller than the 100 percent used by these authors. The next two incentive measures are forward-looking measures. “Peak value” represents the difference between SSW at its peak and SSW at present. It thus differentiates itself from the accrual measure by the fact that it takes into account the entire SSW process, not only the variation from the present to the next period.9 The second forward-looking measure is the concept of “option value” as defined by Stock and Wise (1990), which is based on a utility-maximization framework. The utility function Vt underlying the computation of the option value process can be summarized by the following mathematical expression: r1

S

s
t

s
r

Vt (r)
∑ stY s  ∑ st [kBs(r)], where the first expression on the right-hand side represents the utility derived from labor income Y, and the second expression represents utility derived from retirement income Bs(r);10  is the time-preference rate for which we assume a discount rate of 3 percent, which corresponds to a value of  of approximately .97. The variable  corresponds to a parameter of risk aversion and is set to 
0.75. Finally, k
1.5 expresses the relative weight of utility of retirement income as compared to wage income. The concept of option value Gt (r ∗) is then defined as the difference in utility terms between retiring at the best point in the future (r∗) and now (t). Gt (r ∗)
Vt (r ∗)  Vt (t) 8. In line with the Belgian social insurance legislation, we apply a somewhat different projection mechanism for this income figure than for the one used in the pension computation formula that we discussed in section 1.4. 9. Peak value is equal to accrual if the peak of the SSW process is attained with immediate retirement. 10. The value of the benefit corresponds to the weighted sum of retirement income using the weights of table 1.5.

Micro-Modeling of Retirement in Belgium

59

Summary statistics for both of these forward-looking measures are given in table 1.8. The peak value numbers of table 1.8 for men and women older than fifty-nine strongly resemble those we discussed for the accrual definition in table 1.6. Even at lower ages, we find that these concepts only differ at the top of the distribution. The cause for this finding is the pressure built into the Belgian retirement systems to retire as early as possible. The peak of the SSW variable is often attained by retiring as soon as possible, hence bringing the two incentive measures to equality. For a nonnegligible frac-

Table 1.8

Peak Values and Option Values Peak Value

Age

Obs.

10th %

Median

90th %

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

2,821 2,586 2,305 2,150 2,076 2,045 1,894 1,701 1,409 1,151 1,027 694 473 375 302 229

–10,408 –9,576 –10,433 –9,817 –9,218 –12,164 –10,868 –14,785 –13,757 –6,919 –18,113 –18,405 –18,816 –18,148 –17,999 –20,918

–5,677 –4,749 –4,911 –3,186 –499 –4,761 –4,017 –3,890 –4,204 –646 –9,633 –8,995 –8,348 –7,660 –8,051 –12,376

19,975 19,790 18,842 18,515 17,802 15,790 13,694 12,698 10,144 6,054 –167 –170 –205 –301 –374 –6,596

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

1,464 1,273 1,118 940 869 824 747 621 514 443 391 227 97 78 57 44

–9,327 –5,500 –7,324 –6,641 –6,219 –7,276 –8,408 –9,494 –8,755 –9,027 –18,179 –17,306 –17,430 –17,753 –14,555 –15,120

3,836 4,956 3,557 3,720 3,062 2,009 1,592 1,175 1,132 –12 –10,890 –10,217 –10,315 –9,753 –8,455 –8,899

Option Value SD

10th %

Median

90th %

SD

Men 14,717 14,299 13,869 13,388 12,301 12,676 11,331 11,596 10,066 6,357 7,293 7,243 7,527 7,120 7,421 6,318

2,256 2,192 1,903 1,701 1,463 975 919 674 654 710 –812 –929 –913 –872 –912 –1,037

9,917 9,291 8,186 7,065 6,053 5,117 4,478 3,722 3,106 2,191 644 656 759 847 483 –215

22,959 21,681 19,915 17,602 15,650 13,535 12,954 11,466 10,805 9,597 8,608 10,096 13,011 12,927 8,583 7,253

12,014 10,801 10,498 9,539 8,976 9,318 9,002 9,048 7,905 7,692 8,371 9,755 8,360 7,528 5,918 5,748

Women 26,321 15,449 25,590 14,209 25,138 13,817 24,703 13,216 22,974 12,542 20,097 11,917 17,510 11,269 14,502 10,349 12,123 8,716 6,383 7,883 –6,294 6,453 –6,213 5,754 –5,928 6,700 –5,361 5,293 –4,922 4,553 –3,434 4,259

1,459 1,453 989 956 805 500 233 –137 –222 –324 –1,532 –1,466 –1,264 –1,126 –681 –728

8,730 8,154 7,389 6,746 6,077 5,430 4,661 3,764 2,959 1,959 284 507 1,125 761 1,323 1,561

17,333 16,145 14,897 13,437 12,279 10,739 9,198 7,460 5,530 4,016 2,120 1,920 2,998 2,851 4,152 4,167

7,267 6,336 5,789 5,499 5,268 5,507 4,968 4,236 3,359 3,536 2,886 4,920 5,422 2,762 3,826 5,418

Note: Obs. = observations; SD = standard deviation.

60

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

tion of the population younger than sixty, there are quite large possible gains from continuing to work. Hence, we find rather large standard deviations in the peak value indicator. As for the option value statistics, it is important to recognize the major role played by the utility term that is based on wage income during any additional period of work. As a result, most values are positive. The same qualitative results still hold true if we do some sensitivity analysis by replacing either the value of k by 1, the value of  by 1, or both. 1.6 Regression Results The present section summarizes the regression results we obtained while performing probit estimations with three of the above incentive measures—namely accrual, peak value, and option value. The dependent variable is retirement; it equals one in the case that the individual retires within the year of observation, and zero otherwise. As indicated in section 1.3, we define retired persons as those who have either a pension or early retirement income and have income from work smaller than the threshold, or who have unemployment or disability income and no income from work. All estimations include an intercept term, as well as a differing series of controls. The controls include demographic variables (marital status, a dummy for an active spouse, a dummy variable for dependent children, and the age difference between the individual and the partner). Furthermore, age is inserted for some specifications in the form of a dummy variable for each age, and for other specifications in the form of a linear age variable. Earnings appear in three ways: individual projected annual earnings and average lifetime income (only available for wage earners); projected potential spousal annual earnings (all these variables are in dollars, converted at the exchange rate on 31 December 1999), and lastly, system variables are also used in all the specifications. Furthermore, all of these models contain controls for activity sector (ten dummies), level of schooling (eight dummies), squared earnings and lifetime earnings, an occupational pension dummy, a private retirement savings dummy, a home ownership dummy, real estate income, regions (Brussels, Flanders, or Wallonia), and dummies for the year of the observation (1993, 1994, or 1995). The estimates of them are not reported in the tables for space reasons.11 Tables 1.9 and 1.10 summarize our regression results and are organized as follows: We estimated a total of six probit models, separate for men and women. The six models are the combination of our three dynamic incentive measures with two different specifications of the age variable. The first column for every incentive measure reports the results of a probit model 11. Complete results can be obtained from the authors upon request.

0.0003

–0.0007 (–0.0069) –0.0442 (–0.4590)

0.0039 0.0498 0.0393 0.0039 0.0360

0.0066 0.0012 0.0024

0.1105 0.0870 –0.0525 0.0002 –0.0964

0.0129 –0.0085 0.0036

0.0018

0.2675

SE

–7.9660

Coef.

Age

0.0143 –0.0085 0.0039

0.0886 –0.0530 0.0005 –0.0873

–0.0008 (–0.0076) –0.0428 (–0.4307)

–2.5199

Coef.

0.0068 0.0012 0.0024

0.1049 0.0503 0.0397 0.0039 0.0364

0.0019

0.0003

0.1690

SE

Age Dummies

Accrual

Retirement Probits for Men

Demographic variables Age Married Active spouse Age difference Dependent Income earnings variables Life cycle earnings Earnings (1000s) Spouse earnings (1000s) (continued )

AC, PV, OV (1000s)

Intercept Incentive measures SSW (1000s)

Table 1.9

0.0138 –0.0084 0.0043

0.0039 0.1027 –0.0501 –0.0003 –0.0999

–0.0008 (–0.0079) –0.0380 (–0.3769)

–7.5957

Coef.

Age

0.0066 0.0012 0.0024

0.0499 0.0394 0.0039 0.0361

0.0016

0.0003

0.2698

SE

0.0151 –0.0084 0.0044

0.1050 0.1026 –0.0515 –0.0001 –0.0911

–0.0008 (–0.0082) –0.0364 (–0.3505)

–2.4346

Coef.

0.0068 0.0012 0.0024

0.0053 0.0505 0.0399 0.0039 0.0364

0.0017

0.0003

0.1690

SE

Age Dummies

Peak Value

0.0155 0.0014 0.0037

0.0294 –0.0857 –0.0009 –0.0729

–0.0001 (–0.0008) –0.0392 (–0.4111)

–7.5110

Coef.

Age

0.0066 0.0018 0.0024

0.0488 0.0383 0.0038 0.0352

0.0054

0.0003

0.3387

SE

0.0180 –0.0002 0.0031

0.0342 –0.0839 –0.0015 –0.0754

–0.0001 (0.0001) –0.0327 (–0.3383)

–2.2976

Coef.

0.0068 0.0018 0.0024

0.0495 0.0388 0.0039 0.0355

0.0054

0.0003

0.1709

SE

Age Dummies

Option Value

0.5489 0.0023

Age

0.1913

0.1251 0.1237

— — — — — — — — — — — — — — —

SE

0.5399 0.0164

0.0619 0.2851 0.3115 0.4730 0.6921 0.6771 0.8893 1.0551 0.8427 1.5016 1.4963 1.0288 1.0927 0.9724 1.9490

Coef.

0.2076

0.1270 0.1255

0.0826 0.0778 0.0787 0.0773 0.0720 0.0741 0.0724 0.0735 0.0838 0.0743 0.0804 0.0956 0.1021 0.1126 0.1081

SE

Age Dummies

Accrual

0.6314 –0.0328

— — — — — — — — — — — — — — —

Coef.

Age

0.1901

0.1258 0.1235

— — — — — — — — — — — — — — —

SE

0.6177 –0.0216

0.0523 0.2779 0.3191 0.4516 0.6751 0.6429 0.8546 1.0048 0.7310 1.4439 1.4396 0.9767 1.0380 0.9208 1.9180

Coef.

0.2067

0.1277 0.1254

0.0827 0.0779 0.0788 0.0773 0.0720 0.0740 0.0723 0.0734 0.0833 0.0745 0.0806 0.0958 0.1023 0.1129 0.1083

SE

Age Dummies

Peak Value

0.3850 –0.0408

— — — — — — — — — — — — — — —

Coef.

Age

0.1512

0.1254 0.1244

— — — — — — — — — — — — — — —

SE

0.3858 –0.0396

0.0134 0.2289 0.2087 0.2871 0.5664 0.5003 0.7359 0.8659 0.4475 1.4456 1.4302 0.9868 1.0364 0.9320 1.9801

Coef.

0.1741

0.1281 0.1271

0.0808 0.0763 0.0773 0.0760 0.0718 0.0745 0.0738 0.0760 0.0869 0.0816 0.0876 0.1028 0.1097 0.1198 0.1191

SE

Age Dummies

Option Value

Notes: Coef. = coefficient; SE = standard error. Dashes indicate variables not included in the model. The probability effect, which appears in parentheses below the coefficient, is the expected change in the probability of retirement corresponding to an infinitesimal change in the selected independent variable.

Pseudo R 2

Civil servant Self-employed

— — — — — — — — — — — — — — —

Coef.

(continued)

Age and schemes dummies 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

Table 1.9

0.0003

–0.0005 (–0.0068) –0.0540 (–0.7109)

0.0059 0.0714 0.0559 0.0063 0.0585

0.0106 0.0027 0.0021

0.0877 0.1994 0.0174 0.0202 –0.1540

0.0077 –0.0098 –0.0027

0.0034

0.4077

SE

–6.3141

Coef.

Age

0.0056 –0.0088 –0.0021

0.1854 –0.0154 0.0146 –0.1627

–0.0002 (–0.0021) –0.0409 (–0.5345)

–1.9632

Coef.

0.0107 0.0027 0.0021

0.0727 0.0572 0.0064 0.0589

0.0039

0.0004

0.2505

SE

Age Dummies

Accrual

Retirement Probits for Women

Demographic variables Age Married Active spouse Age difference Dependent Income earnings variables Life cycle earnings Earnings (1000s) Spouse earnings (1000s) (continued )

AC, PV, OV (1000s)

Intercept Incentive measures SSW (1000s)

Table 1.10

0.0081 –0.0090 –0.0021

0.0887 0.2222 –0.0367 0.0206 –0.1539

–0.0003 (–0.0042) –0.0307 (–0.3940)

–6.2897

Coef.

Age

0.0104 0.0027 0.0020

0.0060 0.0708 0.0552 0.0063 0.0586

0.0023

0.0003

0.4100

SE

0.0062 –0.0083 –0.0017

0.2030 –0.0606 0.0145 –0.1687

–0.0001 (–0.0001) –0.0222 (–0.2868)

–1.8504

Coef.

0.0106 0.0027 0.0020

0.0641 0.0725 0.0566 0.0064 0.0590

0.0024

0.0004

0.2497

SE

Age Dummies

Peak Value

0.0069 0.0125 –0.0011

0.0080 0.2269 –0.0784 0.0288 –0.1346

–0.0007 (–0.0089) –0.0793 (–1.0341)

–4.7976

Coef.

Age

0.0106 0.0038 0.0020

0.0709 0.0544 0.0066 0.0580

0.0089

0.0004

0.5161

SE

0.0053 0.0102 –0.0010

0.2079 –0.0920 0.0226 –0.1575

–0.0004 (–0.0048) –0.0651 (–0.8434)

–1.5951

Coef.

0.0107 0.0039 0.0021

0.0729 0.0560 0.0068 0.0587

0.0091

0.0004

0.2554

SE

Age Dummies

Option Value

Notes: See table 1.9.

Pseudo R2

Civil servant Self-employed

Age

0.2478 –0.1986

— — — — — — — — — — — — — — —

Coef.

(continued)

Age and schemes dummies 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65

Table 1.10

0.1644

0.1228 0.1306

— — — — — — — — — — — — — — —

SE

0.1748 –0.2326

0.0512 0.2775 0.3783 0.2818 0.5650 0.4214 0.5806 0.4396 0.3560 1.4321 1.6353 0.4892 0.4537 1.0051 1.3046

Coef.

0.1948

0.1247 0.1327

0.0924 0.0877 0.0898 0.0955 0.0890 0.0944 0.0955 0.1066 0.1175 0.0973 0.1129 0.1813 0.1993 0.1992 0.2113

SE

Age Dummies

Accrual

0.2851 –0.2508

— — — — — — — — — — — — — — —

Coef.

Age

0.1536

0.1224 0.1292

— — — — — — — — — — — — — — —

SE

0.2128 –0.2766

0.0094 0.2391 0.3247 0.2192 0.4942 0.3489 0.5099 0.3334 0.2208 1.4579 1.6646 0.5275 0.4805 0.9965 1.3057

Coef.

0.1918

0.1248 0.1319

0.0923 0.0877 0.0897 0.0951 0.0888 0.0942 0.0951 0.1060 0.1161 0.0973 0.1124 0.1795 0.1976 0.1988 0.2116

SE

Age Dummies

Peak Value

0.1541 –0.2041

— — — — — — — — — — — — — — —

Coef.

Age

0.1365

0.1228 0.1300

— — — — — — — — — — — — — — —

SE

0.1287 –0.2281

–0.0462 0.1767 0.2195 0.0871 0.3306 0.1500 0.2681 0.0456 –0.1246 1.2533 1.4447 0.2841 0.2177 0.6746 1.0055

Coef.

0.1860

0.1252 0.1330

0.0922 0.0876 0.0899 0.0955 0.0906 0.0973 0.1001 0.1128 0.1261 0.1123 0.1243 0.1881 0.2072 0.2108 0.2218

SE

Age Dummies

Option Value

Micro-Modeling of Retirement in Belgium

65

with a linear age trend. The second column then reports the results of a model in which we replace the linear age trend by age dummies. The motivation for this change is to allow for nonlinearities in the systems that our incentive measures do not fully capture. Inspection of the different columns of table 1.9 reveals that the incentive measures are significant when taken individually. Comparing the estimates from the different specifications, we see that this result is pretty robust as these estimates do not depend on the precise functional form of the specification. The SSW has a small negative effect on the probability of retirement. The numbers reported in parentheses indicate the change in the underlying probability function as a result of a small change in the incentive variable. The three dynamic incentive variables, accrual (AC), peak value (PV), and option value (OV), have a strong negative effect on the probability of retiring, as reported in parentheses below the parameter estimates. Also notice the positive effect of the civil-servant-system dummy, which contrasts with a generally insignificant self-employment dummy. Table 1.10 reveals that the dynamic incentive variables (AC, PV, and OV) also display a large degree of significance for women and an even stronger probability effect than for men. However, our estimates seem to indicate that the level of SSW does not have a lot of explanatory power in women’s retirement decisions. Furthermore, the significance of the system dummies is very different from the findings for men. The civil servants dummy is never significant at any conventional level, and the self-employment dummy is always negative, but rarely with a high degree of significance. This seems to indicate that self-employed women represent a somewhat special group that is more reluctant to retire, or alternatively, this finding may be due to the fact that we do not fully capture all the characteristics of the system. As expected, age variables have a strong effect on retirement probabilities, either in the form of continuous or dummy variables. However, age dummies at ages smaller than sixty seem to be much stronger for men than for women, particularly so at the key early retirement ages (fifty-two, fiftyfive, and fifty-eight years old). Women—given their generally incomplete earnings histories—are often simply not eligible for some or all of the early retirement benefits. Alternatively, it may reflect the fact that women’s incomes generally represent a smaller fraction of household resources and hence that women’s behavior is strongly influenced by the decisions that their husbands make. Looking at the pseudo R 2 of the different probit models, we notice that those models that include age dummies uniformly perform better than those simply integrating a linear age trend. However, regarding the question of which dynamic incentive variable has the highest explanatory power, we find that the accrual measure performs best for both men and women.

66

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

At this stage, it is important to check whether our incentive measures (SSW, AC, PV, and OV) capture the entire impact of the numerous benefit provisions in the different retirement income systems available in Belgium. Particularly, and somewhat surprisingly, the replacement of the linear age trend by a dummy-variable model does not seem to have a major impact on the sign, the value, or the significance of the estimates of the incentive variables SSW, AC, PV, and OV for men. The story is slightly different for females as we can observe in table 1.10 where the probability effect of the accrual variables changes slightly more. We interpret these findings as rather comforting as they tend to indicate that our SSW and dynamic incentive variables have a rather stable explanatory power and hence, they capture some of the nonlinearities of the retirement income systems in Belgium. Coming back to the other results presented in tables 1.9 and 1.10, a finding common to all estimations for men and women is that the presence of an active spouse does not have a significant impact on the retirement probability. However, the presence of any additional dependent in the family has a significant negative impact on retirement decisions. Being married also has a positive effect on the probability of retirement, with a particularly high level of significance for women. A similar strongly positive effect is observable with respect to the age difference between spouses, but again only for women.12 The effect of the earnings variables is rather similar for men and women. First of all, average life-cycle earnings (a variable only available for wage earners) positively influence the retirement decision, but the coefficients are not statistically significant in the case of women. On the contrary, projected earnings have, as expected, a negative and significant influence upon retirement when the dynamic incentive variables are AC and PV. In the OV models, however, this effect vanishes naturally, as projected earnings enter directly in the calculation of the incentive measure. Finally, the projected spouse earnings appear to have no impact on individual decisions for either men or women. Figures 1.1 and 1.2 plot the observed hazard rate of departures into retirement by sex and by age on the same scale as the departure probabilities implied by age effects of the age-dummy regression models. One important result is that the dummy effects follow the changes in the empirical hazards very closely. This tends to indicate that our incentive models only explain a fraction of the retirement process and that the dummies are a good complementary tool for capturing some of the nonlinearities that our general SSW computation cannot absorb. One plausible explanation for this finding is that many of the retirement decisions taken in the Belgian companies and public administrations can be seen as mandatory for individuals, while 12. As indicated in table 1.4, the average age difference between women and men within couples is –1.9 years.

Fig. 1.1

The retirement hazard and age dummies for men

Fig. 1.2

The retirement hazard and age dummies for women

Micro-Modeling of Retirement in Belgium

69

they can be seen as collectively voluntary, as trade unions intervene in the negotiations relating to many early retirement schemes. Therefore, incentive measures have only a very limited role to play in the decision on when to stop working. Our findings also back the previously discussed result that the explanatory power of the accrual model is highest for men, while option value does best for women. 1.7 Simulations We present the results of some simulation exercises to better understand the results we found in the preceding section. We focus our attention on two hypothetical policy reforms. The first one consists of an increase by three years of all the key ages in the various retirement and early retirement systems. Even though there is no clear eligibility age for unemployment, sickness, or disability insurance, we suppose that these programs become available three years later than we supposed in the original setup. The second reform consists of a policy that would harmonize the retirement-income systems in all the countries covered in the present collection. This latter “common reform” replaces the myriad of current retirement and early retirement systems by a uniform and unique retirement system. Early retirement would be possible at age sixty at the earliest, while the normal retirement age would be sixty-five. At age sixty-five, every individual would be entitled to retirement income that corresponds to a replacement rate of 60 percent of the pension with respect to the average income over the five years of income between ages fifty-five and fifty-nine. For retirement prior to age sixty-five, a 6 percent actuarial reduction is applied to the amount of benefit entitlement on a lifetime basis. Similarly, late retirement (i.e., after age sixty-five) is rewarded by a lifetime increase in benefits of 6 percent per year of delay. For this second policy simulation, we suppose that there is no unemployment, sickness, or disability retirement pathways available to the individual. For each of these two policy changes, we use three different methodologies and apply those to the three model specifications (AC, PV, and OV) that we derived in the previous section. Hence, we perform a total of eighteen simulations both for men and for women. The first simulation approach, S1, uses the estimates from the models with a linear age trend. Given that our incentive measures SSW and AC, PV, and OV are all derived using age-specific weights, we apply the weights of section 1.5 on age a to the incentive measures at age a  3. Hence, expressed a little differently, we suppose that the age-specific probabilities of replacement-income receipt are shifted up by three years. This change in the weighting will also have implications on retirement through the incentive-variable coefficients. Notice that this change in weighting only matters for the first policy change, as by definition, we impose the absence of any other pathway to retirement in the second policy. In summary, S1 simply consists of a recom-

70

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

putation of SSW at every age under these new rules and a prediction of retirement rates by the application of the new SSW and AC values to our estimated coefficients. The second simulation approach, S2, is the same as the first, except that it uses the model with age dummies included. The impact of this change in the modeling of age does not have a major impact on the coefficients of the incentive variables. However, the age-dummy effects are far from linear, and hence it is possible that these dummies better pick up the nonlinearities in the various retirement and early retirement systems, or alternatively that tastes for leisure are not a linear function of age. Also, as we already mentioned in a preceding section, it is possible that a nonnegligible fraction of retirees are facing compulsory retirement at a given age. However, a comforting finding is that the incentive effects seem to be pretty robust to the change in the modeling of age. This seems to indicate that the dummies do not simply take away explanatory power from the incentive variables, but rather contribute new information of their own to the analysis of the variance. This second simulation approach implicitly privileges the explanation of different leisure tastes over the explanation that the SSW and the related incentive variables do not fully capture the nonlinearities in the system. A third and last simulation approach is labeled “S3.” For policy 1, it uses the model with age dummies, but on top of incrementing the incentive and SSW measures and the eligibility probabilities, it also increments the age dummies by three years. We do not only recompute the values of our SSW measures arising from the change, but we also recompute the value of the age dummies themselves, so that the age-fifty-three dummy takes on the estimated value of the age-fifty dummy, the age-sixty-eight dummy takes on the estimated value of the age-sixty-five dummy, and so forth. This approach takes a view rather opposed to S2 as it implicitly imputes the entire effect on the age dummies to the social security incentives. Clearly, the truth will be between these two extremes. For policy 2 and S3, we proceed in a similar way, but the impact of age dummies are modified in a different way. On the one hand, given that in this policy simulation alternative retirement pathways are assumed out, we apply the age-fifty-one dummy for both men and women to all ages up to age fifty-nine, just prior to the early retirement age.13 On the other hand, 13. The method applied differs from the one used in the other articles of the volume. In those papers, the effects of age dummies before age sixty are imputed following the trend of estimated dummies from age fifty to fifty-four (the age at which the path first breaks in the dummies) whenever this trend is positive. However, this would lead us to unrealistic results. The observed age trend in dummies between the ages of fifty and fifty-four—and hence the extrapolation thereof till the age of fifty-nine—would be highly progressive and extremely powerful, it would occur in a situation in which we have assumed away all early retirement schemes. The reason for this extreme finding is that even people in the lowest age bracket considered face high exit rates from the labor force because of the characteristics of the Belgian retirement income landscape.

Micro-Modeling of Retirement in Belgium Table 1.11

71

Average Retirement Age Simulated Reform

Model

Simulation

Policy 1

Policy 2

S1 S2 S3

58.38 59.49 58.78 61.18

58.38 59.22 59.08 59.11

PV

S1 S2 S3

59.72 59.01 61.34

59.15 59.05 59.05

OV

S1 S2 S3

59.89 59.18 61.47

59.14 59.06 59.34

57.43 58.59 58.18 59.99

57.43 57.33 57.36 57.51

Men Base retirement rate AC

Women Base retirement rate AC

S1 S2 S3

we keep the effect of age-sixty and -sixty-five dummies unchanged, assuming that policy 2 will not affect individual behavior at these particular ages. Finally, using these two dummy values, we imputed the values of the intermediary dummies, from age sixty-one to age sixty-four assuming a smooth path trend. We present the results of these simulations in three broad ways. First of all table 1.11 summarizes the effect on the average retirement age of the three models (AC, PV, and OV) and the three simulations (S1, S2, and S3) for the two policy changes discussed (policy 1 and policy 2). In an attempt not to overcrowd the paper with tables and figures, only the results corresponding to the accrual model for women are reported.14 Second, figures 1.3 to 1.14, panels A, illustrate the hazard rates of departure into retirement under the different specifications as compared to the baseline observed hazard using the underlying specification. The first nine of these graphs summarize the results of the nine simulations for men. The last three figures (figures 1.12, 1.13, and 1.14) are the results of the simulations done for women using the accrual model. Third, figures 1.3 to 1.14, panels B, present the cumulative distribution function (CDF) of departures into retirement. 14. From table 1.10, it appears that the estimation of the accrual model gives the best pseudo R 2.

Fig. 1.3

A

AC, S1: A, Hazard rates for men; B, CDF for men

Fig. 1.3

B

(cont.)

Fig. 1.4

A

AC, S2: A, hazard rates for men; B, CDF for men

Fig. 1.4

B

(cont.)

Fig. 1.5

A

AC, S3: A, Hazard rates for men; B, CDF for men

Fig. 1.5

B

(cont.)

Fig. 1.6

A

PV, S1: A, Hazard rates for men; B, CDF for men

Fig. 1.6

B

(cont.)

Fig. 1.7

A

PV, S2: A, Hazard rates for men; B, CDF for men

Fig. 1.7

B

(cont.)

Fig. 1.8

A

PV, S3: A, Hazard rates for men; B, CDF for men

Fig. 1.8

B

(cont.)

Fig. 1.9

A

OV, S1: A, Hazard rates for men; B, CDF for men

Fig. 1.9

B

(cont.)

Fig. 1.10

A

OV, S2: A, Hazard rates for men; B, CDF for men

Fig. 1.10

B

(cont.)

Fig. 1.11

A

OV, S3: A, Hazard rates for men; B, CDF for men

Fig. 1.11

B

(cont.)

Fig. 1.12

A

AC, S1: A, Hazard rates for women; B, CDF for men

Fig. 1.12

B

(cont.)

Fig. 1.13

A

AC, S2: A, Hazard rates for women; B, CDF for men

Fig. 1.13

B

(cont.)

Fig. 1.14

A

AC, S3: A, Hazard rates for women; B, CDF for men

Fig. 1.14

B

(cont.)

96

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

Our findings tend to indicate that the proposed reforms would have a rather significant impact on the retirement behavior of older Belgians, especially for policy 1. This is not really surprising since the hypothetical changes represent massive shake-ups of the system. Indeed, in a country such as Belgium where the current average observed departure out of the labor force lies well below the early retirement age of the official pension system (sixty for most individuals aside from teachers and similarly privileged occupations), the elimination of all these early exit routes from the labor force has to be seen as an earthquake. This is particularly true if we consider simulations of the S3 type. No matter which incentive variable we use (AC, PV, or OV), we see strong effects on the hazard rate of departures and, as a consequence, strong effects on the cumulative distribution function. The figures clearly display the wide disparity between the hazard functions and the cumulative distribution functions depending on which simulation method (S1, S2, or S3) serves as a basis for the simulation. There is, however, a large degree of similarity between the results we obtain using one of the three simulation methods while changing between any of the three different incentive variables (AC, PV, or OV). Summary measures such as the average retirement age, although useful in their own way, are clearly insufficient for understanding the total change in the retirement patterns. Very similar average retirement ages can be derived from very different hazard processes. The findings also comfort the intuitive idea that the changes in the hazard rate implied by policy 1 should be clearest in S1 and S3, where the jumps at ages sixty are simply shifted up by three years. Of the three methods, S1 is the most conservative in terms of the changes in the hazard rate, which should also not surprise the reader because of the underlying linear age trend that is involved. Not surprisingly, the effect of policies 1 and 2 are the most divergent in S3, where their specificities fully come to bear on the hazard rates. This clearly illustrates the importance of the question of whether we should (S3) or should not (S2) also change the weights on the age dummies in the dummy model. 1.8 Conclusions The rapid aging of the Belgian population creates major problems for the financing of the public retirement and early retirement systems. This is even more so, given the rather impressive decline of labor force participation that we have witnessed in Belgium over the last several decades, which has made the country one of the world leaders in putting its people into retirement at very early ages. Because of the varying departure patterns from the labor force in the different systems, these challenges will also have a very different impact on their viability.

Micro-Modeling of Retirement in Belgium

97

The present paper sets a new standard in analyzing the retirement decisions of Belgian workers because it models and analyzes the impact of incentive variables, such as the present-discounted value of social security entitlements and the change that occurs in the latter when people change their age of retirement. Our paper finds strong evidence that social security accruals are strongly negative for most people aged sixty and above. More strikingly, more than 50 percent of workers face negative accruals as early as age fifty-eight. We find a similar picture using more forward-looking incentive measures that take into account the entire future path of benefit dynamics. Even more importantly, we find a strong and very significant negative impact of these dynamic incentive variables, such as AC and OV, on the decision to work. Hence, we find that workers with smaller rewards or even penalties on continued work do indeed retire earlier from the labor force. These findings are of a crucial importance in the light of reforms to one, several, or all of the Belgian retirement systems. Governments and policymakers cannot simply assume that retirement decisions are static, but rather have to take into account the impact of the SSW and the dynamic incentive measures. We illustrate this logic by applying two rather distinct hypothetical policy proposals for reforming the Belgian retirement landscape, which have in common a reduction of benefit entitlements to improve the chances of survival of the public retirement-income systems. Our simulations show that such reforms have the potential to induce major changes in observed retirement patterns.

References Blöndal, S., and S. Scarpetta. 1998. Falling participation rates among older workers in the OECD countries. Paris: Organization for Economic Cooperation and Development (OECD). Bouillot, L., and S. Perelman. 1995. Evaluation patrimoniale des droits à la pension (Social security wealth estimations for Belgium). Revue Belge de Sécurité Sociale 37 (4): 803–31. Deboosere, P., and S. Gadeyne. 2000. Zijn de regionale sterftepatronen in België te verklaren door individuele socio-economische kenmerken (Do individual socioeconomic characteristics explain regional mortality differences in Belgium)? Free University of Brussels, Social Research Center. Working Paper no. 2000-3. European Economic Community (EEC). 1994. Supplementary pension in the European Union. Development report by the European Commission’s Network of Experts on supplementary pensions. Brussels: EEC. Jousten, A., and P. Pestieau. 2002. Labor mobility, redistribution and pension reform in Europe. In Social security pension reform in Europe, ed. M. Feldstein and H. Siebert, 85–105. Chicago: University of Chicago Press.

98

Arnaud Dellis, Raphaël Desmet, Alain Jousten, and Sergio Perelman

Organization for Economic Cooperation and Development (OECD). 1994. OECD economic survey: Belgium, Luxembourg. Paris: OECD. Pestieau, P., and J. P. Stijns. 1999. Social security and retirement in Belgium. In Social security and retirement around the world, ed. J. Gruber and D. Wise, 37–71. Chicago: University of Chicago Press. Stock, J., and D. Wise. 1990. Pensions, the option value of work, and retirement. Econometrica 58 (5): 1151–80.

2 Income Security Programs and Retirement in Canada Michael Baker, Jonathan Gruber, and Kevin Milligan

2.1 Introduction Government transfers to older persons in Canada are one of the largest and fastest growing components of the government budget. Total expenditures on the three primary transfer programs for older Canadians amounted to $22.7 billion in 1998–1999, which represented 20 percent of program spending in the federal budget of that fiscal year. In 1974–1975, total expenditures were only $3.4 billion, amounting to just 10 percent of program spending. The contributory public pension faces fiscal similar pressures. In 1975, contributions per capita exceeded benefits per capita by roughly $200 (1998 dollars). By 1998, benefits per capita instead exceeded contributions per capita by roughly $200 (Baker and Benjamin 2000). Moreover, without changes to the system, these trends will likely continue in the foreseeable future. The ratio of persons aged sixty-five and over to persons aged twenty to sixty-four is projected to grow from its current level of 19 percent to over 40 percent by the year 2075. As a result, the payroll Michael Baker is professor of economics at the University of Toronto, and a faculty research fellow of the National Bureau of Economic Research (NBER). Jonathan Gruber is professor of economics at the MIT, director of the research program on children and research associate of the NBER. Kevin Milligan is assistant professor of economics at the University of British Columbia, and a faculty research fellow of the NBER. Prepared for the Project on International Social Security Comparisons. We are grateful to Human Resources Development Canada, Statistics Canada and the Social Science and Humanities Research Council (SSHRC; Baker, in particular) for research support; Sue Biscope, Richard Dupuy, Leonard Landry, and Garnett Picot for assistance accessing the data; Terence Yuen for excellent research assistance; and to Paul Finn and participants in seminars at Canadian International Labour Network (CILN), Human Resources Development Canada (HRDC), NBER, and the University of British Columbia for helpful comments. All views expressed in this paper are the authors’ own and do not necessarily reflect the views of HRDC or Statistics Canada.

99

100

Michael Baker, Jonathan Gruber, and Kevin Milligan

tax necessary to finance the public pension for older persons, the Canada and Quebec Pension plan, will grow from 7.0 percent in of wages in 2000 to 9.9 percent by the year 2003. Similar cost increases are in store for the transfer programs for older Canadians, which are financed from general revenues: the Old Age Security demogrant, and the income-tested Guaranteed Income Supplement, and the Spouse Allowance programs. In this context, a notable trend in labor force behavior in Canada is the steady decline in work among many groups of older workers, as documented in figure 2.1. It is particularly striking for older males. Note that the participation rate for forty-five to sixty-four year olds masks a large decline among the older individuals in this group. For example, in 1960, 87 percent of men aged fifty-five to sixty-four were participating in the labor force; by 1999, this proportion had fallen to 61 percent. For females, any trend towards earlier retirement appears to be swamped by the century-long secular increase in the participation of women. These time series span a period in which there were a variety of changes in the structure of income support programs for older persons that has made retirement more attractive and work less attractive. In 1960, for example, workers under the age of seventy were not entitled to any income support upon retirement. By the mid-1990s, however, married low-income workers could receive public retirement benefits that actually exceeded their preretirement incomes (Gruber 1999). Of course, it is difficult to causally relate these time trends; there were many other developments over this time period, such as growing private pension coverage and rising incomes, that may have also contributed to the decline in work among older Canadians. In the United States where there are similar trends, there is an extensive literature that examines the relationship across individuals between social security entitlements and retirement decisions.1 This research mostly suggests that social security incentives play an important role in retirement decisions, but a modest one relative to the time trends. In contrast, there is little complementary work in the Canadian context.2 Recent studies have examined the impacts of changes in some of these programs in isolation. Baker and Benjamin (1999b) analyze the effects of the removal of the earnings tests from the Canada Pension Plan and Quebec Pension Plan (CPP/ QPP) in the 1970s. They also examine the effect of the introduction of an early retirement option to the QPP in 1984 and to the CPP in 1987 (Baker and Benjamin 1999a, 2001). The determinants of the CPP/QPP take up decision are studied by Tompa (1999). Compton (2001) studies the effect of CPP/QPP benefits on retirement, using a short panel of data from the 1. For a review of the literature and some empirical evidence, see Coile and Gruber (2000). 2. Papers by Burbidge (1987) and Pesando and Rea (1977) provide a careful outline of the potential effects of the various Income Security (IS) programs, but no estimate of their empirical magnitude.

Historical trends in the labor force participation of older males and females

Source: Statistics Canada (1983) and Canadian Socio-Economic Information Management System (CANSIM; available at http://dc2.chass.utoronto. ca/cansim2/English/index.html). These series span the Change in the Labor Force questionnaire in 1976.

Fig. 2.1

102

Michael Baker, Jonathan Gruber, and Kevin Milligan

mid-1990s. Finally, Baker (2002) investigates the effects of the introduction of the Spouse Allowance in 1975 on the labor market behavior of eligible couples. A weakness common to all of these papers is that only one component of the IS system is studied in isolation from the others. Our work aims to take a more comprehensive approach by modeling the entire IS system in a unified framework. In a previous paper (Baker, Gruber, and Milligan 2003), we also addressed the incentive effects of Canada’s income security system on retirement. In that paper, we undertook an in-depth examination of the robustness of retirement models to variations in specification and variable definition. Two major findings emerged. First, we found that including rich controls for past and current earnings had a substantial impact on the magnitude of our estimates. Second, our results varied sensibly with changes in specification and variable definition.3 We use these findings to guide us in the choice of specification and variable definition we employ for the present paper. The key to our approach is the building of a comprehensive data set (based on the Statistics Canada Longitudinal Worker File) that has information for a very large sample of older Canadians on their earnings histories, work decisions, marital status and spousal characteristics, and the characteristics of their jobs. These data are employed to construct a simulation model that incorporates the incentives for retirement under the various programs of the Canadian public IS system: the CPP/QPP; the Old Age Security (OAS) system comprising the basic OAS benefit; the Guaranteed Income Supplement (GIS); and the Spouse Allowance (SPA). For each person in our data set, the financial incentives for retirement are computed along two dimensions. First, the present discounted value of all future entitlements to benefits from the different programs of the public IS system is calculated. This measure of income security wealth (ISW) is recalculated for each year the person appears in our data set to reflect the changes to their benefit entitlements. The second dimension is a measure of how ISW evolves through time. An ISW accrual measure can be calculated by comparing the ISW of the person if they retired in the present year to the ISW of the person if they worked an extra year. Several different measures of accrual are contemplated, which alternatively assume that individuals look only one year forward in making their retirement decision and that individuals look forward to some “optimal” retirement date in making their decision. An empirical model of the retirement decision as a function of these incentive variables, as well as a rich set of control variables designed to capture other impacts on retirement, is then estimated. There are two findings of importance. First, for the typical worker, the 3. For example, we found that those with workplace pension plans are less sensitive to public pension incentives.

Income Security Programs and Retirement in Canada

103

public IS system provides increasingly strong disincentives to work after age sixty. Workers actually see the present discounted value of there IS entitlement fall from additional work after age sixty-one, and by age sixtynine, the reduction in IS entitlement amounts to 43 percent of what they would earn in that year. Second, there is a significant impact of these disincentives on work decisions. Using both one-year and more forwardlooking measures, we estimate that workers with larger returns to additional work are less likely to leave the labor force. 2.2 Background The decision to retire in Canada is made in the context of a complicated web of program incentives. Below, we first describe in detail the components of the public IS system in Canada. This is followed by a brief description of private pensions and a summary of the different paths to retirement. 2.2.1 The Old Age Security (OAS) System The oldest component of the IS system for older Canadians is the OAS system, which was put into place in 1952, replacing a provincially run income-tested benefits system that had existed since 1927. This program is available to anyone aged sixty-five or over who meets certain residence requirements.4 The program originally provided benefits to those of age seventy or over, and the age of eligibility was dropped to sixty-five over a fiveyear period beginning in 1966. The OAS pension itself is a uniform demogrant that was equal to $419.92 in March 2000. Individuals who do not fully meet residence requirements may be entitled to a partial OAS benefit. The OAS benefits have been indexed to the consumer price index (CPI) since 1972, and OAS benefits are fully taxable. In addition, there is a clawback of OAS benefits from very high-income individuals; the OAS for an individual is reduced by 15 cents per dollar of personal net income exceeding $53,215. The OAS basic benefit is financed from general taxation revenues. 2.2.2 The Canada Pension Plan and the Quebec Pension Plan (CPP/QPP) The largest component of the income security system is the CPP/QPP. These programs began on 1 January 1966 and are administered separately by Quebec for the QPP and by the federal government for the CPP. 4. Individuals must have been a Canadian citizen or legal resident of Canada at some point before application and must have resided in Canada for at least ten years (if currently in Canada) or twenty years (if currently outside Canada). The benefit is prorated for pensioners with less than forty years of Canadian residence, unless they are “grandfathered” under rules that apply to the persons who were over age twenty-five and had established attachment to Canada prior to July 1977.

104

Michael Baker, Jonathan Gruber, and Kevin Milligan

The plan is financed by a payroll tax of 3.5 percent (in 2000) for both employers and employees. This payroll tax is levied on earnings between the year’s basic exemption ($3,500) and up to the year’s maximum pensionable earnings (YMPE), $37,600 in 2000 (which approximates average annual earnings). The YMPE is indexed to the growth in average earnings in Canada. Eligibility for this plan is conditioned on contributions in at least one calendar year during the contributory period. The contributory period begins at age eighteen, or at 1 January 1966 for those who were older than eighteen on that date. It ends at the commencement of the retirement pension or at age seventy, whichever is earlier. Benefits are then computed in several steps. First, the number of months used to compute the retirement pension is determined by subtracting from the number of months in the contributory period months that the person was (a) receiving a disability pension, (b) rearing small children,5 and (c) between age sixty-five and the commencement of the pension,6 as well as subtracting 15 percent of the remaining months. The last two of these conditions are subject to the provision that it not reduce the contributory period below 120 months after taking into account the allowable offset for months of disability pension receipt. In addition, excess earnings in one month above one-twelfth of the YMPE may be applied to months in the same year where earnings are below one-twelfth of the YMPE. Second, the remaining months of earnings history are converted to current dollars, using the following adjustment factor: the ratio of the YMPE in each year (up to 1998) to the average of the YMPE over the three years prior to (and including) the year of pension receipt. This average was raised to four years for benefits claimed in 1998 and five years for benefits beginning in 1999. Finally, the benefit is computed as 25 percent of the average of this real earnings history. This 25 percent ratio has been in place since 1976; from 1967 to 1976, the program was phased in, with the share of average earnings paid out in benefits rising from 2.5 percent in 1967 to 25 percent in 1976. The maximum monthly retirement benefit is $762.92 in 2000. Until 1984 for the QPP and 1987 for the CPP, benefits could not be claimed before the sixty-fifth birthday, and there was no actuarial adjustment for delayed claiming. Beginning at these times, individuals were allowed to claim benefits as early as age sixty, with an actuarial reduction of 0.5 percent for each month of early claiming (before age sixty-five), and an actuarial increase of 0.5 percent for each month of delayed claiming (after age sixty-five, and up to the age of seventy). 5. This is defined as months in which there was a child less than seven years of age and the worker had zero or below average annual earnings. 6. Periods after age sixty-five to age seventy can be substituted for periods prior to age sixtyfive if this will increase their future retirement pension.

Income Security Programs and Retirement in Canada

105

Since this early retirement provision has been in place, about half the new CPP recipients each year have claimed a retirement benefit before the age of sixty-five. The Office of the Superintendent of Financial Institutions (OSFI) estimated that, after 1991, a CPP pension for someone retiring before the age of sixty-five was, on average, 82 percent of what it would have been had they not opted for early retirement.7 Initially, receipt of benefits between ages sixty-five and seventy under the CPP/QPP was conditioned on low earnings levels, with earnings above these ceilings taxed away at high rates. In 1975 and 1977, these earnings tests were eliminated from the CPP and QPP, respectively. With the introduction of early retirement in the 1980s, workers can only claim early benefits if their annual rate of earnings for the year in which the pension is claimed does not exceed the maximum retirement pension payable at age sixty-five. This earnings test is only applied at the point of application, however; after that point, there is no additional check on the individual’s earnings.8 Moreover, the earnings test does not apply once the individual reaches age sixty-five. The CPP/QPP benefits are based on an individual’s earnings history, and the retirement benefits of one spouse are not linked to that of the other spouse.9 There is, however, interdependence through survivor benefits (as well as the interdependencies through the income-tested programs described below). Spouses are eligible for survivor pensions if the deceased contributor made contributions for the lesser of ten years or one-third of the number of years in the contributory period, and if the spouse is over age forty-five or is disabled or has dependent children. For nondisabled spouses with children, the CPP benefit is prorated downward by age between forty-five and thirty-five.10 For spouses under age sixty-five, the survivor pension is a combination of a flat-rate portion plus 37.5 percent of the earnings-related pension of the deceased spouse. For spouses age sixtyfive and above, the survivor’s pension is equal to 60 percent of the earningsrelated pension. The pension used to calculate the survivor’s benefit is not subject to actuarial adjustment. If the surviving spouse is receiving his or her own CPP disability or retirement pension, then the combination of the earnings-related portion of the two pensions cannot exceed the maximum retirement pension available in the year. Under changes made effective in 1998, the two benefits do not stack up to this ceiling; rather the contributor receives the larger of the two earnings-related portions plus 60 percent 7. Special calculations for the 1992 OAS program evaluation performed by OSFI. 8. There are no restrictions on returning to work after the benefit is being paid. 9. Couples do have the option of sharing their benefits for income tax purposes, since taxation is at the individual level. Each spouse can claim up to half of the couple’s total CPP/QPP pension credits. The exact calculation depends on the ratio of their cohabitation period to their joint contributory period. 10. The QPP rules for younger surviving spouses differ from those of the CPP.

106

Michael Baker, Jonathan Gruber, and Kevin Milligan

of the smaller. Also, if under the age of sixty-five, the survivor receives the flat-rate portion of the survivor benefit or, if a disability pensioner, the (larger) disability flat-rate benefit only. Children of deceased contributors are also entitled to a CPP survivor’s benefit if under age eighteen or a full-time student between eighteen and twenty-five; this benefit is a flat amount. The corresponding QPP benefit ends at age eighteen. There is also a lump-sum death benefit, which is generally equal to one-half of the annual CPP/QPP pension amount up to a maximum ($3,500 in 1997).11 Since 1973, benefits have been legislated to increase annually with the CPI: This annual indexation factor is the ratio of the CPI average over the twelve-month period ending with October of the preceding year to the average of the prior twelve-month period. Benefits are fully taxable by the federal and provincial governments. Another dimension of the CPP/QPP that is potentially important here is the disability-benefit program. This program provides benefits to those workers unable to work due to disability. The basic benefits structure consists of two portions: a flat-rate portion, which is a lump sum paid to all disabled workers, and an earnings-related portion, which is 75 percent of the applicable CPP/QPP retirement pension calculated with the contributory period ending at the date of disability. This program is fairly stringently screened, and fewer than 5 percent of older Canadian men are on CPP/QPP disability. The maximum CPP disability benefit was increased by 30 percent per month in 1987. Earlier disability coverage was also extended to new entrants. Also, persons receiving survivor benefits no longer had their benefits discontinued on remarriage. 2.2.3 The GIS and SPA The GIS is an income-tested supplement available to recipients of OAS that was introduced in 1967. Individuals must reapply for the GIS each year, and the income test for eligibility is repeated. The definition of income for the purpose of income testing is the same as for income tax purposes, with the important exclusion of OAS pension income. Unlike the OAS clawback or CPP/QPP, GIS benefits are based on family income levels. There are separate single and married guarantee levels for the GIS; in 2000 (January to March), these were $499.05 for singles and $325.06 (per person) monthly for the married. Benefits are then reduced at a rate of 50 percent as other income rises, although a couple with one member over age sixty-five and one under age sixty is taxed at only 25 percent with an initial amount of income exempted. The SPA, which was introduced in 1975, is an income-tested monthly 11. Under the 1997 legislation, this maximum is fixed at $2,500 for all years after 1997, and in the case of the QPP, all death benefits are set at this level.

Income Security Programs and Retirement in Canada

107

benefit available to sixty to sixty-four-year-old spouses of OAS recipients and to sixty to sixty-four-year-old widows and widowers. For the spouse of an OAS recipient, the benefit is equal to the OAS benefit plus GIS at the married rate; the OAS portion is reduced by 75 percent of other income until it is reduced to zero, and then the combined GIS benefits of both spouses are reduced at 50 percent, as other income rises. For a widowed spouse, the benefit is equal to the OAS plus GIS at the widowed rate, and is “taxedback” equivalently. Both the GIS and SPA guarantees are also indexed to inflation, and neither source of income is taxable by either the federal or provincial governments. 2.2.4 Other Public Programs In addition to the federal retirement programs, there are a variety of provincial programs that provide supplements to low-income retirees. For example, the Guaranteed Annual Income Supplement for the Aged (GAINS-A) program in Ontario provides $80 per month to Ontario residents who are recipients of the GIS; these benefits are taxed back at 50 percent as other (non-OAS or GIS) income rises. 2.2.5 Private Pension Coverage Another important feature of the retirement landscape is private pensions. Defined-benefit pension plans share many of the same incentive features as public insurance plans. In fact, many Canadian workers are covered by occupational pensions (known as Registered Pension Plans, or RPPs). In 1997, 41.2 percent of paid workers were covered by occupational pensions, with coverage slightly higher for males than for females (Statistics Canada 1999). Eighty-six percent of plan members were in definedbenefit plans, although the share in defined-contribution plans has been growing recently. Defined-contribution plans may also affect retirement through income effects, but there should not be tax or subsidy effects on the work decision since the payout is not dependent on work patterns. One weakness of the data that are employed in this study is a lack of information about private pension plan coverage. As a result, it is only possible to include an indicator for whether the individual is likely to have a pension (based on industry of employment), but not for the retirement incentives inherent in that particular pension plan (as is done, for example, in Gruber and Madrian 1995). The methods and data sources for this imputation are described below. 2.2.6 The Different Paths to Retirement Given the differences in the age of initial eligibility across the different IS programs and the availability of other income-support programs before the age of sixty-five, there are a variety of paths that individuals may follow into retirement. Perhaps the most straightforward is from employment onto IS benefits at age sixty-five or later. At these ages an individual is eli-

108

Michael Baker, Jonathan Gruber, and Kevin Milligan

gible for all the IS programs so the full potential retirement income from public sources will be available. Early entitlement for CPP/QPP benefits is available starting at age sixty. Receipt is conditioned on a one-time retirement test, although beneficiaries are free to work once the test is met. Since other sources of support, such as OAS and GIS, are not available until age sixty-five, benefit income may be augmented by earnings from full- or part-time employment. Income is also potentially available from other social insurance programs, such as employment insurance (EI), although there are conditions (e.g., unemployment), and pension income is deducted from any benefits from this source. Even if early CPP/QPP benefits are not claimed, EI benefits, social assistance benefits, or both are another potential source of support for older workers and thus a path into IS receipt. Also, disabled individuals are eligible for a CPP/QPP pension prior to age sixty that gets automatically converted to a retirement pension at age sixty-five. Finally individuals who participate in RPPs with attractive early retirement packages may start claiming these benefits as a prelude to IS-benefit receipt at later ages. As explained below, our measure of retirement is based on earnings (or the lack thereof), and therefore employment. We have no direct measure of IS-benefit receipt, so alternative definitions of retirement on this basis are not possible. Our data do record EI-benefit receipt, however, so there is some possibility of tracking individuals who use this path to retirement. Data on other forms of income such as an RPP or social assistance are not available, however, so these paths are also not visible. In table 2.1, we provide a view of the employment and program participation of older Canadians using data from the 1998 Individual Files of the Survey of Consumer Finances (Statistics Canada 1998a). Full-time work declines dramatically for both males and females between the ages of fifty and sixty-four. Between the ages of sixty and sixty-four, 34 percent of men and just 13 percent of women are in this category. A constant fraction of males work part time in each age group, but for females, the proportion displays a moderate decline before age sixty-five and dramatic fall off in the oldest age group. The proportions not working, and therefore by some measures retired, rise steadily for either sex with age. In the age group sixty to sixty-four, when early CPP/QPP benefit receipt is available, 60 percent of males and 77 percent of females are not working. In the older age group just 10 percent of males and virtually no females are still employed. The table also reveals that benefits from a variety of programs may support those in the younger age groups that are not working. The proportion drawing a private pension or Registered Retirement Savings plan (RRSP) benefits rises steadily to almost one in three males and one in five females by ages sixty to sixty-four. Income from EI and social assistance flows to a relatively constant proportion (17 percent of males and 13 percent of females) between the ages of fifty and sixty-four. The popularity of the early retirement option of the CPP/QPP program for both sexes is apparent: At

Income Security Programs and Retirement in Canada Table 2.1

109

Labor Market Participation and Program Participation in 1997–98 Age 50–54

55–59

60–64

65+

0.76 0.04 0.21

0.60 0.05 0.35

0.34 0.05 0.60

0.07 0.04 0.89

0.00 0.05 0.05 0.10 0.06

0.00 0.07 0.15 0.09 0.08

0.03 0.44 0.30 0.06 0.12

0.97 0.92 0.58 0.01 0.08

0.50 0.14 0.36

0.33 0.13 0.54

0.13 0.10 0.77

0.01 0.02 0.97

0.00 0.06 0.03 0.07 0.05

0.00 0.09 0.10 0.07 0.06

0.15 0.40 0.20 0.04 0.09

0.97 0.74 0.35 0.00 0.11

Males Labor market participation in April 1998 Working full time Working part time Not working Program participation in 1997 Received OAS/GIS/SPA benefits Received CPP/QPP benefits Received private pension/RRSP benefits Received employment insurance benefits Received social assistance benefits

Females Labor market participation in April 1998 Working full time Working part time Not working Program participation in 1997 Received OAS/GIS/SPA benefits Received CPP/QPP benefits Received private pension/RRSP benefits Received employment insurance benefits Received social assistance benefits

Source: Statistics Canada (1998a). Notes: The statistics on labor market participation are for the reference week of April 1998 used by the survey. The statistics for program participation are for the reference year (1997).

least 40 percent of both males and females between the ages of sixty and sixty-four receive this sort of income. The statistics also show that females are far more likely to take advantage of the SPA program than males and thus receive OAS, GIS, and SPA income between the ages of sixty and sixty-four. This message here, therefore, is that in the late 1990s a majority of older Canadians are not working by ages sixty to sixty-four. In fact a significant minority are not working by ages fifty-five to fifty-nine. Income support at these younger ages may be coming from private pensions and other social insurance programs. In their early sixties, a significant number of Canadians also avail themselves of the early retirement option in the public pension program. 2.3 Data There are few Canadian data sets that provide both large sample sizes of older individuals and the information necessary to calculate their in-

110

Michael Baker, Jonathan Gruber, and Kevin Milligan

centives to retire. This has hindered research on retirement in Canada. To overcome this obstacle, the analysis here makes use of data from a number of sources. These data provide the most comprehensive setting available in which to study the incentives of the Canadian IS system on retirement. The primary data source is the Longitudinal Worker File (LWF) developed by the Business and Labour Market Analysis (BLMA) Division of Statistics Canada.12 It is a 10 percent random sample of Canadian workers for the period 1978–1996. These data are the product of information from three administrative data files: The T-4 file of Revenue Canada, the Record of Employment (ROE) file of Human Resources Development Canada, and the Longitudinal Employment Analysis Program (LEAP) file of BLMA. The T-4 tax forms are issued annually by employers for any employment earnings that exceed a certain annual threshold; trigger income tax contributions to Canada’s public pension plans, or EI premiums; or both.13 The earnings information from this source has several advantages over its counterparts in survey data and other administrative files. Most importantly, it is based on employers’ reports under the provisions of the income tax laws. Therefore, the earnings variable should be free of the measurement error often observed in survey data. Employers issue ROE forms to employees in insurable employment14 whenever an earnings interruption occurs. Earnings interruptions result from events such as strikes, layoffs, quits, dismissals, retirement, and maternity or parental leave. The reason for the interruption is recorded on the ROE form. Finally, the LEAP is a longitudinal data file on Canadian businesses at the company level. It is the source of information on the company size and industry of the jobs in which employees work. The LWF data provide information on the (T-4) wages and salaries and 3-digit industry codes for each job an individual holds in a given year; their age and sex;15 the province and size (in terms of employees) of the estab12. The construction of the database is described in Picot and Lin (1997) and Statistics Canada (1998b). Our description draws heavily on these sources. 13. The data include incorporated self-employed individuals who pay themselves a salary, but not other self-employed workers. The federal program that provides insurance against unemployment changed names from “unemployment insurance” to “employment insurance” in 1996. Throughout this paper, we use employment insurance (EI) when referring to this program. 14. Over the sample period, insurable employment covers most employer-employee relationships. Exclusion includes self-employed workers, full-time students, and employees who work less than fifteen hours per week and earn less than 20 percent of maximum weekly insurable earnings (20 percent • $750
$150 in 1999). Individuals working in insurable employment pay EI contributions on their earnings and are eligible for EI benefits subject to the other parameters of the EI program. 15. Information on the age and sex of individuals is taken from the T-1 tax returns, which individuals file yearly. To obtain this information, therefore, it is necessary that they filed a tax return at least once in the sample period.

Income Security Programs and Retirement in Canada

111

lishment for which they work;16 and their job tenure starting in 1978. Because T-4s are also issued for EI income, we also observe any insured unemployment, maternity, or sickness spells. For current purposes, the prime advantage of the LWF data are the earnings histories stretching back to 1978. These were extended further to 1975 for each individual using the T-4 earnings files for these years. For the purposes of calculating CPP/QPP entitlement these histories are still nine years short, however, as these programs started operating in 1966. Our methods of backcasting the missing years are described below. The focus of the analysis is the period 1985–1995. Separate samples of males and females aged fifty-five through sixty-nine in 1985 are drawn, and then younger cohorts of individuals are added as they turn fifty-five in the years 1986–1991.17 Agricultural workers and individuals in other primary industries are excluded.18 The sample is selected conditional on working so that the incentives for retirement conditional on being in the labor force are examined. Work is defined as positive T-4 earnings in two consecutive years. If an individual has positive earnings in one year and zero earnings in the next, the year of positive earnings is considered the retirement year.19 Given that our data run to 1996, this means that the last year for which an observation can be formed is 1995, since we need to see one year forward to determine retirement. Since T-4s are not issued to the unincorporated self-employed, this definition of retirement will also capture any persons moving from paid employment into this sector.20 Only the 16. The records of the LWF data are at the person, year, and job level. For some calculations it is necessary to aggregate the data to the person and year level. In years in which an individual has more than one job, there will be multiple measures of tenure, industry, firm size, and in some cases province. In these cases, the characteristics of the job with the highest earnings for the year are used. 17. Individuals with missing age, sex, or province variables are excluded from the sample. 18. We make this exclusion because our definitions of retirement are based on earnings, and the earnings streams for these workers, given high rates of self-employment and special provisions in the EI system for fishers and other seasonal workers, are difficult to interpret. For example, individuals in these industries who were too young to collect IS benefits are observed with years of very small earnings (in the hundreds of dollars) and no (or sporadic) evidence of EI benefits. One possibility is that they are primarily unincorporated self-employed, and therefore the majority of their earnings are unobserved. 19. Baker, Gruber, and Milligan (2003) check the robustness of the results under two different definitions of retirement. First, an unemployment insurance (UI)-based definition encompasses UI benefits along with labor market earnings. Second, an earnings-based definition labels someone retired if their earnings fall below a minimum earnings threshold. The results with both of these definitions are broadly consistent with those presented here. 20. While older individuals do work in unincorporated self-employment, the proportion doing so remains fairly constant over our sample period. For males, Canadian census data (Individual Files for 1981, 1986, 1991, and 1996) reveal that the proportion of the population of sixty to sixty-four year olds working in this sector is 0.08–0.09 (0.04 for ages sixty-five and older) in Quebec and 0.13–0.16 (0.06–0.08 for ages sixty-five and older) in the rest of Canada between 1980 and 1995. For females, the statistics are 0.01 (0.00–0.01 for ages sixtyfive and older) in Quebec and 0.20–0.40 (0.01 for ages sixty-five and older) in the rest of Canada.

112

Michael Baker, Jonathan Gruber, and Kevin Milligan

first observed retirement for each individual is considered. If a person reenters the labor market after a year of zero earnings, the later observations are not used. Finally, individuals are only followed until age sixtynine. The retirement of an individual who has positive earnings in every year up to this age is not observed since it presumably occurs after the age of sixty-nine. The working sample, therefore, is a panel data set for the years 1985 through 1995 of individuals between the ages of fifty-five and sixty-nine who worked in 1985 or in the year they turned fifty-five, whichever is later.21 The marital status and any spouses of individuals in our sample are identified using information from the T-1 family file maintained by Statistics Canada; T-4 earnings histories for the period 1975–1995 are then constructed for the spouses, again through reference to the T-4 earnings files for these years. An important piece of information for calculating retirement incentives that is not available in the LWF data is participation in a RPP. We estimate the probability of RPP coverage by 3-digit industry codes22 using crosssectional samples of males or females from the 1986–1990 Labour Market Activity Survey (LMAS) and the 1993–1995 Survey of Labour and Income Dynamics (SLID). In these surveys, individuals are asked if they participate in any RPP. These probabilities are then imputed to individuals in the LWF, matching on industry codes. Probabilities for the years 1991 and 1992 are simple linear interpolations. The sample definitions for these additional data sources are described in the appendix. 2.4 Earnings and Nonlabor Income Projections The following analysis involves constructing each sample individual’s entitlement to IS benefits at any given age, as well as estimates of future entitlements. To calculate the CPP/QPP component, we require a full earnings history from 1966, the year in which the program started. As noted above, our earnings information only extends back until 1975. In estimating future entitlements, we must project future earnings to construct the relevant earnings history. Therefore, both earnings backcasts and forecasts are needed for these calculations. After experimenting with a number of projection methods, earnings are 21. The ROEs were considered as an alternative source of information on when individuals retired. It was found, however, that generally less than one-third of individuals who retired in the earnings sense (e.g., had zero T-4 earnings), also had “retirement” coded on their last ROE. “Still working” or “unknown” were the most common codes for those in the complementary group. The ROEs, therefore, would appear to impose a restrictive definition of retirement that has an unknown basis. 22. Some industries are aggregated to obtain sufficient sample sizes. Unfortunately, the sample sizes of these data sets would not permit us to calculate these probabilities exclusively for older individuals.

Income Security Programs and Retirement in Canada

113

forecasted by applying a real growth rate of zero percent per year to the average of an individual’s observed earnings in the three years preceding the retirement year. Within-sample evaluation revealed that this method is a better predictor (in a mean-squared error sense) of future earnings than methods involving a projection equation that included demographic variables, lagged earnings, and individual fixed effects. To backcast the missing earnings data, cohort-specific earnings growth rates calculated from the 1972, 1974, and 1976 Census Family files of the Survey of Consumer Finance (SCF; Statistics Canada, various years)23 were applied to a three-year average of an individual’s last valid earnings observations in the LWF sample. This allows us to construct earnings histories back to 1971. For the remaining five years, earnings growth rates implied by the cross-sectional profile from the 1972 SCF were used, appropriately discounted for inflation and productivity gains using the industrial composite wage for the period 1966–1970.24 The GIS and SPA and OAS components of IS benefits are fully or partly means tested. Our data set contains no information on nonlabor income, although these are clearly a crucial input to calculating entitlement to these benefits. To project nonlabor income, we construct age profiles of familylevel income by sex, region, and industry and sex, region, and maritalstatus cells for individuals in and out of the labor market, respectively.25 The data for these profiles are from the 1986 and 1991 Census Family files of the Canadian Census (Statistics Canada, various years). The measure of nonlabor income that we use includes both investment income and income from private pensions. Details on the formal definition of the measure are provided in the appendix. When entitlement is projected in future retirement years, it is necessary to impute the level of nonlabor income an individual will receive at different ages when they are retired. To do this, we use the age profile for this income for individuals out of the labor market in the relevant sex, region, and marital-status cell. Likewise, for individuals who continue to work past age sixty-five (sixty for the SPA), it is necessary to impute their level of nonlabor income to calculate the benefits they might draw from OAS, GIS, or SPA. To do this, the age profile for employed individuals in the relevant sex, region, and industry cell is used. The sample and cell definitions that are employed are also described in the appendix. Both projected earnings and nonlabor income are net of federal and provincial income taxes. Also deducted are the employee’s portions of the CPP/QPP payroll tax that they would pay if they worked. In either case, the 23. We use samples of paid workers with positive earnings in the relevant birth cohorts. 24. The data on the industrial composite wage are from Statistics Canada (1983). The obvious limitation of this backcasting approach is that we will not predict absences from the labor market, which may be important at younger ages. 25. The age profiles are appropriately inflated by the CPI for use in future years.

114

Michael Baker, Jonathan Gruber, and Kevin Milligan

parameters of the tax system are held constant in real terms for all future years. 2.5 Construction of the Incentive Measures 2.5.1 Benefit Entitlements The retirement incentives inherent in the various programs of the Canadian IS system for seniors are calculated regarding the OAS, the GIS and SPA, and the CPP/QPP. The first step is to calculate an individual’s entitlement in any given year. This will involve both their own entitlements to each of the programs as well as the entitlements of any spouse. The OAS benefit is the most straightforward as it is a uniform benefit available to anyone who is sixty-five years or older. Two possible complications are the residency requirements and the clawback of benefits from high-income recipients. The residency requirement for this benefit is not implemented, as there is no information on the place of birth or year of arrival in Canada of individuals in the sample. The clawback provisions (starting in 1989) are fully implemented, however, based on projections of labor and nonlabor income. Either the GIS entitlement, SPA entitlement, or both are functions of the age requirements, described previously, and family income. The ages of individuals and any spouses are directly observable in the data. The income test on benefits is again fully incorporated based on projections of labor and nonlabor income. As discussed above, nonlabor income is projected using census data and matched to our data. For each individual, the OAS and GIS and SPA benefit entitlement with and without the imputed level of nonlabor income is calculated. The two results are then averaged using as weights the cellspecific probability that nonlabor income is positive. The calculation of CPP/QPP entitlement involves constructing an individual’s and their spouse’s earnings history over the contributory period. Given the age range in the sample, this is the period starting in 1966. The direct observations on T-4 earnings back to 1975 and predicted earnings in the period 1966–1974 are used. The dropout provisions for years between the sixty-fifth birthday and the commencement of retirement and for lowearnings months up to 15 percent of the contributory period are fully implemented. Disabilities or time spent in care of children are not observed, however, and therefore deletions for these reasons are not captured.26 This information in tandem with earnings projections for future years permits 26. Note that the dropout provisions for childcare came in to effect in 1977 under the QPP and 1978 under the CPP. The childbearing years of many females in our sample will have been prior to these dates.

Income Security Programs and Retirement in Canada

115

the construction of average pensionable earnings (APE) at all future retirement dates for any given individual. The reforms of the CPP/QPP system over the period are also accounted for, including the introduction of early retirement to the CPP, the retirement test on benefit receipt at ages sixty to sixty-four, and the actuarial adjustment to benefits for initiating benefit receipt at ages other than sixty-five, all in 1987. 2.5.2 Spousal Behavior A complete model of family labor supply is beyond the scope of this paper. The simplifying assumption that the spouse starts collecting any entitlement at the earliest age possible under the current rules of IS programs is made: For most of the sample period, this is age sixty-five for OAS and GIS, age sixty for the SPA, and age sixty for the CPP/QPP. For CPP/QPP benefits prior to age sixty-five and any income-tested benefit, the assumption implies a cessation of the spouse’s employment (i.e., retirement). Gruber (1999) and Baker and Benjamin (2000) provide estimates of age and employment profiles as well as employment hazards (the conditional probability of labor market exit) for older men and women over the sample period. This evidence provides some justification for this assumption about labor market exit rates in our analysis of the male sample in which spouses are females. On the other hand, this assumption may prematurely remove the male spouses of individuals in our sample of females from the labor market. This is unlikely to have a large effect on our estimates, as the independence across spouses in determination of most of the benefits means that spousal retirement is only a minor contributor to IS incentive calculations. 2.5.3 The Present Discounted Value of Income Security Wealth (ISW) Once these calculations of entitlement for each of the programs are made, the expected net present value of the family’s ISW associated with each retirement date is constructed. For single workers, this is the sum of future benefits discounted by time preference and survival probabilities. For married workers, we account for the likelihood of the joint survival of worker and the spouse, and the survivor provisions of the CPP/QPP and SPA, as described in more detail in Gruber (1999). We use a real discount rate of 3 percent and survival probabilities from the age- and sexspecific Canadian life tables from Statistics Canada (Statistics Canada 1984). 2.5.4 The One-Year Accrual Calculation We compute a number of different incentive variables using these estimates of the present discounted value (PDV) of ISW at all future retirement dates. The first is the one-year accrual of ISW resulting from an ad-

116

Michael Baker, Jonathan Gruber, and Kevin Milligan

ditional year of work. In the Canadian system, an additional year of work can raise ISW through the dropout provisions of the CPP/QPP, and it can either raise or lower ISW through the actuarial adjustment of benefits.27 The first of these factors is fairly small. In the Canadian system, the contributory period is a fixed age interval, so that other things equal, the marginal year replaces only 15 percent of a low-earnings year.28 Furthermore, this benefit is attenuated in the period examined here, by the real decline in the YMPE in the early 1970s. Initially set to match average wages, the YMPE declined dramatically in the initial years of the program, falling to 67 percent of the industrial composite wage in 1973.29 In 1975, both the CPP and QPP were amended to allow the YMPE to rise at a rate of 12.5 percent per annum until equality with average wages was reattained, but this did not occur until 1987. The upshot is that even individuals with low wages would have made the maximum contribution to the system in the 1970s. Therefore, a marginal year in the late 1980s and early 1990s would not necessarily dominate earlier years when the relative YMPE was much lower. Starting at age sixty (in years 1987 and forward for the CPP), an additional year of work also implies a delay in claiming and, thus, both an (upward) actuarial adjustment in benefits and reduction in the years of potential receipt. The actuarial adjustment between ages sixty and seventy is a linear 6 percent per annum. Whether this provides a net increment or decrement to ISW depends on the size of the adjustment relative to the expected number of years of remaining lifetime over which benefits will be collected. Given the linear nature of the adjustment, it will clearly become more and more unfair with age. This adjustment also interacts with the income testing of the GIS and SPA program. Low-income individuals may get some of the actuarial reduction in CPP/QPP benefits for early retirement back starting at age sixty-five through qualification for a higher GIS benefit. This further increases the disincentives for additional work after age sixty for those who are likely to be on the GIS program. Another way of looking at this is that the actuarial increase in benefits for delaying retirement may reduce entitlement to means-tested benefits starting at age sixty-five. For these individuals, therefore, the effective actuarial adjustment is less than 6 percent per year and therefore, much more likely to be unfair. 27. Here we use the value of the accrual, rather than normalizing the accrual by earnings to form an implicit tax or subsidy, as is done in Gruber (1999). We do this because we are controlling for earnings itself in the regression model, so that we, in essence, capture both pieces of the incentive to work (earnings and ISW accrual) separately. 28. This contrasts, for example, with the U.S. Social Security system where the substitution is one for one: For those with less than thirty-five years of work, the marginal year replaces a zero in the social security (SS) calculation; for those with thirty-five years or more of work, it replaces a full low-earnings year. 29. The YMPE equaled 99.8 percent of the industrial composite wage in 1966.

Income Security Programs and Retirement in Canada

117

2.5.5 The Peak Value Calculation Forward-looking measures of retirement incentives that involve the future path of ISW are also considered. The simple measure of one-year accrual only accounts for the immediate benefit to working an additional year. But an additional year of work also sustains the option of retiring at an even later date. The value of this choice can be important if there are large nonlinearities in the accrual profile. For example, if there is a small negative accrual at age fifty-nine, but a large positive accrual at age sixty, it would be misleading to say that the system induces retirement at age fiftynine; the disincentive to work at that age is dominated by incentives to work at age sixty. One way of capturing this possibility is to use the peak value calculation suggested by Coile and Gruber (2000). Rather than taking the difference between ISW today and next year, peak value takes the difference between ISW today and in the year in which the expected value of ISW is maximized. This measure therefore captures the trade-off between retiring today and working until a year with a much higher ISW: the option dollar value of continued work. In years beyond the year of peak expected value ISW, this calculation collapses to the simple one-year accrual variable. 2.5.6 The Option Value Calculation If a utility function that captures work preferences can be appropriately defined, then an approach that compares the utility of retirement at future dates is preferable. To explore this approach, the option value calculation of Stock and Wise (1990) is used. Here the utility of retiring today is compared to its value at the optimal retirement year in the future. The calculation uses a specification of the individual’s indirect utility function: (1)

R1

T

s
t

s
t

Vt (R)
∑ ps|t d st (ys ) g  ∑ ps|t d st [k  Bs(R)]g,

where R is the retirement date, d is the discount rate, p is the probability of being alive at some future date conditional on being alive today, y is income while working, B is retirement benefits, g is the parameter of risk aversion, k is a parameter to account for disutility of labor (k  1) T is maximum life length. In this model, additional years of work have three effects. First, they raise total wage earnings, increasing utility. Second, they reduce the number of

118

Michael Baker, Jonathan Gruber, and Kevin Milligan

years over which benefits are received, lowering utility. Third, they may raise or lower the benefit amount depending on the shape of the benefit function B(R). The last two effects receive greater weight than the first due to the disutility of labor. The optimal year for retirement is the year in which the utility gained from additional earnings is outweighed by the utility lost from the reduction in retirement income. The option value is the difference in utility from retirement at the optimal date and retirement today. Relative to peak value, option value has one major advantage and several disadvantages. The advantage is that the reference year in the peak value calculation (the year in which ISW is maximized) is arbitrary; there is no particular reason why this should be the year to which a given worker compares this year’s ISW in making their retirement decision. The option value approach more carefully specifies the optimal retirement date, and as such provides an economic basis for the reference year. Offsetting this advantage, however, are a number of disadvantages. The option value approach requires a particular specification of the indirectutility function and evaluation of its structural parameters. Also, earnings enter directly into the utility calculation and thus will drive some part of the variation of the option value across individuals. If earnings are in turn correlated with some unobserved component of tastes for retirement, the identification of the option value effects can be undermined. To implement this approach, values of k (the parameter for the disutility of labor), d (the discount rate), and g (the parameter of risk aversion) are taken from the literature. Following Stock and Wise (1990), k
1.5 and g
0.75, while d
0.03, following Coile and Gruber (2000). Sensitivity analysis suggests that the results are not dramatically different for sensible variations in these parameter values. 2.5.7 Sample Estimates of the Different Incentive Measures In table 2.2, we provide information on the distribution of the one-year accrual measure, by age, for the male sample. The median ISW rises to a peak at age sixty-one, then starts on a smooth descent. The median oneyear accrual is positive to age sixty, but becomes increasingly negative thereafter. The initial positive accrual is due to the dropout provisions, which work in favor of the worker with the median PDV of ISW. This effect is attenuated with age, however, as the implied larger CPP/QPP entitlement reduces GIS and SPA entitlement through the income test. The negative accruals start at age sixty-one as the early retirement provisions of the CPP/QPP come into play. Important here is that the linear CPP/QPP actuarial adjustment becomes increasingly unfair as the person delays retirement. There is an additional consideration for individuals who will eventually claim on GIS benefits (45 percent of OAS pensioners received GIS benefits in 1990). The higher CPP/QPP benefits gained by delaying retire-

Income Security Programs and Retirement in Canada Table 2.2

Age 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

119

The Distribution of the One-Year Accrual, by Age (male sample)

N

Median ISW ($)

Median ($)

10th Percentile ($)

90th Percentile ($)

57,387 61,167 63,818 65,091 65,541 60,051 52,539 45,898 39,711 33,776 27,118 13,932 9,008 6,812 5,480

107,533 108,702 109,531 110,065 110,385 110,603 110,633 110,384 109,737 108,684 107,287 104,356 101,022 97,304 93,265

1,169 829 534 320 218 30 –249 –648 –1,053 –1,397 –2,931 –3,334 –3,718 –4,040 –4,340

547 0 0 0 0 –1,259 –1,491 –1,987 –2,479 –2,903 –4,694 –5,095 –5,403 –5,749 –6,026

1,736 1,775 1,821 1,873 1,933 1,620 1,270 853 505 209 –838 –1,252 –1,532 –1,826 –2,131

SD

Median Tax Rate 1

Median Tax Rate 2

467 655 729 775 807 1,103 1,068 1,108 1,217 1,335 1,518 1,514 1,497 1,513 1,523

–0.049 –0.038 –0.027 –0.018 –0.012 –0.002 0.014 0.037 0.063 0.086 0.188 0.237 0.298 0.366 0.425

–0.042 0.005 0.036 0.037 0.038 0.077 0.085 0.085 0.096 0.186 0.367 0.413 0.396 0.327 0.340

Source: The numbers reported are the result of the ISW calculation described in the text. Median tax rate 1 is calculated from the analysis sample. Median tax rate 2 is from Gruber (1999). Notes: N = number of observations; ISW = Income Security Wealth; SD = standard deviation. All dollar values in 1998 U.S. dollars. Definitions of the different measures of ISW accrual are provided in the text.

ment, either through improving the earnings history or the actuarial adjustment, are offset by reduced income-tested GIS benefits at older ages. That is, the means-tested aspects of the income security system essentially “tax back” the actuarial adjustment, reducing the incentive to work past age sixty. The net effect of these factors is increasingly negative, as the median accrual falls from –$249 to –$1,397 between ages sixty-one and sixtyfour. The median accrual rises in absolute value at age sixty-five as OAS and GIS benefits come on line (there are SPA benefits in this range as well, given that the spouses of these male workers are typically several years younger). This jump reflects the fact that additional earnings after sixty-five will decrease the OAS, GIS, and SPA benefits through the income test for many workers. From age sixty-six to age sixty-nine, the accrual becomes more negative quickly, reflecting the increasingly unfair actuarial adjustment of CPP/QPP benefits and the fact that continued work sacrifices GIS benefits, through the income test, and OAS benefits, if earnings are high enough, through the income tax clawback. Overall, the loss in ISW wealth in table 2.2 is substantial between ages sixty-one and sixty-nine: The sum of the median accrual over these ages is –$21,709. In the next to last column the median tax or subsidy rates are reported. This is calculated as the median ratio of the one-year accrual to current af-

120

Michael Baker, Jonathan Gruber, and Kevin Milligan

Table 2.3

The Distribution of the Peak Value and Option Value Accrual, by Age (male sample) Peak Value

Age

Median ($)

10th Percentile ($)

90th Percentile ($)

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

2,997 1,998 1,164 807 577 47 –247 –647 –1,053 –1,397 –2,931 –3,334 –3,718 –4,040 –4,340

1,012 120 0 0 0 –1,258 –1,491 –1,987 –2,479 –2,902 –4,694 –5,095 –5,403 –5,749 –6,026

9,363 8,283 7,061 5,592 4,188 2,929 1,766 1,149 654 278 –838 –1,252 –1,532 –1,826 –2,131

Option Value

SD

Median ($)

10th Percentile ($)

90th Percentile ($)

SD

3,423 3,399 3,082 2,642 2,183 2,100 1,798 1,621 1,503 1,374 1,563 1,549 1,520 1,529 1,523

16,804 14,839 12,965 11,131 9,396 8,151 6,993 5,878 4,820 3,808 2,698 1,695 977 516 210

7,814 7,040 5,980 4,686 3,555 2,755 2,108 1,424 835 381 97 –2,186 –4,177 –4,865 –5,257

26,397 23,719 21,106 18,657 16,296 14,319 12,324 10,471 8,707 6,996 5,556 4,538 3,400 2,270 1,138

10,764 9,983 9,484 8,771 8,011 6,903 6,790 6,333 5,841 5,141 4,675 5,142 4,886 4,427 3,151

Source: The numbers reported are the result of the ISW calculation described in the text. Notes: N = number of observations; SD = standard deviation. All dollar values in 1998 U.S. dollars. Definitions of the different measures of ISW accrual are provided in the text.

ter tax earnings. After the initial period of subsidy, the tax rate becomes positive at age sixty-one. By age sixty-nine, the median tax rate is about 43 percent. These figures are somewhat lower than the estimates from Gruber (1999), presumably reflecting the fact that the dropout provisions have greater value here because we use real rather than simulated earnings histories. That is, if the real earnings history is more variable than a simulated earnings history, there will be more value to replacing lower-earnings years that will in turn increase the incentive to continue working. The median accrual masks considerable variation in the one-year accrual across individuals. For example, the standard deviation averages $1,122 across age groups. The accrual at the ninetieth percentile does not turn negative until age sixty-five. Presumably few of these individuals would qualify for GIS due to private pensions and savings. Many should also be in the clawback range for the OAS. As a consequence, we might not expect age sixty-five to be so pivotal for these individuals. That said, average nonlabor income is imputed to individuals, and this will be more inappropriate for people in the tails of the income distribution. Corresponding information for the peak value accrual is provided in table 2.3. Not surprisingly, the main difference from the one-year accrual is at ages fifty-five to fifty-nine. The median accrual is larger at these ages, but the change is fairly modest. For example, the median one-year accrual at

Income Security Programs and Retirement in Canada Table 2.4

Age 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

121

The Distribution of the One-Year Accrual, by Age (female sample)

N

Median ISW ($)

Median ($)

10th Percentile ($)

90th Percentile ($)

SD

Median Tax Rate 1

43,062 43,889 44,160 43,528 42,611 37,306 31,320 26,480 22,364 18,646 15,072 8,298 5,564 4,196 3,312

102,235 102,908 103,471 103,997 104,481 104,929 105,036 105,007 104,700 104,086 103,255 100,865 98,234 95,409 92,371

673 563 526 484 448 107 –30 –307 –614 –831 –2,391 –2,631 –2,825 –3,038 –3,321

116 79 20 0 0 –980 –1,130 –1,460 –1,815 –2,198 –3,962 –4,364 –4,748 –5,130 –5,389

1,426 1,408 1,445 1,483 1,510 1,455 1,115 777 560 403 –642 –797 –899 –1,157 –1,406

475 492 531 568 597 933 895 915 996 1,090 1,278 1,365 1,467 1,533 1,585

–0.049 –0.045 –0.045 –0.045 –0.046 –0.011 0.003 0.030 0.061 0.087 0.241 0.301 0.357 0.423 0.481

Source: The numbers reported are the result of the ISW calculation described in the text. Median tax rate 1 is calculated from the analysis sample. Notes: See table 2.2.

age fifty-seven is $534, while the median peak value accrual is $1,164. Correspondingly, adding together the median one-year accruals in table 2.3 between ages fifty-five and sixty, the distance to the peak is not that large. The primary inducement to continued work at older ages is the dropout provisions of the CPP/QPP, which, as previously explained, are modest and attenuated in the period that we examine due to the real decline in YMPE over the 1970s. That is, continued work may not qualify the individual for a larger CPP/QPP entitlement. Furthermore, the CPP/QPP is only one of three components of ISW. Therefore, we might expect the financial option value of continued work to be modest at older ages.30 Note that once age sixty is reached, the peak value calculation is the same as the oneyear calculation for most individuals, since they have already reached their peak. Table 2.3 also contains information on the option value accrual. Here the accrual is positive throughout the age range, reflecting the fact that the median optimal age of retirement by this measure is at age seventy or seventy-one. The magnitudes of these numbers are difficult to interpret as they are in units of utility. In table 2.4 we present corresponding information on the one-year ac30. In contrast, Coile and Gruber (2000) report large differences between one-year and peak value accrual for the United States. This is not surprising because, as previously explained, the dropout provisions of the U.S. Social Security system can lead to large changes in SS wealth with work at older ages.

122

Michael Baker, Jonathan Gruber, and Kevin Milligan

crual in the female sample. The age profile of the one-year accrual largely reflects the same factors as the profile for males (e.g., dropout provisions of the CPP/QPP and the straight-line actuarial adjustment of CPP/QPP benefits). One might reason that females’ lower earnings entitle them to smaller CPP/QPP benefits, and therefore their ISW entitlement should be smaller. This effect is attenuated by the large relative decline in the YMPE over the 1970s and the fact that CPP/QPP is only one of three components of the IS package. Another consideration is that the longer lifespan of females means that the actuarial adjustment for delayed receipt of CPP/QPP benefits will be fairer for this group. We can see this in the smaller proportionate changes in the accrual over the age profile. For males, the median accrual increases (in absolute value) by $1,427 between ages sixty and sixty-four and by $1,409 between ages sixty-five and sixty-nine. For females, the corresponding changes are $938 and $931 respectively. Also it is important to remember that the sample individuals are selected by conditioning on positive earnings in the first year the individual enters the sample. These, therefore, are a select sample of females who worked at older ages, yet belong to birth cohorts that historically have not had high participation rates.31 The peak value and option statistics presented in table 2.5 are also very similar to their counterparts for males. Again the early peak in ISW and the lack of any strong variation in accrual mean there are only modest differences between the one-year and peak value calculations. In figures 2.2 and 2.3 we graph the age profiles of the median of the various measures of accrual. The relative levels are meaningless, as the option value is measured in utility units. A comparison of the age profiles of the different measures of accrual, however, is meaningful, highlighting the differences among the measures. For both sexes the one-year and peak value have very similar age profiles. The median accruals decline over the full age range, with increases in the rate of decline noticeable at ages sixty and sixty-five. The difference in the peak value measure is entirely in the age range fifty-five to fifty-nine. The option value calculation provides a very different profile, as accrual declines continuously at a decreasing rate over the age range. 2.6 Empirical Framework and Estimation Results 2.6.1 The Empirical Framework The regression equation relates the retirement decisions of individuals to their demographic and economic characteristics as well as their ISW. 31. The participation rate of forty-five- to sixty-four-year-old females was 41 percent in 1976, 48 percent in 1986, and 58 percent in 1996 (the source is CANSIM, available at http:// dc2.chass.utoronto.ca/cansim2/English/index.html).

Income Security Programs and Retirement in Canada Table 2.5

123

The Distribution of the Peak Value and Option Value Accrual, by Age (female sample) Peak Value

Age

Median ($)

10th Percentile ($)

90th Percentile ($)

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

2,684 2,205 1,736 1,235 741 194 –14 –303 –612 –831 –2,391 –2,631 –2,825 –3,038 –3,321

708 399 182 44 0 –979 –1,130 –1,460 –1,815 –2,198 –3,962 –4,364 –4,748 –5,130 –5,389

7,407 6,639 5,771 4,671 3,562 2,769 2,019 1,382 944 506 –642 –795 –899 –1,157 –1,406

Option Value

SD

Median ($)

10th Percentile ($)

90th Percentile ($)

SD

2,683 2,576 2,345 2,041 1,785 1,821 1,586 1,399 1,284 1,164 1,363 1,451 1,519 1,557 1,585

10,589 9,431 8,253 7,135 6,038 5,230 4,531 3,795 3,118 2,427 1,626 960 576 283 114

2,994 3,153 2,859 2,377 1,836 1,421 1,149 828 517 242 –1,691 –2,309 –3,560 –4,343 –4,867

19,840 17,416 15,323 13,323 11,498 10,193 8,910 7,624 6,322 5,203 3,991 3,057 2,245 1,521 747

7,051 6,254 5,484 4,786 4,257 3,859 3,613 3,262 2,787 2,543 2,491 2,248 2,351 2,404 2,278

Source: See table 2.3. Note: See table 2.3.

The ISW plays a dual role in the decision. First, higher levels of ISW have wealth effects that cause individuals to retire earlier; more wealth through IS programs will lead to increased consumption of all goods, including leisure. Second, however, higher accruals of ISW from additional work should have a substitution effect that leads to later retirement; if there is a large financial incentive to additional years of work, then individuals will retire later. Therefore, equations are estimated of the form (2)

R it
0  1ISWit  2ACCit  3AGEit  4EARNit  5APEit  6SPEARNit  7SPAPE  8RPPit  9Xit  it ,

where

• R it is a variable which equals 1 in the year of retirement and 0 otherwise;

• ISWit is the expected PDV of ISW in year t; • ACCit is one of the measures of accrual outlined above (the simple one-year accrual, the peak value accrual, or the utility-based option value accrual); • AGEit represents a set of dummy variables for each age in our sample and a measure of the difference in ages across spouses;

The age profile of the median accrual, peak value, and option value in the male sample

Note: The profiles are graphs of the numbers reported in tables 2.2 and 2.3.

Fig. 2.2

The age profile of the median accrual, peak value, and option value in the female sample

Note: The profiles are graphs of the numbers reported in tables 2.4 and 2.5.

Fig. 2.3

126

Michael Baker, Jonathan Gruber, and Kevin Milligan

• EARNit and APEit represent cubics in measures of the individual’s projected earnings in year t and their APE (for CPP/QPP calculations); • SPEARNit and SPAPEit are the corresponding variables of any spouse; • RPPit is the measure of the probability of RPP coverage at the 3-digit industry level,32 • Xit are a set of additional control variables, including a dummy variable for marital status; a quadratic in tenure on the job and a dummy variable that equals one if tenure is censored at 1978; a quadratic in the individual’s and their spouse’s labor market experience measured as the number of years of positive T-4 earnings between 1975 and year t; eleven industry dummies; and dummies for six categories of establishment size and province and year effects. To capture potential nonlinear relationships between earnings and retirement decisions, we include a full set of interactions between the cubics in EARNit and APEit , and SPEARNit and SPAPEit . The equations are estimated separately for males and females as a probit. As mentioned at various points of this discussion, the Canadian IS system went through a number of reforms in our period of analysis. This is a distinct advantage of evaluating the retirement incentives of the Canadian IS system relative to other countries. These policy interventions potentially provide more credible identifying variation for the parameters of the ISW incentive variables than the variation across individuals due to differences in earnings histories, family circumstances, and tastes. This advantage is highlighted in the reforms of the CPP system that do not have a counterpart in the QPP system in the period 1985–1995. In this case, the residents of Quebec provide a control group for the effects of the reform. Of particular importance here is the introduction of early retirement to the CPP system in 1987. A similar reform of the QPP was accomplished in 1984.33 2.6.2 Sample Characteristics In tables 2.6 and 2.7, some average demographic and job characteristics for the male and female samples are presented, which are calculated over all the observations.34 The average age in our male sample is almost sixty years old. Fifty-six percent of the sample observations are for married men. Average projected earnings for the males are $19,503, while the average APE is $19,847. The corresponding averages for their spouses are 32. The standard errors here are potentially biased due to a correlation of the error term across individuals within the three-digit industry code (the “grouped data problem”). Correcting for this bias would lead to larger estimated standard errors on the parameter on RPP. 33. Another reform during our sample period is the introduction of the clawback of OAS benefits in 1989. This applied to individuals in all parts of the country. 34. An alternative would be to calculate the means over individuals.

Income Security Programs and Retirement in Canada Table 2.6

127

Summary Statistics for the Male Sample

Retired Probability of RPP Married Tenure Tenure Censored Experience Spouse’s Experience Age Age Difference Projected Earnings Projected Spousal Earnings APE Spouse’s APE No. of observations No. of individuals

Mean

SD

0.148 0.582 0.558 8.763 0.441 15.235 5.144 59.779 2.067 $19,503 $3,033 $19,847 $5,393 607,329 121,204

0.355 0.256 0.497 4.500 0.496 5.162 6.848 3.375 3.820 29,088 7,732 4,525 7,996

Notes: SD = standard deviation. The reported statistics are means (averages) calculated over all observations in the male and female datasets, respectively (rather than over all individuals). All dollar values are in 1998 U.S. dollars. Definitions of all variables are provided in the appendix. Table 2.7

Summary Statistics for the Female Sample

Retired Probability of RPP Married Tenure Tenure Censored Experience Spouse’s Experience Age Age Difference Projected Earnings Projected Spousal Earnings APE Spouse’s APE No. of observations No. of individuals

Mean

SD

0.151 0.428 0.404 8.660 0.374 14.450 5.279 59.488 –0.684 $11,458 $4,050 $13,871 $7,500 389,808 77,845

0.358 0.262 0.491 4.423 0.484 5.551 7.187 3.365 2.718 8,433 12,897 6,924 10,189

Note: See table 2.6.

$3,033 and $5,393 respectively.35 Finally, the average probability of RPP coverage across observations is 58 percent. 35. Note that the averages for the spouses are much less than the averages for the males. This is part because we calculate these averages over all males, including those who have no spouse or whose spouse does not work. In these cases, spousal earnings will be 0, thus lowering the average.

128

Michael Baker, Jonathan Gruber, and Kevin Milligan

The average age in our female sample is 59.5 years. Both the married and RPP proportions are lower than in the male sample at 0.40 and 0.43, respectively. Average projected earnings are $11,458, while the average APE is $13,871. It is interesting to note that these female workers have similar tenure (length of time with the current employer) and experience (number of years in the labor market) as their male counterparts, but this is likely because our measures are only since 1978 (tenure) and 1975 (experience); over the full working life, presumably these means are much lower for women. The reason that the means of spousal earnings are lower than own earnings for women is because average spousal earnings are calculated using zeros for nonmarried working women. In figures 2.4 and 2.5, we present estimates of the retirement hazard for males and females in our sample, which is calculated across all birth cohorts in the sample and all years. The hazard for each sex displays a distinct jump at age fifty-nine to sixty, which is the point of first eligibility for CPP/ QPP benefits. This is also the age at which individuals are first eligible for the SPA. It then increases modestly at ages sixty-one through sixty-four. Finally, there is a spike at age sixty-five. Relative to the profile for males, the hazard for females is slightly higher at ages younger than sixty, rises more quickly and higher at fifty-nine to sixty, remains above its male counterpart until age sixty-four, and has a smaller spike at age sixty-five. 2.6.3 Retirement Regression Results In table 2.8 we present our basic regression results. There are three major groupings (represented in columns) corresponding to the three incentive measures used—accrual, peak value, and option value. In the first column of each grouping we report results from the specification that include linear controls for age, while the second column is for the specification that includes age dummies.36 Since the coefficients from these probit regressions are difficult to interpret directly, we report the effect of a $10,000 change in ISW or a $1,000 change in the accrual measure. Our results are uniformly supportive of an important role for IS incentives in determining retirement. In every column, there is a positive and significant coefficient on ISW, and a negative and significant coefficient on the accrual measure. When age dummies are excluded, we find that a $10,000 rise in ISW raises the odds of retirement by 1.4 to 2 percentage points, from a base of 14.8 percent. When age dummies are included, the effect of ISW falls considerably, however, so that a $10,000 rise in ISW raises retirement rates by only 0.23 to 0.51 percentage points. The incentive measure effects also are reduced when age dummies are 36. See Baker, Gruber, and Milligan (2003) for other specifications. Controls for past and current earnings change the magnitude of the estimated parameters substantially. This emphasizes the importance of including rich controls as we do in this chapter.

The retirement hazard and age dummies for males

Source: Authors’ calculations from the analysis sample of males (see the appendix); regression results from table 2.8.

Fig. 2.4

The retirement hazard and age dummies for females

Source: Authors’ calculations from the analysis sample of females (see the appendix); regression results from table 2.9.

Fig. 2.5

Table 2.8

Retirement Probits (male sample) Accrual Model

ISW $10,000 change ACCRUAL $1,000 change RPP MARRIED TENURE TENURESQ TENURE CENS EXP EXP SQ SPOUSE EXP SPOUSE EXPSQ AGE AGEDIFF AGE56 AGE57 AGE58 AGE59 AGE60 AGE61 AGE62 AGE63 (continued )

Peak Value Model

Option Value Model

Linear Age

Age Dummies

Linear Age

Age Dummies

Linear Age

Age Dummies

0.093 (0.003) 1.97 –1.183 (0.023) –2.21 0.125 (0.013) –0.539 (0.020) –0.029 (0.002) 0.002 (0.0001) 0.026 (0.006) –0.022 (0.002) 0.000 (0.0001) –0.027 (0.003) 0.001 (0.0001) 0.013 (0.001) 0.003 (0.001)

0.025 (0.004) 0.51 –0.798 (0.028) –1.52 0.123 (0.013) –0.182 (0.022) –0.031 (0.002) 0.002 (0.0001) 0.026 (0.007) –0.020 (0.002) 0.000 (0.0004) –0.028 (0.001) 0.001 (0.0001)

0.077 (0.003) 1.60 –0.577 (0.017) –1.11 0.124 (0.013) –0.454 (0.020) –0.034 (0.002) 0.002 (0.0001) 0.028 (0.006) –0.018 (0.002) 0.000 (0.0001) –0.029 (0.003) 0.001 (0.0001) 0.033 (0.001) 0.001 (0.001)

0.022 (0.004) 0.45 –0.345 (0.018) –0.67 0.123 (0.013) –0.173 (0.022) –0.033 (0.002) 0.002 (0.0001) 0.027 (0.007) –0.018 (0.002) 0.000 (0.0001) –0.029 (0.003) 0.001 (0.0001)

0.067 (0.004) 1.38 –0.284 (0.011) –0.56 0.124 (0.013) –0.411 (0.021) –0.030 (0.002) 0.002 (0.0001) 0.034 (0.006) –0.014 (0.002) 0.0000 (0.0001) –0.027 (0.003) 0.001 (0.0001) 0.039 (0.001) –0.001 (0.001)

0.012 (0.004) 0.23 –0.315 (0.009) –0.61 0.126 (0.013) –0.110 (0.023) –0.029 (0.002) 0.002 (0.0001) 0.032 (0.007) –0.015 (0.002) 0.000 (0.0001) –0.028 (0.003) 0.001 (0.0001)

–0.003 (0.001) –0.036 (0.011) –0.053 (0.011) 0.002 (0.012) 0.050 (0.012) 0.199 (0.012) 0.162 (0.014) 0.162 (0.015) 0.171 (0.016)

–0.003 (0.001) –0.036 (0.011) –0.059 (0.012) –0.015 (0.012) 0.021 (0.012) 0.189 (0.013) 0.162 (0.14) 0.179 (0.015) 0.203 (0.016)

–0.004 (0.001) –0.050 (0.011) –0.095 (0.012) –0.079 (0.012) –0.071 (0.013) 0.101 (0.013) 0.066 (0.014) 0.074 (0.015) 0.089 (0.016)

132 Table 2.8

Michael Baker, Jonathan Gruber, and Kevin Milligan (continued) Accrual Model Linear Age

AGE64

Linear Age

0.309 (0.017) 0.914 (0.020) 0.527 (0.021) 0.190 (0.024) 0.100 (0.025) 0.063 (0.027)

AGE65 AGE66 AGE67 AGE68 AGE69 Pseudo R 2 OTHER CONTROLS

Age Dummies

Peak Value Model

0.103 Yes

0.116 Yes

Age Dummies

Option Value Model Linear Age

0.353 (0.017) 1.028 (0.019) 0.658 (0.021) 0.338 (0.023) 0.262 (0.024) 0.238 (0.026) 0.100 Yes

0.115 Yes

Age Dummies 0.227 (0.018) 0.928 (0.019) 0.555 (0.020) 0.229 (0.022) 0.148 (0.024) 0.118 (0.025)

0.099 Yes

0.116 Yes

Notes: AGEDIFF (the difference in ages between the individual and his spouse) is coded 0 for singles. The statistics reported in the rows “$10,000 change” and “$1,000 change” are the percentage point change in the probability of retirement for the indicated change in ISW or accrual. Other control variables are 11 industry dummies; dummies for 6 categories of establishment size; province and year effects; a cubic in both own and spouse’s predicted earnings and APE; and a full set of interactions between these cubics. The estimated parameters on these additional variables are not reported. Standard errors are in parentheses.

included, but they are less sensitive than is ISW. For the one-year accrual, we find that a $1,000 rise leads to a 2.21 percentage point decline in retirement without age dummies and a 1.52 percentage point decline with age dummies. The effects of a change in peak value are roughly half as large. The effects of option value, a 0.6 percentage point decline in retirement for a $1,000 increase, are essentially invariant to the inclusion of age dummies. The control variables themselves have their expected signs. Of particular interest is that the effect of our measure of RPP-coverage probability is positive and statistically significant. This is consistent with a wealth effect as result of the additional savings represented by the RPP entitlement. Marriage and larger age differences (Agediff) between spouses have a negative relationship with the probability of retirement. Conditional on age, both experience and tenure reduce the probability of retirement, although in each case at a decreasing rate. Older men are more likely to retire. There is a clear pattern of highly significant age dummies from age sixty onwards. The estimates of the year effects (not reported) reveal that most of the time effects are cyclical. Also, the probability of retirement displays a vague U shape with establishment size: larger probabilities in the smallest and

Income Security Programs and Retirement in Canada

133

largest establishments. The higher probability of retirement in the largest establishments may be partly an (unobserved) RPP effect. It is important to note that the ISW and accrual variables may not fully capture the impacts of the IS system. This is illustrated in figures 2.4 and 2.5, in which we report the baseline retirement hazard, along with the estimated pattern of age dummies, from our three different models that include age dummies.37 What is immediately apparent is that while the IS variables in our model account for some of the age effects in the hazard, the basic profile of the effects is still very evident. There are two conclusions one can draw from this observation. The first is that we have fully captured the impact of the IS system on behavior, and the age-specific pattern of retirement that remains is due to nonlinear changes in the taste for leisure with age or institutions, such as mandatory retirement, that are not otherwise captured in our model. The second is that the spike at age sixty-five is capturing the more fundamental aspects of the response to the IS system that are not captured by our regressors, so that we are underestimating the full impact of the system on retirement decisions. As we discuss below, which interpretation is correct has important implications for assessing the policy implications of our findings. Table 2.9 and figure 2.5 contain parallel results for women. The findings for ISW and for one-year accrual are remarkably similar to those for men. This is striking and may suggest that, conditional on being in the labor force at older ages, women can be viewed as responding to financial incentives in the same way as do men. This finding is consistent with Coile’s (1999) findings for the United States. The effects are a bit smaller for women, however, for the peak and option value measures. 2.7 Policy Simulations The preceding results are difficult to interpret in a vacuum. Thus, for both evaluation and comparison with the other results in this volume, we consider the implications of our findings for two significant reforms to the Canadian system. The first is an increase of three years in the age of both early and normal entitlement for IS programs, holding all other aspects of the system constant. The second is a shift to a new, simpler system that features only one program and has an early retirement age of sixty and a normal retirement age of sixty-five. This system provides for all persons, at age sixty-five, a benefit equal to 60 percent of their projected age sixty earnings; we use projected earnings so that we have a value even for those who retire before age sixty. There is an actuarial reduction of 6 percent per year for 37. The age dummies are constructed as the incremental probability implied by the probit coefficient for each age. The probability is calculated at the mean of the other variables.

Table 2.9

Retirement Probits (female sample) Accrual Model

ISW $10,000 change ACCRUAL $1,000 change RPP MARRIED TENURE TENURESQ TENURE CENS EXP EXP SQ SPOUSE EXP SPOUSE EXPSQ AGE AGEDIFF AGE56 AGE57 AGE58 AGE59 AGE60 AGE61 AGE62 AGE63 AGE64

Peak Value Model

Option Value Model

Linear Age

Age Dummies

Linear Age

Age Dummies

Linear Age

Age Dummies

0.091 (0.005) 1.96 –1.074 (0.033) –2.06 0.164 (0.016) –0.254 (0.020) –0.009 (0.002) 0.001 (0.0001) –0.038 (0.008) –0.038 (0.002) 0.001 (0.00004) –0.010 (0.007) 0.000 (0.007) 0.010 (0.002) 0.015 (0.001)

0.022 (0.005) 0.45 –0.653 (0.040) –1.27 0.161 (0.017) –0.079 (0.021) –0.008 (0.002) 0.001 (0.0001) –0.037 (0.008) –0.038 (0.002) 0.001 (0.00004) –0.010 (0.007) 0.000 (0.007)

0.082 (0.005) 1.74 –0.306 (0.022) –0.61 0.160 (0.016) –0.247 (0.020) –0.012 (0.002) 0.001 (0.0001) –0.033 (0.008) –0.039 (0.002) 0.001 (0.00004) –0.009 (0.007) 0.000 (0.007) 0.030 (0.002) 0.010 (0.001)

0.030 (0.005) 0.61 –0.089 (0.022) –0.18 0.158 (0.017) –0.114 (0.021) –0.008 (0.002) 0.001 (0.0001) –0.032 (0.008) –0.038 (0.002) 0.001 (0.00004) –0.009 (0.007) 0.000 (0.007)

0.076 (0.005) 1.63 –0.116 (0.018) –0.24 0.157 (0.016) –0.238 (0.021) –0.012 (0.002) 0.001 (0.0001) –0.030 (0.008) –0.038 (0.002) 0.001 (0.00004) –0.009 (0.007) 0.000 (0.007) 0.037 (0.002) 0.009 (0.001)

0.017 (0.006) 0.34 –0.138 (0.018) –0.28 0.159 (0.017) –0.070 (0.022) –0.007 (0.002) 0.001 (0.0001) –0.031 (0.008) –0.037 (0.002) 0.001 (0.00004) –0.010 (0.007) 0.000 (0.007)

0.004 (0.001) –0.010 (0.013) –0.010 (0.013) 0.007 (0.013) 0.063 (0.014) 0.227 (0.015) 0.146 (0.016) 0.159 (0.018) 0.169 (0.019) 0.250 (0.020)

0.003 (0.001) –0.010 (0.013) –0.012 (0.013) –0.001 (0.014) 0.049 (0.014) 0.234 (0.015) 0.157 (0.016) 0.183 (0.018) 0.204 (0.019) 0.294 (0.021)

0.002 (0.001) –0.015 (0.013) –0.023 (0.013) –0.017 (0.014) 0.028 (0.015) 0.214 (0.015) 0.136 (0.017) 0.161 (0.018) 0.180 (0.020) 0.268 (0.021)

Income Security Programs and Retirement in Canada Table 2.9

135

(continued) Accrual Model Linear Age

AGE65

Linear Age

0.844 (0.023) 0.456 (0.025) 0.169 (0.028) 0.053 (0.030) 0.046

AGE66 AGE67 AGE68 AGE69 Pseudo R 2 OTHER CONTROLS

Age Dummies

Peak Value Model

0.107 Yes

0.117 Yes

Age Dummies

Option Value Model Linear Age

0.970 (0.022) 0.598 (0.024) 0.324 (0.027) 0.222 (0.029) 0.232 0.104 Yes

0.116 Yes

Age Dummies 0.947 (0.022) 0.572 (0.024) 0.292 (0.027) 0.184 (0.029) 0.188

0.104 Yes

0.116 Yes

Note: See table 2.8.

early claiming and an actuarial increase of 6 percent per year for later claiming. For each reform, we consider nine different simulations: There are three simulation approaches for each of the three incentive measures (accrual, peak value, and option value). The first simulation approach (which we label “S1”) uses the estimates from the model estimated with a linear control for age. In this case, the simulation operates by incrementing the incentive measures only according to the policy change. Thus, we simply recompute IS wealth at each age under these new rules, and then predict retirement rates by applying these new ISW and accrual values to our estimated coefficients. The second simulation approach (we label “S2”) is the same as the first, but it uses the model with age dummies included. The difference between these approaches is highlighted by the dramatic reduction in the ISW coefficient when the linear age term is replaced by age dummies. There are two possible reasons for this reduction. The first is that tastes for leisure do not increase linearly with age and that the age dummies more appropriately pick up the nonlinearities. Since ISW is nonlinear in a corresponding manner with age, the inclusion of age dummies reduces the coefficient significantly. Alternatively, the age dummies could pick up the effects of other institutions, such as RPPs or mandatory retirement, in certain sectors. The second possibility, however, is that the jump in retirement at age sixty-five is due to the IS system itself, and not inherent tastes for leisure, so that including age dummies, to some extent, robs the ISW term of its true explanatory value.

136

Michael Baker, Jonathan Gruber, and Kevin Milligan

Table 2.10

Average Retirement Rates in Simulations Simulated Reform Plus Three Years

Base retirement rate Accrual S1 S2 S3 Peak value S1 S2 S3 Option value S1 S2 S3 Base retirement rate Accrual S1 S2 S3

Common System

Males 0.150

0.150

0.096 0.125 0.105

0.326 0.200 0.174

0.104 0.129 0.105

0.293 0.188 0.158

0.114 0.133 0.112

0.264 0.176 0.152

Females 0.151

0.151

0.107 0.134 0.119

0.222 0.160 0.143

We are statistically unable to distinguish these views, and the S2 approach errs on the side of the first explanation (nonlinear tastes for leisure) over the second (IS-related, nonlinear age effects). Thus, we also pursue a third simulation (S3) approach using the model with age dummies, but incrementing the age dummies at ages sixty and older by three years for the first policy change. That is, we not only recompute the values of our IS measures arising from the change, but we also recompute the value of the age dummies themselves, so that the age-sixty dummy takes on the estimated value of the age-sixty-three dummy, the age-sixty-one dummy takes on the estimated value of the age-sixty-four dummy, and so forth. This approach errs on the side of the second explanation by attributing all of the age dummy effects to the IS system. Clearly, the truth will be between these two extremes. We present the results of these eighteen simulations in three forms. First, in table 2.10, we show the effects of each reform on average retirement rates in our sample. Then, in each of the figures 2.6–2.17 we show the impact of each reform on hazard rates and then on the cumulative probability of retirement, each relative to the relevant model’s predicted baseline. We estimate all eighteen models for men; for women, we just present the results from the accrual models. Our findings suggest that these reforms would have very significant impacts on the retirement decisions of older Canadians. Starting with the first

Income Security Programs and Retirement in Canada

137

A

B

Fig. 2.6 S1 on males using accrual estimates: A, Simulated hazard; B, Cumulative probability

reform, we find that there is a substantial reduction in retirement when the early and normal eligibility ages are increased. For example, in table 2.10 the average retirement rate is lower regardless of the simulation method or model specification. The reduction ranges between 11 percent (option value, model S2) and 36 percent (accrual value, model S1). The pattern of results is fairly similar across the three types of accrual measures, although the largest effects are for the accrual measure and the smallest are for option value. Comparing across the three simulation methods, we find the

138

Michael Baker, Jonathan Gruber, and Kevin Milligan

A

B

Fig. 2.7 S2 on males using accrual estimates: A, Simulated hazard; B, Cumulative probability

largest effects from model S1, which yields findings very close to model S3, and the results are significantly mitigated under model S2. Thus, either controlling linearly for age or assuming that the age dummies capture IS effects yields similar results. The distribution of the changes across the age profile is revealed in figures 2.6 through 2.17. With the accrual model, regardless of the simulation method, much of the reduction in retirement occurs in the years just before the new normal retirement age of sixty-eight. For example, in figure 2.6,

Income Security Programs and Retirement in Canada

139

A

B

Fig. 2.8 S3 on males using accrual estimates: A, Simulated hazard; B, Cumulative probability

panel A, the jump in the baseline hazard at age sixty-five shifts over to age sixty-eight under the reform, although there are also sizable reductions in the hazard at earlier ages. In figures 2.7 and 2.8, panels A, the fall in the hazard is even more concentrated at ages sixty-four through sixty-seven. The smaller overall impact of the reform under S2 would appear to result from the greater congruence of the baseline and simulated hazards at earlier ages and the strong reduction in the hazard at age sixty-five. The agesixty-five spike in the hazard remains under the reform because we do not shift the age dummies. Therefore, much of the strength of this reform would appear to result from the change in behavior between the old and

140

Michael Baker, Jonathan Gruber, and Kevin Milligan

A

B

Fig. 2.9 S1 on males using peak value estimates: A, Simulated hazard; B, Cumulative probability

new normal retirement ages. Other than that, the reform leads to a simple shift out in the cumulative density over the retirement interval (sixty to sixty-nine) suggesting that the same incentives at work on a higher base. A very similar story is told in the results for females (figures 2.15–2.17). The peak value model simulations reveal a similar story in S1 (figure 2.9, panel A), but in S2, the reduction in the hazard appears more evenly distributed across ages sixty through sixty-six. Again, with the peak value model, the sharp shift to the right in the hazard in S3 (figure 2.10, panel A) is a result of the shifting of the age dummies. Using the option value model, the effects are similar in direction, but more modest in magnitude. This re-

Income Security Programs and Retirement in Canada

141

A

B

Fig. 2.10 S2 on males using peak value estimates: A, Simulated hazard; B, Cumulative probability

flects the smaller estimated coefficients for the incentive variables in the option value regressions. This evidence, therefore, suggests that a reform to change eligibility ages could have a very large impact on retirement rates under the Canadian IS system. However, the size of the shift in the predicted hazards depends strongly on whether one assumes the age dummies should or should not be shifted. The second reform, moving to a single program with program parameters very similar to the CPP/QPP, has just the opposite effect on average re-

142

Michael Baker, Jonathan Gruber, and Kevin Milligan

A

B

Fig. 2.11 S3 on males using peak value estimates: A, Simulated hazard; B, Cumulative probability

tirement rates (table 2.10): They increase. The effects here are much more sensitive to the simulation approach employed. For model S1, the retirement rate more than doubles using the one-year accrual model and almost doubles in the peak value and option value models. For models S2 or S3, however, the effects are much more modest, with roughly a 20 percent rise in retirement under model S2 and only about 10 percent under S3. There are at least two effects at work here. First the levels of ISW are uniformly higher under the reform. Since ISW is positively related to retirement this would raise retirement rates at all ages. The second is that accrual effects arising, for example, from the actuarial adjustment of benefits, are

Income Security Programs and Retirement in Canada

143

A

B

Fig. 2.12 S1 on males using option value estimates: A, Simulated hazard; B, Cumulative probability

potentially stronger here as they apply to the entire benefit (recall that in the current system these only arise for the CPP/QPP component of the total retirement benefit). At the median, this will lead to larger negative accruals after age sixty than is the case under the current system, which will also increase the hazard at these ages. Again the figures provide the details. Regardless of the model used, in S1 there are enormous increases in the retirement hazard at all ages. In each case the entire age profile of the hazard shifts upwards 0.1 or more. The shifts in the hazards in S2 and S3 are far more modest, although the wide distribution of the effect across the age profile remains.

144

Michael Baker, Jonathan Gruber, and Kevin Milligan

A

B

Fig. 2.13 S2 on males using option value estimates: A, Simulated hazard; B, Cumulative probability

In each of the models, the difference between the reform and baseline hazards is fairly uniform between ages fifty-five and fifty-nine. In this age range, only the higher level of ISW under the reform comes into play: Greater wealth leads to higher retirement rates. The effects are notably smaller using the option value model (figures 2.12, 2.13, and 2.14, panels A) reflecting the smaller coefficient on the level of ISW in the option value regression. After age sixty, the distance between the reform hazard and the baseline hazard increases monotonically. The negative accruals resulting from the unfairness of the actuarial adjustment are exaggerated under the reform

Income Security Programs and Retirement in Canada

145

A

B

Fig. 2.14 S3 on males using option value estimates: A, Simulated hazard; B, Cumulative probability

due to the 60 percent replacement rate compared to the 25 percent replacement rate in the CPP/QPP. At higher ages, the unfairness of the adjustment increases, which explains the growth of the gap between the hazards. In the simulation using the accrual model (figures 2.6, 2.7, and 2.8, panels A), this growth in the gap between the (common) reform and baseline hazards is most stark. This is a result of the much larger estimated parameter on the accrual incentive measure compared to the estimated coefficients on the peak and option value measures. Overall, the common reform simulations uncover two important messages. First, the wealth effects of moving to a more generous IS system

146

Michael Baker, Jonathan Gruber, and Kevin Milligan

A

B

Fig. 2.15 S1 on females using accrual estimates: A, Simulated hazard; B, Cumulative probability

could have a large impact on retirement rates. Second, any unfairness in actuarial adjustments would have a much larger impact on retirement rates if all Canadian IS benefits were subject to the adjustments rather than just the CPP/QPP. 2.8 Conclusions The aging of the Canadian population, coupled with a trend towards earlier retirement, places financial stress on the Canadian IS system. It is

Income Security Programs and Retirement in Canada

147

A

B

Fig. 2.16 S2 on females using accrual estimates: A, Simulated hazard; B, Cumulative probability

important, therefore, to understand how this complicated system affects retirement decisions. Other papers have suggested some role for IS programs, but no previous paper has taken a comprehensive look at how this panoply of programs affects retirement in Canada. This paper accomplishes this task, using an excellent data source matched to a rich simulation model that allows us to assign IS entitlements to our sample workers. Also, a variety of parameterizations of the incentives for retirement are considered. We have two findings of importance. First, for the typical worker, the IS system provides increasingly strong disincentives to work after age sixty.

148

Michael Baker, Jonathan Gruber, and Kevin Milligan

A

B

Fig. 2.17 S3 on females using accrual estimates: A, Simulated hazard; B, Cumulative probability

Workers actually see the present discounted value of there IS entitlement fall due to additional work after age sixty-one, and by age sixty-nine the reduction in IS entitlement amounts to 43 percent of what they would earn in that year. Second, there is a significant impact of these disincentives on work decisions. Using both one-year-forward measures and measures that account for the entire future path of incentives, we estimate that workers with larger returns to additional work are less likely to leave the labor force. This finding in turn has implications for policy evaluation. Evaluations of changes to the Canadian IS system cannot be done assuming static retirement decisions; these evaluations must build in the type of dynamic retirement behavior that we observe. We illustrate these effects through two

Income Security Programs and Retirement in Canada

149

reforms and show that these changes can have important effects on the retirement decisions of older Canadian workers.

Appendix Data Descriptions and Sample Definitions Longitudinal Worker File (LWF) Data: List of Variables from LWF data AGEDDIFF: a variable recording the difference in age between an individual and their spouse (in years) AGE55–AGE69: a dummy variable that equals 1 if an individual is the indicated age and 0 otherwise (AGE55 is the excluded variable) APE: a variable recording an estimate of an individual’s current APE Experience: a variable recording the number years since 1975 that an individual has had positive T-4 earnings Married: a dummy variable that equals 1 if the individual is married and 0 otherwise RPP: a variable that ranges between 0 and 1 recording the proportion of workers in an individual’s 3-digit industry that is a member of an RPP Tenure: a variable recording the number of years since 1978 that an individual has been with the current firm Tenure Censored: a dummy variable that equals 1 if an individual has been with their current firm continuously since 1978 Y85–Y95: a dummy variable that equals 1 in the indicated year and 0 otherwise (Y90 is the excluded variable) S04–S500p: a dummy variable that equals 1 for the indicated size of the workforce at the place of work and 0 otherwise; categories are 0–4, 5–19, 20–49, 50–99, 100–199, 200–499, and 500 (S5099 is the excluded variable) IND1–IND10: a dummy variable that equals 1 for the indicated industry of employment and 0 otherwise (IND1 is the excluded variable). The ten are IND1: manufacturing (standard industrial classification [SIC] 100 to 399) IND2: construction (SIC 400 to 449) IND3: storage and transportation (SIC 450 to 499) IND4: wholesale trade (SIC 500 to 599) IND5: retail trade (SIC 600 to 699) IND6: finance, insurance, and real estate (SIC 700 to 769) IND7: business services (SIC 770 to 809) IND8: government services (SIC 810 to 849) IND9: education, health, and social services (SIC 850 to 909) IND10: accommodation, food, and other services (SIC 910 to 999)

150

Michael Baker, Jonathan Gruber, and Kevin Milligan

NF: a dummy variable that equals 1 if resident of Newfoundland and 0 otherwise PEI: a dummy variable that equals 1 if resident of Prince Edward Island (PEI) and 0 otherwise NS: a dummy variable that equals 1 if resident of Nova Scotia and 0 otherwise NB: a dummy variable that equals 1 if resident of New Brunswick and 0 otherwise QU: a dummy variable that equals 1 if resident of Quebec and 0 otherwise ON: a dummy variable that equals 1 if resident of Ontario and 0 otherwise (excluded variable) MB: a dummy variable that equals 1 if resident of Manitoba and 0 otherwise SA: a dummy variable that equals 1 if resident of Saskatchewan and 0 otherwise AB: a dummy variable that equals 1 if resident of Alberta and 0 otherwise BC: a dummy variable that equals 1 if resident of British Columbia and 0 otherwise TERR: a dummy variable that equals 1 if resident of Yukon or Northwest Territories and 0 otherwise Reconciliation of Sample Sizes from Longitudinal Worker File Data Labour Market Activity Survey (LMAS) and Survey of Labour and Income Dynamics (SLID) Data Cross-sectional samples from the 1986–1990 LMAS and the 1993–1995 SLID (see table 2A.1) address males or females, ages twenty-three to sixtynine who are paid workers in jobs in the month of September of the indicated year. The RPP coverage probabilities are calculated by three-digit industry code. Probabilities for 1991–1992 are simple linear interpolations of the 1990 and 1993 data. Census Family Files of the Canadian Census The data are from the 1986 and 1991 public-use micro-data files. In each year, males or females who are fifty-four and older are selected. Nonlabor income is defined as the sum of “investment income of census family or non-family person” plus “retirement pensions and other money income of census family or non-family person” plus “retirement pensions and other money income of census family or non-family person” (recorded separately as “Retirement Pensions, Superannuations and Annuities of census family or non-family person” and “Other Money Income of census family or non-family person” in the 1991 sample). Separating individuals who work (weeks and earnings greater than 0) and don’t work (weeks and hours equal to 0), the probability that nonlabor income is positive, and its conditional mean are calculated for the following cells:

Income Security Programs and Retirement in Canada Table 2A.1

151

Sample Selection Observations

Base sample from LWF Missing earnings data Primary industries Already retired at age 55 Final sample

Individuals

Males

Females

Males

Females

770,989 20,306 44,158 99,196 607,329

523,332 10,165 23,146 100,203 389,808

141,182 2,563 6,444 14,244 117,931

103,726 1,378 4,387 20,579 77,382

Notes: The base sample is all individuals aged 55 to 69 in 1985, plus the cohorts of individuals who turn age 55 in years 1986 through 1991. The deletions for missing earnings data are due to nonsensical earnings records for some individuals in Quebec in 1992 (e.g., some individuals with earnings of $46 billion). An attempt was made to replace these records with information from an alternative T-4 data set for this year. This was not successful in all cases, however, which led to the deletion of all observations for these individuals. The deletions for employment in primary industries are explained in the text. The sample is conditioned on employment at age 55, so individuals with 0 earnings at this age are deemed to have already retired and are thus deleted from the data set.

Males who are employed: by region (East, Ontario, or West), by industry (manufacturing; construction; transportation and communications; wholesale and retail trade; financial, insurance, and real estate [FIRE] and business services; government, health, and education services; or accommodation, food, beverage, and other services), and by age (54–55, 56–57, . . . , 60–61, 62–64, 65); Males who are not employed: by region (East, Ontario, or West), by marital status (married; spouse’s age  1; married, spouse’s age
age  1/ – 1; married, spouse’s age age  1; not married), and by age (54–60, 61– 63, 64–66, . . . , 73–75, 76); Females who are employed: by region (East, Ontario, or West), by industry (manufacturing; construction; transportation and communications; wholesale and retail trade; FIRE and business services; government, health, and education services; or accommodation, food, beverage, and other services), and by age (54–55, 56–57, . . . , 60–61, 62–64, 65); Females who are not employed: by region (East, Ontario, or West), by marital status (married, spouse’s age  age –1; married, spouse’s age
age  1/ – 1; married, spouse’s age age  1; not married), and age (54–60, 61–63, 64–66, . . . , 73–75, 76–80, 81).

References Baker, Michael. 2002. The retirement behavior of married couples: Evidence from the spouse’s allowance. Journal of Human Resources 37 (Winter): 1–34. Baker, Michael, and Dwayne Benjamin. 1999a. Early retirement provisions and the

152

Michael Baker, Jonathan Gruber, and Kevin Milligan

labour force behavior of older men: Some evidence from Canada. Journal of Labour Economics 17 (October): 724–56. ———. 1999b. How do retirement tests affect the labour supply of older men? Journal of Public Economics 71 (January): 27–51. ———. 2000. Public pension programs and attachment to the labour force. In Adapting public policy to a labour market in transition, ed. F. St-Hilaire and W. C. Riddell, 287–315. Montreal: Institute for Research on Public Policy (IRPP). ———. 2001. Working time over the life-cycle: Do public pensions matter? In Working time in comparative perspective. Vol. 2, Life-cycle working time and nonstandard work, ed. Susan Houseman and Alice Nakamura, 181–216. Kalamazoo, Mich.: W. E. Upjohn Institute. Baker, Michael, Jonathan Gruber, and Kevin Milligan. 2003. The retirement incentive effects of Canada’s income security programs. Canadian Journal of Economics 36 (2): 261–90. Burbidge, John. 1987. Social security in Canada. Canadian Tax Paper no. 79. Toronto: Canadian Tax Foundation. Coile, Courtney. 1999. Retirement incentives and couples’ retirement decisions. Ph.D. diss. MIT, Department of Economics, Cambridge, Massachusetts. Coile, Courtney, and Jonathan Gruber. 2000. Social security and retirement. NBER Working Paper no. 7830. Cambridge, Mass.: National Bureau of Economic Research, August. Compton, Janice. 2001. Determinants of retirement: Does money really matter? Department of Finance Working Paper no. 2001–02. Department of Finance, Government of Canada. Gruber, Jonathan. 1999. Social security and retirement in Canada. In Social security and retirement around the world, ed. Jonathan Gruber and David Wise, 73– 100. Chicago: University of Chicago Press. Gruber, Jonathan, and Brigitte Madrian. 1995. Health insurance availability and the retirement decision. American Economic Review 85 (4): 938–48. Pesando, James E., and Samuel A. Rea. 1977. Public and private pensions in Canada: An economic analysis. Toronto: University of Toronto Press. Picot, Garnett, and Zhengxi Lin. 1997. Are Canadians more likely to lose their jobs in the 1990s? Analytical Studies Branch Research Paper no. 96. Ottawa: Statistics Canada. Statistics Canada. 1983. Historical statistics of Canada. 2nd ed. Catalogue no. E86103. Ottawa: Statistics Canada. ———. 1984. Life tables, Canada and provinces 1980–1982. Catalogue no. 84-532. Ottawa: Statistics Canada. ———. 1998a. Survey of consumer finances, individual files (1997 income year). Ottawa: Statistics Canada. ———. 1998b. Permanent layoffs, quits and hirings in the Canadian economy. Catalogue no. 71-539-XIB. Ottawa: Statistics Canada. ———. 1999. Pension plans in Canada. Catalogue no. 74-401-XPB. Ottawa: Statistics Canada. ———. Various years. Canadian census. Ottawa: Statistics Canada. Stock, James H., and David A. Wise. 1990. Pensions, the option value of work, and retirement. Econometrica 58 (5): 1151–80. Tompa, Emile. 1999. Transitions to retirement: Determinants of age of social security take up. Social and Economic Dimensions of an Aging Population (SEDAP) Working Paper no. 6. Ontario: SEDAP.

3 The Impact of Incentives on Retirement in Denmark Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

3.1 Introduction In Denmark, the official retirement age of sixty-seven years—from 2004, sixty-five years—is high in an international context. A main point in the Danish policy reform discussions is, however, the widening gap there has been between the official retirement age and actual behavior concerning exit from the labor force. The average actual retirement age has been declining to around sixty-one years old, a decline that gained momentum especially after the introduction of a labor market policy program for early retirement in 1979.1 Before that time, early retirement with a publictransfer income was only possible for health reasons. Both the official and the actual average retirement ages in Denmark are still high compared with most other Organization for Economic Cooperation and Development (OECD) member countries. Nevertheless, policy options and analyses in the pension area are central to current debates of economic policy in Denmark (e.g., see Socialkommissionen 1993; Finansministeriet 1996; Finansrådet 1998; and Economic Council 1998). Two Paul Bingley is associate professor at the National Center for Register-Based Research (NCRR) at the University of Aarhus. Nabanita Datta Gupta is associate professor of economics at the Aarhus School of Business. Peder J. Pedersen is professor of economics at the University of Aarhus. All three authors are affiliated with the Center for Integration and Marginalisation (CIM), Aarhus, and the Center for Labour Market and Social Research (CLS), Aarhus. The authors gratefully acknowledge financial support from the Danish Social Science Research Council, the Danish National Research Foundation and the University of Aarhus Research Fund. 1. Average retirement differs depending on individual background characteristics. In 1997, it was sixty-three years for employed men, fifty-nine years for men who entered retirement from unemployment, sixty-one years for employed women, and fifty-eight years for women who left the labor force from unemployment (Statistics Denmark 1999).

153

154

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

main factors behind the high priority of this area are the projected change in demography and the public finance prospects from this change regarding publicly funded social security programs for retirement. The current level of public-sector-financed income transfers to social security retirement programs is around 8–9 percent of the gross national product (GNP). Public expenditures on health and other services to elderly people should be added to this amount. Projected changes in the age composition of the population along with changes in retirement behavior have potentially large consequences for public-sector finances. In the Danish context, the full impact on the tax-GNP ratio has been estimated to be an incremental increase of around 4–5 percent in a number of recent studies (cf. the references at the beginning of this paragraph). The structure of the remainder of this chapter is as follows. Section 3.2 gives an introduction to the policy environment, both the current as well as the recent, ongoing state policy reforms. Furthermore, section 3.2 introduces some quantitative aspects of the retirement landscape in Denmark. Section 3.3 describes the contents and structure of the micro-data we analyze. Section 3.4 contains the earnings histories and projections needed as inputs in the micro-analyses. In section 3.5, we describe the construction of the social security incentive measures and some of the complications met in this area. Section 3.6 contains the results from a great number of estimations of the impact from the incentive measures described in section 3.5 on retirement behavior. In Section 3.7, the estimation results are used to simulate the expected outcome in the Danish context from a common set of strategic reforms of retirement policy. Finally, section 3.8 concludes the chapter. 3.2 The Policy Environment Until recently, the foundation of Danish policy in the field of pensions has been the so-called universal or Beveridge-type system of eligibility to a pay-as-you-go (PAYG) system for old age pension (OAP) that has age and residence in the country for a sufficient number of years as the only criteria. The expenditures are financed from general tax revenues. Until a few years ago, a base amount was paid to everybody sixty-seven years or older, while a supplementary amount was means tested against other income in the household. From 1993, the base amount has been means tested against earnings from work, but not against capital income. The other main social security program is the route to early retirement through social disability pension (SDP). Initially, eligibility to this program for people younger than sixty-seven years was dependent on health criteria. More recently, there has been a gradual widening of the scope to include also social criteria for eligibility. The program is quite complex, including a number of different benefit levels and with an intricate combina-

The Impact of Incentives on Retirement in Denmark

155

tion of various tax treatment and different rules regarding means testing for the many components making up this program. When people covered under this program reach the age of sixty-seven years, they enter OAP.2 Participation in the SDP program has been fairly stable, reflecting among other things the opening up of other routes—non–health related, however—to early retirement. Like OAP, the SDP is financed from general tax revenues. A dominant part of Danish social security is defined benefits, fully financed from general tax revenues. There is however a small area—of growing importance in the future—of defined-contribution programs (following the so-called Bismarck model). The oldest, mandatory public pension program is the Additional Labour Market Pension (ATP) program, which has been operative since 1963.3 The flat-rate contributions depend on the number of hours of work. Until now, this program has been of minor importance compared with OAP. Beginning in 1999, the program has expanded, as 1 percent of all wages are contributed to a so-called Special Pension program (SP). Future individual benefits from this program are, however, dependent on the accumulated amount of work and not on the individual earnings history throughout the working life.4 The accumulated benefits are paid out as a ten-year annuity from the age of sixty-seven or sixty-five. The historical development in the important area outside social security per se is made up by a multitude of labor market pensions and individual tax-subsidized arrangements and began in 1849 with the introduction of a defined-benefit system of pensions, the tjenestemandspension (public employee pension; PEP) for some public-sector employees. About one hundred years later in the 1950s and 1960s, the build up of funded pension programs (Labour Market Pension programs; LMP) began for other groups of public-sector employees and for academics, both in public and private employment. Until the late 1980s, this resulted in a coverage of about onethird of the labor force with supplementary labor market pension programs (PEP and the funded LMPs). From the beginning of the 1990s, funded LMPs have been extended to cover an increasing part of the public sector and blue-collar workers in manufacturing, trade, services, construction, and transport in the private sector. The typical structure is defined contributions with 3 percent of earnings from the worker and 6 percent from the employer. Certainly the majority and presumably all of the programs have an early retirement option before the OAP age, typically from 2. Here and at later occasions, sixty-seven is changed to sixty-five from 2004 for everybody younger than sixty in 1999, according to a policy reform enacted in 1999. 3. The program shares the name with a corresponding program in Sweden, but the scope in Denmark has been much more limited so far. 4. This new instrument seems to have an impact on both the lifetime allocation of income and on the distribution of income between individuals.

156

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

the age of sixty and depends on the specific age in most of the programs with an actuarially fair reduction. Since 1979, the dominant path to early retirement in Denmark has been through the efterløn (post-employment wage program; PEW). Officially, this is a labor market policy program intended to redistribute a (presumed) restricted number of jobs from older workers with physically demanding jobs to unemployed young workers. In practice, the program has functioned as a reduction in the pension age for broad groups in the labor market. Entry can occur either from a job or from unemployment. Eligibility starts at age sixty and extends to sixty-six, depending on long-term labor market attachment and is documented by membership in an unemployment-insurance fund. A further requirement is that labor supply while in receipt of pension must not exceed 200 hours per year. The rules regarding eligibility have been tightened on a number of occasions. Benefits in PEW are related to unemployment-insurance benefits, either to maximum benefits or 82 percent of the maximum. The main source of funding for the programs is from general tax revenues. A reform was enacted in 1998–1999 to be phased in during the coming years. The contents and the purpose of the reform are taken up below. In 1992, a transitional benefits program (TBP) was introduced. Eligibility to this new early retirement program was conditional on being fifty-five to fifty-nine years old, a member of an unemployment-insurance fund, and to have been unemployed for at least twelve out of the last fifteen months. From the beginning of 1994, the program was extended to cover fifty to fifty-four year olds with the same labor market criteria as for the fifty-five to fifty-nine year olds. Benefits in the program are 82 percent of maximum unemployment-insurance benefits and the duration is until the person enters PEW at the age of sixty years. Participation in the program greatly exceeded government expectations, and entry was terminated in the beginning of 1996 with a strong cyclical upturn underway. 3.2.1 Work and Retirement In this section, we illustrate first the current state of labor market attachment among the fifty and older age group in Denmark. Table 3.1 summarizes the participation rate by age intervals, while figure 3.1 shows the profile for single-year age groups separately for men and women in 1998. It is evident that the largest decline occurs at age sixty, where many become eligible for PEW. Figure 3.1 illustrates the almost-linear declines during the fifties, with the steepest slope for women. In spite of the declining participation rates from the age of fifty, the level, especially from sixty to sixty-seven, is high in international comparison. The Economic Council (1998) replicates the approach taken in Gruber and Wise (1999; available as a working paper in 1997) and calculates a measure of unused capacity for Denmark. Based on the share of men aged fifty-five to sixty-five outside

The Impact of Incentives on Retirement in Denmark Table 3.1

157

Labor Force Participation Rate, 50+ Group (1998) Age 50–54 55–59 60–62 63–66 67–70 70+

Participation Rate 82.3 72.1 38.9 20.0 14.7 5.4

Source: Statistics Denmark (1999).

Fig. 3.1

Participation rates, 1998

the labor force, they found that only Sweden, the United States, and Japan have lower levels of unused capacity than Denmark. Consequently, one interesting question is why the participation rate is so high in Denmark for the sixty to sixty-six age group, considering the availability of a multitude of exit routes from the labor force. A possible explanation is that financial incentives to retire early are strong for people with fairly low wages. Regarding the extent to which early retirement creates financial incentives to continue working during ones’ sixties,5 out-of-work compensation is typically fairly low relative to earnings for people with higher wages. In Denmark, net compensation is between 70 percent for the low-wage group and 40 percent for the high-wage group. In Germany, the reverse is true—that is, the differences are much smaller between compen5. This is illustrated in Ministry of Economics (2000, 46), which shows for a number of countries the net compensation from public pensions relative to earnings for people with low, average, and high wages—more precisely, for people with 75, 100, and 150 percent of the earnings of an average production worker.

158

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

Fig. 3.2 Average retirement hazards for women and men, fifth and tenth decile in the income distribution

sation in the three wage-earner groups, but net compensation is highest for the high-wage group. So, given the structure of public pensions in Denmark, an obvious interpretation is the high-wage earners work to a greater degree later in life simply to accumulate supplementary pension income in order to raise net compensation in retirement above the 40 percent level reached through the public system. Other indicators in the same direction can be found in Economic Council (1998, 93) and in Pedersen (1998, 175). The Economic Council relates the actual shares in early retirement to age and to the degree of financial compensation in different educational groups. Using 1996 data, the degree of compensation, as well as the share in early retirement, are consistently lower for groups with higher education. Pedersen analyzes the expected retirement age in a panel survey, supplemented with register data on incomes, and finds a significant positive impact (i.e., higher expected age of retirement) from gross earnings for those with supplementary pension schemes, while no effect is found for those without any supplementary arrangements. The dichotomy regarding early retirement with one group generally retiring as soon as possible, and another with fairly high labor force participation beyond the age of sixty, can be studied further using panel data. An illustrative example is shown in figure 3.2. This is drawn from a 2 percent

The Impact of Incentives on Retirement in Denmark Table 3.2

159

Structure of the Working and the Nonworking Parts of the Population, 50+ Group (1998) Working Population (%) Nonworking Population (%)

Age 50–54 55–59 60–66 67–74 75+

SelfEmployed

Wage Earner, Full time

Wage Earner, Part time

Total

Unemployed

Other

Total

11.7 14.1 26.0 47.2 65.2

81.2 78.0 58.5 7.9 2.5

7.1 7.9 15.5 44.9 32.3

100 100 100 100 100

20.7 17.9 3.0 — —

79.3 82.1 97.0 100.0 100.0

100 100 100 100 100

Source: Calculations based on Statistics Denmark (1999) Note: Dashes indicate that data is not applicable.

representative sample of the population followed since 1980. We suppress the fact that retirement behavior may vary over time in our present illustration in order to get enough observations to relate retirement behavior to individual income distribution at the age intervals that are most relevant regarding retirement. Each person in the sample contributes one income observation for each age at which they are still a labor market participant. Take, for example, observations relating to age sixty, where we compute decile points in the (real) income distribution for the subsample observed since 1980 that is sixty years old and still in the labor force at some point in time. We then calculate the share retiring for each decile of income distribution for labor market participants at each age between fifty-nine and sixty-nine. (Calculations are made separately for women and men.) A small selection of the great number of age- and income-related retirement frequencies are shown in figure 3.2. Tentatively, they seem to support the hypothesis that the relatively high participation rate in Denmark of the sixty and older age group is concentrated mainly among males at the top of the income distribution. Next, the structure in 1998 of the working and nonworking population fifty years and older is shown in table 3.2. For the working group, the structural composition divides—not surprisingly—at the age of sixty. The share of wage earners working full time goes down 20 percentage points, and from age sixty-seven, more than 90 percent of those still working are either self-employed or part-time workers. The cross-sectional nature of the table does not allow any conclusions regarding part-time work as a stepping stone from full-time employment to retirement.6 The right-hand panel in table 3.2 shows that the expansion of routes to early retirement has meant practically an end to unemployment in the sixty and older group. A major 6. Casual evidence does not support this conjecture (i.e., the stepping-stone idea). Further analyses drawing on panel data will enable conclusions to be drawn on this point.

160

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

Table 3.3

Proportions of the Nonworking Population 50–69 Years, by Dominant PublicSector Benefit Program Category (1998)

Age

UIB

LMP

TBP

PEW

SWB

SDP

OPA

None

Total

50–54 55–59 60–66 67–69

20.7 17.9 3.0 —

3.8 1.9 0.1 —

16.5 24.0 — —

— — 52.6 —

3.5 1.4 0.4 0.3

42.1 43.6 35.6 —

— — — 97.8

13.5 11.1 8.2 1.9

100 100 100 100

Source: Calculations based on Statistics Denmark (1999). Notes: UIB = unemployment-insurance benefits; LMP = different labor market programs; TBP = Transitional Benefit Program; PEW = post-employment wages; SWB = social welfare benefits; SDP = social disability pension; OAP = national old age pension. A case where a labor market pension is the income source will, in principle, fall in the category “None.” The coverage of labor market pensions and private schemes is, however, imperfect in the table. Dashes indicate that data is not applicable.

part of the income for people fifty years and older, outside of labor force participation or unemployment, comes from the broad range of mainly publicly financed programs. The distribution in 1998 by different sources of income ranked according to the dominant program in each individual case is shown in table 3.3.7 The dominant source of income for nonworkers is SDP until the eligibility age of sixty for PEW. The lower share receiving TBP in the fifty to fiftyfour group simply reflects that entry to this program was terminated in early 1996, so the youngest recipients in table 3.3 are fifty-two years old. The residual group is fairly small. Because the options cover both the (shrinking) group of housewives without an independent income and people with an early receipt of labor market pension or private means as their dominating source of income, it is obvious that early retirements options outside the (mainly) social security programs are still a long way from maturity. Next, we consider some aspects of the transitions to a condition outside the labor force. Presenting a comprehensive empirical picture of all the possible routes would be a major undertaking, so the present intention is to concentrate on some of the major transitions. A sketch of these is presented in table 3.4. Before age sixty, SDP is the only exit route with an income compensation, apart from the temporary TBP for long-term unemployed, open for two and four years, respectively, in the mid-1990s. In the age interval of sixty to sixty-six, transitions can be to SDP (although not to the highest level) from the same states of origin as for younger people. Next, the very important PEW program is available—conditional on eligibility—for entry from either a job or from unemployment. The stock of people covered 7. Note that the unemployed in table 3.2 are implicitly defined as those who receive unemployment benefits. However, some unemployed receive social welfare benefits (SWB), and some do not receive any benefits.

The Impact of Incentives on Retirement in Denmark Table 3.4

161

Typical Exit Routes from the Labor Force by Age Intervals


50 years

50–59 years

To: SDP From: Employment; unemployment; sickness benefits; welfare benefits

60–66 years

Same option and same rules regarding SDP as 49 years and younger group

To: SDP Same as younger but excluded from highest level SDP

To: TBP a From: Long term unemployment

To: PEW From: Employment; unemployment; TBP (100%) To: Early option, labor market pension From: Employment

67+ To: OAP From: All states

To: LMP, PEP From: All states

Note: See table 3.3 for explanation of abbreviations. a Program for long-term unemployed, open 1992–96, transition from age 60 to PEW.

Fig. 3.3

Percent of labor force exits from employment

by TBP transit automatically to PEW at age sixty. Finally, most—or all— labor market pension schemes have, as mentioned, an early retirement option. People with this option who are at the same time eligible for the PEW will most likely have an incentive to enter the PEW and postpone use of labor market pension schemes due to means-testing rules in the PEW program. People who are not eligible for the PEW and who retire early on a labor market pension scheme will normally transit directly from a job. From age sixty-seven, people still in a job (or the very few in unemployment), as well as people in all kinds of early retirement programs, transit to the OAP. For people initially in the labor force and fifty to sixty-six years old in 1997, figure 3.3 shows the share of the work-originating exits to retirement

162

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

Fig. 3.4 The absolute number of transactions to PEW and the share from employment

to be around two-thirds until the PEW option at age fifty-nine, at which point the share from employment jumps to three-fourths. The share at age sixty-one is a low point and is related to the specific transition patterns to PEW (cf. figure 3.4 below). For those sixty-two years and older who still are in the labor force, the share coming from employment is somewhat higher, which is not surprising given the strong selection mechanisms into programs for the unemployed. Figure 3.4 concentrates on exits to PEW. The absolute number of exits to PEW, independent of state of origin, peaks when the minimum possible age of eligibility is reached. The only exception to the much lower number at higher ages is the local peak at sixty-two years. This presumably reflects the option of being eligible for maximum unemployment benefits for the whole period remaining until OAP if entry to PEW is postponed until sixty-two. Figure 3.4 also shows the share by age of exits to PEW coming from a job. This is seen to be fairly stable between 70 and 80 percent, except for the dip at age sixty-one that could reflect the incentive for people in long-term unemployment, when turning sixty, to delay entry to PEW, which might imply a reduction of their income up to 82 percent of maximum unemployment benefits. Instead, in some cases, an option is to remain unemployed and collect maximum unemployment insurance benefits until the end of benefit eligibility and then exit to PEW. Looking at transitions to PEW from employment by industry, we see the profile for 1997–1998 shown in figure 3.5. The transition rates shown are calculated as the flow to PEW relative to the stock of employed people sixty to sixty-six years old in each of the industries. The highest values are found

The Impact of Incentives on Retirement in Denmark

Fig. 3.5

Transition rate to PEW by industry, 1997–98

Table 3.5

Status Prior to Entry to SDP

163

Status Immediately Before Entry (%)

Earlier Status Employed Unemployed Welfare benefits Sickness benefits Other Total

Employed

Unemployed

Welfare Benefits

Sickness Benefits

Other

Total

8 0 0 0 0 9

0 0 0 1 1 2

0 10 7 3 8 28

22 21 3 1 2 50

1 0 2 0 8 11

31 32 12 5 20 100

Source: Calculations based on Statistics Denmark (1999).

for manufacturing and for public and private services (i.e., industries with a high share of unskilled workers). Many of them will have a fairly high potential rates of compensation upon entering the PEW.8 The low share from agriculture reflects both the dominance of self-employment in this age group and the condition that a farmer must sell the farm to be eligible for the PEW. Another low share is found for financial services, probably because this is a group in which most employees have access to labor market pensions. The industry group “other” includes workers with a more marginal attachment to the labor force. A higher-than-average share of these probably do not fulfill the requirements for eligibility to PEW. Regarding the transition to SDP, table 3.5 shows the status immediately 8. Many in these groups will also have had physically demanding, repetitive jobs, or both (i.e., belong to the groups for whom the PEW was originally designed).

164

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

before the transition (i.e., as a minimum, the latest month) and shows also the earlier status, or the one preceding the state immediately before entry to SDP. Looking at the immediately preceding status, we find—not surprisingly, as an SDP application takes time to process—that only 11 percent are in the labor force. But by extending the perspective to include people coming from unemployment, we find that nearly two-thirds come from the labor force. For people coming from employment, both immediately before and as the earlier status, it seems fairly obvious that the route to SDP is health related. 3.2.2 Policy Reforms—Pressure and (Beginning) Action The background to policy reforms is similar in Denmark to most other OECD countries, even though the pressure for reforms may be less. The setting is the well-known combined effects from the projected aging of the population, the declining average retirement age, and the transition from a predominantly PAYG system to predominantly funded systems. During the next forty years, the number of people within the potential labor force age range relative to people sixty-five years and older is projected to decline from a 4-1 ratio to a 2-1 ratio. At the same time, a significant decline has occurred in the actual average retirement age during the last quarter of a century, widening the gap between actual average retirement ages and the hitherto very high official pension age of sixty-seven years. The core of the problem is, like in other OECD countries, an expected increase in public expenditures for pensions and for health-related care for elderly people, and at the same time, the decline in average retirement ages means an erosion of real production and the tax base in the economy. Denmark shares the problems in this area along with other OECD countries, but the specific problems related to demography are expected to be relatively smaller. Roseveare et al. (1996) project an increase in taxes relative to GNP conditional on demographic factors until the year 2030. Italy, Germany, Japan, and the Netherlands are on top in this ranking with projected increases in the range of 9–11 percentage points. Sweden, Denmark, Norway, the United Kingdom, and Ireland are at the opposite end with projected increases in the range of 2–4 percentage points. Nevertheless, when the problems are considered very important by Danish policy makers, it reflects among other things that the fairly small impact on taxes as a share of GNP (tax/GNP) must be seen in relation to a level of the tax/GNP that—along with the level in Sweden—is the highest among the OECD countries. It will be ten years before the demographic change affects Denmark. Nevertheless, reform discussions and initiatives have been on the policy agenda already for some time. The current reform plans are summarized in

The Impact of Incentives on Retirement in Denmark

165

the present section. The general strategy seems to consist of three main areas.

• A gradual shifting of the weight from PAYG to funded programs • Policy changes to reverse the decline in actual average retirement ages • More broad policy changes to mobilize an increase in labor supply from nonparticipants in different public-sector income-transfer programs Obviously, the last part of the strategy has the double benefits of reducing public expenditures at the same time as expanding the tax base. Regarding the OAP, the general trend—in accordance with the strategy outlined previously—has been a reduction of the universal base amount and an increase in the means-tested supplementary pension. As the labor market pension schemes reach maturity, an increasing number of pensioners will be entitled only to the base amount. The universal base amount was independent of other income until changes in 1982 and 1993 introduced means testing against earnings, but not against capital income. This seems to reflect the philosophy of those years of trying to combat unemployment by reducing labor supply. The major current change regarding OAP is the reduction of the eligibility age from sixty-seven to sixty-five. Superficially, this seems surprising considering the nature of the problems regarding the pension burden. It must however be interpreted in the light of, first, the widening gap between the high official retirement age and the declining actual average age. Then, a major share of people sixty-five and sixty-six years old receive the PEW, which is higher than OAP, so public expenditures may decline.9 Finally, means testing of the base amount against earnings is changed in a way to make gradual retirement more attractive. Reduction of the base amount begins at a higher level of earnings than previously and the rate of reduction of the base amount is reduced from 60 to 30 percent, implying that the base amount will only be fully phased out at an earnings level well above the average earnings of skilled workers. The number of people entering the SDP annually has gone down since 1988. Especially in the most recent years, the number of new entrants has been at a very low level. The background is the spending of more resources on rehabilitation to reverse the expansion of SDP, and in 1998–1999, an increasing number of “flex-jobs” on special conditions have been supplied as an alternative to SDP. Critics of the recent reforms have claimed that the strong decline in admission to SDP mostly reflects a tightening by the administration of requirements for eligibility. The policy changes have been in effect for too short a period to reach any conclusion as to their effect yet. 9. A counter effect might come from people who have not received any income transfers, but who now become eligible for OAP two years earlier.

166

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

The PEW, introduced in 1979, has become the dominant route to early retirement after the age of sixty. Until now, the only financial incentive to delay entry has been eligibility to maximum unemployment-insurance (UI) benefits until the age of sixty-seven if entry was postponed until the age of sixty-three. A reform of the PEW was enacted to gradually have effects onward from July 1999. The main point in the reform is to create incentives to postpone entry to the program until the age of sixty-three. Combined with the reduction of the OAP age to sixty-five, the long-run effect is intended to be a significant decline in the stock of people in the PEW program. As mentioned in the introduction, the reform tightens the entry requirement for PEW regarding the years of membership of an UI fund. Benefits are changed from a system in which they were reduced as a function of time spent in the PEW to a system in which benefits depend on age at entry being sixty to sixty-one or sixty-two or older. In the first case, benefits are 91 percent of the maximum UI benefits until OAP eligibility, while in the second case they are 100 percent of maximum UI benefits. Furthermore, benefits are means tested against income from all other pension schemes for people aged sixty to sixty-one years. After age sixty-two, means testing is only against the income from pension schemes with monthly payments originating in earlier employment. Until the reform, this was the rule for everybody aged sixty to sixty-six. Hitherto, participants in the program were not allowed to work more than 200 hours annually. The reform replaces this with a flexible scheme, always resulting in a financial incentive to work to some extent. Eligibility to PEW in the new system will be terminated if the person works thirty hours or more per week. This compares with the earlier system in which no reduction occurred as long as annual labor supply was below 200 hours and entitlement was permanently and completely lost if a recipient worked more than 200 hours. Furthermore, the reform introduces a tax premium for people who are entitled to PEW, but who go on working after age sixty-two. Working until the new OAP eligibility age of sixty-five would entitle the person to a tax premium at a level of 50 percent of the annual income of an unskilled worker. In summary, the main objective of the PEW reform is to reverse the decline in the average actual retirement age and increase labor supply, especially among people in their early sixties. The reform has only been in effect for a short period of time, so evidence concerning the impacts on behavior is not yet available. Danø, Ejrnæs, and Husted (2000) have entered the main elements of the PEW reform into a quantitative study of retirement in Denmark in order to simulate the effects. The major impact from the reform is a big reduction in the implicit tax on continued work, especially for sixty-one year olds and for people sixty-three to sixty-four. Based on the model, the prediction is that the reform will result in increases in the aver-

The Impact of Incentives on Retirement in Denmark

167

age retirement age for eligible individuals in the 1–1.5 year range for men and 1–2 years for women. Christensen and Datta Gupta (1999, 2000) have compared the labor supply and government budgetary effects of a delay in PEW eligibility age versus a reduction in PEW benefits and have found that delaying PEW eligibility from sixty to sixty-two will lead to a greater increase in the average retirement age (0.6 years for men and 0.75 years for women) and greater per capita savings than a reduction in benefits. The main points in current policy reforms are thus to delay early retirement and to increase labor supply. This combines with the ongoing longrun shift of emphasis from PAYG to funded pension systems. 3.3 Data Overview The data for this study are drawn from the Integrated Database for Labour Market Research (IDA; Statistics Denmark 2003), which is combined longitudinal data for persons and establishments in the period 1980– 1995 and is compiled by Statistics Denmark. The IDA links information on employees and establishments drawn from central administrative registers that contain labor market information on the entire Danish population. Thus, there is no sample survey component to this data. As far as persons can be followed over time, IDA contains annual labor market information on all individuals in Denmark. We use a 2 percent extract of IDA (the population is all public- and private-sector workers) and restrict our retirement estimations to all older workers aged fifty to seventy in the period 1980–1995 who have been observed in the labor force at any time. Although early retirement eligibility begins at age sixty, we start our analysis at fifty, as this is the age at which public or private retirement income for disability or illness can first be expected for those in the labor market. We condition on labor force attachment before fifty. The sample sizes for the estimations are 210,073 female person-years10 and 224,621 male person-years. The definition of retirement is based on the receipt of benefits such that a person is considered retired if the main source of income during the year comes from a pension. In addition, we also condition on the absence of labor earnings and employer pension contributions during the year. Several key public pension benefit programs are considered in addition to the pure–social security program or OAP. These are, respectively, the PEW, SDP, and PEP. We also include the TBP, which was a special earlyretirement pension option available to the long-term unemployed (aged fifty-five to fifty-nine initially, and later also including fifty to fifty-four years) that was in effect between 1992 and 1996. 10. “Person-years” represents the volume of observations. For instance, 5 women observed for 5 years will constitute 25 female person-years.

168

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

The IDA sample contains information on an individual’s demographics, years of education, earnings, and other income and assets. The main limitations from the point of view of a retirement study are lack of information on private pensions and health. Mortality information is available, and survivor probabilities are assigned on the basis of gender- and age-specific national average life tables for ages fifty to ninety-nine (death is assumed at age one hundred). In the Danish case, there is only a limited use for spousal information, as calculation of benefits is almost entirely on an individual basis and does not depend on the spouse’s entitlement, although adjustments are made to the benefit level for marital status in some of the programs. Neither dependent benefits nor survivor benefits are relevant in the Danish social security context. We determine eligibility for OAP on the basis of age, and eligibility to PEW on the basis of age and retrospective information on membership in an unemployment insurance fund. Potential entry in to the special TBP can also be determined on the basis of age, UI fund membership, and information on individual unemployment degree in the two years before potential eligibility. Eligibility to disability pension is determined probabilistically, based on population averaged by year, age, and gender participation rates in each of the three levels of the disability program. Without access to health information, it is difficult to make a more precise assignment to disability program eligibility. Finally, eligibility to PEP is observed directly in the data. But while we can effectively analyze retirement decisions of those who would be eligible for the major social security benefits programs, it is not possible on the basis of these data to determine eligibility to employer-provided private pension plans. 3.4 Sample Descriptives and Earnings Projections Sample descriptives are reported in tables 3A.11 and 3A.12 in appendix A. The gender-specific means are shown both for the full sample (table 3A.11) and by PEW eligibility (table 3A.12). Means are taken across all person-year observations within each gender group. On average, males are around 56.6 years of age and females a few months younger. Around 17 percent of the males in the sample are eligible for PEW, although only 9 percent of females have accumulated the necessary years of membership in an UI fund. In terms of incentive measures, the average accrual for males is around $1,745, while for females it is $1,161. Both peak and option value figures are about 1.4 times higher for males than females on average. About 9 percent of the males and 10 percent of the females are retired. Average annual earnings for males are $35,970 while the corresponding figure for females is only a little more than half, $19,368. In terms of occupational ranking, males are more likely to be in skilled blue-collar managerial positions (especially top- and middle-management positions), and females are

The Impact of Incentives on Retirement in Denmark

169

more likely to be in white-collar and the unclassified occupational category. When disaggregating by PEW status in table 3A.12, it can seen that males who are eligible for PEW are on average two years older, have slightly higher annual earnings ($36,406 versus $35,625), and are much more likely to be in skilled and unskilled blue-collar jobs and less likely to be in top- and middle-management jobs, compared to males not eligible for PEW. In table 3A.12, females eligible for PEW are not that different compared to uneligible females, except in terms of higher earnings ($23,138 versus $17,692), higher representation in white-collar and unskilled bluecollar, and lower representation in the unclassified occupational category. Annual earnings are top coded at the real 99.9 percent level. The (log real) annual earnings regressions are based on a sample of workers aged forty-nine to seventy in all the available years (1980–1995). A simple, agequadratic, individual fixed-effects specification is chosen for the log earnings regressions, with separate regressions run by gender. In the Danish case, as retirement benefits are largely independent of past earnings, the benefit calculations in most cases depend only on most recent earnings and not on earnings histories or projections. The only program that ties benefits to past record is the PEP. For the purposes of earnings projections of an individual’s expected earnings stream from work, a flat real ageearnings profile is applied when going forward from the predicted earnings in the last observed year of work.11 3.5 Construction of Incentive Measures Central to the analysis are the construction of social security wealth (SSW) and the calculation of the various incentive measures, which are based on SSW and other data components. The construction of SSW is complicated by the fact that there has been extensive reform of pension programs in Denmark over the 1980–1995 period. Also, in the Danish context, there are multiple policies, such as SDP and TBP, for older workers, that need to be integrated (see the discussion in section 3.2). We make the assumption that all expected future changes in eligibility, generosity of programs, or both are unknown at each age. In other words, the agent assumes in each calendar year, that the future social security system will remain as it is in that calendar year. Table 3.6 summarizes the main elements of the programs and their respective eligibility criteria. We adopt a weighted-average income flow measure, in which the weights are determined by empirical take-up rates of disability pension. These 11. The within estimator R 2 is equal to 0.1733 for males, and R 2 is equal to 0.1395 for females. All estimated coefficients have the expected signs and are highly significant. The ageearnings profiles are concave and reach maximums at age 52.5 for males and at age 53.1 for females.

170

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

Table 3.6

Criteria for Program Eligibility

Program

Eligibility Age

Determines Eligibility Based On

PEW (post-employment wage)

60–66

Age and insurance fund information

TBP (Transitional Benefits Program)

55–59 long-term unemployed from 1992 to 1996; 50–59 from 1994 to 1996

Age and insurance fund and unemployment info in two years prior to age

PEP (public employees pension)

60–69

Employer pension contributions over required period

SDP (social and disability pensions)

18–66

Probability: Based on observed participation rates by age, year, and gender in each of the three levels of the disability program

OAP (old age pension)

67

Age 67

weights are calculated from the age, year, and gender disability rates within the IDA sample. Thus, at each potential retirement age, the individual is assumed to be eligible for disability pension with probability p and for one of the other nondisability social security retirement programs with probability 1 – p. Within disability, the level of disability (highest incapacity, middle incapacity, and lowest incapacity) is also determined on the basis of population age- and gender-specific participation rates. Within the nondisability social security programs, we allow one income flow for each potential retirement age from fifty to sixty-nine depending on eligibility. This is consistent with the Danish system in which individuals cannot draw multiple pensions. The possible programs are PEW, the TBP, PEP, or no pension. From the official retirement age of sixty-seven and onward, all individuals (independent of potential disability status) are entitled to OAP and possibly PEP in addition, if eligible.12 Thus, suppose that with probability p, a person goes on disability pension at age a, then the individual’s income flow at age a, IFLOWR, is given by or IFLOWRTBP or IFLOWRPEP IFLOWRa  (1  p)(IFLOWRPEW a a a  pIFLOWRSDP , 49  a  66 a and IFLOWRa  IFLOWROAP , a

a  67,

12. Part of them will have supplementary income from labor market pensions or from private arrangements that are not included in the present data and therefore are not covered in the study.

The Impact of Incentives on Retirement in Denmark

171

where SSW is the present discounted value of future income flow. Our SSW measure incorporates both broad changes in pension policy that took place in the 1980–1995 period, as well as the detailed year-to-year reforms in the generosity, taxation, and eligibility rules. The major reforms in this period include the sweeping SDP reform of 1984 in which the so-called widow’s pension and special pension rules for single women was cancelled. In its place, the disability pension was broadened to include pensions based on social criteria. A special retirement window, the TBP, was available starting in 1992 (and ending in 1996) for the long-term unemployed between the ages of fifty-five and fifty-nine—later also between fifty and fifty-four—who would fulfill the eligibility criteria for the PEW program by age sixty. Finally, in 1993, there were benefit increases dependent on delayed retirement and other changes in compensation in the PEW program. Tables 3A.1–3A.10 summarize the incentive calculations based on our construction of SSW on the micro-data. Tables 3A.1–3A.5 present results for the males, and tables 3A.6–3A.10 are the corresponding calculations for the females. For each of the preceding samples, we present calculations made for the overall sample, as well as specifically by the type of pension program (PEW or not). Three incentive measures are considered: accrual, peak value, and option value. The accrual is the accumulation in SSW if the worker postpones retirement by a year. The option value is the difference between the maximum of the expected present values of retiring at each future age minus the present value of immediate retirement. Thus, if the option value is positive, the individual continues to work, otherwise he retires. We assume a discount factor  is 0.97, the risk aversion parameter  is 0.75, and the utility of a dollar of income while retired relative to a dollar of income while working equal to 1.5, using values previously estimated for the United States by Stock and Wise (1990). The peak value is simply the option value under the assumptions of no risk aversion, and the same utility from a dollar while working and a dollar while retired—i.e., (, k) is equal to (1, 1)—or what has been labeled the “financial option value” by Coile and Gruber (2000). We consider the worker at the median, tenth percentile, and ninetieth percentile who retires between the ages of forty-nine and seventy. In general, working another year can affect SSW in the following ways.

• By working another year, the worker who is less than sixty-seven may forgo a year of benefits through either SDP or early retirement (TBP plus PEW, or early PEP) if he meets eligibility criteria. This will lower net SSW. • The worker who considers retiring at fifty-nine (instead of sixty) would be entitled to receive early retirement benefits starting the next year only, which would lead to a big increase in SSW at retirement age sixty. • Between the retirement ages of sixty and sixty-seven, working another

172

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

year can lead to accumulating an additional year of “pension rights” in the PEP program (if the worker is eligible) that increase the pension that can be drawn after sixty-seven and therefore SSW. • At retirement age sixty-seven, we would expect another spike in SSW, reflecting the availability of OAP. • Finally, if a worker chooses to work after sixty-seven, this would mean forgoing a year of social security benefits available to all, which would mean a shorter interval over which social security can be collected and therefore lower net SSW. Table 3A.1 reports median SSW and median, tenth percentile, and ninetieth percentile accrual figures for the males, together with the standard deviation of accrual and the rates of taxes and subsidies, which is the change in SSW relative to what the worker could have earned over the coming year. Median SSW increases slowly up to retirement age sixty (last work age fifty-nine) and declines thereafter. This is also seen in the median accrual, which is positive up to age sixty (the first age of eligibility of early retirement through PEW) and then turns negative. The negative accrual clearly reflects the actuarial unfairness of the benefits system such that the gain in wage earnings from postponing retirement are more than offset by the loss in future social security benefits. The corresponding subsidies are in the range of 10–15 percent before retirement age sixty and taxes of 15–35 percent after sixty. Thus, this would indicate that there are strong incentives to retire through PEW in the Danish system. However, at retirement age sixty-seven, accrual turns positive again and the tax becomes a subsidy of 40 percent. This sudden reversal in sign at retirement age sixty-seven is possibly a result of pooling together workers of different eligibilities (namely those eligible for PEW and those not eligible for PEW and only eligible for either OAP or PEP) who respond differently to program incentives. Thus, a sizable fraction of our sample (mostly low-wage earners) retire through PEW, while the remainder (highwage earners) have incentives to remain in the labor force until age sixtyseven (see the discussion in section 3.2). This is also confirmed in part by considering separately the tenth and ninetieth percentile in the SSW distribution. At the tenth percentile in table 3A.1, accruals follow the expected pattern, positive up to retirement age sixty and negative thereafter. At the ninetieth percentile, accrual is always positive, but a has a double-peaked profile at last work ages fifty-nine and sixty-six (the relatively larger of the two), reflecting the large gain in SSW by continuing to work until program eligibility. Variability in SSW (standard deviation) increases with age, reaching a maximum at retirement age sixty-seven, as some are eligible for PEP in addition to OAP and others not. In table 3A.2, we present peak value and option value calculations at the median, tenth percentile, and ninetieth percentile for the males. Peak and

The Impact of Incentives on Retirement in Denmark

173

option value are measured in utility terms and are therefore not comparable to accrual figures. Both are positive but declining throughout the age interval, indicating a propensity to continue working even at older ages. Thus, retirement incentives, as measured by option and peak value, seem to be weak in the Danish system. Although, it should be pointed out that these figures are typically higher at the ninetieth and lower at the tenth percentile relative to the median. Program-induced incentives are difficult to detect when the sample is pooled irrespective of program eligibility. This is because, as described in section 3.2, a significant fraction retires through the PEW program at age sixty or shortly thereafter, while the remainder retires at age intervals up to sixty-seven. The two groups are heterogeneous in terms of their work history and expected replacement rate of income through social security, and therefore large selection effects may be present when pooling the data. In tables 3A.3–3A.5, we regenerate the incentive calculations for the subsamples eligible for PEW and those uneligible. The latter is relatively more heterogeneous, consisting of individuals retiring through disability, PEP, or, in most cases, only through OAP (compare standard deviations of the incentive measures and SSW across the two groups). We report only results at the median here. Looking at the age distribution of accrual by PEW eligibility in table 3A.3, it is clear that strong early retirement incentives are present for those who are eligible for PEW. In this case, accrual is positive until retirement at age sixty and turns sharply negative thereafter. Work is encouraged (subsidies of 9–17 percent) in the forty-nine to sixty retirement age group, and work is heavily taxed after that with taxes ranging from about 10–40 percent. For those not eligible for PEW, the story is not as clear due to the mixing of different groups (see previous discussion), but the tax on work at older ages is clearly lower than for the PEW-eligible persons ranging from about 1–6 percent. Tables 3A.4 and 3A.5 report the distribution of the forward-looking measures separately by PEW eligibility. Both peak value and option value measures are positive and declining throughout for both samples and are roughly comparable in magnitudes. Tables 3A.6–3A.10 present similar calculations for the female sample. Several features are worth noting in comparison to the male sample. Median SSW is higher for women at retirement ages forty-nine to fifty-five and sixty-seven to seventy. In the earlier age groups, women have higher rates of participation in disability and therefore higher SSW (as well as pre-1984 access to widow’s pension). In the highest age group, the working women present in the sample may not be representative in terms of their access to different pension schemes. In the normal retirement age interval, men have higher SSW. At the median, one-year accruals are smaller for women than men at all ages, indicating a smaller incentive to keep working before sixty and a

174

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

smaller incentive to retire after sixty. This is also seen in the relatively smaller tax and subsidy rates for women. The female sample is also more homogeneous in terms of the variability of SSW, accrual, peak value, and option value (compare standard deviations in tables 3A.6 and 3A.7 to those in 3A.1 and 3A.2). Peak and option value are smaller for women, although the same order of magnitude (compare tables 3A.2 and 3A.7). The 10-90 split is not as clear-cut in the case of the female sample. While at the tenth percentile, accruals turn negative from retirement age fifty-eight and onward at the ninetieth percentile, accruals turn negative at retirement age sixty-one but positive again at age sixty-seven. This indicates that some selection effects may be present due to pooling together of heterogeneous groups who may be retiring at different times due to program-induced incentives (see table 3A.6). Females eligible for PEW (like males) also exhibit a clear preference for early retirement, although in this case, retirement begins for the median worker at age sixty-one, perhaps due to the delayed retirement incentive present in PEW. (PEW benefits are 90 percent of previous wage up to a maximum UI-benefit ceiling for the first two and a half years on PEW starting at age sixty, and 80 percent thereafter until age sixty-six. If retirement is first taken at age sixty-three, benefits can be collected at the higher 90 percent rate for the rest of the remaining period under PEW.) Median SSW for females on PEW is higher than that for males at all retirement ages, perhaps indicating that the average female on PEW also has access to other types of social insurance. Retirement income for this sample is also more variable (see table 3A.8). Women eligible for PEW also exhibit slightly smaller (but still positive and declining with retirement age) peak value and option value than men, indicating a lower tendency towards postponing current retirement (see tables 3A.9 and 3A.10). In the Danish case, therefore, the driving force behind the age-incentive structure seems to be program rules and eligibility. As social security is funded through general taxes, no payroll tax–induced work disincentives can be expected. Except for the delayed retirement incentive in the PEW described above, no actuarial adjustments are made to benefit levels for later retirement through OAP (a “waiting supplement” was phased out in 1983). The PEP program is actuarially adjusted, as a reduction in benefit amount is present for retirement before age sixty-seven, with the reduction depending on age at retirement and number of pension years that have been accumulated up to the point of retirement. However, as only a fraction of the labor force is eligible for this program in our sample period, this effect is unlikely to be reflected in the data. The old-age social security system in Denmark works mainly as a safety net, providing a basic guaranteed income at older ages. This redistributive function explains why low-wage workers face a fairly high replacement rate while higher-wage workers resort to labor market pensions and private pensions to make up for the

The Impact of Incentives on Retirement in Denmark

175

rather low compensation through the social security system (cf. the discussion in section 3.2). 3.6 Estimation We estimate an option value–based model of retirement (see Stock and Wise 1990) in which an individual compares the expected value of retiring today to the expected value of continuing to work today and postponing retirement until the best possible future age of retirement. Thus, for the individual, the value of working from age t to age r – 1, and then retiring at age r is given by r1

S

St

Sr

Vt (r)  ∑ StUW (YS)  ∑ StUR [BS(r)], where UW is the instantaneous utility of work that takes as its argument a worker’s income, and UR is the instantaneous utility of retirement and is a function of retirement benefits that depend on the worker’s retirement age, and  is the discount factor. Let EtVt (r) be the expected value at age t of working through r – 1 and retiring at r, and let EtVt (t) be the expected value of current retirement, and let r ∗ be the value of r that maximizes EtVt (r). Then, the individual will postpone retirement if the option value Gt (r∗) is positive. Gt (r∗)  EtVt (r ∗)  EtVt (t) 0 Otherwise, the person retires at age t. Utility specifications allow for risk aversion through the parameter . The parameter k measures the utility of a dollar of income obtained while retired relative to the utility of a dollar of income obtained while working. Thus, if k 1, then a dollar of income while retired gives greater utility than a dollar of income accompanied by work. The utility functions are UW (YS )  YS  S , UR (BS (r))  [kBS (r)]   S . A simple alternative to a full-blown maximum likelihood estimation of the option value model is to specify retirement in terms of the gain from continuing to work, computed on the basis of some assumed values for , k, and , which can then be used to calculate income.13 Assuming that retirement depends on this calculated option value as well as other determinants (observed and unobserved), a probability model of retirement (for example, probit) can be specified as 13. We assume that the value of k is 1.5 and  is 0.75, as found by Stock and Wise (1990) for the United States.

176

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

Pr(retire in year t)  Pr[  Gt (r∗)  X  εt 0], where the dependent variable is binary and takes the value 1 if retired, 0 if not, and where G() is the option value of postponing retirement (in other cases the peak value or the accrual measure) calculated under the assumed parameter values, and X is a vector of additional variables. Thus, any changes in social security plan provisions that would affect income and benefits and hence the retirement decision are reflected in changes in G() and captured by the parameter . The larger the option value, the greater is the gain from postponing retirement—that is,  should be negative in the probit regression. The sample for the probit regression is all individuals aged forty-nine to seventy in our data, which is 2 percent of the population of workers. Everyone is assumed to have retired before age seventy-one. Observations on individuals are pooled in the analysis, and individuals are retained up to and including the year of retirement. The results of this probit regression analysis can be found in table 3A.13, where the first panel presents the results for the male sample for each incentive measure (accrual, peak value, and option value), and the second panel shows the corresponding results for females. These regressions use as controls a full set of demographics: sector, earnings, and higher powers in earnings. For each incentive measure, we report two alternative specifications with a linear age term or with agespecific dummies. The social security measure is also included in addition to the incentive measure in all the specifications in order to capture the wealth effects of the system. The incentive measures can then be thought of as capturing the substitution effects. In terms of goodness of fit as measured by the pseudo-R 2, all three models yield comparable fits in the range of 0.24–0.37. The option value gives the best fit when age enters linearly, while the accrual model explains the data the best when age dummies are used in place of a linear age term. This is true for both males and females. For all incentive measures, the specification with age dummies yields a better fit, for both males and females. In all cases, fit is marginally better for males than females. All coefficients are highly significant, except the coefficient on SSW in two out of the twelve cases. For the males, the coefficient on the incentive measure is estimated negative (as it should be) in all cases except in the accrual model with linear age. The coefficient on the SSW variable is positive in all cases. Thus for males, the greater the incentive measure (particularly the forward-looking ones), the greater is the incentive to continue working, and the higher the accumulated SSW, the higher is the probability of retirement. For females, the coefficient on SSW is in all cases positive, and the coefficient on incentive is negative, except in the accrual model with linear age. Again, for the most part, retirement wealth and incentive measures have the expected effects.

The Impact of Incentives on Retirement in Denmark

177

Besides improving fit, for both males and females, adding age dummies seems to yield signs and significance that conform to prior expectations. In terms of the estimated coefficients on the incentive measures, the peak and option value give the conventional signs in all cases, while the accrual model does so only in the case with age dummies. The retirement hazard and the profile of the age dummies from the estimations for women and men are shown in figure 3B.1 in appendix B. The age dummies reflect very clearly the main rules in the major retirement programs, that is, the spikes at sixty and at sixty-seven reflecting the PEW entitlement and the universal entitlement to OAP, respectively. In the next section, we carry out some simulations of the impacts of policy changes on retirement probabilities. 3.7 Policy Simulations In this section, we describe the results arising from two kinds of policy simulations.14 First, in what we label the three-year reform, we raise the age of first eligibility of all early retirement programs and normal retirement by three years. Thus, the age of first eligibility to PEW and early retirement through PEP is increased by three years (from sixty to sixty-three), and the age of first eligibility of TBP-UI is also delayed by three years, from fiftyfive to fifty-eight. The normal OAP retirement age is increased by three years. The age- and gender-specific probability of disability for those aged sixty to sixty-two is assumed to be that probability observed in the data at age fifty-nine, and for those aged sixty-three to seventy the age- and gender-specific probabilities are those observed in the data for individuals three years younger.15 In the second policy simulation, which we call the common reform, we simulate the response to imposing a single unified plan with the following features

• An early entitlement age of sixty • A normal retirement age of sixty-five • A replacement rate of 60 percent of projected age-sixty earnings if claimed at sixty-five

• A 6 percent per year actuarial reduction for claiming before sixty-five • A 6 percent per year actuarial increase for claiming after sixty-five • No other pathways to retirement (i.e., SDP or UI) For each type of policy reform, we model the effect of age on retirement in three different ways. The first approach is to implement the reforms on 14. Benchmark program rules relate to the most recent year—1995. 15. At the same time, we adjust the age at which the supplement for delayed retirement to PEW is effective from sixty-three to sixty-six.

178

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

the model with a linear age specification by recomputing SSW and applying the new SSW and accrual values to the estimated coefficients from that model and thereby deriving new retirement rates. This is referred to as simulation type S1. Note that the linear age specification will imply that people will have a greater preference for retirement as they age, as preferences for leisure rise smoothly with age. As a variant, we try using age dummies in place of the linear age in the estimation in order to capture any nonlinearities in the effect of age, thus allowing for spikes at the early and normal retirement ages (or at other program-induced variations) and nonmonotonicities in the demand for leisure at various ages. However, the estimated age dummies are not applied in the simulation. This is done in simulation type S2. As S2 most likely attributes too much of an effect to the age dummies (by including all program-induced spikes) and too little to the SSW measure, we try a variant to S2, which we label S3, that compensates for this. Under this approach, we use the estimated age indicators in a way to simulate the change in retirement due to the program. For example, in the case of the three-year reform, S3 shifts the profile for share in retirement at different ages (age-retirement) forward by three years by assigning the age-sixty-three dummy to age sixty and the age-sixty-four dummy to age sixty-one, and so forth. In the case of the common reform, age dummies are adjusted in a way as to retain the effect on retirement age of early and normal retirement and the expected increased desire for leisure at older ages but also to take out the effect of other retirement programs.16 Thus, by shifting the age dummies, S3 loads all of the estimated effect to the program changes, while S2 likely minimizes the effect of program changes. Therefore, S2 and S3 can be thought of as bounding the prediction, while S1 falls in the middle. We expect that in the Danish case, like most other countries, S3 will give the largest estimated effect because the age of eligibility is key to the retirement decision, as evidenced in section 3.5, and S2 will give the smallest effect. Results of the simulation exercise appear in table 3A.14 and in figures 3B.2–3B.10 for males and figures 3B.11–3B.19 for females (see appendixes A and B). There are nine figures in all for each gender (three simulation types for each of the three incentive measures). Each figure has two portions, the first showing the predicted hazards for the baseline, the threeyear reform, and the common reform and the latter showing the cumulative probabilities, or the proportion of the sample that retires at each age. Table 3A.14 compares baseline average retirement age to average predicted retirement age under each of the two reforms, applying each type of simulation to each of the three incentive measures. Looking at the results 16. The actual adjustment involves identifying the spike at the first eligibility age (not due to early retirement) and linearizing the age-retirement profile up to the early retirement age on the basis of the implied annual growth rate in retirement attributed to this program. Also, the retirement profile is made linear between normal and early retirement ages.

The Impact of Incentives on Retirement in Denmark

179

for males, we see that while the baseline retirement age ranges from between 60.27 to 60.77 years, the three-year reform leads to predicted retirement ages between 60.28 and 62.59 years, while the common reform results in average retirement ages ranging from about 60.38 to 61.23 years. In fact, under the three-year reform, all specifications predict higher retirement ages compared to the baseline, the only exception being the accrual case with the S1 assumption. Under the common reform, five of nine cases predict a higher average retirement age than baseline. Simulated average retirement ages are close for the peak value and option value estimates and predict a 0.9–3.1 percent rise in the average retirement age under the threeyear reform, and a modest fall in the average retirement age of about 1 percent in some cases (S3), while a slight rise in the average retirement age of about 0.6–1 percent in some cases (S2 and sometimes S1) under the common reform. All specifications show, however, that there is a definite difference in the Danish case as to whether or not age is entered linearly or captured through dummies and whether or not variation induced by other program eligibility is smoothed out in the age-dummy specification (i.e., between S2 and S3). The effects also vary depending on whether the incentive measure adopted is forward-looking or myopic. However, there is not a lot of difference between the two forward-looking measures. Very similar patterns are evidenced for the women in table 3A.14. While table 3A.14 shows the effects of each reform on the average retirement ages in our sample, to get an idea of how the changes resulting from each of the reforms are distributed across the age profile, we compare baseline retirement age hazards to simulated hazards for each specification, for each reform type, for each gender. Looking at figures 3B.2–3B.10 for males, we see that regardless of the incentive measure being considered, the baseline hazard lies between the common reform (higher hazard) and the three-year reform (lower hazard). This is true irrespective of whether age is entered linearly or captured through dummies. That is, retirement is predicted to increase with the common reform but decrease under the three-year reform. This is due to the generosity of the proposed common reform compared to the existing system, in terms of a higher average replacement rates and the earlier availability of normal social security benefits. Note that the S2 and S3 hazards in general show much more variability because of the use of age indicators, while the S1 hazard by definition rises smoothly with age. Interestingly, the baseline hazard reproduces the early and normal peaks at first availability ages (sixty and sixty-seven) in the accrual case only, while in the peak and option value cases, the early peak occurs first at age sixty-one, probably reflecting actuarial adjustment of PEW in a setting with forward-looking behavior. Comparing the relative effects of the different assumptions to our pri-

180

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

ors—the expectation that S3 should give the biggest effect and S2 the smallest effect, while S1 should lie in the middle—we see that at least for the peak and option value incentive measures, S3 predicts the greatest decrease in retirement, S2 the smallest, and S1 an intermediate decrease. For example, looking at the cumulative retirement probabilities for the male sample, the decline in the proportion of fifty to sixty-five year olds who would retire following the three-year reform compared to the baseline is approximately 13 percentage points under S1, 6–7 percentage points under S2 and 17–18 percentage points under S3 in peak and option value cases. Only in the accrual case does S1 give a smaller impact (in fact, a 1 percentage point increase in the proportion of fifty to sixty-five year olds who retire) than S2 (7 percentage point decrease) and S3 (15 percentage point decrease), but recall that the accrual model with the S1 assumption gave the worst fit and a positive sign on the incentive measure. Looking at the effects of the common reform on the cumulative distribution functions, again we see that S3 gives the greatest impact (in this case, increasing retirement), while S2 and S1 predict much smaller effects but still imply increased retirement, and S1 is bounded by S2 and S3 for all three incentive measures. For example, in the case of the forward-looking incentive measures, S1 predicts a 6 percentage point increase in retirement in the fifty to sixty-five age range, S2 predicts a 5 percentage point increase, and S3 predicts a full 27 percentage point increase. In the case of the accrual measure, S1 predicts an 8 percentage point increase in retirement, S2 actually predicts a 2 percentage point decrease in retirement, while S3 predicts a 23 percentage point increase in the proportion of fifty to sixty-five year olds who retire. In general, comparing across both types of reforms, all three incentive measures exhibit similar age-retirement hazards, although peak and option value estimates, both being forward-looking, are closer. In fact, the treatment of age and type of reform seems to make a bigger difference to the predicted retirement patterns than the type of incentive measure being considered. This derives from the fact that, in the Danish case, very similar coefficients were estimated on SSW and the incentive measure (except for accrual and linear age) in the probit regressions. Comparing the relative impacts of the reforms across the age profile (hazards), we again see that quite substantial effects are associated with both the reforms under S3 assumptions and more modest impacts with S1 and S2 assumptions. For example, using option value estimates, the S2 specification produces hazards that are bunched fairly tightly around the underlying baseline hazard (less than 0.05 difference in hazards), with the common hazard lying slightly above and the three-year hazard slightly below the baseline. Moving to S3, we see that by making the profile linear between the early (sixty-one) and normal (sixty-seven) peaks, the common reform distributes the mass of retirement between these two ages so that

The Impact of Incentives on Retirement in Denmark

181

there is dramatic upward shift of the hazard between sixty-two and sixtysix. In the three-year case, the simulated hazard is shifted to the right so that the early and normal peaks now appear at ages sixty-four and seventy. Similar results are obtained in the peak value case. Figures 3B.11–3B.19 show the comparable predictions for the females. For the females, we focus only on the figures pertaining to the option value incentive measure (3B.17–3B.19), as this model gives both precise estimates and an equally good fit for males and females. The results here are fairly similar to those for the males. The baseline retirement rate of the fifty to sixty-five year-old group is much higher than in the male sample, confirming the greater incidence of early retirement among women in Denmark. In fact, almost 92 percent are retired under S1 and 78 percent under S2 and S3. The comparable figures for the males (option value) are 79 percent (S1) and 72 percent (S2 and S3). The relative effects of the two reforms are the same for men and women, although the predicted effects of the reforms on the female sample are slightly smaller. Under the three-year reform, S1 predicts a 7 percentage point decrease, S2 predicts a 4.5 percentage point decrease, and S3 predicts a 13.5 percentage point decrease in the proportion of fifty to sixty-five year-olds who retire. Under the common reform, retirement is predicted to increase, just as for the males. Looking at the cumulative retirement probabilities in the fifty to sixty-five age group, S1 predicts a 2 percentage point increase, S2 predicts a 1.3 percentage point increase, and S3 predicts a 21 percentage point increase. To summarize, the three-year reform would delay retirement while the common reform would bring it forward, and the actual extent of these changes is quite sensitive to the treatment of the age effect. That is, the specification with age dummies (in which the program effect is maximized) shows large effects using both types of reforms, whereas the linear age specification and the simple age-dummy specification produce more modest effects. As pointed out earlier, the age-dummy specification in general produces a better fit of the data and produces signs and significance that conform more closely with prior expectations than the linear age model, so that simulated effects arising from this model can be considered particularly relevant. 3.8 Conclusions We model the impacts of incentives on retirements on micro-data using a 2 percent sample of the Danish population. An interesting aspect of retirement in Denmark is that, compared to most other OECD countries, participation is high in the sixty to sixty-six age group, despite the availability of a multitude of exit routes from the labor market. A possible explanation is that the financial incentives to retire early are strong among low-wage earners. Whereas for high-wage earners, the replacement rate is

182

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

typically low, creating financial incentives to continue working in the sixties. We construct a measure of SSW that integrates multiple policies, such as OAP, early retirement, disability, and retirement options for the longterm unemployed. The implicit tax rates derived from this measure confirm our expectations that there are tax rates on continued work for the lowest decile after the start of early retirement eligibility, while work is encouraged at older ages for the highest decile. Estimations of a probability model of retirement show that the provisions of the social security system play an important role in determining retirement behavior in Denmark. In particular, forward-looking incentive measures, such as peak and option value, have the expected significantly negative effect on the probability of retirement, both with a specification in which age is entered linearly as well as with a full set of age dummies. In the Danish data, for both males and females, the option value gives the best fit when age enters linearly, while the accrual model fits the data the best when age dummies are used instead. Furthermore, the specification with age dummies generally fits the data better than the specification with the linear age term. A simulation of changes in the social security system reveals that while a policy reform designed to raise the first eligibility age of early and normal retirement will increase the average retirement age, a common policy reform that replaces all programs by a single, unified plan with a 6 percent actuarial adjustment will reduce it. However, the simulated impacts vary according to the treatment of age and, to a lesser extent, to the incentive definition that is used. Assuming forward-looking behavior and allowing for variation in program availability and in age-specific demands for leisure that affect the ageretirement profile, we find quite a large delaying impact of the three-year reform on the retirement of men and women in Denmark.

The Impact of Incentives on Retirement in Denmark

183

Appendix A Table 3A.1

Age 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

Distribution of One-Year Accrual (males)

Median SSW ($)

Median ($)

10th Percentile ($)

90th Percentile ($)

SD

Median Tax Rate

143211.98 147333.55 151533.75 155283.28 159540.84 164108.61 168756.02 173405.06 178059.63 182432.17 185462.44 188715.20 172423.08 154982.63 138475.11 125049.11 108636.64 91859.85 90699.14 81051.21 75878.41 73303.66

— 3953.19 4077.58 4092.34 4257.56 4375.28 4352.42 4533.42 4712.42 4670.63 4875.97 4899.05 –11974.49 –11541.14 –7624.74 –4436.73 –4943.47 –3456.88 16647.82 –2563.63 –2678.45 –2607.57

— 3025.39 3002.89 2988.55 2784.72 2218.49 2118.35 1571.86 1357.88 1074.46 971.26 818.62 –14443.27 –15033.03 –15450.25 –12360.71 –12102.69 –11265.78 –11969.81 –3751.34 –3525.67 –3226.49

— 4380.95 4606.97 4636.06 5058.67 5153.81 5261.59 5248.23 5576.45 5560.48 5627.75 5962.23 594.59 497.30 334.98 774.72 463.85 413.28 25003.32 723.87 535.21 663.25

0.00 1124.24 1209.03 1314.57 1438.41 1538.61 1647.33 2440.39 2648.25 2813.21 3030.70 3210.62 6131.92 6237.18 6295.76 4954.03 5032.07 4795.36 15860.04 3238.97 3044.00 2807.40

— –0.103 –0.108 –0.108 –0.115 –0.119 –0.121 –0.124 –0.129 –0.131 –0.138 –0.151 0.298 0.257 0.239 0.149 0.144 0.135 –0.401 0.063 0.054 0.053

Note: SD = standard deviation. Dashes indicate that the data is not applicable.

0.00 95998.62 92267.14 88351.88 84418.66 79861.38 75674.55 71604.65 68729.24 65952.61 62820.40 55800.30 51582.30 46135.71 40393.09 34650.58 28972.44 23536.92 16994.13 12199.02 6125.14 0.00

49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

0.00 80207.07 76198.95 71896.80 67066.88 62170.80 57368.09 52396.02 47245.83 41598.39 36082.00 33419.48 31308.11 29344.13 26730.68 23767.39 20461.15 16960.61 12201.09 10204.67 5070.02 0.00

10th Percentile 0.00 145999.91 141785.91 139451.25 134649.56 128902.25 123826.98 117655.74 113015.88 107051.13 100789.26 90516.85 83921.17 76389.12 67905.03 59001.59 49717.47 40086.25 30184.55 24699.81 13109.13 0.00

90th Percentile

Peak Value

0.00 40843.25 39948.32 37140.73 36229.84 37320.10 36425.18 34233.61 33959.56 31543.21 33899.89 29623.71 30681.21 31733.36 33429.75 31969.22 28625.23 26708.35 28742.81 40530.54 39472.77 0.00

SD

The Distribution of the Peak Value and Option Value (males)

Note: SD = standard deviation.

Median

Age

Table 3A.2

0.00 136639.25 131129.72 125379.41 119710.90 113073.91 107064.45 101246.27 97687.55 93078.26 88576.70 78556.51 72670.40 65101.30 56913.44 48785.67 41163.05 33678.94 23838.55 17208.75 8608.13 0.00

Median 0.00 112674.36 106778.58 100552.45 93524.03 86394.81 79573.41 72568.01 65061.86 56989.33 49269.13 45356.14 42822.30 40064.57 36737.39 32901.03 28308.22 23583.20 16614.45 14261.82 7067.56 0.00

10th Percentile

0.00 210627.08 205131.20 202187.06 196106.69 187321.00 179689.70 170776.56 164401.78 155513.56 146412.72 131018.98 121491.27 110832.55 98234.39 85495.06 71998.82 58247.19 43517.37 36154.59 19322.92 0.00

90th Percentile

Option Value

0.00 53502.30 52914.45 50997.67 50782.11 51132.07 50469.34 48791.08 48481.36 46182.51 47545.41 42109.18 41779.68 41041.54 40336.55 37596.95 32790.45 29678.56 30849.18 42067.59 40356.11 0.00

SD

144659.48 148416.05 152511.17 156464.36 160282.16 164649.48 168975.81 173441.63 178123.92 183098.19 186957.84 190882.63 177716.50 162633.58 145355.69 132083.97 115132.77 105053.26 94835.80 89012.65 86178.20 82333.38

49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

— 3951.20 4062.42 4081.00 4347.88 4478.36 4390.75 4640.38 4763.59 4670.63 4901.69 5111.80 –12565.91 –12920.03 13232.73 –10163.13 –10597.73 –10617.13 –11090.40 –2662.57 –2860.09 –2949.91

Median ($) 0.00 344.38 333.20 319.39 336.07 349.55 385.83 372.50 403.73 1137.67 1777.05 2391.11 2552.22 2825.74 2970.29 2513.27 2575.84 2824.10 2164.53 1187.99 1004.80 718.25

SD — –0.092 –0.097 –0.100 –0.109 –0.116 –0.121 –0.127 –0.135 –0.140 –0.151 –0.173 0.402 0.410 0.424 0.320 0.349 0.352 0.470 0.495 0.105 0.136

Median Tax Rate

Note: SD = standard deviation. Dashes indicate that the data is not applicable.

Median SSW ($)

PEW Eligibility = 1

Distribution of One-Year Accrual, by PEW Eligibility (males)

Age

Table 3A.3

143192.61 147043.41 151145.36 155283.28 159312.33 163899.83 168639.20 173241.13 177554.11 178639.50 168343.95 82661.48 82116.45 79906.70 77147.30 71531.92 71335.55 68730.88 90262.70 78423.75 74588.34 73303.66

Median SSW ($) — 3953.19 4096.52 4107.78 4251.70 4347.45 4351.59 4506.92 4648.31 4582.16 4106.97 1305.92 79.44 –638.31 –404.95 –187.68 –708.75 –493.66 19970.09 –2432.77 535.21 663.25

Median ($) 0.00 1253.12 1386.79 1560.87 1768.42 1963.10 2194.05 3412.96 3932.27 4292.50 4633.19 4735.91 1336.81 1413.56 1794.21 2185.37 2618.84 3081.32 12409.89 3848.87 3438.39 3056.65

SD

PEW Eligibility = 0

— –0.107 –0.112 –0.113 –0.120 –0.122 –0.122 –0.119 –0.120 –0.108 –0.101 –0.064 –0.002 0.015 0.018 0.013 0.028 0.024 –0.941 0.034 –0.009 –0.029

Median Tax Rate

— 3951.20 4062.42 4081.00 4347.88 4478.36 4390.75 4640.38 4763.59 4670.63 4901.69 5111.80 –12565.91 –12920.03 –13232.73 –10163.13 –10597.73 –10617.13 –11090.40 –2662.57 –2860.09 –2949.91

49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

0.00 78345.64 74113.05 69634.33 65004.21 60227.28 55231.69 50384.81 45773.62 40690.94 35835.87 32659.98 30549.24 28553.48 25982.44 23013.68 19560.62 16065.14 12091.20 10098.49 4122.83 0.00

10th Percentile 0.00 176147.94 167060.50 156936.94 147801.02 138147.38 129593.52 120200.23 111405.98 103956.83 94915.47 84743.34 78251.56 72595.62 64703.67 56181.68 47947.02 38995.49 29509.38 22370.22 12474.96 0.00

90th Percentile 0.00 45123.92 47347.44 39092.98 38122.80 37026.28 35796.77 33184.37 32137.32 29120.59 30413.52 26483.39 27570.57 28467.25 25323.01 25227.28 27193.10 25033.66 19036.69 30098.33 17237.35 0.00

SD

Note: SD = standard deviation. Dashes indicate that the data is not applicable.

Median ($)

PEW Eligibility = 1

Distribution of One-Year Peak Value, by PEW Eligibility (males)

Age

Table 3A.4

0.00 93398.78 89014.29 84780.95 80176.71 75546.09 70907.43 65888.84 61839.89 57854.36 60428.18 56940.30 50124.45 44578.53 38858.59 33908.68 28933.37 23771.93 16737.10 11865.01 6024.20 0.00

Median ($) 0.00 80543.23 76507.66 72554.68 68106.31 63394.08 58740.35 54051.69 49130.46 43009.28 36704.02 35784.70 33217.43 30637.03 27558.50 24462.02 21212.44 17663.21 12233.29 10470.12 5394.47 0.00

10th Percentile

0.00 130055.45 125727.99 123755.31 121284.07 118573.37 116972.76 114079.93 114277.72 114273.30 114487.33 106598.52 94862.23 84686.77 73859.42 63792.27 51649.46 40956.27 30411.20 25534.32 13389.75 0.00

90th Percentile

PEW Eligibility = 0

0.00 38186.21 35393.29 34821.63 33840.25 36634.32 36295.34 34870.02 36074.86 35222.58 40090.95 36387.98 36077.74 36262.29 41597.45 37632.30 29722.87 27719.20 32164.07 44656.77 44542.27 0.00

SD

0.00 173164.97 165041.58 156052.89 146681.97 136647.53 126502.70 117515.35 107913.00 98503.56 88962.55 78168.72 72889.39 66018.77 58147.71 49485.38 41076.35 32252.06 24927.49 19023.69 9534.70 0.00

49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

0.00 109844.87 103622.81 96966.55 90449.73 83333.59 76162.58 69023.02 62395.09 55114.11 48182.57 43747.91 41011.63 38321.23 35041.63 31150.89 26527.32 21846.35 16492.87 13950.99 5631.07 0.00

10th Percentile

Note: SD = standard deviation

Median ($) 0.00 256692.00 244802.53 229619.72 216334.69 201684.09 188317.13 174969.17 162233.48 151050.78 138127.57 122543.40 113329.52 105151.66 93562.16 81182.95 69439.78 56461.01 42518.27 32759.92 18343.10 0.00

90th Percentile

PEW Eligibility = 1

0.00 62429.79 63675.74 55882.99 54735.04 52773.96 50940.24 47998.75 46407.99 43287.80 43475.67 38453.14 38136.07 37805.96 33178.30 31613.55 31914.52 28675.94 21569.37 31559.80 17951.85 0.00

SD

Distribution of One-Year Option Value, by PEW Eligibility (males)

Age

Table 3A.5

0.00 132535.16 126079.45 119963.83 113270.29 106557.66 99814.75 92595.00 87097.19 81633.84 85324.74 80333.11 71365.18 63371.10 54960.82 48360.00 41198.29 33916.35 23521.34 16749.21 8477.30 0.00

Median ($) 0.00 113170.81 107252.49 101639.92 94999.22 88176.17 81714.34 75190.66 68359.71 60069.43 50914.33 50223.87 46587.25 43120.09 38642.64 34495.96 29713.63 24747.58 16712.52 14670.22 7534.30 0.00

10th Percentile 0.00 187325.45 180955.42 178338.11 175599.36 171888.94 169587.75 165075.78 165928.45 164939.20 166952.78 155518.91 138393.75 122544.48 107002.02 92004.36 75033.62 59479.99 44294.38 37143.47 19758.03 0.00

90th Percentile

PEW Eligibility = 0

0.00 48273.08 45702.46 46187.25 46270.81 48458.29 48894.87 48919.37 50854.90 50618.36 54766.46 49830.68 48017.63 45532.73 47910.70 42773.06 33449.05 30278.17 34202.50 46254.99 45558.05 0.00

SD

Table 3A.6

Age 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

Distribution of One-Year Accrual (females)

Median SSW ($)

Median ($)

10th Percentile ($)

90th Percentile ($)

SD

Median Tax Rate

158875.05 163250.86 167168.27 170994.97 174423.78 177859.48 180839.23 183404.34 185599.83 187222.30 187168.89 186051.48 127766.96 121948.33 110386.89 102262.80 97501.88 91118.83 101632.31 97780.57 94683.20 92215.84

— 3778.25 3629.19 3474.39 3494.22 3567.14 3430.91 3137.95 2974.00 2525.63 2036.66 1506.25 –3049.71 –2988.79 –3318.54 –3635.62 –4370.25 –5056.95 15095.95 –759.34 –226.01 525.07

— 2148.21 1824.48 1609.57 1463.80 1282.34 1127.76 597.71 38.72 –782.88 –1361.98 –1685.02 –13152.69 –13205.45 –13535.91 –11325.13 –11328.67 –9928.78 –9779.96 –3172.29 –3189.26 –3237.77

— 4875.30 4587.98 4594.31 4563.17 4799.66 4745.67 4329.91 4210.06 4010.81 3870.95 3618.73 –463.41 –354.95 –1270.05 –1862.73 –2338.30 –2427.97 21295.80 769.04 619.16 557.57

0.00 1037.87 1113.12 1190.89 1287.50 1409.93 1482.65 2210.97 2362.08 2439.66 2464.56 2632.93 5261.08 4920.24 4588.36 3625.02 3425.55 3139.08 11449.18 1958.09 1838.10 1872.84

— –0.171 –0.162 –0.156 –0.153 –0.154 –0.157 –0.138 –0.128 –0.117 –0.104 –0.085 0.256 0.262 0.296 0.341 0.401 0.492 –1.219 0.060 0.022 0.039

Note: SD = standard deviation. Dashes indicate that the data is not applicable.

0.00 73254.02 69649.49 65982.53 62128.19 58428.43 54644.05 50663.78 46736.94 42540.49 38600.55 36496.20 33361.45 30578.87 27333.56 24231.50 21017.32 17670.24 12678.47 8968.79 4617.27 0.00

49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

0.00 61960.35 58796.41 55040.00 51109.70 47364.20 43568.14 39793.71 36120.79 32016.75 27549.10 25742.69 23470.67 21480.78 19283.95 17308.75 15066.95 12728.39 8799.90 7935.75 4075.60 0.00

10th Percentile 0.00 112371.68 110292.34 108609.36 105843.20 102967.57 98454.13 94639.33 89422.42 84019.50 78090.82 69307.23 63965.91 55934.20 47550.36 39679.36 32693.70 25299.68 17848.43 16927.49 9296.94 0.00

90th Percentile

Peak Value

0.00 27539.52 26863.52 26497.97 26571.26 25994.87 25323.36 24578.72 24333.65 23542.80 22781.16 21364.57 20029.00 18084.79 16255.56 14639.42 13604.44 10970.26 8957.09 9001.85 10243.72 0.00

SD

The Distribution of the Peak Value and Option Value (females)

Note: SD = standard deviation.

Median

Age

Table 3A.7

0.00 101847.41 96553.86 91467.79 86076.74 80706.56 75264.63 69711.48 64094.30 57857.30 52620.56 49632.83 45501.77 41686.16 37496.23 33519.20 29275.98 24805.88 17450.93 12369.13 6351.05 0.00

Median 0.00 85326.71 80760.09 75187.83 69855.80 64267.14 59140.05 54018.93 48811.80 43055.67 36919.44 34339.11 31282.97 28474.38 25536.77 23095.54 20210.93 17309.74 11609.68 10854.78 5578.86 0.00

10th Percentile 0.00 158978.27 155347.13 153139.88 150690.16 146891.14 140455.75 134767.80 127007.38 119284.64 111517.75 98676.51 91188.53 79550.38 67198.67 55972.61 46135.87 34991.93 25237.64 24225.58 13371.85 0.00

90th Percentile

Option Value

0.00 37911.62 37746.70 37792.40 38093.64 37550.34 36779.75 35880.15 35413.09 34389.09 33535.62 30339.47 27903.97 24660.69 21961.68 19176.94 17112.18 13523.97 10983.92 10945.89 10813.16 0.00

SD

172279.63 177029.50 181382.31 185597.08 188909.05 192193.84 194583.20 195308.52 198435.34 200314.03 203418.72 206200.86 192540.77 179060.16 167233.81 156591.31 144520.44 128108.98 116280.27 109058.48 105627.53 102025.34

49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

— 4460.31 4049.09 3919.50 3905.28 3934.22 3883.86 3678.28 3525.64 3244.08 2871.03 2483.27 –11902.61 –11825.50 –11921.13 –10842.59 –11294.75 –10593.06 –11698.32 –2955.10 –3137.51 –3149.94

Median ($) 0.00 604.53 582.60 598.04 525.73 606.89 674.48 625.23 746.27 1325.41 1939.07 2576.18 3015.18 3247.32 3234.71 2554.46 2579.68 2729.40 2665.10 871.67 806.96 762.73

SD — –0.182 –0.171 –0.168 –0.164 –0.170 –0.174 –0.166 –0.165 –0.155 –0.143 –0.136 0.631 0.620 0.655 0.647 0.699 0.700 0.798 0.159 0.145 0.223

Median Tax Rate

Note: SD = standard deviation. Dashes indicate that the data is not applicable.

Median SSW ($)

PEW Eligibility = 1

Distribution of One-Year Accrual, by PEW Eligibility (females)

Age

Table 3A.8

158328.14 162345.77 165855.72 169137.67 172100.97 174129.33 167957.14 131674.58 129489.27 125704.91 123376.51 111139.90 110361.95 106783.88 103802.38 99920.37 92382.86 86098.36 101194.30 93152.98 93254.95 91445.42

Median SSW ($) — 3675.14 3547.07 3404.56 3399.58 3400.56 2953.94 1832.60 1177.84 616.63 22.14 –728.80 –1203.05 –1767.21 –2601.23 –3199.66 –4012.11 –4802.48 16528.70 –599.68 489.46 525.07

Median ($) 0.00 1049.65 1175.02 1268.78 1406.89 1554.19 1605.95 2513.86 2675.07 2613.96 2284.70 1843.88 1404.68 1523.60 1687.35 1985.50 2141.76 2313.25 8366.21 1980.12 1805.61 1800.51

SD

PEW Eligibility = 0

— –0.169 –0.161 –0.153 –0.147 –0.145 –0.146 –0.113 –0.094 –0.054 –0.003 0.075 0.119 0.187 0.243 0.290 0.378 0.474 –1.288 0.043 –0.036 –0.042

Median Tax Rate

0.00 95605.45 89993.39 85124.35 79637.95 74240.77 69141.58 64009.86 58591.16 53902.96 49437.16 43040.61 40287.81 35618.43 30636.93 25993.58 22048.64 17477.33 13933.08 11174.88 6215.48 0.00

49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

0.00 61851.51 58593.09 55176.91 51485.41 47896.21 44378.41 41007.68 37317.43 33559.91 29708.43 27068.70 24534.68 22493.97 20065.31 17436.18 15006.32 12168.59 9339.72 7869.88 4050.70 0.00

10th Percentile

Note: SD = standard deviation.

Median ($) 0.00 145391.44 139655.84 131970.47 125261.01 117891.38 110513.95 103323.50 96275.00 88738.91 81505.33 72860.01 68137.14 62048.95 54638.42 48090.66 40418.93 33120.81 26988.82 20382.88 11054.96 0.00

90th Percentile

PEW Eligibility = 1

0.00 36232.81 35068.63 32860.33 32095.00 29661.91 28383.11 26437.79 25501.71 24622.58 23339.75 21451.06 21224.38 18935.78 18961.02 16974.29 17605.03 17178.88 13166.59 19949.90 25030.39 0.00

SD

Distribution of One-Year Peak Value, by PEW Eligibility (females)

Age

Table 3A.9

0.00 71556.85 67843.20 64146.85 60318.45 56359.88 52269.15 48260.95 43883.57 39665.29 35120.91 33019.02 31219.29 29197.65 26721.78 24057.83 20979.04 17676.40 12592.52 8911.81 4605.22 0.00

Median ($) 0.00 61978.05 58837.06 54958.70 51072.74 47289.96 43335.41 39422.82 35670.38 31194.91 26341.29 24903.47 22975.66 20994.07 18969.50 17277.93 15066.95 12885.25 8718.13 7935.84 4075.60 0.00

10th Percentile 0.00 98637.68 93019.09 87709.34 82094.46 77045.84 71744.05 67785.00 66006.34 68237.04 68177.75 53421.48 47691.91 41995.66 35364.19 31779.15 26656.53 21916.19 15043.08 14885.39 7503.13 0.00

90th Percentile

PEW Eligibility = 0

0.00 23526.33 22109.54 21813.02 21754.12 21809.87 21190.17 21074.98 21398.66 20657.67 20418.56 19944.98 17938.27 16882.36 14472.03 13595.73 12248.40 9211.92 8156.84 5120.08 6166.64 0.00

SD

0.00 134147.50 126867.57 119724.80 111835.73 104005.23 96553.91 89138.45 81585.05 74443.11 67990.88 58596.03 54959.26 48577.20 41542.93 35145.30 29766.93 23838.88 19334.79 15646.53 8438.62 0.00

49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

0.00 84864.45 80297.89 75187.83 69973.96 64642.27 59764.71 54854.74 49697.86 44514.33 39140.73 35462.63 31976.84 29388.59 26237.23 22630.57 19502.27 15993.09 12504.49 10781.68 5616.31 0.00

10th Percentile

Note: SD = standard deviation.

Median ($) 0.00 210744.33 201390.69 190477.56 181278.67 170298.03 159382.20 148559.20 138484.11 127242.63 116923.89 104068.16 97238.21 88718.13 77959.25 68281.29 58052.09 47005.20 38511.39 29834.01 16152.96 0.00

90th Percentile

PEW Eligibility = 1

0.00 51883.65 50393.89 48030.68 47014.70 44171.09 42322.17 39629.84 38025.50 36475.11 34818.37 31663.36 30716.46 27458.41 26512.71 23110.82 22722.51 21140.82 16072.26 22360.39 26552.87 0.00

SD

Distribution of One-Year Option Value, by PEW Eligibility (females)

Age

Table 3A.10

0.00 99599.39 94146.25 88794.46 83296.77 77532.70 72063.32 65810.00 59888.31 54155.42 47858.95 45194.18 42727.07 40151.02 36823.88 33330.85 29266.82 24852.68 17326.89 12194.15 6339.21 0.00

Median ($) 0.00 85378.34 80847.03 75198.69 69863.76 64168.23 58816.43 53723.98 48564.03 42300.46 35531.66 33557.15 30913.91 28236.46 25409.79 23229.79 20336.63 17597.40 11485.98 10854.78 5578.86 0.00

10th Percentile

0.00 139409.00 131850.97 124107.88 115976.00 108908.18 101332.74 96194.48 92764.79 93404.23 93691.13 73047.31 65565.77 56804.77 48300.81 42249.71 35439.90 29147.52 20307.77 21804.79 10975.90 0.00

90th Percentile

PEW Eligibility = 0

0.00 31308.58 30268.22 30260.30 30345.94 30417.57 29784.21 29835.38 30222.60 29669.51 29586.44 26594.31 23756.51 21812.04 18981.63 17419.69 15173.68 11387.22 9978.14 7334.26 6567.07 0.00

SD

87.33411 56.62542 0.165964 1 1744.804 67913.57 96548.69 0.0881529 35970.25 0.0235063 0.1174734 0.1501997 0.1138985 0.1263061 0.2233629 0.2452531

Mean 4.755396 5.369312 0.3720491 0 6282.143 45786.08 63356.46 0.2835178 20782.79 0.1515052 0.3219843 0.3572678 0.3176887 0.3321948 0.4165005 0.4302381

SD 80 49 0 1 –19090.28 –51380.59 –41800.23 0 4755.71 0 0 0 0 0 0 0

Min. 95 70 1 1 33712.56 15151.64 15178.19 1 160952.2 1 1 1 1 1 1 1

Max.

Notes: Obs. = observations; SD = standard deviation; Min. = minimum; Max = maximum.

224,621 224,621 224,621 224,621 207,483 219,827 213,140 224,621 224,621 224,621 224,621 224,621 224,621 224,621 224,621 224,621

Obs.

Males

Sample Descriptives for Males and Females

Year Age PEW eligible Male Accrual Peak value Option value Now retired Earnings Top management Middle management Lower management White-collar Skilled blue-collar Unskilled blue-collar Unclassified occupation

Table 3A.11

210,073 210,073 210,073 210,073 193,414 205,664 199,806 210,073 210,073 210,073 210,073 210,073 210,073 210,073 210,073 210,073

Obs. 87.11039 56.45707 0.0937817 0 1161.496 50183.04 69841.08 0.1020026 19368.03 0.0008188 0.0159706 0.0790487 0.2159107 0.0055409 0.2353468 0.4473635

Mean 4.745704 5.306651 0.2915254 0 4972.877 32924.22 46402.92 0.3026525 12007.51 0.0286024 0.1253621 0.2698154 0.4114536 0.0742311 0.4242164 0.4972229

SD

Females

80 49 0 0 –18489.91 –81760.66 –74182.77 0 1572.583 0 0 0 0 0 0 0

Min.

95 70 1 0 34111.59 671790.4 673198.2 1 157701.3 1 1 1 1 1 1 1

Max.

145,422 145,422 145,422 145,422 130,920 142,457 138,547 145,422 145,422 145,422 145,422 145,422 145,422 145,422 145,422 145,422

Year Age PEW eligible Male Accrual Peak value Option value Now retired Earnings Top management Middle management Lower management White-collar Skilled blue-collar Unskilled blue-collar Unclassified occupation

Notes: See table 3A.11.

125,510 125,510 125,510 125,510 111,233 122,835 119,203 125,510 125,510 125,510 125,510 125,510 125,510 125,510 125,510 125,510

Obs.

87.55998 56.17505 0 0 1519.716 45547.52 63103.7 0.1067032 17692.1 0.0008664 0.0168131 0.0763227 0.1709576 0.0045248 0.1793951 0.5511202

88.47216 55.6616 0 1 3265.991 65849.21 93407.71 0.0790216 35625.5 0.248745 0.1399888 0.1456697 0.1143574 0.1023504 0.1601068 0.3126524

Mean

4.918789 5.666062 0 0 4248.528 30903.36 42810.15 0.3087367 11523.15 0.0294228 0.1285713 0.2655147 0.3764732 0.0671142 0.3836842 0.4973816

4.852045 5.720727 0 0 4599.857 47543.12 64586.35 0.2697735 21936.92 0.1557433 0.3469768 0.3527761 0.3182461 0.3031098 0.3667065 0.4635759

SD

PEW Eligibility = 0

Sample Descriptives, by PEW Eligibility

Year Age PEW eligible Male Accrual Peak value Option value Now retired Earnings Top management Middle management Lower management White-collar Skilled blue-collar Unskilled blue-collar Unclassified occupation

Table 3A.12

95 70 0 1 33172.56 15151.64 15178.19 1 160952.2 1 1 1 1 1 1 1 95 70 0 0 34111.59 671790.4 673198.2 1 157701.3 1 1 1 1 1 1 1

Females 80 49 0 0 –12924.88 –81760.66 –74182.77 0 1572.583 0 0 0 0 0 0 0

Max.

Males 80 49 0 1 –19090.28 –51380.59 –41800.23 0 4755.71 0 0 0 0 0 0 0

Min.

64,651 64,651 64,651 64,651 62,494 63,207 61,259 64,651 64,651 64,651 64,651 64,651 64,651 64,651 64,651 64,651

99,111 99,111 99,111 99,111 96,250 96,992 93,937 99,111 99,111 99,111 99,111 99,111 99,111 99,111 99,111 99,111

Obs.

86.09912 57.09143 0.3047285 0 411.0516 60630.65 85078.74 0.0914294 23137.75 0.0007115 0.0140756 0.0851804 0.3170253 0.0078266 0.3612009 0.2139797

85.89295 57.84597 0.3761338 1 –13.18183 70527.97 100534.5 0.0997165 36406.83 0.0217736 0.0889609 0.1559363 0.1133174 0.1566426 0.3034678 0.1599015

Mean

4.157165 4.325368 0.4602959 0 6155.572 34887.7 50431.44 0.288221 12218.44 0.0266649 0.1178035 0.2791522 0.465321 0.0881221 0.4803524 0.4101158

4.209636 4.607424 0.4844167 0 7406.575 43318.34 61530.46 0.2996231 19213.28 0.1459441 0.2846887 0.3627967 0.316982 0.3634653 0.4697578 0.3665166

SD

PEW Eligibility = 1

80 49 0 0 –18489.91 –70460.58 –63700.79 0 2048.208 0 0 0 0 0 0 0

80 49 0 1 –16842.5 –40433.14 –32911.92 0 5609.921 0 0 0 0 0 0 0

Min.

95 70 1 0 7572.984 520721.4 62470.1 1 10838.3 1 1 1 1 1 1 1

95 70 1 1 7792.203 14405.81 15164.11 1 160952.2 1 1 1 1 1 1 1

Max.

0.29 –44879.02

Pseudo R 2 Log-likelihood

Pseudo R 2 Log-likelihood

SSW Coefficients Standard error Marginal effect IMEASURE Coefficients Standard error Marginal effect

–0.0355246 0.0008928 –0.0025742

0.0036759 0.000608 0.0002882

0.3189 –43.938,555

–0.0337874 0.001249 –0.0037461

0.0312452 0.000776 0.0037686

0.2359 –49.291,676

0.0031201 0.0001504 0.0003459

0.0034211 0.0001374 0.0004126

0.3658 –40085.73

0.0000568 0.0001361 4,12E-06

Age Dummies

0.0023437 0.000122 0.0001838

Linear Age

Accrual Incentive

Retirement Probits

SSW Coefficients Standard error Marginal effect IMEASURE Coefficients Standard error Marginal effect

Table 3A.13

0.2364 –51.215,638

–0.0177367 0.0002482 –0.0018442

0.003988 0.0001358 0.0004147

Females

0.3104 –44.763,012

–0.0159361 0.0002211 –0.0009874

0.0021998 0.0001238 0.0001363

Males

Linear Age

0.2835 –4805.4

–0.0044163 0.0004025 –0.0005069

0.0031184 0.0001482 0.0003579

0.321 –44.077,778

–0.0050171 0.000312 –0.0003915

0.0009605 0.0001306 0.000075

Age Dummies

Peak Value Incentive

0.2368 –49.653,267

–0.0128501 0.0001804 –0.0013209

0.0037542 0.0001368 0.0003859

0.3117 –43.327,098

–0.0112939 0.0001574 –0.0006887

0.0020257 0.0001252 0.0001235

Linear Age

0.2835 –46.618,008

–0.0041298 0.0002887 –0.0004711

0.0030627 0.00015 0.0003494

0.3217 –42.699,919

–0.0041719 0.0002204 –0.0003228

0.0009254 0.0001325 0.0000716

Age Dummies

Option Value Incentive

Table 3A.14

Average Retirement Age in Simulations Simulated Reform

Baseline

Plus Three Years

Difference from Baseline

Common Reform

Difference from Baseline

Males Accrual S1 S2 S3 Peak value S1 S2 S3 Option value S1 S2 S3

60,44432 60,27340 60,27340

60,28469 60,63449 62,28802

–0.15963 +0.37009 +2.01462

6,38432 60,95542 59,60625

–0.06000 +0.68202 –0.66715

60,76490 60,68337 60,68337

62,12958 61,23892 62,57302

+1.36468 +0.55555 +1.88965

61,18836 61,08496 60,09098

+0.42346 +0.40159 –0.59239

60,77458 60,69852 60,69852

62,16372 61,32813 62,59626

+1.38914 +0.62961 +1.89774

61,23221 61,13278 60,14049

+0.45763 +0.43426 –0.55803

Females Accrual S1 S2 S3 Peak value S1 S2 S3 Option value S1 S2 S3

60,44432 60,27340 60,27340

60,28469 60,64349 62,28802

–0.15963 +0.37009 +2.01462

60,38432 60,95542 59,60625

–0.06000 +0.68202 –0.66715

60,76490 60,68337 60,68337

62,12958 61,23892 62,57302

+1.36468 +0.55555 +1.88965

61,18836 61,08496 60,09098

+0.42346 +0.40159 –0.59239

60,77458 60,69852 60,69852

62,16372 61,32813 62,59626

+1.38914 +0.62961 +1.89774

61,23221 61,13278 60,14049

+0.45763 +0.43426 –0.55803

Appendix B A

B

Fig. 3B.1

The retirement hazard and age dummies: A, Males; B, Females

Fig. 3B.2

A

S1 males using accrual estimates: A, Hazard; B, CDF

Fig. 3B.2

B

(cont.)

Fig. 3B.3

A

S2 on males using accrual estimates: A, Hazard; B, CDF

Fig. 3B.3

B

(cont.)

Fig. 3B.4

A

S3 on males using accrual estimates: A, Hazard; B, CDF

Fig. 3B.4

B

(cont.)

Fig. 3B.5

A

S1 on males using peak value estimates: A, Hazard; B, CDF

Fig. 3B.5

B

(cont.)

Fig. 3B.6

A

S2 on males using peak value estimates: A, Hazard; B, CDF

Fig. 3B.6

B

(cont.)

Fig. 3B.7

A

S3 on males using peak value estimates: A, Hazard; B, CDF

Fig. 3B.7

B

(cont.)

Fig. 3B.8

A

S1 on males using option value estimates: A, Hazard; B, CDF

Fig. 3B.8

B

(cont.)

Fig. 3B.9

A

S2 on males using option value estimates: A, Hazard; B, CDF

Fig. 3B.9

B

(cont.)

Fig. 3B.10

A

S3 on males using option value estimates: A, Hazard; B, CDF

Fig. 3B.10

B

(cont.)

Fig. 3B.11

A

S1 on females using accrual estimates: A, Hazard; B, CDF

Fig. 3B.11

B

(cont.)

Fig. 3B.12

A

S2 on females using accrual estimates: A, Hazard; B, CDF

Fig. 3B.12

B

(cont.)

Fig. 3B.13

A

S3 on females using accrual estimates: A, Hazard; B, CDF

Fig. 3B.13

B

(cont.)

Fig. 3B.14

A

S1 on females using peak value estimates: A, Hazard; B, CDF

Fig. 3B.14

B

(cont.)

Fig. 3B.15

A

S2 on females using peak value estimates: A, Hazard; B, CDF

Fig. 3B.15

B

(cont.)

Fig. 3B.16

A

S3 on females using peak value estimates: A, Hazard; B, CDF

Fig. 3B.16

B

(cont.)

Fig. 3B.17

A

S1 on females using option value estimates: A, Hazard; B, CDF

Fig. 3B.17

B

(cont.)

Fig. 3B.18

A

S2 on females using option value estimates: A, Hazard; B, CDF

Fig. 3B.18

B

(cont.)

Fig. 3B.19

A

S3 on females using option value estimates: A, Hazard; B, CDF

Fig. 3B.19

B

(cont.)

234

Paul Bingley, Nabanita Datta Gupta, and Peder J. Pedersen

References Christensen, B. J., and Datta Gupta, N. 1999. The effects of pension systems reform on retirement and government finances: Predictions using Danish data on married couples. Aarhus School of Business. Working Paper no. 99-01, January. ———. 2000. Effekten af pensionsreform på danske ægtepars udtræden af arbejdsmarkedet (The effect of pension system reform on the retirement behavior of married couples). Nationaløkonomisk Tidsskrift 138 (2): 222–42. Coile, C., and Gruber, J. 2000. Social security incentives for retirement. NBER Working Paper no. 7651. Cambridge, Mass.: National Bureau of Economic Research, April. Danø, A. M., M. Ejrnæs, and L. Husted. 2000. Hvordan påvirker efterlønsreformen de ældres tilbagetrækningsalder (How does the reform of PEW influence the retirement age)? Nationaløkonomisk Tidsskrift (138 (2): 205–21. Economic Council. 1998. The Danish economy, Autumn 1998 (in Danish), Copenhagen: Economic Council. Finansministeriet (Ministry of Finance). 1996. Finansredegørelse 1996 (Fiscal Report 1996). Copenhagen, Ministry of Finance. Finansrådet (Financial Council). 1998. Opsparing og fremtidig velfærd (Savings and future welfare). Copenhagen: Financial Council. Gruber, J., and D. A. Wise. Eds. 1999. Social security aknd retirement around the world. Chicago: University of Chicago Press. Ministry of Economics. 2000. A Stable Pension System (in Danish). Copenhagen: Ministry of Economics. Pedersen, P. J. 1998. De ældre og arbejdsmarkedet (The elderly and the labor market). In Arbejde, incitamenter og ledighed (Work, incentives and unemployment), ed. N. Smith, 15–78. Aarhus: Aarhus University Press. Roseveare, D., W. Leibfritz, D. Fore, and E. Wurzel. 1996. Aging populations, pension systems and government budgets: Simulations for 20 OECD countries. Economics Department Working Paper no. 168. Paris: OECD. Socialkommissionen (Danish Government Social Commission). 1993. De ældre— En belysning af ældregenerationens forsørgelse, rapport nr. 5 (An analysis of the provision for the elderly). Copenhagen: Social Commission. Statistics Denmark. 1999. Statistical reports. Labour market series no. 37. Copenhagen: Statistics Denmark. ———. 2003. Integrated database for the labour market. Copenhagen: Statistics Denmark. Stock, J. H., and D. A. Wise. 1990. Pensions, the option value of work, and retirement. Econometrica 58 (5): 1151–80.

4 Estimating Models of Retirement Behavior on French Data Ronan Mahieu and Didier Blanchet

4.1 Introduction The pension debate in France has essentially focused, until recently, upon ways of financing retirement, with a strong opposition between supporters of maintaining the quasi-exclusivity for pay-as-you-go (PAYG) financing and supporters of a progressive introduction of funded pensions, in addition to the existing PAYG basic and complementary schemes. However, attention has shifted recently toward another variable of adjustment to the new demographic context, which is the age at retirement or, more widely, the age at exit from the labor force (Bommier, Magnac, and Roger 2001). The mean age at retirement in France is in the lower tail of the European distribution and has kept on diminishing for the past twenty years for two main reasons.

• The incentive structure of the pension system, itself creates issues, especially since the introduction of the retraite à 60 ans. Until 1982, the first age of eligibility to social security (SS) benefits was sixty-five, and it was shifted to sixty for all wage earners in 1983. Retirement before reaching a total tenure of 37.5 years remained strongly penalized in the régime général, which covers about 65 percent of wage earners, but this constraint did not bind most older male workers (since they often Ronan Mahieu manages the Department of Statistics and Research of the Caisse Nationale des Allocations Familiales (a subdivision of the French social security system that pays family benefits). Didier Blanchet is head of the Department of Employment and Labor Income Statistics at the National Institute of Statistics and Economic Studies (INSEE), Paris. We thank the Direction de la Recherche, des Etudes, de l’Evaluation et des Statistiques (DREES) for access to data from the Echantillon Interrégime de Retraités. We also thank Béatrice Sedillot, Thierry Magnac, and Guy Laroque for their useful comments. The usual disclaimer applies.

235

236

Ronan Mahieu and Didier Blanchet

began to work about age fifteen) and thus did not prevent a continuous decline in participation rates. • The relative generosity of unemployment benefits or early retirement provisions before this age of sixty is the second reason. As unemployment rose in the 1970s, the generosity of these schemes was expanded, allowing people aged between sixty and sixty-five to retire. When the early retirement age was set at sixty, unemployment or early retirement provisions were targeted at people aged fifty-five to fifty-nine, whose participation rates also sharply decreased. Age at retirement appears to be the key economic variable for potential adjustments. A first step in the direction of increasing employment rates was the 1993 reform of the régime général, which planned a progressive strengthening of the conditions giving access to “normal” (full-rate) retirement at age sixty: The previous condition was the accumulation of at least 150 quarters (or 37.5 years) of contributions to pension schemes. This threshold progressively increases to reach 160 quarters (i.e., forty years) from 2003. One of the propositions discussed in the Charpin report (Charpin 1999; ordered by the prime minister) is to go further in the same direction, raising this threshold to 170 quarters. Of course, modifying this state of affairs raises many issues. Early retirement policies are historically a response to an employment shortage, and it is often feared that less permissive policies may worsen the situation on the labor market. Conversely, policies of early withdrawal from the labor force have never proved to be of any help in mitigating employment problems (unemployment reached a peak at 12.4 percent of the labor force in 1997, which was high compared to other European countries). Another issue is identifying the best way to induce people to leave the labor force later. A first possibility is coercion. A second one preferably relies on incentives, with the idea to compensate for the desired increase of the average retirement age by the introduction of more flexibility in this retirement age (Taddei 2000). This is specifically the option proposed by the Charpin report, which suggests to compensate for the strengthening of conditions necessary to get a normal pension at age sixty with a reduction of penalties associated with either anticipating or postponing retirement. The French system is characterized by a strong deviation from marginal actuarial fairness, and the proposition consists in bringing it closer to this rule. This context calls for closer inspection on what factors determine retirement behavior, from both demand and supply sides. The analysis presented below will essentially focus on supply-side effects, although we shall try, systematically, to remind the reader of the importance of the demand side. We shall proceed in five steps. We shall first recall the features of the

Estimating Models of Retirement Behavior on French Data

237

French pension system, which will be necessary to understand the rest of the paper (section 4.2). The analysis will concentrate on two subpopulations: wage earners belonging to the private sector and civil servants. We shall give a detailed account of rules governing pensions for these two categories, including possibilities of early exit from the labor force, through either early retirement or specific features of unemployment insurance. We shall then describe data sets that are used and explain why we focused the analysis on three specific cohorts: cohort 1930 for workers in the private sector and cohorts 1930 and 1932 for civil servants (section 4.3). We shall then move to a descriptive analysis of incentives to withdraw from the labor force that applied to these cohorts, given their specific histories. This analysis completes the one performed earlier on this specific sample by Blanchet and Pelé (1999; section 4.4). These incentives will then be introduced in probit models of withdrawal from the labor force (section 4.5). Section 4.6, at last, will present simulations derived from these models. 4.2 Basic Facts about the French Pension System 4.2.1 The General Structure The French system is complex, but its structure can nevertheless be summed up quite simply. For a large part of the population (wage earners in the private sector), pensions rely on two compulsory pillars.

• The basic general scheme (SS) offers benefits corresponding to the share of gross wages below a SS ceiling (€2,352 per month in 2002). In 1992, 70.5 percent of people aged sixty or older received benefits from this general scheme. On the contributors’ side, in the same year, the general scheme gathered 64.8 percent of the labor force. • Complementary schemes are organized on a occupational basis. They consist of a large number (about 180) of specific schemes that are federated in two main organisms, ensuring interscheme demographic compensation: the Association Générale des Institutions de Retraite des Cadres (AGIRC) for executive workers, which applies only to the fraction of their wages over the SS ceiling, and the Association des Régimes de Retraite Complémentaire (ARRCO) for other workers’ and executives’ wages below the ceiling. In 1972, contributing to a complementary scheme became compulsory. Today, complementary schemes provide 40 percent of pensions for wage earners in the private sector. Receiving a complementary pension is conditioned on receiving SS benefits. Besides this simple two-pillar structure, the complexity of the French system is essentially due to the existence of a large number of exceptions to

238

Ronan Mahieu and Didier Blanchet

this general rule of organization. These exceptions are the result of two factors. When SS was created in 1945, people who already benefited from more generous dispositions refused to join the new system (for instance civil servants or people employed in large state-owned companies). Conversely some categories chose cheaper systems offering lower protection because they thought that a large part of their retirement needs was likely to be covered by other sources, such as professional assets for the selfemployed. Besides the two-pillar system constituted by the general scheme and ARRCO and AGIRC, there are a multiplicity of specific schemes (e.g., those for civil servants and the self-employed) applying specific rules. In particular, it must be observed that civil servants are not really covered by an autonomous pension system, since their pensions are directly paid on the state budget. For all categories of people, there is, at last, a system of old-age minimum allowance (minimum vieillesse), which is a means-tested allowance available for people aged sixty-five or older. The population benefiting from this minimum pension has regularly declined in the past, due to the increasing maturity of normal pensions. It is now slightly below 1 million, compared to 2.55 million in 1959 (Commissariat Général du Plan 1995). The following analysis will deal with two subpopulations: wage earners from the private sector and civil servants. We now give more details about the computation of pensions for these two categories. 4.2.2 Wage Earners in the Private Sector: Rules for the General Regime The basic general scheme offers contributory benefits corresponding to the share of wages below the SS ceiling. The SS benefits are proportional to the number of quarters of contribution to the system (truncated to Nmax quarters), and to a reference wage that, until 1993, has been the average wage of the D best years of the pensioners’ career (past nominal wages being reevaluated at time of liquidation according to a set of retrospective coefficients). The equation giving the initial pension level is therefore (1)




N of quarters, truncated to Nmax Pension
   Nmax



 (average wage of the D best years) with the proportionality coefficient  being itself modulated. It is maximal when the pensioner leaves at age sixty, with Nmax quarters of contributions or more to all pension schemes: In that case, its value is set at 50 percent, and this exactly ensures a replacement rate of the reference wage (not necessarily the last wage) equal to 50 percent. The same value of  also applies whatever the number of years contributed when the individual leaves at age sixty-five. In all other cases, the coefficient is reduced (table 4.1)

Estimating Models of Retirement Behavior on French Data Table 4.1

239

Value of
Depending on Age at Receipt of First Benefit and N, Number of Quarters of Contribution to the General Regime Age (%)

N

60

61

62

63

64

65

32.5 33.5 34.5 35.5 36.5 37.5

25 30 35 40 45 50

30 30 35 40 45 50

35 35 35 40 45 50

40 40 40 40 45 50

45 45 45 45 45 50

50 50 50 50 50 50

• Either by 1.25 percentage point for each quarter missing to reach the value of Nmax quarters;

• Or by 1.25 percentage point for each quarter missing to reach age sixty-five, the adjustment actually applied being the one that leads to the most favorable outcome for the pensioner (see table 4.1). For cohorts born before 1934, Nmax
Nmax
150. Access to the full rate is also possible before sixty-five for people having less than 150 quarters if they are considered as disabled or suffer from handicap. Values of Nmax and D are currently changing, while Nmax remains set at 150. As mentioned in the introduction, the value of Nmax should reach 160 quarters when the 1993 reform fully produces its effects (cohorts born from 1943). The same reform also scheduled a progressive increase of D, up to twenty-five years (to be reached for cohorts born from 1948). But for the cohorts considered here, the rules are the ones that prevailed between 1983 (when the possibility of retiring at age sixty was generalized) and 1993— that is, Nmax equals 150 (37.5 years) and D equals ten years. This system means that the number of years of contributions affects the pension level in two ways, which may imply, in some cases, a very strong dependency between the age at retirement and the level of SS benefits. To provide a full understanding of this interaction, table 4.2 shows the consequences of this system, with pre-1993 parameters, for three reference cases with individuals arriving at age sixty with, respectively, twenty-five, thirty, and thirty-five years of contribution.

• The first individual has to wait until age sixty-five to get retirement at

a full rate  (50 percent). Nevertheless, their pension is reduced by the fact that they only have 120 quarters of contribution at this age. Their replacement ratio is therefore only equal to four-fifths (120 quarters divided by 150) of the maximum replacement ratio, which is equal to 50 percent. Note that, at each age lower than sixty-five, the downward adjustment of  is here computed on the basis of the number of years

240

Ronan Mahieu and Didier Blanchet

Table 4.2

Replacement Rate Provided by the General Regime and the Civil Servants Regime for Three Reference Cases

Tenure (years)

 (General Retime) (%) (1)

 (Civil Servants) (%) (2)

No. of Years/ 37.5 (3)

Replacement Ratio (General Regime) (%) (1)  (3)

Replacement Ratio (Civil Servants) (%) (2)  (3)

60 61 62 63 64 65

25 26 27 28 29 30

25 30 35 40 45 50

75 75 75 75 75 75

Individual A 0.667 0.693 0.720 0.747 0.773 0.800

16.7 20.8 25.2 29.9 34.8 40.0

50.0 52.0 54.0 56.0 58.0 60.0

60 61 62 63 64 65

30 31 32 33 34 35

25 30 35 40 45 50

75 75 75 75 75 75

Individual B 0.800 0.827 0.853 0.880 0.907 0.933

20.0 24.8 29.9 35.2 40.8 46.7

60.0 62.0 64.0 66.0 68.0 70.0

60 61 62 63 64 65

35 36 37 38 39 40

37.5 42.5 47.5 50 50 50

75 75 75 75 75 75

Individual C 0.933 0.960 0.987 1.000 1.000 1.000

35.0 40.8 46.9 50.0 50.0 50.0

70.0 72.0 74.0 75.0 75.0 75.0

Age

needed to reach age sixty-five, rather than the number of quarters missing to reach a value of N equal to 150, since the rule consists in applying the most advantageous of the two adjustments. • The second individual also has to wait until age sixty-five to get the full rate , but benefits at this age are at a higher replacement rate, equal to fourteen-fifteenths (120 quarters divided by 150) of the maximum replacement ratio of 50 percent. In this case again the downward adjustment before age sixty-five is based on the number of years needed to reach this age of sixty-five. • The third person will not have to wait until age sixty-five. They benefit from the maximum replacement rate as soon as they reach a cumulated number of years of contributions equal to 150 (i.e. at sixty-twoand-one-half years). If they decide to leave between age sixty and this sixty-two-and-one-half years, the downward adjustment will then be computed according to the number of years missing to reach the total of 150 contributed quarters, rather than the number of years needed to reach age sixty-five, since the first rule is now the most generous. Note

Estimating Models of Retirement Behavior on French Data

241

also that, for this person, working past sixty-two-and-one-half years does not increase their SS entitlements. Some additional observations must be added to this presentation of the general scheme.

• Some people were successively affiliated to different schemes, especially in older cohorts (for instance, people transiting from agriculture or self-employment to the status of wage earner in the industry or in services). These people will cumulate two basic pensions: one from their initial scheme and one from the general scheme. The latter one will be proportional to the number of years spent in this scheme, according to equation (1), yet coefficient  will be evaluated taking into account the total number of years contributed, regardless of the scheme. Reductions of , furthermore, do not apply in a certain number of cases: veterans, disabled workers, and female workers with twenty-four contributed years who have raised three children. • Equation (1) also implies that pensions, at the time they are claimed, are computed in current euros. They are then reevaluated each year on a discretionary basis. During the 1970s and early 1980s, the general policy was to overindex these pensions (with respect to the average gross wage), in order to suppress the initial gap between standards of living of workers and pensioners. Since the mid-1980s, the practice has consisted rather in an indexation on prices. This practice has been confirmed by the 1993 reform. • When the average annual wage (D best years) falls below a floor (about €12,000 in 2000), it is raised to the level of that floor for individuals who can claim a full-rate pension. These provisions (the minimum contributif ) mainly concern women who had part-time jobs or whose careers were short and whose annual earnings are thus very low. They involve an additional strong incentive to postpone retirement until the full-rate threshold. • For women, Nmax and Nmax are increased by two years for each child they bred. Moreover, people (either men or women) who bred at least three children enjoy a 10 percent increase in their basic pension. 4.2.3 Wage Earners from the Private Sector: Complementary Schemes These schemes are almost fully contributory and are organized in a defined-contribution way (although they are not funded). Workers accumulate “points” during their careers, which are the pension’s basic unit of calculation.

• The points are accumulated during workers’ careers in proportion to their contributions: The contribution rate is fixed, and €1 contributed

242

Ronan Mahieu and Didier Blanchet

in year t is considered as equivalent to the formal buying of 1 / PPt points, where PPt is the purchase price of one point (the official term for this purchase price is salaire de référence). • The pension is then equal to the total number of points accumulated over the pensioner’s career, multiplied by a coefficient V (the official term being valeur du point), which is fixed every year. For a pensioner who started working at time t0 and stopped at time t1, the pension level at time t can therefore be written as t1 (t)w(t) Pension
V(t1)  ∑ , PP(t) t
t0

(2)

where (t) and w(t) are respectively the contribution rate and the worker’s wage at time t. As explained before, only a fraction of the wage is taken into account for computing contributions and points accumulated each year:

• For executives, contributions are collected by ARRCO for the part of the wage below the ceiling, and by AGIRC for the segment of the wage comprised between the SS ceiling and four times the ceiling; and • For nonexecutives, the wage is truncated to three times the SS ceiling and contributions are collected by ARRCO. Concerning retirement age in these complementary schemes, normal retirement theoretically remains at age sixty-five even after the 1983 reform, which introduced retirement at age sixty in the general scheme. For retirement below sixty-five, a quasi-actuarial adjustment is supposed to be applied. But since the 1983 reform, this adjustment is not applied to people who fulfill the conditions for a basic retirement at full rate (more than 37.5 years of contribution). 4.2.4 Civil Servants Civil servants have a unique pension scheme, directly financed by the state budget. As a general rule, claiming the pension is possible at age sixty if people have at least fifteen years of service. A rather large minority, however, can leave beginning at age fifty-five: primary-school teachers, policemen, prison officers, and the like. For women who have bred at least three children, the age condition is completely relaxed (but the fifteen-years-ofservice condition remains valid). The benefit formula is (3)




N of quarters, truncated to Nmax Pension
0.75   Nmax



 (last gross wage, excluding bonuses). The pension is a proportion of the last gross wage. Note that this gross wage excludes bonuses, which represent up to 50 percent of the total net in-

Estimating Models of Retirement Behavior on French Data

243

come for some specific categories (i.e., the ones with the highest incomes): These bonuses remain insignificant for most civil servants working for the Education Department, which is the largest employer. The key variable is the number of years a civil servant worked. Each year entitles him to a 2 percent annuity (table 4.2), the sum being truncated to 75 percent. Once this basic annuity is computed, some other periods may be taken into account: The most important provision is the additional year given to women for each child they bred. Each additional year also yields an additional 2 percent annuity that may increase the basic annuity up to 80 percent. Finally, people (either men or women) who bred at least three children enjoy a substantial increase in their pension. This increase is 10 percent if they have bred three children and 5 percent for every additional child. These provisions are roughly the same as in the private sector. Note that this system strongly differs from the general regime as regards incentives to retire early: Let us consider the example of people reaching the legal minimum age of retirement with only 32.5 years contributed and who decide to claim immediately for their benefits. The civil servant’s replacement rate is 65 percent (instead of 75 percent for a complete career). The private sector wage earner’s replacement rate (basic pension only) is 21.7 percent (instead of 50 percent for a complete career). 4.2.5 Other Regulations Concerning Age at Retirement: Mandatory Retirement and Eligibility to Early Retirement Mandatory retirement as such only exists for civil servants or within specific schemes. The age for mandatory retirement is generally sixty-five, with some exceptions either below that age (e.g., militaries, etc.) or above (very limited categories are allowed to work until age sixty-eight, such as academics). In the private sector, a firm is not allowed to layoff a worker according to any age criterion. Yet it is allowed to do so when this worker reaches the conditions to get a full-rate SS pension. Given the employment context of the 1990s—and the relatively large wage gap between elder and younger workers—it is quite likely that firms will quasi-systematically make use of this possibility. A consequence, which shall be recalled later when interpreting results, is that decisions to retire at the age where people get the full rate may be interpreted as demand-side as well as supply-side decisions. Supply- and demand-side aspects are also strongly intertwined for all forms of early retirement. Early retirement developed in France in several steps. We shall only describe rules set in after the 1983 reform, that is, after the generalization of possibilities to retire at age sixty. There are two main paths to early exit from the labor force:

• One is through unemployment insurance. People falling into unemployment are entitled to a compensation for a limited period of time,

244

Ronan Mahieu and Didier Blanchet

and the level of unemployment benefits, from 1992 to 2001, was decreasing with the duration of unemployment. But these rules do not apply to people losing their jobs past a certain age (fifty-seven until mid-1993, when it was raised to fifty-eight) who can benefit from a full compensation until they are able to benefit from a normal SS pension at a full rate. This system is not officially described as an early retirement system, and people cannot enter into it completely freely: They can do so only if they have been explicitly laid off by their employers. Yet this system is more or less equivalent to an early retirement scheme; • The second path for early exit is the Fonds National pour l’Emploi (FNE; National Fund for Employment). The level of early retirement benefits is roughly similar to the level of unemployment benefits. People benefiting from this system can leave the labor force around fifty-eight with benefits maintained until access to a full-rate pension in the general regime. The difference with the former path is that this system is under direct control by the state: Access to the FNE only concerns workers laid off in the context of a social plan negotiated between the firm and the state, with some compensations offered by the firm (for instance, a commitment to hire new young workers). 4.3 Data Description and Scope of the Present Study 4.3.1 Empirical Observations and Research on Labor Force Trends at Older Ages How do these institutional rules affect aggregate labor force participation at older ages? In 1998 (table 4.3), employment rates reached 75 percent for people aged fifty to fifty-four, but sharply decreased thereafter to 53 percent for the fifty-five to fifty-nine age group and only 12.4 percent (most of them being self-employed) for the sixty to sixty-four age group. Participation rates are close to zero after sixty-five. Very few self-employed retire before sixty, but exit rates are high from fifty-five for wage earners. Table 4.3

Labor Market Participation, by Age Group Employed

Age

Cohort

Public Sector

Private Sector

Self-Employed

Not Working

50–54 55–59 60–64 65–69

1944–48 1939–43 1934–38 1929–33

22.6 13.3 2.5 0.0

33.4 21.9 3.5 0.3

18.8 18.0 6.4 1.2

25.2 46.8 87.6 98.5

Source: INSEE, 1998 Financial Assets Survey.

Estimating Models of Retirement Behavior on French Data

245

As mentioned in the introduction, for men, this is the result of a large decrease in labor force participation after age fifty over the past twenty years. The share of men employed at ages fifty-five, sixty, and sixty-five decreased from 83.4 percent, 47.0 percent, and 14.7 percent, respectively, in 1983 to 78.5 percent, 32.1 percent, and 4.9 percent in 1998. Nonetheless, the regular decrease in male employment rates appears to have slowed down since 1997, due to the economic recovery. For women, current figures are the result of the combination between this tendency to earlier exits from the labor force and the impact of the long-run increase in overall labor force participation between successive cohorts. The decline at ages fifty-five to fifty-nine and sixty to sixty-four has been lower than for men: from respectively 29.1 percent and 9.4 percent in 1983 to 25.9 percent and 5.9 percent in 1998. And the trend remained positive in the age group fifty to fiftyfour: from 52.2 percent to 57.9 percent over the same period. About 8 percent of the population received public benefits (mainly unemployment benefits) between fifty and fifty-four in 1998 (table 4.4). This figures reaches 23.7 percent between fifty-five and fifty-nine, due to unemployment, early retirement (in the private sector), or SS benefits (for a strong minority of civil servants). Between sixty and sixty-four, 72.7 percent of the population receives public benefits (mainly SS benefits). Previous research on retirement behavior in France is relatively scarce, partly because economists lacked suitable data until appropriate administrative files were built. Moreover, individuals were so heavily constrained by SS incentives that explaining actual behaviors did not require a sophisticated approach (in econometric terms, for instance). In the first part of this project, Blanchet and Pelé (1999) showed that incentives to retire at the full rate were very strong, and Pelé and Ralle (1997), using a lifecycle model (based on an intertemporal budget constraint), demonstrated that retiring at the full rate was consistent with a rational utility-maximizing behavior. Of course, retirement cannot entirely be explained by SS incentives: Analyzing early retirement behaviors in France as a three-player game (the firm, the employee, and the government) may be of great interest, but once again, the lack of appropriate firm data did not allow for a comprehensive analysis of individual behaviors concerning early retirement. Table 4.4

Part of the Population Receiving Public Benefits, by Age Group

Age

Cohort

SS Benefits

ER Benefits

UI Benefits

Total

50–54 55–59 60–64 65–69

1944–48 1939–43 1934–38 1929–33

1.3 9.4 68.9 86.2

0.0 6.0 1.4 0.0

7.2 8.3 2.4 0.0

8.5 23.7 72.7 86.2

Source: INSEE, 1998 Financial Assets Survey. Notes: ER = early retirement; UI = unemployment insurance.

246

Ronan Mahieu and Didier Blanchet

4.3.2 The Data Set: The Echantillon Interrégime de Retraités Few systematic data sets exist in France concerning the economic situation of retired people. Income surveys only give instantaneous and imperfect pictures of transfer-income benefits to retirees: They do not allow the reconstitution of past labor income that would allow the evaluation of what these transfers would have been if pensioners had made other choices concerning their age at retirement. Some other specific surveys were also realized to analyze the transition between activity and retirement (e.g., a questionnaire on this topic was added to the periodic Labor Force Survey in 1996). These surveys are especially useful for analyzing the variety of institutional paths from full-time activity to retirement (Heller 1985; Caussat and Roth 1997; and Burricand and Roth 2000) and provide some interesting information on standards of living before retirement. However, these surveys do not provide precise information on past wages and thus do not allow the computation of financial incentives to retirement. This is the reason why another approach has been developed since 1984 that consists of matching administrative data collected from all pension schemes that exist in France. In practice, the only large-scale survey that is available and appropriate for the current study is a specific panel, the Échantillon Interrégime De Retraités (hereafter referred as EIR). The panel has been initially developed by the Service des Statistiques et des Systèmes d’Information (SESI),1 the statistical unit within the Ministry of Social Affairs, in connection with the INSEE. For the first run in 1988, four cohorts of pensioners were selected: those born in 1906, 1912, 1918, and 1922. A total sample of 20,000 people belonging to these four cohorts was drawn by INSEE. Their national identification numbers were transmitted by INSEE to all existing pension schemes (more than 120 basic schemes and about 180 complementary schemes). All these pension schemes then had to search for these individuals in their records. If they were present, the information about their pension entitlements was then transmitted to the SESI, who then carried out the matching, for all individuals of the sample, of the information returned by all existing pension schemes. The operation was renewed in 1993 and 1997. Each time, the same samples were redrawn for the cohorts included in the previous studies (and enlarged to compensate for mortality), and new cohorts added to the panel: cohort 1926 in 1993, cohorts 1930, 1932, 1934, 1936, 1938, 1940, and 1942 in 1997 (table 4.5). Since 1990, an additional matching has also been introduced with information from other administrative sources:

• The annual declarations of social data (DADS), made each year by firms, that allow retrieval of the wages of the sample participants over 1. See Dangerfield and Prangère (1996). Since 1998, the SESI has been integrated into a new department, the DREES, within the Ministry of Social Affairs.

Estimating Models of Retirement Behavior on French Data Table 4.5

247

The Structure of the Interregime Panel of Pensioners Pensions (if any) Observed In:

Cohort 1906 1912 1918 1922 1926 1930 1932 1934 1936 1938 1940 1942 a

1988 ✕ ✕ ✕ ✕

1993

1997

Wages and/or UI/ER Benefits Observed From: a

✕ ✕ ✕ ✕

✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕ ✕

Age 59 → retirement Age 55 → retirement Age 53 → retirement Age 51 → retirement Age 49 → retirement Age 47 → retirement Age 45 → retirement Age 43 → retirement

One year missing (1990).

the years before retirement if these people were wage earners in the private sector or in state-owned companies; • The wage files from the State Service for former civil servants; and • Files from the Union Nationale pour l’Emploi dans l’Industrie et le Commerce (UNEDIC), the French system of unemployment insurance, for people in unemployment before retirement, allowing therefore the incorporation of the form of early retirement benefits offered by the UNEDIC and the FNE (see previous discussion). This matching, however, does not allow a full reconstitution of past careers for these pensioners. The DADS, in particular, generally do not go back further than 1985, with one additional missing year in 1990. This matching, for this reason, has not been done for cohorts 1906, 1912, and 1918, for whom it would have been irrelevant. Table 4.5 sums up the structure of data available in the panel. Our question has been to explore how these data could be best used for the estimation of model of retirement behavior for France. The choices which have been made resulted from two constraints:

• The need, conversely, to have people for whom the situation before retirement has been observed over a significant period, in order to be able to extrapolate what their standard of living would have been in case they would have retired later than they actually did; and • The need to limit ourselves to cohorts for whom entry into retirement can be considered as fully completed. As detailed in the next subsection, our method for reconstructing individual pension entitlements under alternative retirement ages essentially relies on the pension level obtained at the actual retirement age. Of course, one possibility would have been, for people not yet retired, to evaluate entitlements on the

248

Ronan Mahieu and Didier Blanchet

basis of past working records. But the length of our wage records was too short for such a reconstitution, and, for this population, our files did not provide any proxy at all for the key variable, which is the number of quarters of past contributions to SS. The first constraint clearly ruled out cohorts 1906 to 1922. We also considered that wage data were too short on average for cohort 1926 (only two years of wages being observed for an individual of this cohort retiring in 1986). The second constraint, on the opposite, ruled out cohorts 1934 to 1942. Even if a significant share of these cohorts was retired in 1997, we would have missed the fraction retiring at sixty-five, which is precisely the fraction that brings the variance necessary to identify models. We considered that the same problem existed for workers from the private sector in cohort 1932. So that, for this category, we finally restricted ourselves to cohort 1930. For civil servants, however, we decided to use both cohorts 1930 and 1932, in order to increase somewhat the sample size, considering that the selection bias on cohort 1932 was lower than for the private sector and given an average age at retirement, which is lower in the public than in the private sector. Concerning the key question of the definition of retirement, our data allowed two possible choices: either the age when people definitely leave the labor market or the age when people claim SS benefits. But this latter definition is not the most interesting from an economic point of view, since a huge majority of people in the private sector claim SS benefits as soon as they reach the full-rate age. It is more interesting to analyze the impact of SS provisions (and, if possible, preretirement or unemployment provisions) on the decision to definitely leave the labor market. We therefore decided to model the last year of recorded past employment using DADS data. This, of course, implies a restriction to people who are in paid employment in 1985, which limits our sample a bit further. 4.3.3 Reconstructing Wages and Pension Levels What are the prerequisites concerning wage data in order to evaluate incentives to retire at different ages? A priori, wage data are needed for two things:

• Full wage histories are necessary to know how pension entitlements change with age at retirement, and

• A projection of wages is also necessary to evaluate earnings foregone in case of exit from the labor force. As stated before, our data do not go back earlier than the mid-eighties, so wage histories in our sample are strongly truncated. As mentioned, a reconstitution of wages for earlier time periods could have been attempted, but the specific rules concerning the computation of pensions imply that

Estimating Models of Retirement Behavior on French Data

249

such a retropolation did not appear to be necessary once we restricted ourselves to cohorts for whom at least one observation concerning the level of pensions was available. The strategy is the following: We know, for these people, their exact age at retirement r and the basic and complementary pensions obtained at this age Pb (r) and Pc (r). We have to compute what pension entitlements would have been in case of claiming pension at ages r  greater or lower than r. Concerning the basic pension, if we go back to equation (1) and if we consider that delaying or anticipating retirement would only have a marginal impact on the average wage of the ten best years,2 the impact of a change in r is only to change N, the total number of quarters of contribution (if not truncated to 150), and to change the coefficient , which for 60 r 65 or N 150 quarters, is reduced by 5 percent for each year of anticipation. The result of these changes on Pb (r) is quite easy to compute. Information on wages is here superfluous. Concerning complementary schemes, information on wages becomes necessary, but we do not need more than information on wages at later ages. Let us consider the case of an individual whose complementary pension only depends on ARRCO. From equation (2), we get that the variation of the expected pension level at t, if working until age r 1 instead of r would have been V(t)
V(t)( g r 1)w( g r 1)/PP( g r 1) for an individual born in g, plus the eventual application of the reduction coefficient for those people not fulfilling the condition for the maximum value of the annuity rate  in the general scheme. This computation, too, does not require any retrospective information concerning wages. For civil servants, the only necessary information is the last wage: There is no need of past wages. The only requirement concerning wages, therefore, is the extrapolation of notional wages for periods later than people’s actual retirement ages. This extrapolation was made for workers from the private sector using wage equations more fully described in the appendix. For civil servants, we limited ourselves to extrapolations of observed wages, indexed only on prices. 4.3.4 Other Data Computing the actual value of future pension benefits required some additional information regarding people’s own mortality risk, as well as the presence of a spouse and this spouse’s mortality risk, assuming that individual evaluations of benefits include the evaluation of survival benefits if the individual dies before their spouse. 2. This assumption is especially plausible given the truncation of wages. For people above the ceiling, the average wage of the ten best years will be generally equal to this ceiling, and one more year of work will generally not change this. This does not hold for people below the ceiling, but these people’s careers are generally flatter, so that the same approximation may remain valid.

250 Table 4.6

Ronan Mahieu and Didier Blanchet Descriptive Statistics on the Sample Variable Sex (Male = 0, Female = 1) Age Married Widowed Single Wage (€) Private Executive Technician Employee Skilled blue-collar Unskilled blue-collar Public Category A Category B Category C Total tenure (years)

Mean Value 0.439 57.4 0.753 0.113 0.125 20,095 0.126 0.154 0.215 0.189 0.105 0.119 0.055 0.037 36.4

Note: Sample size = 9,884 observations corresponding to 2.202 individual paths.

Mortality rates for people in the sample used are differentiated by sex, age, and socioprofessional group. One point must be noted here: Since the sample is conditioned on surviving until the age of sixty-four or sixty-six (depending on the cohort), a selection bias may result if there is a correlation between mortality and the retirement decision. If people with bad health status and a higher mortality risk tend to more frequently anticipate the claiming of their benefits, there will be a tendency to overestimate the actual age at retirement. Concerning information on spouses, unfortunately, the EIR did not produce any reliable information, even limited to the indication of the presence of a spouse. The reason is that information on marital status, in these files, is only updated when it becomes necessary, that is, generally at the pensioner’s death (if survival benefits are to be paid). We restricted ourselves to a model of retirement choice where only personal entitlements are taken into account, rather than attempting a reconstitution of variables relating to spouse’s presence, age, and status. The final sample consists in 9,884 observations (table 4.6) corresponding to 2,202 individuals still employed at fifty-five (who are thus observed on average between four and five years before they retire), 75% of which are employed in the private sector (with a majority of men). Note that the average tenure at fifty-five is pretty high (over thirty-six years) and close to the tenure required to reach the full rate at sixty: This reflects the fact that most people from the sample are entitled to full SS benefits as soon as sixty (especially men, see figure 4.1).

Estimating Models of Retirement Behavior on French Data

Fig. 4.1

251

Distribution of tenure at age 55

Table 4.7

Pathways to Retirement in the Sample Retiree Category Private Sector

Civil Servants

Pathway

Men

Women

Total

Total

Directly to SS ER then SS UI then SS

57.4 20.7 21.9

60.8 18.4 20.8

58.7 19.8 21.5

100.0 0.0 0.0

Source: Authors’ calculations from EIR, cohort 1930; people still working at 55 (data source).

Analyzing pathways to retirement is straightforward for civil servants (they have no other choice than waiting until the minimum age to claim SS benefits, unless they chose to consume their savings). In the private sector (table 4.7), about 60 percent of people still working at fifty-five do not receive public benefits other than SS benefits. The remaining 40 percent are roughly equally divided between people retiring through unemployment and early retirement schemes. Table 4.8 provides information on the level of the parameter . A very tiny minority of men (0.3 percent) claim SS benefits at reduced rate, whereas the figure grows to 4.4 percent for women. About 4 percent of men and women are considered as disabled are thus allowed to claim full-rate SS benefits at sixty (even if their tenure is below 150 quarters). Also, 3.7 percent of men and 10.7 percent of women are “unfit” to do a job and thus benefit a full-rate pension at sixty. Others (over 80 percent of the sample)

252 Table 4.8

Ronan Mahieu and Didier Blanchet Level of the Pension Rate (
) when People Claim SS Benefits (private sector; %)

Full rate Normal conditions “Unfit” for a job Disabled Reduced rate

Men

Women

92.0 3.7 4.0 0.3

81.0 10.7 3.9 4.4

reach the full rate in normal conditions. In the public sector, there is no such incentive to postpone claiming SS benefits after the minimum age (mostly sixty) since  is set to 75 percent whatever the total tenure. Nonetheless, it is worthwhile to note that the retirement rate for the civil servants that reach age sixty with 150 quarters or more is 69 percent, whereas it drops to 53 percent for those who reach age sixty with less than 150 quarters. Moreover, the mean wage of civil servants who keep on working after sixty is €32,000 (instead of €23,400 for those who quit at sixty). Remember that highly paid civil servants have, on average, lower replacement rates since a large part of their wage consists of bonuses. At first glance, civil servants also seem sensitive to SS incentives (despite their weakness), but these preliminary observations must be confirmed by a deeper analysis. 4.4 A Descriptive Analysis of Incentives to Retire 4.4.1 Definition of Incentive Variables Two kinds of models will be applied to the analysis of labor force participation rates of older workers. In a first step, we shall introduce simple measures of SS incentives to retire in probit models to describe the choice to retire at age t for individuals still in the labor force at this age. For an individual aged t, we first compute SS wealth at age t. The value of this social security wealth (SSW) will depend on the age t greater than or equal to t at which this individual will decide to retire. Also, Bs (t) is the expected level of pension at age s for an individual who retired at age t; if (s/t) is the probability of surviving up to age s for an individual ages t, and if T, at last, is the maximal age at death, we write:




T s SSWt,t
∑ st  Bs (t) t s
t

From this value, we derive the pension accrual at age t that is the algebraic increase in SSW, which results, at age t, from the postponement of retirement by one year, that is,

Estimating Models of Retirement Behavior on French Data

253

Accrual t
SSWt,t 1  SSWt,t . The accrual will be our first measure of SS incentives. The tax rate is directly derived from the accrual: It captures the fact that a negative accrual involves an implicit tax on continued work, because a part of the expected wage (if the agent postpones retirement) is taxed through the decrease in the SSW. The tax rate thus writes accrual t Tax ratet
. Etwt 1 An alternative measure is also directly derived from the definition of SSW. This variable is the peak index, which is the difference between the maximum of the SSWs associated to all possible ages at retirement beyond the current year, and SSW in case of an immediate retirement. Peakt
max[SSWt,s ]  SSWt,t st 1

It assumes a less myopic behavior by the individual, who considers not only the potential gain in SSW resulting from delaying retirement by one year, but also gains that may be derived from retiring in any subsequent year. However, as with all measures derived from SSW, a limitation of this index is that it does not take into account the comparison that the individual can make between pension benefits and the level of his labor income. It assumes that the retirement decision is only affected by variations of pension entitlements. This limitation will be corrected in the following estimation by the introduction of wages as covariates in probit models, but it is more satisfactory to introduce incentive measures which introduce this comparison between benefit and wage levels in a less ad hoc way. This is the case if we start from a model which fully includes expected flows of utility derived either from labor or retirement income. The model used is the Stock and Wise (1990) option value model. Let us again consider an individual still in the labor force at age t. If they expects to retire at age r, they can expect a flow of labor incomes of (Yt , . . . Yr–1 ) until retirement and then a flow of pension benefits (Br(r), Br 1(r), . . . , Bs (r), . . .). It is assumed that this individual derives an indirect utility Uw from his labor income and an indirect utility Ur from pension benefits. Time discounting occurs at rate . For an age at retirement equal to r, the expected utility at age t is therefore r1

T

s
t

s
r

Vt (r)
∑ stEt [Uw (Ys )] ∑ stEt{Ur [Bs (r)]}, with Uw (Ys )
Y s , Uw(Bs )
[kBs ] .

254

Ronan Mahieu and Didier Blanchet

Note that this specification does not consider the possibility of smoothing income flows through private savings, an assumption that will be essentially valid for low- or medium-income workers. Given this definition of utility, we assume that the individual decides to retire if the resulting expected utility is higher than the maximum value of utilities expected for all other possible choices r  t. If we write Gr(r)
Vt (r)  Vt (t), the individual chooses to remain in the labor force if Gt(r ∗) greater than 0 where r ∗
Arg max Vt (r). rt 1

The equation Gt (r∗) greater than 0 is called the option value of postponing retirement in order to express that, given the irreversibility of retirement, remaining in the labor force offers the option to leave the labor force at a later age under better conditions. Stock and Wise (1990) performed a full maximum likelihood estimation of the model on U.S. data that yielded equals 0.97,  equals 1.25, and  equals 0.6. Our own estimation of the model on French data led us to adopt the following parameterization: equals 0.97,  equals 1.6, and  0.25. These values imply some risk aversion and a moderate preference for leisure: In the context of a oneperiod model, a value of  equal to 1.6 means that an individual would demand a leisure income equal to 62.5 percent of his labor income to accept not to work. 4.4.2 Including Incentives Linked to Unemployment Benefits and Early Retirement We next present evaluations of incentives that (imperfectly) take into account the additional incentives imbedded in unemployment insurance (UI) and early retirement schemes. Assume, as a first step, that an individual is actually free to choose one of these means of early exit from the labor force. We can therefore compute three values for the SSW: the one computed above on the basis of normal pension entitlements only and the values if we assume that the individual begins by spending a few years in unemployment or in the early retirement scheme and then moves on to normal retirement once he is entitled to the full-rate SS. For instance, for an individual aged fifty-five, we compute the following values, depending on age t at which they will leave the labor force: SSW155,t
ΣsT
60 s–t (s/t)B Pension (t) is the individual only relies on his s normal pension; s–t s–t UI T Pension SSW255,t
Σ59 (t) for a transition s
t (s/t)Bs (t) Σs
60 (s/t)Bs through UI;

Estimating Models of Retirement Behavior on French Data

255

s–t s–t ER T Pension SSW355,t
Σ59 (t) for a transition s
t (s/t)Bs (t) Σs
60 (s/t)Bs through early retirement.

Benefits B une and B pre s s are computed as a fraction of the last wage by direct application of official rules. We then compute a weighted average of these three SSWs. Weights are a function of the sector of activity and reflect take-up probabilities. We tested other covariates, like sex and professional status (executives versus blue-collars versus white collars), but their coefficients are mostly insignificant if sector dummies are included. This strategy is consistent with the results of previous studies that show that the sector of activity predicts access to early retirement or unemployment schemes far better than qualification or social group (Colin Iéhlé, and Mahieu 2000). As a general rule, the probability of facing a period of unemployment or early retirement at the end of one’s career is markedly higher (at least for cohorts born around 1930) in industries. In particular, it is the automobile industry that concentrates the highest risks: At fifty-five, there was a 60 percent probability of entering into unemployment or early retirement for a wage earner in the automobile industry. The reason is that, around 1985, some sectors—including the automobile industry—benefited from exceptions allowing a lower age at entry in the Allocation Spéciale du Fonds National pour l’Emploi (ASFNE; fifty-five years instead of fifty-six years and two months). We finally compute incentives (accrual, peak value, and option value) with this weighted SSW. 4.4.3 Incentive Analysis Private Sector Individuals are followed between ages fifty-five and sixty-five. We ask what has been the structure of incentives to retire in this sample and how have these incentives determined actual retirement decisions? Given the discrepancy between SS incentives in the private sector and for civil servants, it seems appropriate to give separate results for each sector. In the private sector, the median SSW for males regularly increases with age after fifty-five to reach €200,000 at sixty (table 4.9). This is due partly to the increase in pension entitlements while tenure grows, but also to a selection bias: People who quit before the early retirement age of sixty (who mostly receive unemployment or early retirement benefits) get, on average, lower wages than those who stay on the labor market. Since pension entitlements are strongly correlated with labor income, this partly explains the age profile of median SSW. The median accrual is positive (though diminishing) until fifty-nine, with a relatively large dispersion: Those who already have a tenure above 37.5 years have very low accruals (see the tenth

256 Table 4.9

Ronan Mahieu and Didier Blanchet Retirement Incentives for Men in the Private Sector (€ ) Accrual Tax Rate

Age 55 56 57 58 59 60 61 62 63 64 65

N

SSW Median

Median

1,077 862 712 622 539 501 85 84 60 48 38

152,476 161,350 170,499 180,368 189,460 199,327 247,305 244,555 220,831 196,433 163,729

8,438 4,945 3,621 1,745 2,937 –11,734 –11,335 –11,634 –11,956 –10,654 –10,810

10th 90th Percentile Percentile 2,751 1,631 1,047 –529 647 –19,845 –27,410 –29,285 –30,532 –31,313 –40,138

22,324 19,877 18,243 12,812 14,955 –4,198 4,141 2,996 2,464 2,998 –4,610

SD

Median

Previous

19,714 18,634 8,031 7,227 8,419 9,122 15,528 14,760 13,525 12,391 11,194

–0.54 –0.31 –0.21 –0.09 –0.15 0.73 0.35 0.45 0.46 0.50 0.75

–0.91 –0.97 –0.46 0.04 0.05 0.67 0.60 0.63 0.56 0.56 0.52

Peak Value

Median 55 56 57 58 59 60 61 62 63 64 65

1,077 862 712 622 539 501 85 84 60 48 38

12,792 6,962 4,819 3,339 2,958 –11,734 –11,335 –11,634 –11,956 –10,654 –10,810

10th 90th Percentile Percentile 4,120 1,680 1,057 –529 647 –19,845 –27,410 –29,285 –30,532 –31,313 –40,138

53,414 43,187 36,087 24,602 16,144 –4,198 4,937 4,143 2,858 2,998 –4,610

Option Value

SD

Median

45,792 48,979 18,072 16,306 18,491 18,010 30,251 24,864 15,841 13,056 11,194

46.46 36.52 28.73 20.16 11.56 –0.57 1.50 0.51 0.20 0.00 –0.68

10th 90th Percentile Percentile 38.94 30.00 24.27 16.91 9.84 –1.86 –2.02 –2.05 –2.21 –2.46 –3.81

64.70 54.61 44.25 32.63 20.76 5.56 20.18 14.67 11.78 8.00 1.91

SD 15.66 14.33 9.98 9.12 9.96 8.87 12.32 9.83 7.24 4.71 2.50

Note: N = number of observations; SD = standard deviation.

percentile), while those whose tenure is below 37.5 years have very high accruals (see the ninetieth percentile). Median tax rates between fifty-five and fifty-seven exhibit quite large subsidies to work (from 20 percent to 50 percent) that nonetheless remain low compared to the base case computed by Blanchet and Pelé (1999; last column)—about 90 percent at ages fifty-five to fifty-six. This discrepancy results from the characteristics of their base case: a man who works continuously from age twenty. As a result, the base case exhibits very large subsidies until he reaches the full rate (at fiftyeight). After that age, the very low increase in pension entitlements (essentially through complementary schemes) cannot even compensate the loss of payroll taxes, which explains the slight tax on continued work at fiftyeight to fifty-nine (4–5 percent) in this previous study. After age sixty, the median SSW suddenly increases by about 25 percent,

Estimating Models of Retirement Behavior on French Data

257

which reflects the fact that people who keep on working after the early retirement age are often better educated and thus better paid than the average. Of course, the median SSW declines from age sixty-one since median pension entitlements grow very slowly (most of these people could claim a full-rate SS benefit) while these people give up one year of benefits. The median accrual thus remains strongly negative (about –11,000) while a minority of men have positive accruals (see the ninetieth percentile) because they have a total tenure below 37.5 years. The median tax rate is lower at ages sixty-one to sixty-four (below 50 percent than at sixty (73 percent) due to this selection bias: People still on the labor market after sixty are either better paid than the average (and the denominator of the tax rate is higher) or they cannot get a full-rate SS pension (and their tax rate is negative). But tax rates reach 75 percent at age sixty-five (and accruals are systematically negative) since everybody may claim a full-rate SS pension from that age. If we turn to peak value (PV) measures (table 4.9), the results are very close to those obtained with accrual measures: After age sixty, accrual and PV are negative (increases in pension entitlements cannot compensate the loss of one year or more of benefits since most men already have the full rate), and thus accrual and PV are mostly the same. Concerning option value (OV) measures, we now have further explanations for the behavior of men who keep on working after age sixty, although they are entitled to a full-rate SS pension. We mentioned that these people were quite well paid, which was suggested by tax rate measures but not by accrual measures (the median value of accrual was strongly negative). Here, the median OV value is positive at ages sixty-one to sixty-four, whereas it was negative at sixty. The OV measures include data on expected wages. Given the structural form of the utility used to build OV measures, the OV is decreasing with the replacement rate. Since the replacement rate is lower for highly paid employees, they face an incentive to postpone retirement, and their OV is thus positive. The distribution of OV is thus consistent with behaviors observed in the data. Women in the private sector basically face similar incentives (table 4.10). The only serious difference is that a larger proportion of them cannot claim a full-rate SS pension at sixty since female careers are shorter than male careers (although this fact is somewhat weakened by our sample selection: Women who were still working at fifty-five had a larger tenure than the average). As a result, after age sixty, the median accrual is positive and the median tax rate exhibits a strong subsidy on continued work (about 30–40 percent). Note that the ninetieth percentile for accrual is particularly high at sixty-four: This results from the minimum pension provisions for lowwage earners (minimum contributif ) that are available at sixty-five for people with short careers. The PV, in general, is larger than the accrual for these women with short careers: This accounts for the fact that they face a strong incentive to postpone retirement by several years.

258 Table 4.10

Ronan Mahieu and Didier Blanchet Retirement Incentives for Women in the Private Sector (€) Accrual Tax Rate

Age 55 56 57 58 59 60 61 62 63 64 65

N

SSW Median

Median

679 563 474 404 343 325 94 94 80 65 52

98,405 108,031 114,757 121,668 130,117 139,462 108,838 118,268 112,247 112,234 119,365

5,042 3,818 3,271 1,987 2,873 –4,734 3,686 3,967 2,901 5,273 –5,188

10th 90th Percentile Percentile 2,469 1,955 1,305 –147 771 –12,774 –10,766 –11,117 –10,315 –9,625 –12,036

16,636 15,536 14,891 11,636 11,604 7,965 10,444 9,157 7,265 29,056 –1,895

SD

Median

Previous

6,055 6,001 5,379 5,012 5,098 8,094 8,048 7,818 7,813 15,150 5,398

–0.44 –0.41 –0.32 –0.21 –0.26 0.45 –0.44 –0.36 –0.31 –0.47 0.41

–0.91 –0.97 –0.46 0.04 0.05 0.67 0.60 0.63 0.56 0.56 0.52

Peak Value

Median 55 56 57 58 59 60 61 62 63 64 65

679 563 474 404 343 325 94 94 80 65 52

20,501 14,649 10,400 6,891 4,139 –4,734 13,830 13,228 8,471 5,273 –5,188

10th 90th Percentile Percentile 4,579 2,539 1,379 –127 771 –12,774 –10,766 –11,117 –10,315 –9,625 –12,036

53,505 47,250 42,504 38,170 37,370 35,149 33,704 30,485 28,572 29,056 –1,895

Option Value

SD

Median

20,424 18,120 17,195 16,303 15,620 18,851 18,392 16,384 15,579 15,151 5,398

55.18 45.12 34.91 25.14 14.67 1.60 12.55 10.20 7.19 4.03 0.51

10th 90th Percentile Percentile 40.55 31.06 24.74 17.60 9.93 –1.20 0.52 0.26 –0.39 –0.72 –0.60

76.58 65.88 57.52 48.43 39.47 28.71 32.39 29.55 30.24 25.92 1.92

SD 15.22 14.00 14.01 13.66 13.60 13.97 13.72 12.09 11.49 9.85 1.13

Note: N = number of observations; SD = standard deviation.

Civil Servants If we turn to male civil servants (table 4.11), the median SSW is much higher than in the private sector since it reaches €375,000 at sixty (instead of €200,000). This discrepancy is first due to higher wages on average (civil servants are often better educated, and a large proportion of them are teachers). But the higher life expectancy at sixty for civil servants also plays a role in this gap. The male civil-servants sample may be divided between those are entitled to a pension from fifty-five and those who are entitled from sixty. In the first subsample, the accrual is always negative from age fifty-five: the increase in the pension for each additional year is too weak to compensate the loss of one year of pension. The tenth percentile of accrual is thus negative from fifty-five. But, since a majority of civil servants cannot claim SS benefits before sixty, the median accrual remains positive

Estimating Models of Retirement Behavior on French Data Table 4.11

259

Retirement Incentives for Male Civil Servants (€) Accrual

Age 53 54 55 56 57 58 59 60 61 62 63 64 65

N

SSW Median

Median

88 89 174 157 132 128 126 123 46 34 22 18 13

266,692 280,429 296,910 307,998 332,907 350,753 360,674 375,387 409,197 409,402 416,091 400,195 462,172

5,694 2,845 4,506 2,728 4,793 3,142 3,510 –19,292 –22,290 –21,885 –28,092 –25,848 –33,517

10th 90th Percentile Percentile 3,659 1,774 –14,276 –17,613 –2,957 –5,134 –3,050 –29,377 –31,295 –32,314 –36,585 –38,011 –44,240

7,444 4,215 7,622 4,719 7,504 5,307 9,161 –12,145 –11,669 –9,008 –11,335 –10,560 –13,348

SD

Tax Rate Median

1,776 1,468 9,241 8,990 5,975 5,903 5,863 7,697 8,346 8,892 10,489 9,755 12,272

–0.27 –0.13 –0.21 –0.12 –0.19 –0.13 –0.14 0.66 0.59 0.50 0.67 0.56 0.61

Peak Value

Median 53 54 55 56 57 58 59 60 61 62 63 64 65

88 89 174 157 132 128 126 123 46 34 22 18 13

14,754 10,487 10,817 9,344 10,143 6,280 3,510 –19,292 –22,290 –21,885 –28,092 –25,848 –33,517

10th 90th Percentile Percentile 6,966 2,346 –14,276 –17,613 –2,957 –4,847 –3,050 –29,377 –31,295 –32,314 –36,585 –38,011 –44,240

40,109 33,463 29,933 24,291 20,620 14,002 9,161 –12,145 –11,669 –9,008 –11,335 –10,560 –13,348

Option Value

SD

Median

13,121 12,525 17,298 15,547 10,504 8,820 5,863 7,697 8,346 8,892 10,489 9,755 12,272

76.37 65.81 56.33 46.84 36.56 24.80 12.86 –0.21 0.02 0.10 –0.90 –0.39 –14.76

10th 90th Percentile Percentile 23.62 11.90 –0.60 –0.96 31.19 21.59 11.13 –1.66 –1.26 –1.04 –2.24 –1.32 –15.82

87.67 76.52 64.56 53.52 41.15 28.39 15.70 2.23 3.10 1.65 0.79 0.56 –11.70

SD 28.24 28.87 28.22 20.99 9.59 6.04 3.15 1.93 2.23 2.43 2.31 1.27 2.01

Note: N = number of observations; SD = standard deviation.

until fifty-nine. Conversely, from sixty, the accrual is negative for all. Median tax rates exhibit a slight subsidy to continued work until fifty-nine (10–20 percent) and then become clearly positive (above 50 percent). Nonetheless, the very low decrease in the median tax rate after age sixty reflects the fact that those who postpone retirement after sixty often have higher wages and lower replacement rates (since a larger proportion of their labor income consists in bonuses that give no additional rights for pensions). This is confirmed by the slight increase in the OV measure between ages sixty and sixty-one (table 4.11), But this result is less robust than in the private sector. The OV measure becomes extremely negative at

260 Table 4.12

Ronan Mahieu and Didier Blanchet Retirement Incentives for Female Civil Servants (€) Accrual

Age 53 54 55 56 57 58 59 60 61 62 63 64 65

N

SSW Median

Median

122 123 223 188 161 149 140 134 45 33 19 12 6

280,498 294,718 301,786 300,239 278,203 277,062 278,215 286,881 353,164 339,196 328,787 258,821 256,774

5,946 2,816 2,891 1,937 3,704 2,623 4,049 –12,655 –13,358 –13,079 –16,543 –15,383 –16,423

10th 90th Percentile Percentile –7,517 –10,512 –16,901 –19,122 –14,030 –16,519 –13,116 –20,744 –23,275 –21,312 –29,611 –26,945 –31,058

7,415 4,323 6,879 4,448 6,537 5,003 9,377 –2,619 –3,503 –2,821 –2,989 –2,548 –3,554

SD

Tax Rate Median

5,896 6,450 9,632 10,122 8,692 8,637 8,159 6,962 8,107 8,183 8,064 9,323 9,961

–0.33 –0.16 –0.16 –0.13 –0.23 –0.16 –0.24 0.61 0.49 0.46 0.68 0.63 0.68

Peak Value

Median 53 54 55 56 57 58 59 60 61 62 63 64 65

122 123 223 188 161 149 140 134 45 33 19 12 6

11,507 5,415 2,891 4,854 9,319 6,135 4,049 –12,655 –13,358 –13,079 –16,543 –15,383 –16,423

10th 90th Percentile Percentile –7,517 –10,512 –16,901 –19,122 –14,030 –16,519 –13,116 –20,744 –23,275 –21,312 –29,611 –26,945 –31,058

40,915 33,879 29,366 23,567 19,866 14,084 9,377 –2,619 –3,503 –2,821 –2,989 –2,548 –3,554

Option Value

SD

Median

17,456 16,875 17,977 16,920 13,335 11,496 8,159 6,962 8,107 8,183 8,064 9,323 9,961

69.40 60.01 49.77 41.90 33.22 23.03 12.20 0.11 0.44 0.46 –0.37 –0.50 –12.36

10th 90th Percentile Percentile 0.56 –0.22 –1.03 –1.25 –0.39 –0.44 0.62 –1.84 –1.02 –0.83 –1.29 –1.08 –14.54

86.71 75.23 62.91 51.52 39.80 28.21 16.89 5.67 6.04 5.09 4.35 2.24 –8.84

SD 33.17 31.06 28.96 22.95 16.04 10.53 5.56 3.14 3.20 2.42 1.85 1.32 2.02

Note: N = number of observations; SD = standard deviation.

sixty-five because of mandatory retirement: If people postpone claiming their pension, they do not enjoy any labor income flows. The results are similar for female civil servants (table 4.12). 4.5 Econometric Analysis We now analyze the decision to retire with probit models including incentive variables among regressors. Control variables are age, tenure, socioprofessional group, sector dummies, expected earnings and its square, and linear age or age dummies. Estimations are performed separately for

Estimating Models of Retirement Behavior on French Data Table 4.13

261

Probit Models (men, no full-rate dummy)

SSW (10,000) Standard deviation Implied probability Incentive variable (10,000) Standard deviation Implied probability Projected earnings (1,000) Square of projected earnings Age (linear) 55 56 57 58 59 60 61 62 63 64 65 Pseudo R 2

Accrual Model (linear age)

Accrual Model (age dummies)

PV Model (linear age)

PV Model (age dummies)

OV Model (linear age)

OV Model (age dummies)

–0.017** 0.003 0.001 –0.498** 0.031 0.001 0.115 0.022** 0.009

–0.006 0.003 0.075 –0.212** 0.035 0.001 –0.117 0.014**

–0.008** 0.003 0.013 –0.240** 0.020 0.001 –0.114 0.037** 0.061**

–0.002 0.003 0.465 –0.088** 0.020 0.001 –0.228 0.019**

–0.015** 0.003 0.001 –0.041** 0.003 0.001 0.892** –0.124 –0.120**

–0.013** 0.004 0.001 –0.035** 0.003 0.001 0.766** –0.112

0.159**

REF 0.159** –0.046 –0.080 –0.537** 1.214** –0.023 0.256** 0.011 0.143 1.293** 0.240**

0.132**

REF 0.173** –0.043 –0.060 –0.554** 1.387** 0.167 0.450** 0.228 0.352 1.524 0.237**

0.148**

REF –0.052 –0.473** –0.755** –1.527** 0.133 –1.107** –0.860** –1.112** –1.039** 0.070 0.256**

**Coefficients are significant at the 5 percent level.

men and women. In the following tables, SSW and incentive variables (IV) amounts are given in € ten thousands and annual wages in € thousands. We did not include lifetime earnings as suggested in the template since we lacked appropriate data. Models run on men or women have a quite satisfactory predictive power (tables 4.13 and 4.14): Pseudo R 2 values range from 12.5 percent to 16.2 percent with linear-age specifications, and from 23.7 percent to 26.6 percent with the age-dummies specifications. The SSW variable is always significant with linear-age specifications. In models with the age-dummy specifications, it is significant only with the OV as IV. The coefficient is always negative: The larger the SSW, the more the individual postpones retirement. This result may be surprising since an increase in SSW may be seen as a wealth effect and, thus, as an incentive to increase the consumption of leisure, which requires retiring earlier. Another possible explanation is that highly paid people have more interesting jobs than blue-collar workers and thus quit later. The IV is always strongly significant with the expected negative sign. Note that models with OV measures provide more robust coefficients for the IV, which is consistent with our expectations: This variable contains a richer set

262 Table 4.14

Ronan Mahieu and Didier Blanchet Probit models (women, no full-rate dummy)

SSW (10,000) Standard deviation Implied probability Incentive variable (10,000) Standard deviation Implied probability Projected earnings (1,000) Square of projected earnings Age (linear) 55 56 57 58 59 60 61 62 63 64 65 Pseudo R 2

Accrual Model (linear age)

Accrual Model (age dummies)

PV Model (linear age)

PV Model (age dummies)

OV Model (linear age)

OV Model (age dummies)

–0.013** 0.006 0.034 –0.530** 0.038 0.001 0.647 –2.075

–0.005 0.006 0.374 –0.336** 0.041 0.001 –0.123 –1.174 0.090** REF –0.047 –0.025 –0.098 –0.677** 1.242** –0.068 0.345** 0.181 0.503** 2.207** 0.266**

–0.010** 0.006 0.028 –0.212** 0.021 0.001 0.770 –1.047

–0.007 0.007 0.265 –0.136** 0.022 0.001 0.019 –0.567 0.094** REF –0.047 –0.048 –0.111 –0.743** 1.318** –0.017 0.368** 0.216 0.352** 2.241** 0.258**

–0.019** 0.007 0.004 –0.024** 0.002 0.001 1.086 –0.432

–0.019** 0.007 0.008 –0.020** 0.003 0.001 1.227 –0.814 0.006 REF –0.140 –0.256** –0.445** –1.224** 0.758** –0.618** –0.274 –0.439** –0.330 1.561** 0.263**

0.162**

0.130**

0.125**

**Coefficients are significant at the 5 percent level.

of information than accrual or PV measures. The level of the coefficients does not crucially depend on the age specification (–0.041 with linear age versus –0.035 with age dummies for men [table 4.13] and –0.024 with linear age versus –0.020 with age dummies for women [table 4.14]). With accrual or PV measures, the coefficient of the incentive variable is divided by 2 or 3 if we turn from linear-age specifications to the age-dummy specifications. Moreover, the coefficient of SSW is more robust in models with OV measures (–0.015 with linear age versus –0.013 with age dummies), whereas it becomes insignificant in models with age dummies and accrual or PV measures. Projected earnings are insignificant in models run on women samples. They have a significant positive impact on retirement for men. This result may be surprising at first glance since people with high earnings have, on average, lower replacement rates (which increases the price of leisure). Two possible explanations can be mentioned: First, highly paid people may have saved a lot in the past and thus may quit earlier since their retirement income has a significant nonpension component. Second, this may reflect demand-side effects on the labor market. Among blue-collars, those with the highest wages (and who may be paid far above their marginal produc-

Estimating Models of Retirement Behavior on French Data

263

tivity) might be more likely to be laid off by firms (through unemployment or early retirement schemes) than those with the lowest wages. The estimated effect of age variables is uncertain. In linear-age specifications, we obtain significant positive coefficients in models run with accrual or PV measures (the only exception being the accrual model for women where the coefficient is insignificant). This result was expected as the disutility of work is assumed to increase with age. In models run with OV measures, the coefficient is insignificant for women and significantly negative for men, which may be puzzling. A reason for this may be found in unobserved individual heterogeneity on the disutility of work. Consider two populations that only differ in their disutility of work. Those who have a large disutility of work quit early, say at sixty. At sixty-one, the OV measure of the remaining population is lower than the OV measure of the whole population at sixty (there remain fewer years of potential continued work), but is underestimated since the computation does not account on the endogenous selection on the disutility of work. Nonetheless, the observed retirement rate will be lower at sixty-one than at sixty since the considered population has a low disutility of work. If the economist does not observe the disutility of work, the coefficients of linear-age variables will be significantly negative. The models run with age dummies exhibit an irregular profile with strong spikes at sixty and sixty-five. Three reasons may be mentioned to explain these spikes: First, our variables do not perfectly capture incentives associated with SS rules. Second, these spikes may reflect demand-side effects: our incentives sum up supply-side effects, but employers are allowed to lay off workers as soon as they reach the full rate, mostly at sixty or sixty-five. Third, people may be induced to retire at sixty or sixty-five by a sort of social habit: sixty has been the normal retirement age since 1983, and sixty-five was the normal retirement age until 1983. We also computed models with an early retirement, for civil servants, or a full-rate, for the private sector (early retirement/full-rate) dummy (tables 4.15 and 4.16): for civil servants this dummy is set to 1 if the agents’ age is the minimum legal age to claim SS benefits (fifty-five or sixty, depending on his occupation) and to 0 otherwise. In the private sector, since claiming SS benefits is strongly discouraged below the full rate (virtually nobody claims SS benefits at a reduced rate, table 4.8), the dummy is set to one if the agent reaches the first year he can claim full-rate SS benefits. This dummy is strongly significant with linear age specifications and increases the pseudo R 2 that now reaches from 14.0 percent to 19.0 percent. But with age dummies specifications, it does not really improve the pseudo R 2 and is not significant in models run on males. To sum up, these estimations provide a satisfactory description of retirement behavior: The coefficients on incentive variables are always significant with the expected sign. The impact of SSW on the retirement decision

264 Table 4.15

Ronan Mahieu and Didier Blanchet Probit Models (men, full-rate dummy)

SSW (10,000) Standard deviation Implied probability Incentive variable (10,000) Standard deviation Implied probability Early retirement/full rate dummy Standard deviation Implied probability Projected earnings (1,000) Square of projected earnings Age (linear) 55 56 57 58 59 60 61 62 63 64 65 Pseudo R 2

Accrual Model (linear age)

Accrual Model (age dummies)

PV Model (linear age)

PV Model (age dummies)

OV Model (linear age)

OV Model (age dummies)

–0.016** 0.003 0.001 –0.463** 0.031 0.001

–0.006 0.003 0.077 –0.208** 0.035 0.001

–0.007** 0.003 0.024 –0.211** 0.021 0.001

–0.002 0.003 0.471 –0.085** 0.020 0.001

–0.014** 0.004 0.001 –0.038** 0.003 0.001

–0.013** 0.004 0.001 –0.036** 0.003 0.001

0.651** 0.095 0.001 0.112 0.021**

0.142 0.108 0.187 –0.120 0.014** 0.006 REF 0.169** –0.037 –0.071 –0.529** 1.199** –0.006 0.274 0.029 0.164 1.217** 0.240**

0.753** 0.094 0.001 –0.118 0.033**

0.155 0.108 0.148 –0.231 0.019** 0.057** REF 0.184** –0.033 –0.049 –0.543** 1.369** 0.184 0.468** 0.245 0.373** 1.438** 0.237**

0.639** 0.095 0.001 0.822** –0.115

–0.113 0.112 0.001 0.791** –0.115 –0.111** REF –0.066 –0.490** –0.778** –1.556** 0.119 –1.146** –0.901** –1.153** –1.084** 0.106 0.257**

0.164**

0.140**

0.153**

**Coefficients are significant at the 5 percent level.

is more difficult to analyze. The explanatory power of these models is pretty good since pseudo R 2 values range from 12.5 percent to 26.6 percent, which is relatively satisfactory with individual data. The PV models are clearly dominated by accrual models. The comparison of OV models with accrual models is less straightforward: On the one side, OV models mostly provide lower pseudo R 2 values than accrual models, but on the other side, the coefficient on the IV appears to be more robust to any change in specification (linear age versus age dummies, for instance) in OV models than in accrual models. These estimations were performed with incentive variables including part of the incentives associated with early retirement or unemployment schemes. But our specification very poorly accounts for the discrepancy in access probabilities to these schemes. In particular, we use sector-specific access probabilities, whereas access probabilities are basically firm spe-

Estimating Models of Retirement Behavior on French Data Table 4.16

265

Probit Models (women, full-rate dummy)

SSW (10,000) Standard deviation Implied probability Incentive variable (10,000) Standard deviation Implied probability Early retirement/full rate dummy Standard deviation Implied probability Projected earnings (1,000) Square of projected earnings Age (linear) 55 56 57 58 59 60 61 62 63 64 65 Pseudo R 2

Accrual Model (linear age)

Accrual Model (age dummies)

PV Model (linear age)

PV Model (age dummies)

OV Model (linear age)

OV Model (age dummies)

–0.014** 0.006 0.018 –0.475** 0.039 0.001

–0.006 0.006 0.305 –0.329** 0.041 0.001

–0.016** 0.006 0.014 –0.183** 0.022 0.001

–0.008 0.007 0.212 –0.132** 0.022 0.001

–0.019** 0.007 0.005 –0.020** 0.002 0.001

–0.019** 0.007 0.010 –0.019** 0.003 0.001

0.999* 0.095 0.001 1.167 –2.539

0.308** 0.112 0.006 –0.026 –1.236 0.072** REF –0.011 0.007 –0.064 –0.063** 1.194** –0.037 0.371** 0.212 0.517** 1.958** 0.267**

1.086** 0.094 0.001 1.292 –1.616

0.328** 0.111 0.003 0.150 –0.653 0.076** REF –0.009 –0.013 –0.074 –0.704** 1.266** 0.014 0.394** 0.248 0.371** 1.981** 0.260**

1.056** 0.093 0.001 1.370 –0.875

0.226** 0.112 0.044 1.188 –0.809 0.001 REF –0.108 –0.221** –0.403** –1.173** 0.751** –0.565** –0.222 –0.383 –0.281 1.416** 0.263**

0.190**

0.164**

0.160**

**Coefficients are significant at the 5 percent level.

cific. But we lack appropriate firm data to model this phenomenon. We performed some other regressions including only SS incentives into incentive variables (access probabilities are set equal to 0). The predictive power of these models is not really weaker than the predictive power of the first regressions: Our attempt to account for incentives associated with early retirement or unemployment schemes is not conclusive.

4.6 Simulations Two reforms are simulated in this section.

• The three-year-increase reform shifts the minimum age to claim SS benefits to sixty-three. The full rate is obtained if people have worked at least 162 quarters (instead of 150) or if they are sixty-eight. Access

266

Ronan Mahieu and Didier Blanchet

probabilities to unemployment or early retirement schemes are incremented by three years (access to these schemes is therefore impossible before fifty-eight. • The common reform allows people to claim SS benefits from sixty. Claiming SS benefits at sixty-five provides a 60 percent replacement rate. The pension is decreased or increased by 6 percent per year below or above sixty-five, respectively. Access to early retirement or unemployment schemes is impossible. Figure 4.2 displays actual retirement rates for men and women and the profile provided with the OV model if the age dummies are the only source of variation: The data show pretty high retirement rates until fifty-nine (about 15 percent) and large spikes at sixty and sixty-five—but the latter only concerns a minority since most people quit before sixty-one. For each reform and each specification (accrual, PV, and OV), three simulations are performed (S1, S2, and S3):

• S1 is the simulation performed with linear age; • S2 uses the results of the model run with age dummies without modifying the dummies;

• S3 uses the results of the model run with age dummies, but we modify the dummies in a “plausible” way. In the three-year increase reform, dummies are incremented by three years so that, for example, the agefifty-eight dummy in the simulation is the estimated age-fifty-five dummy. In the common reform, dummies are the estimated dummies at sixty and at sixty-five (spikes at sixty and sixty-five are assumed to account for early and normal retirement). Between sixty-one and sixty-four, the dummy is the average of the estimated dummies at fiftyeight and fifty-nine. 4.6.1 Three-Year-Increase Reform Model S1 Figures 4.3, 4.4, and 4.5 show the S1 results. Simulations performed with the accrual specification show a translation to the right of the graph: Before the minimum age to claim SS benefits—now sixty-three—the average increase in the pension caused by a one-year delay remains almost unchanged (it is unchanged for short careers, which cannot reach the full rate in any case, and for long careers, which already have 162 quarters). The age-sixty spike is thus moved to sixty-three, but the average age of retirement is only increased by 0.42 years since retirement rates between fiftyfive and sixty-two remain pretty high (see table 4.17). This reflects the fact that we were not able to accurately model eligibility to early retirement benefits: As a result, the constant in the probit regression is likely to be

A

B

Fig. 4.2

Male retirement rates: A, Actual retirement rates; B, Dummies profile

Note: In this figure, we compare actual retirement rates for men with the simulated effect of age dummies on retirement rates. We fixed all other variables (SSW, option, earnings, etc.) to their mean sample value, and we use the estimated coefficients of the OV model (with age dummies) to assess the effect of age dummies. This helps us to understand which part of the fit is explained by the dummies.

268

Ronan Mahieu and Didier Blanchet

A

B

Fig. 4.3

S1 men: A, Accrual, hazard; B, Accrual, cumulative

overestimated and the coefficient on accrual underestimated, and the simulated impact of the reform is weak since the constant remains unchanged in the simulation. The effect before sixty is a bit larger with the PV specification: At age fifty-five, people realize that they may enjoy early retirement or unemployment benefits if they wait until fifty-eight (since access to these schemes is now impossible before age fifty-eight), which pulls down retire-

Estimating Models of Retirement Behavior on French Data

269

A

B

Fig. 4.4

S1 men: A, PV, hazard; B, PV, cumulative

ment rates (this fact is not captured by the accrual specification). The average retirement age grows by 1.39 years. The effect is the largest with the OV specification. Two reasons for this may be mentioned: First, the OV measure is more appropriate since it includes not only the SSW, but also the wage component. This is an incentive to delay more retirement, since people do not want to stay several years without wage or pension income: Retirement rates are thus lower before sixty-three. Second, the problematic decreasing age profile, as previously discussed, explains that the age-sixty-three spike (three-year-increase reform) is lower than the age sixty spike (baseline). This latter fact explains

270

Ronan Mahieu and Didier Blanchet

A

B

Fig. 4.5

S1 men: A, OV, hazard; B, OV, cumulative

why the average retirement age increases by 3.14 years (a figure above 3 is of course doubtful). Model S2 Figures 4.6, 4.7, and 4.8 show the S2 results (age dummies not incremented). Since age dummies remain unchanged, retirement rates remain close to the baseline case for accrual or PV specifications (with a slight decrease in the age-sixty spike since people do not reach the full rate at that

Estimating Models of Retirement Behavior on French Data Table 4.17

Average Retirement Age, Baseline, and Simulations Men

Baseline Accrual S1 S2 S3 PV S1 S2 S3 OV S1 S2 S3

271

Women

Three-Year Increase Reform

Common Reform

Three-Year Increase Reform

Common Reform

58.64

58.64

58.85

58.85

59.06 58.85 59.16

57.60 58.23 58.32

— — —

— — —

60.03 59.17 59.64

57.68 58.27 58.40

— — —

— — —

61.78 61.39 60.50

60.20 59.88 59.95

60.43 60.17 59.97

59.32 59.14 59.28

Note: Dashes indicate that data is not available.

age). We do not observe any spike at sixty-three (although a majority of people can claim full rate SS benefits from sixty-three) since the age-sixtythree dummy is quite low. The increase in the average retirement age is thus only 0.21 and 0.53 years with the Accrual PV specifications, respectively. But the decrease in the age-sixty spike (and even in retirement rates before sixty) is far larger with the OV specification since the model captures the fact that people have no income for three years if they stop immediately. The average retirement age increases by 2.75 years. Model S3 Figures 4.9, 4.10, and 4.11 show the S3 results. Since age dummies are incremented by three years, retirement rates are more or less shifted by three years to the right for accrual or PV specifications. The increase in the average retirement age is 0.52, 1.00, and 1.86 years with the Accrual, PV, and OV specifications, respectively. Once again, the effect is larger with the OV specification because this more forward-looking measure accounts for the fact that people include not only the SSW but also the wage component in making their decisions. One unexpected result should be pointed out: In OV specifications, the increase in the average retirement age is larger with model S2 than S3 (it is the opposite with accrual or PV specifications). This is the result of the globally decreasing age profile in OV specifications (not only with linear age, but also with age dummies). In model S2, the simulated retirement rate at fifty-eight is computed with the age-fifty-eight dummy, which is quite low, and thus explains a very low probability of exit. In model S3, this sim-

272

Ronan Mahieu and Didier Blanchet

A

B

Fig. 4.6

S2 men: A, Accrual, hazard; B, Accrual, cumulative

ulation is performed with the age fifty-five dummy, which is much larger, and the simulated retirement rate is thus higher. 4.6.2 Common Reform Model S1 Accrual and PV specifications provide very flat retirement rates (figures 4.3 and 4.4). This is the expected outcome of introducing an actuarially fair

Estimating Models of Retirement Behavior on French Data

273

A

B

Fig. 4.7

S2 men: A, PV, hazard; B, PV, cumulative

pension scheme: the level of SSW is more or less the same, whatever the age of people who claim SS benefits. The accrual and PV measures are thus close to zero, and the annual retirement rates constant at pretty high levels: the average retirement age decreases by 1.04 and 0.96 years) with the accrual and PV specifications, respectively. The OV measure seems more appropriate: Even if the age profile of SSW is flat, people may prefer to have a wage income until sixty instead of nothing if they quit earlier (figure 4.5). The retirement rate thus increases with age but is lower (by about 10 points) than in the baseline case since people are sensitive to SS incentives. After age sixty, retirement rates

274

Ronan Mahieu and Didier Blanchet

A

B

Fig. 4.8

S2 men: A, OV, hazard; B, OV, cumulative

are flat (the increase in the pension in case of continued work is offset by the loss of leisure). The average retirement age increases by 1.56 years, which is more likely than the results obtained with accrual or PV specifications. Model S2 As in the 3-year increase reform case, retirement rates remain relatively close to the baseline case, but at a slightly lower level with accrual and PV specifications (figures 4.6 and 4.7). In the baseline case, a minority of people could enjoy in increase in their SSW if they stayed one year or more

Estimating Models of Retirement Behavior on French Data

275

A

B

Fig. 4.9

S3 men: A, Accrual, hazard; B, Accrual, cumulative

on the labor market. Things are now different: SSW is more or less constant and retirement rates thus increase. The average retirement age decreases by 0.41 and 0.37 years with the accrual and PV specifications, respectively. The results are somewhat different with the OV specification: Retirement rates before sixty now decrease, in comparison with the baseline case, since people prefer to avoid a situation where they have no income until sixty (figure 4.8). The average retirement age increases by 1.24 years, which is the expected sign, but the effect remains moderate.

276

Ronan Mahieu and Didier Blanchet

A

B

Fig. 4.10

S3 men: A, PV, hazard; B, PV, cumulative

Model S3 Results are very close to those obtained with model S2 since age dummies are not changed between fifty-five and sixty (figures 4.9, 4.10, and 4.11). Age dummies are somewhat lower than in model S2 between sixtyone and sixty-four, but this only concerns the minority of people who are still working at sixty-one. With accrual and PV specifications, the average retirement age decreases by 0.32 and 0.24 years, respectively. With the OV specification, it increases by 1.30 years. 4.6.3 Comparison of the Different Specifications—The Case of Women The current structure of SS incentives induces high retirement rates when people get the full rate (mostly at sixty). At that age, the level of ac-

Estimating Models of Retirement Behavior on French Data

277

A

B

Fig. 4.11

S3 men: A, OV, hazard; B, OV, cumulative

crual suddenly falls for most agents. This explains the very high explanatory power of the accrual specification (greater than the explanatory power of the OV specification, although the latter approach is richer). But while analyzing the impact of changes in the computation of pensions, the OV specification proved to be more appropriate. In particular, enforcing the common reform (with an increase of incentives to postpone retirement after sixty for a majority and the suppression of early retirement schemes) should logically induce a decrease in retirement rates before sixty. This is not the case with accrual or PV specifications. We thus chose to present simulations performed with the OV specifica-

278

Ronan Mahieu and Didier Blanchet

A

B

Fig. 4.12

S1 women: A, OV, hazard; B, OV, cumulative

tion for women (figures 4.12, 4.13, and 4.14). For women, the only drawback of the OV specification should disappear. With men the negative coefficient on linear age (and the decreasing profile of age dummies) were puzzling and led to an excessive simulated impact of the three-year increase reform (see previous). This should not be the case for women since the coefficient on linear age is insignificant in the OV specification. The three-year increase reform would involve an increase by 1.58, 1.32, and 1.12 years of the average retirement age for women with model S1, S2, and S3, respectively. With the common reform, as expected, the effect is

Estimating Models of Retirement Behavior on French Data

279

A

B

Fig. 4.13

S2 women: A, OV, hazard; B, OV, cumulative

smaller: The average retirement age increases by 0.47, 0.29, and 0.43 years with model S1, S2 and S3. 4.7 Conclusion This project allowed us to assess the sensitivity of individual retirement behavior to the structure of SS incentives. Of course, the robustness of these results may be in question for mainly two reasons:

• First, precisely estimating this sensitivity is difficult since the strength of current incentives in France deeply limits the heterogeneity of behaviors in available data;

280

Ronan Mahieu and Didier Blanchet

A

B

Fig. 4.14

S3 women: A, OV, hazard; B, OV, cumulative

• Second, current behaviors strongly reflect demand-side effects on the labor market that our supply-side approach cannot capture. In particular, access to early retirement or unemployment schemes is very imperfectly modeled. Nonetheless we performed various regressions that provide a quite satisfactory description of the behaviors observed in our sample. Sensitivity analysis (through the simulation of alternative rules for the computation of pensions) showed that the OV model (which accounts not only for the pure SSW, but also for the wage component of incentives) seems more appropriate to simulate policy changes. The simulation of individual behaviors under such changes showed a relatively important sensitivity to policy

Estimating Models of Retirement Behavior on French Data

281

changes since, for example, an increase by three years of all important parameters (minimum age to claim SS benefits and tenure required to get the full rate) could induce an increase by between one and two years of the average retirement age (results obtained with the OV specification).

Appendix Wage Projections As stated above, our only requirement for wages is their extrapolation at later ages, starting at age fifty-five. This is done through the estimation of wage equations. Even if the analysis will only apply to cohort 1930, wage equations have been estimated using all the information available, that is, data taken from DADS for all cohorts of the EIR (1926, 1930, 1932, 1934, 1936, 1938, 1940, and 1942): This corresponds to 207,433 observations (33,535 individuals) over the time period 1985–1996 (with one missing point in 1990). Of course, wage equations are estimated on annualized wages, not on the basis of wages effectively earned by people belonging to these cohorts (to avoid the downward bias that would result from exits from the labor force during the year). This is done by dividing wages effectively earned by duration of pay and remultiplying the result by 360. This does not avoid a certain number of imperfections: Some people go on receiving very small wages at later ages, sometimes after the liquidation of their pensions (this is allowed within rather narrow limits). When the duration of pay is poorly registered (for instance, when it is declared to be 360 days while it is obviously less), this method leads to a very low level of the annualized wage which is going to bias the estimation of wage equations. To avoid these imperfections, a somewhat arbitrary test is applied: Observations with an annualized wage lower than 90 percent of the annual minimum wage (salaire minimum interprofessionnel de croissance; SMIC) for a full-time job have been dropped from the sample. Once this is done, there remains 175,109 observations (29,483 individuals), that is, a sample size reduced by 15.6 percent. Estimations are then performed on 1985–1995: This reduces once more the size of the sample, which is now composed of 165,530 observations. Observations for 1996 are left for out-of-sample tests of the accuracy of projections, the criterion for these tests being the mean or median squared error. Four alternative methods have been attempted for the projection of wages.

282

Ronan Mahieu and Didier Blanchet

Method A1 is a simple extrapolation of the last known wage. It is assumed that wages are constant in real terms (i.e., a 2 percent increase in nominal terms between 1995 and 1996). Method A2 estimates wage equations without individual fixed effects. Wages are then projected, at the individual level, by applying the variation predicted by wage equations to the last known wage for 1995. Two subpossibilities have been explored for the estimation of wage equations: Method A21 estimates the wage equation on first order differences; and Method A22 estimates on levels (using logs). Method B estimates wage equations on levels but with individual fixed effects. The estimated average age profile is flattened after age fifty, and individual effects are added to this profile to give expected wage profiles for individuals after this age. Method C is a variant of the simple method A1 and the extrapolation of the average wage over the last three years. The equations adjusted for methods A21, A22, or B are for the models in levels Yt
F(Xt, age, agesq, Yt1 , Yt2 , Yt1  age, Yt2  age, Yt1  agesq, Yt2

 agesq, timet), or for models on differences Yt
F(Xt , age, agesq, Yt1 , Yt2 , Yt1  age, Yt2  age, Yt1

 agesq, Yt2  agesq, timet ), with Yt being the log of wages, Xt a set of explanatory variables including a dummy for people living in or around Paris, a dummy for the socioprofessional group in four groups, and a dummy for the kind of activity (in sixteen groups). Equations have been estimated separately for each gender. A problem is that the estimations of these equations require the knowledge of four successive years if the model is in differences, or three successive years if the model is in levels. Four successive years are also needed with method C. Yet many individuals are not observed during four consecutive years. In particular, cohort 1930, the one to which the model is going to be applied, is not observed before age fifty-five. Of course, at least when the model is in differences, the knowledge of past wages is not absolutely necessary if we want to use the model for projections, since we can limit ourselves to applying the age profile derived from the equation. But it is nevertheless a problem to drop people with short wage records from the sample: It may bias results since these people are less likely to have had good performances in terms of wage progression. Models first have been estimated as initially proposed (on a subsample of people working three consecutive years), and then reestimated by pro-

Estimating Models of Retirement Behavior on French Data

283

gressively dropping variables related to past wages. For instance, for the model in levels, we have successively estimated Yt
F(Xt , age, agesq, Yt1 , Yt1  age, Yt1  agesq, timet ), and Yt
F(Xt , age, agesq, timet ). Table 4A.1 shows the resulting quality of wage predictions for all five methods with a varying number of lags (irrelevant cases are left empty). Results are given both for the total population, and for full-time workers only. The simplest of all methods, method A1 does not perform so badly. Its variant method C does not add any improvement. Method B does not perform very well, especially for men. It is difficult to compare methods A21 and A22. Wage equations estimated on levels perform significantly better if the criterion if the median of squared errors. Another criterion for selection is the examination of resulting age profiles for wages. The profiles are Table 4A.1

Out of Sample Wage Projections for 1996 Men (no. of lags) 0 year

A1 (last wage) A21 (wage equation, differences) A22 (wage equation, levels) B (wage equations, with fixed effects) C (average wage of last 3 years)

A1 (last wage) A21 (wage equation, differences) A22 (wage equation, levels) B (wage equations, with fixed effects) C (average wage of last 3 years)

1 year

All Workers  55 0.1450 0.0037 0.1454 0.1230 0.0054 0.0056 0.1447 0.1447 0.0036 0.0052 0.1465 0.1520 0.0241 0.0133

2 years

Women (no. of lags) 0 year

1 year

2 years

0.0813 0.0028 0.0928 0.0042 0.0937 0.0039 0.1380 0.0029

0.1011 0.0045 0.0904 0.0017 0.0931 0.0041

0.0785 0.0023 0.0865 0.0042 0.0892 0.0031

0.0984 0.0039 0.0859 0.0020 0.0785 0.0037 0.1322 0.0023

a

0.0959 0.1206 0.0059 0.1443 0.0040 0.1436 0.0083 0.1510 0.0042

Full-Time Workers  55 b 0.1246 0.0026 0.1239 0.1005 0.1010 0.0053 0.0051 0.0048 0.1235 0.1257 0.1260 0.0032 0.0041 0.0029 0.1177 0.1333 0.1179 0.0189 0.0132 0.0068 0.1365 0.0031

0.0018 0.0956 0.0034 0.0953 0.0021 0.1031 0.0125

0.0899 0.0014 0.0885 0.0030 0.0888 0.0015 0.0953 0.0018

Note: Estimation period: 1985–1995, 1990 missing. Blank cells indicate that data is not relevant. a Mean squared errors. b Median squared errors.

284

Ronan Mahieu and Didier Blanchet

convex and rapidly increasing when differences are modeled. They look more satisfactory when wage equations are estimated on levels. On the whole, the choice is between methods A1 and A22. The same conclusion can be reached if we restrict ourselves to the subsample of full-time workers. Method A22, without lagged income, is the one which has been finally preferred.

References Blanchet, D., and L. P. Pelé. 1999. Social security and retirement in France. In Social security and retirement around the world, ed. D. A. Wise and J. Gruber, 101– 33. Chicago: University of Chicago Press. Bommier, A., T. Magnac, and M. Roger. 2001. Départsen retraite: Évolutions récentes et modèles économiques (Retirement behavior: Recent evolutions and economic models). Revue Française d’Economie 16 (July): 79–124. Burricand, C., and N. Roth. 2000. Les parcours de fin de carrière des générations 1912–1941: l’Impact du cadre institutionnel (The 1912–1941 generations’ ends of careers: The impact of the institutional framework). Economie et Statistique 335:63–79. Caussat, L., and N. Roth. 1997. De l’emploi à la retraite: Générations passées et futures (From employment to retirement: Past and future cohorts). Revue Française des Affaires Sociales 51:177–201. Charpin, J. M. 1999. L’avenir de nos retraites (The future of our pensions). Paris: La Documentation Française. Colin, C., V. Iéhlé, and R. Mahieu. 2000. Les trajectories de fin de carriére des salariés du seceur privé (The end of workers’ careers in the private sector). Dossiers Solidarité et Santé no. 3:9–27. Commissariat Général du Plan. 1995. Perspecives à long terme des retraites (Longterm prospects for pensions). Paris: La Documentation Française. Dangerfield, O., and D. Prangére. 1996. Des retraites aux retraités: L’échantillon interrégime de retraités, opération 1993, document méthodologique (From pensions to pensioners: The interregime sample of pensioners, 1983 edition, methodological document). Documents Statistiques no. 255. Paris: Service des Statistiques et des Systèmes d’Information (SESI). Gruber, J., and D. Wise. 1999. Introduction and summary. In Social security and retirement around the world, ed. J. Gruber and D. Wise, 1–35. Chicago: University of Chicago Press. Heller, J. L. 1985. La préretraite, choix ou contrainte (Preretirement: Choice or constraint). Economie et statistique 193–194:97–109. Pelé, L. P., and P. Ralle. 1997. Âge de la retraite: Les aspects incitatifs du régime général (Retirement age: Incentive aspects of the general regime). Working Paper no. G9718. Paris: National Institute of Statistics and Economic Studies (INSEE). Stock, J. and D. Wise. 1990. Pensions, the option value of work, and retirement. Econometrica 58 (5): 1151–80. Taddei, D. 2000. Retraites choisies et progressives (Chosen and progressive pensions). Rapport du Conseil d’Analyse Economique. Paris: La Documentation Française.

5 Micro-Modeling of Retirement Decisions in Germany Axel Börsch-Supan, Reinhold Schnabel, Simone Kohnz, and Giovanni Mastrobuoni

5.1 Introduction Germans retire early. Average retirement age is about fifty-nine-andone-half years, half a year younger than the earliest eligibility age for old-age pensions and more than five years younger than the “normal” retirement age in Germany. Early retirement is a well-appreciated social achievement among Germans, but it is costly. Since life expectancy at age sixty is about seventeen years, a year of early retirement corresponds to more than 5 percent of pension expenditures. This paper is part of a multistage research project on the causes for and the effects of early retirement.1 Its significance stems from the mounting strain on the German public pension system. The German public pension or, as it is known in German, “public retirement insurance,” was the first formal pension system when it was installed over one hundred years ago and has been a model for many social security systems in the world. It has been very successful in providing a high and reliable level of retirement income over the past one hundred years. It has survived, although under severe modifications, through World Wars I and II, the Great Depression, and, most recently, the German unification. However, times have changed. According to recent polls, most young Axel Börsch-Supan is director of the Mannheim Research Institute for the Economics of Aging, professor of economics at the University of Mannheim, and a research associate of the National Bureau of Economic Research (NBER). Reinhold Schnabel is professor of economics at the University of Essen. Simone Kohnz is currently a Ph.D. student at the University of Munich and was research fellow at the Mannheim Research Institute for the Economics of Aging in 2002. Giovanni Mastrobuoni is a Ph.D. student in Economics at Princeton University. 1. See the country chapters in Gruber and Wise (1999) for the first stage.

285

286

Fig. 5.1

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

Average age at first receipt of public pensions, 1960–1998

Source: VDR (1999), male workers only.

people do not believe that they will receive a pension that will suffice for their old-age consumption, and the number of employees that are using the few existing loopholes to escape the otherwise mandatory retirement insurance system has increased dramatically. Adding to this nervousness, Germany has experienced two major pension reforms in 1992 and 2001, each of them dubbed “century reforms,” and a constant flurry of minor changes between 1992 and 2001. The German public pension model is under siege, and there appear to be two main culprits for this: negative incentive effects of the system, among them the incentives to retire early that have reduced the number of contributors and increased the number of beneficiaries (the “system-dependency ratio”) since 1972, and the aging population, which will rather dramatically increase the system-dependency ratio beginning in 2015 and onward. This paper is not the forum in which to discuss population aging and its implications on the pension system.2 Rather, we focus on the incentive effects to retire early. Figure 5.1 depicts the evolution of average retirement age among German men from 1960 through 1998, once disaggregated by old-age pensions and disability pensions, and once total. The most obvious feature is the sudden change after 1972, when the retirement age drops sharply for both old-age and disability pensions. Within a few years, the average retirement age for old-age pensions dropped by about three years and has then stabilized. For disability pensions, we see a steady decline since 1972 that has not stopped yet. Composition effects— mainly caused by the tighter disability rules—have led to a consolidation of the total retirement age at about fifty-nine-and-one-half years. The year 1972 marks the first major pension reform after the current pay-as-you-go (PAYG) public pension system was installed in 1957. This reform introduced a “flexible” retirement age without actuarial adjust2. See Börsch-Supan (1998, 2000a) and Schnabel (1998) for descriptions of the problems, and Birg and Börsch-Supan (1999) and Börsch-Supan (2001a) for concrete reform proposals.

Micro-Modeling of Retirement Decisions in Germany

287

ments of pension benefits. Without going further into details—see BörschSupan and Schnabel (1998) for a more detailed description and analysis— figure 5.1 appears to be prima facie evidence for the incentives which pension rules create to retire early.3 Several formal econometric analyses based on micro-data have studied the incentive effects of the nonactuarial adjustment on early retirement (Börsch-Supan 1992; Schmidt 1995; Siddiqui 1997; and Börsch-Supan 2000c, 2001b). These studies employ variants of the micro-econometric option value analysis developed by Stock and Wise (1990). Börsch-Supan (2000c) derives from the estimates that the 1992 reform will increase the average retirement age only by about half a year and will reduce retirement before age sixty from 32 percent to about 28 percent, while a switch to a system with actuarially fair adjustment factors would shift the retirement age by about two years. Indeed, these estimates are well in line with the drop illustrated in figure 5.1. Börsch-Supan (2001b) shows that, in effect, these estimates are robust even when much more sophisticated specifications are applied. This paper builds on these econometric analyses. Its main purpose is to provide further econometric evidence for the strength of the incentive effects to retire early, based on micro-data. It adds to the existing literature in at least four respects. First, this paper uses definitions and specifications that are comparable to the other countries in this volume. Second, the paper extends the comprehensive treatment of retirement as an option with several pathways in Börsch-Supan (2001b) beyond the standard old-age and disability pension. Third, the paper exploits as much of the sample variation as possible; specifically, we include civil servants in our estimations. Fourth and finally, we apply a “family approach” to retirement options and compute the joint incentives for husband and spouse.4 The paper is structured as follows. Sections 5.2 and 5.3 describe the institutional background for private-sector and civil servants’ pensions. Section 5.4 presents data and variable specifications, section 5.5 contains our estimation results; section 5.6 explores what these estimates mean, simulates a set of pension reform steps and concludes. 5.2 Private-Sector Pensions In this section we describe the German public retirement insurance (Gesetzliche Rentenversicherung or GRV), which covers about 85 percent of the German workforce. Most of these are private-sector workers, but the GRV also includes those public-sector workers who are not civil servants. 3. A competing explanation is that labor demand effects are due to rising unemployment. See Riphahn and Schmidt (1995) and Börsch-Supan (2000c), who show that there is no evidence in favor of this. 4. See Coile (1999) for the significance of this extension.

288

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

Civil servants, about 7 percent of the workforce, have their own pension system, described in section 5.3. The self-employed, about 9 percent of the work force, are mainly self-insured although some of them also participate in the public retirement insurance system. For the average worker, occupational pensions do not play a major role in the German system of old-age provision, neither do individual retirement accounts, but there are important exceptions from this general picture. Broadly speaking, the German system is a monolith. The following descriptions focus on the institutional rules that applied during our sample period 1984–1997 (dubbed “1972 legislation,” although there have been several administrative adjustments since 1972). There have been two major pension reforms in 1992 and 2001. At several places, notably the last subsection, we briefly sketch their implications. These reforms, however, did not affect the persons in our sample. 5.2.1 Coverage and Contributions The German PAYG public pension system features a very broad mandatory coverage of workers. Only the self-employed and, until 1998, workers with earnings below the official minimum-earnings threshold (Geringfügigkeitsgrenze, which is 15 percent of average monthly gross wage; below this threshold are about 5.6 percent of all workers) are not subject to mandatory coverage. Roughly 70 percent of the budget of the German public retirement insurance is financed by contributions that are administrated like a payroll tax, levied equally on employees and employers. Total contributions in 2000 are 19.3 percent of the first DM 8,600 of monthly gross income (the upper-earnings threshold, Beitragsbemessungsgrenze, is about 180 percent of average monthly gross wage).5 Technically, contributions are split evenly between employees and employers. While the contribution rate has been fairly stable since 1970, the upper-earnings threshold has been used as a financing instrument. It is anchored to the average wage and has increased considerably faster than inflation. Private-sector pension benefits are essentially tax free. Pension beneficiaries do not pay contributions to the pension system or to unemployment insurance. However, pensioners have to pay the equivalent of the employees’ contribution to the mandatory medical insurance. The equivalent of the employers’ contribution to health insurance is paid by the pension system. The remaining approximately 30 percent of the social security budget are financed by earmarked indirect taxes (a fixed fraction of the valueadded tax and the new “eco-tax” on fossil fuel) and a subsidy from the fed5. This is for West Germany only; it is DM 7,200 in East Germany. One DM has a purchasing power of approximately $0.50.

Micro-Modeling of Retirement Decisions in Germany

289

eral government. The subsidy is also used to fine-tune the PAYG budget constraint, which has a minimal reserve of one month worth of benefits. 5.2.2 Benefit Types The German public retirement insurance provides old-age pensions for workers aged sixty and older; disability benefits for workers below age sixty, which are converted to old-age pensions at age sixty-five at the latest; and survivor benefits for spouses and children. In addition, preretirement (i.e., retirement before age sixty) is possible through several mechanisms using the public transfer system, mainly unemployment compensation. We begin by describing old-age pensions. 5.2.3 Eligibility for Benefits and Retirement Age for Old-Age Pensions Eligibility for benefits and the minimum retirement age depend on which type of pension the worker chooses. The German public retirement insurance distinguishes five types of old-age pensions, corresponding to normal retirement and four types of early retirement. This complex system was introduced by the 1972 social security reform. One of the key provisions was the introduction of flexible retirement after age sixty-three with full benefits for workers with a long service history. In addition, retirement at age sixty with full benefits is possible for women, unemployed, and older disabled workers. Older disabled workers refers to those workers who cannot be appropriately employed for health or labor market reasons and are age sixty or older. There are three possible ways to claim old-age disability benefits. One has to either (a) be at least 50 percent physically disabled; (b) pass a strict earnings test; or (c) pass a much weaker earnings test. The strict earnings test is passed if the earnings capacity is reduced below the minimum-earnings threshold for any reasonable occupation (about 15 percent of average gross wage; erwerbsunfähig or EU). The weaker earnings test is passed when no vacancies for the worker’s specific job description are available, and the worker has to face an earnings loss of at least 50 percent when changing to a different job (berufsunfähig or BU). As opposed to the disability insurance for workers below age sixty (see later discussion), full benefits are paid in all three cases. Figure 5.2 shows the uptake of the various pathways,6 including the disability pathway described below (adding up to 100 percent on the vertical axis) and their changes over time (marked on the horizontal axis), mostly in response to reforms, benefit adjustments, and administrative rule changes, particularly the tightening of the disability screening process. This figure shows the multitude of possible pathways. A major undertaking of this paper is to take account of this diversity. According to the 1992 social security reform and its subsequent modifi6. See Jacobs, Kohli, and Rein (1990) for this concept.

290

Fig. 5.2

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

Pathways to retirement, 1960–1995: Males

Source: Börsch-Supan and Schnabel (1999).

cations, the age limit for types of early retirement will gradually be raised to age sixty-five. These changes will be fully be phased in by the year 2004. The only distinguishing feature of types B and C of “early retirement” will then be the possibility to retire up to five years earlier than age sixty-five if a sufficient number of service years (currently thirty-five years) has been accumulated. As opposed to the pre-1992 regulations, benefits will be adjusted to a retirement age below age sixty-five in a fashion that will be described below. 5.2.4 Benefits Benefits are strictly work related. The German system does not have benefits for spouses, like in the United States.7 Benefits are computed on a lifetime basis and adjusted according to the type of pension and retirement age. They are the product of four elements: (a) the employee’s relativeearnings position; (b) the years of service life; (c) adjustment factors for pension type and (since the 1992 reform) retirement age; and (d) the average pension. The first three factors make up the personal pension base, while the fourth factor determines the income distribution between workers and pensioners in general. The employee’s relative-contribution position is computed by averaging their annual relative-contribution positions over the entire earnings history. In each year, the relative-contribution position is expressed as a multiple of the average annual contribution (roughly speaking, the relativeincome position). A first element of redistribution was introduced in 1972, when this multiple could not fall below 75 percent for contributions before 1972, provided a worker had a service life of at least thirty-five years. A similar rule was introduced in the 1992 reform: For contributions between 1973 and 1992, multiples below 75 percent are multiplied by 1.5 up to the 7. There are, of course, survivor benefits.

Micro-Modeling of Retirement Decisions in Germany

291

maximum of 75 percent, effectively reducing the redistribution for workers with income positions below 50 percent. Years of service life are years of active contributions plus years of contribution on behalf of the employee and years that are counted as service years even when no contribution were made at all. These include, for instance, years of unemployment, years of military service, three years for each child’s education for one of the parents, some allowance for advanced education, and so forth, thus introducing a second element of redistribution. The official government computations, such as the official replacement rate (Rentenniveau), assume a forty-five-year contribution history for what is deemed a normal earnings history (Eckrentner). In fact, the average number of years of contributions is about thirty-eight years. Unlike the United States, there is neither an upper bound of years entering the benefit calculation, nor can workers choose certain years in their earnings history and drop others. Since 1992, the average pension is determined by indexation to the average net labor income. This solved some of the problems that were created by indexation to gross wages between 1972 and 1992. Nevertheless, wage, rather than cost of living, indexation makes it impossible to finance the retirement burden by productivity gains. The average pension has provided a generous benefit level for middleincome earnings. The net replacement rate for a worker with a forty-fiveyear contribution history is 70.5 percent in 1998. For the average worker with thirty-eight years of contributions, it is reduced in proportion to 59.5 percent. Unlike the United States, the German pension system has very little redistribution, as is obvious from the benefit computation.8 The low replacement rates for high incomes result from the upper limit to which earnings are subject to social security contributions—they correspond to a proportionally lower effective contribution rate. Before 1992, adjustment of benefits to retirement age was only implicit via years of service. Because benefits are proportional to the years of service, a worker with fewer years of service will get lower benefits. With a constant income profile and forty years of service, each year of earlier retirement decreased pension benefits by 2.5 percent and vice versa. The 1992 social security reform will change this by the year 2004. Age sixty-five will then act as the pivotal age for benefit computations. For each year of earlier retirement, up to five years and if the appropriate conditions in table 5.1 are met, benefits will be reduced by 3.6 percent (in addition to the effect of fewer service years). The 1992 reform also introduced rewards for later retirement in a systematic way. For each year of retirement postponed past the minimum age indicated in table 5.1, the pension is increased by 6 percent in addition to the natural increase by the number of service years. Table 5.2 displays the retirement-age-specific adjustments for a worker 8. See Casmir (1989) for a comparison.

292

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

Table 5.1

Old-Age Pensions (1972 legislation)

Pension Type

Retirement Age

Years of Service

A: Normal B: Long service life (“flexible”) C: Women D: Older disabled

65 63

5 35

60 60

15 35

E: Unemployed

60

15

Additional Conditions

Earnings Test No Yes

10 of those after age 40 Loss of at least 50% earnings capability 1.5 to 3 years of unemployment (has changed several times)

Yes (Yes) Yes

Notes: This legislation was changed in the reform of 1992. It was effective until 1998. Table 5.2

Adjustment of Public Pensions, by Retirement Age (as percentage of pension one would obtain if retired at age 65) Germany

Age 62 63 64 65 66 67 68 69

Pre-1992 100.0 100.0 100.0 100.0 107.2 114.4 114.4 114.4

a

United States

Post-1992 89.2 92.8 96.4 100.0 106.0 112.0 118.0 124.0

b

Pre-1983 80.0 86.7 94.4 100.0 103.0 106.0 109.0 112.0

c

Post-1983 77.8 85.2 92.6 100.0 105.6 111.1 120.0 128.9

Actuarially c

Fair e 80.5 86.3 92.8 100.0 108.1 117.2 127.4 139.1

Source: Börsch-Supan and Schnabel (1999). a GRV 1972–92. b GRV after 1992 reform has been fully phased in. c U.S. Social Security (OASDHI) until 1983. d U.S. Social Security after 1983 reform has been fully phased in. e Evaluated at a 3 percent discount rate with 1992–1994 mortality risks of West German males and an annual increase in net pensions of 1 percent.

who has earnings that remain constant after age sixty. Table 5.2 relates the income for retirement at age sixty-five (normalized to 100 percent) to the income for retirement at earlier or later ages, and compares the implicit adjustments after 1972 with the total adjustments after the 1992 social security reform is fully phased in. As references, the table also displays the corresponding adjustments in the United States and actuarially fair adjustments at a 3 percent discount rate.9 9. The actuarially fair adjustments equalize the expected social security wealth for a worker with an earnings history starting at age S equals 20. A higher discount rate yields steeper adjustments.

Micro-Modeling of Retirement Decisions in Germany

293

While neither the German nor the U.S. system were actuarially fair prior to the reforms, the public retirement system in Germany as enacted in 1972 was particularly distortive. There was less economic incentive for Americans to retire before age sixty-five and only a small disincentive to retire later than at age sixty-five after the 1983 reform, while the German social security system tilted the retirement decision heavily towards the earliest retirement age applicable. The 1992 reform has diminished but not abolished this incentive effect. 5.2.5 Disability and Survivor Benefits The contributions to the German retirement insurance also finance disability benefits to workers of all ages and survivor benefits to spouses and children. In order to be eligible for disability benefits, a worker must pass one of the two earnings tests mentioned earlier for the old-age disability pension. If the stricter earnings test is passed, full benefits are paid (EU); if only the weaker earnings test is passed and some earnings capability remains, disability pensions before age sixty are only two-thirds of the applicable old-age pension (BU). In the 1970s and early 1980s, the German jurisdiction has interpreted both rules very broadly, in particular the applicability of the first rule. Moreover, jurisdiction also overruled the earnings test (see following discussion) for earnings during disability retirement. This lead to a share of EU-type disability pensions of more than 90 percent of all disability pensions. Because both rules were used as a device to keep unemployment rates down, their generous interpretation has only recently lead to stricter legislation.10 Survivor pensions are 60 percent of the husband’s applicable pension for spouses that are age forty-five and over or if children are in the household (große Witwenrente), otherwise they are 25 percent (kleine Witwenrente). Survivor benefits are a large component of the public pension budget and of total pension wealth as will be shown in section 5.3. Certain earnings tests apply if the surviving spouse has her own income, e.g., her own pension. This is only relevant for a very small (below 10 percent) share of widows. Male and female survivors are treated symmetrically only recently. As mentioned before, the German system does not have a married-couple supplement for spouses of beneficiaries. However, most wives acquire their own pension by active and passive contribution (mostly years of advanced education and years of child education). 5.2.6 Preretirement In addition to benefits through the public pension system, transfer payments (mainly unemployment compensation) enable what is referred to as preretirement. Labor force exit before age sixty is frequent: About 45 per10. See Riphahn (1995) for an analysis of disability rules.

294

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

cent of all men call themselves retired at age fifty-nine. Only about half of them retire because of disability; the other 50 percent make use of one of the many official and unofficial preretirement schemes. Unemployment compensation has been used as preretirement income in an unofficial scheme that induced very early retirement. Before workers could enter the public pension system at age sixty, they were paid a negotiable combination of unemployment compensation and a supplement or severance pay. At age sixty, a pension of type E (see table 5.1) could start. As the rules of type-E pensions and the duration of unemployment benefits changed, so did the unofficial retirement ages. Age fifty-six was particularly frequent in West Germany, because unemployment compensation is paid up to three years for elderly workers; it is followed by the lower unemployment aid. Earlier retirement ages could be induced by paying the worker the difference between the last salary and unemployment compensation for three years and, after these three years, by paying the difference (in yearly income) between the last salary and unemployment aid—it all depended on the “social plan,” in which a firm would negotiate with the workers before restructuring the work force. In addition, early retirement at age fifty-eight was made possible in an official preretirement scheme (Vorruhestand ), in which the employer received a subsidy from unemployment insurance if a younger employee was hired. While the first (and unofficial) preretirement scheme was very popular and a convenient way to overcome the strict German labor laws, few employers used the official second scheme. 5.2.7 Retirement Behavior The retirement behavior of entrants into the German public retirement insurance system has been summarized by figures 5.1 and 5.2. For West Germany, the average retirement age in 1998 was 59.7 years for men and 60.7 years for women. In the East, the average retirement age was 57.9 years for men and 58.2 years for women. The fraction of those who enter retirement through a disability pension has declined (see figure 5.2) and was 29 percent in 1998. Only about 20 percent of all entrants used the normal pathway of an old-age pension at age sixty-five. The most popular retirement age is age sixty. 5.2.8 Pension Reform During and since our sample period, there have been two major pension reforms in 1992 and 2001 and many smaller adjustments in-between. The main changes in the 1992 reform anchored benefits to net, rather than to gross, wages. This implicitly has reduced benefits since taxes and social security contributions have increased, reducing net wages, relative to gross wages. This mechanism is particularly important when the population aging will speed up. The other important change in 1992 was the introduction

Micro-Modeling of Retirement Decisions in Germany

295

of adjustments to benefits in some (not all) cases of early retirement and a change in the normal retirement age for women. (They have been described in subsection 5.2.4.) They will be fully effective in 2017 and reduce the incentives to retire early. However, they are still not actuarially fair even at very low discount rates.11 The 2001 reform is intended to change the monolithic German system of old-age provision to a genuine multipillar system. Benefits will gradually be reduced by about 10 percent, lowering the replacement rate with respect to the average net earnings from 72 percent in 1997 to 64 percent in 2030. The effective benefit cuts are even larger since the credit of earnings points for education and training will be greatly restricted. On the other hand, a redefinition of the official replacement rate minimizes the perception of these cuts because, defined as such, the new replacement rate will be 67 percent with respect to a smaller net earnings base. The resulting pension gap of slightly less than 20 percent of the current retirement income is supposed to be filled with occupational and individual pensions. This new pillar is not mandatory, but the required private savings will be subsidized or tax privileged. The 2001 reform does not change the normal retirement age or the adjustment factors concerning the early retirement age that provide the large incentives to retire early, which is the main subject of this paper. 5.3 Public-Sector Pensions There are two types of workers in the public sector: civil servants and other public-sector workers. As already mentioned, the latter are part of the same system as the private-sector workers described in the previous section. In addition, they participate in a supplemental system that resembles occupational pensions elsewhere and raises the pensions of public-sector workers to the level of civil servants. Civil servants do not pay explicit contributions for their pensions, as the other employees in the private- and public-sectors do.12 Instead, the gross wage for civil servants is lower than the gross wage of other public-sector employees with a comparable education. Civil servants acquire pension claims that are very generous compared to workers in the private sector. 5.3.1 Eligibility: Pathways to Retirement for Civil Servants There are three pathways for civil servants: the standard, the early, and the disability retirement option. The standard retirement age is sixty-five. Before 1 July 1997 the early retirement age for civil servants was sixty-two 11. Not even at zero is it actuarially fair. 12. Civil servants are also exempt from unemployment-insurance contributions since civil servants have a lifetime job guarantee. The government pays a certain fraction of health expenses of the civil servant and their dependents (ranging from 50 to 80 percent). The rest has to be covered by private insurance.

296

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

and thus one year less than the early retirement age in the social security system. In 1997, early retirement age was raised to sixty-three. Discount factors for early retirement are phasing in linearly between the years 1998 and 2003 and will reach 0.3 percentage points per month of early retirement, the same as in the private sector and substantially smaller than actuarially fair. Since our sample covers the years 1984 to 1997, these changes of rules do not play a role in our analysis.13 Filing for disability is a third pathway to retirement for civil servants. In the case of disability, a civil servant receives a pension that is based on their previous salary. The replacement rate depends on the number of service years reached before disability retirement and the number of service years that could potentially have been accumulated up to age sixty. For those who did not reach the maximum replacement rate before disability, one additional year of service raises the replacement rate by only 1/3 percentage point per year. 5.3.2 Computation of Pensions The standard pension benefit for civil servants is the product of three elements: (a) the last gross earnings level; (b) the replacement rate as function of service years, and (c) the new adjustment factors to early retirement. As described previously, this third component does not affect our sample persons. There are three crucial differences between civil servants’ pensions and private-sector benefits. First, the benefit base is gross income, rather than net income. In turn, civil servants’ pensions are taxed like any other income. Finally, the benefit base is the last salary rather than the lifetime average. In the following, we concentrate on describing how the system worked for the sample period 1984–1997. Benefits are anchored to the earnings in the last position and then updated annually by the growth rate of the net earnings of active civil servants. If the last position was reached within the two years preceding retirement, the pension is based on the previous lower position. Due to the difference in the benefit base, gross pensions of civil servants are approximately 25 percent higher (other things being equal) than in the private sector. The maximum replacement rate is 75 percent of gross earnings which is considerably higher than the official replacement rate of the private-sector system, which is around 70 percent of net earnings. The replacement rate depends on the years of service. High school and college education, military service, and other work in the public sector are also counted as service years. For retirement after June 1997, the college education credit is limited to three years. 13. Very specific rules apply to some civil servants. For example, the regular retirement age for police officers is age sixty; for soldiers it is even lower and depends on their rank.

Micro-Modeling of Retirement Decisions in Germany

297

Before 1992, the replacement rate was a nonlinear function of service years. The replacement rate started at a value of 35 percent for all civil servants with at least five years of service. For each additional year of service between the tenth and the twenty-fifth year, the increment was 2 percentage points. From the twenty-fifth to the thirty-fifth year the annual increment was 1 percent. Thus, the maximum replacement rate of 75 percent was reached with thirty-five service years under the old rule. This is much more generous than the private-sector replacement rate of 70 percent that requires forty-five years of service. For persons retiring after 1 January 1992, the replacement rate grows by 1.88 percentage points for each year of service. Thus, the maximum value is reached after forty years of service. However, there are transitional modifications to that simple rule. First, civil servants who reach the standard retirement age (usually age sixty-five) before 1 January 2002 are not affected at all. Second, for younger civil servants, all claims that have been acquired before 1992 are conserved. These persons gain 1 additional percentage point per year from 1992 onward. All persons who have acquired twenty-five service years before 1992 have reached 65 percentage points and also would have gained only 1 additional point per year under the old rule. Only persons with less than twenty-five service years in 1991 can be made worse off by the reform. The new proportional rule only applies if it generates a higher replacement rate than the transitional rule. Our calculations of pension wealth use these institutional changes, but only a few special cases are affected. The generosity of gross pensions received by civil servants vis-á-vis the private-sector workers is only partially offset by the preferential tax treatment of private-sector pensions. Since civil servants’ pensions are taxed according to the German comprehensive income taxation, the net replacement rates of civil service pension recipients depends on their position in the highly progressive tax schedule. In general, the net replacement rate, with respect to the preretirement net earnings, is higher than 75 percent and thus considerably more generous than in the private sector. 5.3.3 Incentives to Retire In our sample, most civil servants have reached the maximum replacement rate by the age of fifty-four. Persons who have started to work in the public sector before the age of twenty-three have reached a replacement rate of 75 percent, when taking into account the disability rules. This also holds for civil servants, who—like professors—receive lifetime tenure late in their life cycle. For those groups, the starting age is usually set at twenty-one. Additional years of service beyond the age of fifty-four increase pensions only if the civil servant is promoted to a position with a higher salary. Retirement incentives therefore strongly depend on promotion expectations. For persons who cannot expect to be promoted after age fifty-four, the

298

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

pension accrual is zero or very small. For those who have already reached the replacement rate of 75 percent, the accrual of the present discounted pension wealth is negative. Since the replacement rate is 75 percent of the gross earnings in the last position before retirement, the negative accrual of postponing retirement by one year is simply 75 percent of the last gross earnings. This is equivalent to a 75 percent tax on earnings. For persons who expect to climb another step in the hierarchy, the gross wage increase is, on average, 10.5 percent. This raises the pension by approximately 10 percent. In order to cash in the higher pension, the civil servant has to defer retirement by at least one year.14 In this extreme case the social security wealth increases 10 percent through the effect of higher pensions and decreases by 5 percent through the effect of pension deferral. In this extreme case, the pension accrual is positive. If the civil servant has to wait several years for the next promotion (or for the promotion to have an effect on pension claims), the accrual of working becomes negative. The dependency on promotion expectations makes modeling the incentive effects for civil servants very hard, since the researcher needs information on the career prospects of the respondent. We do not have such information in our data and must therefore ignore the effect of potential promotions. 5.3.4 Retirement Behavior The retirement behavior of civil servants reflects the very generous disability and early retirement rules. The average retirement age for civil servants in the year 1993 was age 58.9 and thus about one year lower than in the private sector (see section 5.2.7). Disability is the most important pathway to retirement for civil servants—40 percent of those who retired in the year 1993 used disability retirement. Almost one-third used the early retirement option at the age of sixty-two. Only about 20 percent of civil servants retired at the regular retirement age of sixty-five. 5.4 Data and Variable Specification Our main data source is the German Socio-Economic Panel (GSOEP), described subsequently. The remaining subsections are devoted to the variable construction, notably the definition of retirement status, which acts as our dependent variable, and the incentive variables, which act as our main explanatory variables. Aggregate information is provided by the German retirement insurance organization (Verband deutscher Versicherungträger or VDR), which publishes annual statistics on average earnings, system entries, retirement age, and the like (Rentenversicherung in Zeitreihen), and by 14. For the higher earnings to take effect on pensions, it is usually required to work several years after the promotion.

Micro-Modeling of Retirement Decisions in Germany

299

the Labor Ministry (Bundesministerium für Arbeit und Sozialordnung; BMA 1999). 5.4.1 The German Socio-Economic Panel The GSOEP is an annual panel study of some 6,000 households and some 15,000 individuals. The data are gathered by the German Institute for Economic Research (DIW). The GSOEP is a panel survey of private households. Its design closely corresponds to the U.S. Panel Study of Income Dynamics (PSID).15 The GSOEP includes carefully designed household weights that match the data with the German Mikrozensus. The panel started in 1984; we use fourteen annual waves through 1997. In 1997, the GSOEP had four subsamples: (a) West German citizens (9,000 persons in 1984); (b) Foreign workers from Spain, Italy, Greece, Turkey, and former Yugoslavia residing in West Germany (3,000 persons in 1984, oversampled); (c) East German citizens (4,000 persons sampled from 1991 onward); and (d) Germans who have remigrated (mainly from Romania and the former Soviet Union; 1,000 persons sampled in 1995). We draw our working sample from samples (a) and (b) since the labor supply patterns of East Germans and remigrants are substantially different from residents in West Germany so that pooling these samples is not warranted.16 We constructed a equal-sided unbalanced panel of all persons aged fiftyfive through seventy from subsamples (a) and (b) for whom earnings data is available.17 This panel includes 2,223 individuals with 14,401 observations. Average observation time is six-and-one-half years. The panel is left censored, as we include only persons who have worked at least one year during our time window in order to reconstruct an earning history. There is only a little right censoring due to missing interviews. Specifically, foreign workers often leave Germany after retirement. However, since this affects only a few cases, we did not model this censoring. The sample contains private-sector workers, civil servants and other public-sector workers, and the self-employed. The GSOEP data provide a detailed account of income and employment status. Since the GSOEP performs personal interviews with each member aged seventeen and older in the household, we have the same information on husbands and spouses. The personal information includes labor market status, gross and net income, hours worked, education, and marital status 15. Burkhauser (1991) provides an English-language description, code books, and links to an internationally accessible GSOEP version. Börsch-Supan (2000b) discusses the merits and limits of the GSOEP data for studies of retirement behavior. 16. Schmähl (1991) provides a narrative of the transition. 17. We excluded East Germany because its retirement patterns are dominated by the transition problems to a market economy. See Börsch-Supan and Schmidt (1996) for a comparison.

300 Table 5.3

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni Descriptive Statistics of Main Variables

Variable Age Health Married College Skilled Homeowner No wealth Financial assets Experience Former self-employed Former civil servant Children in household

Valid Observations

Mean

Standard Deviation

Minimum

Maximum

14,401 14,401 14,401 14,401 14,401 14,398 14,312 14,401 14,401 14,401 14,359 14,401

59.77 8.09 86% 11% 86% 52% 11% 22% 450.29 9% 8% 33%

4.88 3.05 34% 31% 58% 50% 31% 42% 96.01 29% 27% 47%

53 0 0 0 0 0 0 0 0 0 0 0

70 10 1 1 2 1 1 1 646 1 1 1

Source: GSOEP, working sample of males, 1984–97 (available at http://www.diw-berlin.de/ gsoep).

but only a subjective indicator of health (plus disability status, and number of doctor and hospital visits). The GSOEP also has a very detailed labor market calendar that provides monthly information on the labor market status (full-time, part-time, retired, unemployed, and education) and its corresponding income for each sample person. This detailed information during the sample period is augmented by a retrospective history of labor force participation that starts with age fifteen. It carries the annual labor market status (full-time, part-time, unemployed, out-of-labor force, and so forth) but has no retrospect earnings information. Table 5.3 presents the descriptive statistics of the most common socioeconomic variables in our working sample. 5.4.2 Construction of Earnings Histories Since the benefit formula for private-sector pensions depends on earnings points computed from relative-income positions, and since civil servants’ pensions depend on the last salary, we do not need a complete earnings history of our sample persons. Information on the earnings position in each year relative to the aggregate average of that year is sufficient. We have this information for the sample period but not for earlier years. We therefore estimate the average relative earnings position (EP) using all nonretired, full- or part-time workers in the sample who have a positive wage. We fit a fixed-effects model for EP. The fixed effects absorb the constant covariates (e.g., education, marital status, and race). All aggregate yearspecific covariates drop out since we estimate the relative earnings position. This procedure makes the most efficient use of our earnings data. In the forward projection, we need a forecast of the absolute earnings

Micro-Modeling of Retirement Decisions in Germany

301

level. In retrospect, we recover this by multiplying EP, the average relative earnings position, with the aggregate level of earnings, which we take from the VDR statistics. For future years, we assume a 1 percent real wage growth, corresponding to the average over the last twenty-five years. 5.4.3 Definition of Retirement Status The definition of the retirement status is problematic. Retirement definitions commonly employed in the literature include the retirement status self-reported by the respondent. Few work hours or the receipt of retirement benefits are among other definitions. In many countries (e.g., the United States; see Rust 1990), these definitions do not coincide for a large fraction of old-age workers. The problem is somewhat less severe in Germany, although there are some differences such as the more distinct spikes at the legal ages (described in table 5.1), as can be seen from figure 5.3. The persons in our sample appear to have a very general notion of retirement since, when asked about their labor market status, they consider the receipt of benefits from preretirement schemes as well as from the formal retirement programs as retirement. It seems as if they consider different programs as close substitutes. For instance, persons who receive severance pay from their former employers plus unemployment compensation generally claim to be retired. Moreover, our sample persons rarely report significant hours worked after the receipt of pension benefits. Our first measure of retirement (definition I) is thus the self-assessment as retired, and our results presented below are based on this definition. One additional reason for treating this as retirement is the fact that, after giving up the career job, there is no choice left. For instance, persons in preretirement schemes are automatically shifted from unemployment benefits to old-age pensions of type E (see table 5.1) at age sixty. We also tried out other definitions.18 For instance, we know whether or not persons received formal pension benefits. A definition based on this excludes some forms of early retirement (definition II). We then add persons to definition I who receive formal pensions but do not consider themselves as being retired (e.g., many of the self-employed). This definition III (the joint set of I and II) is the broadest definition. 5.4.4 Handling of Multiple Retirement Programs At least theoretically, a worker at age fifty-five has the choice between three retirement programs:

• Old-age pensions starting at age sixty, • Disability pensions, and • Preretirement schemes. 18. Using one of the other measures does not change the qualitative results. We find that the first measure of retirement works best.

Retirement status by alternative definitions

Source: GSOEP, working sample of males, 1984–1997 (available at http://www.diw-berlin.de/gsoep).

Fig. 5.3

Micro-Modeling of Retirement Decisions in Germany

303

The set of choices is actually larger because some of these programs have several branch programs (e.g., within old-age pensions there are unemployment, long-service life, and so on), as depicted in figure 5.2. We refer to these choices as pathways, as we have done in figure 5.2. It is important to notice that all of these pathways pay the same benefit once a person is eligible.19 In practice, there is no free choice since most of these pathways are subject to eligibility criteria. Among those, we distinguish between “strict eligibility rules” that are tied to objective variables, such as age, gender, and previous contribution history, and “soft eligibility rules” that are subject to discretionary decisions, notably the determination of a workers’ disability status.20 In the construction of social security wealth and the incentive variables (see later discussion), we need to compute expected pension benefits, which depend on the choice of pathway. We used two methods. The first method considers only strict eligibility, implicitly assuming that every individual who wants to obtain a disability pension will eventually be granted one. Hence, expected benefits at a given age are zero if the person is not eligible to any of the pathways, otherwise the (common) benefit for that given age is assumed. For example, those self-employed who pay voluntary contributions are only eligible for early retirement—namely disability—if they have contributed continuously since 1984 (the date of a major reform of voluntary participation), otherwise they can retire at the age of sixty-three at the earliest.21 In the latter case, the pension will be zero for all retirement ages below sixty-three. The second method weights the benefits by its observed frequency. Let’s suppose, the observed frequency of disability status at age fifty-nine is 33 percent, and the sample person is not eligible for any other pathway at that age. Then expected benefits at age fifty-nine for this person will be a third of the (common) benefit level. Börsch-Supan (2001b) explores the sensitivity of estimation results to these two methods, and provides an instrumental-variable interpretation of the second method. This second method is our method of choice and the only one reported in this paper. 5.4.5 Construction of Social Security Wealth A key statistic to measure the incentives to retire early is the change in the net present value of all future benefits when retirement is postponed. In a slight misuse of terminology, we call the net present value of all future benefits “social security wealth” (SSW) for both private-sector and civil 19. Strictly speaking, preretirement programs can have any benefit level because they are negotiated between workers and employers. In practice, however, the outcome of these negotiations is guided by the public insurance benefits. 20. Disability depends on health as well as labor market characteristics. 21. See Schnabel (1999) for details.

304

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

servants’ pensions. If SSW declines because the increase in the annual pension due to postponement of retirement is not large enough to offset the shorter time of pension receipt, workers have a financial incentive to retire earlier. We define SSW as the expected present discounted value of benefits (YRET) minus applicable contributions that are levied on gross earnings (c
YLAB). Seen from the perspective of a worker who is S years old and plans to retire at age R, SSW is 

R1

tR

tS

SSWS (R)  ∑ YRETt (R)
at
tS  ∑ c
YLABt
at
tS. SSW: net present discounted value of retirement benefits S: planning age R: retirement age YLABt: gross labor income at age t YRETt (R ): net pension income at age t for retirement at age R ct : contribution rate to pension system at age t at : probability to survive at least until age t given survival until age S : discount factor  1/(1  r) We choose the usual discount rate of 3 percent. Conditional survival probabilities are computed from the standard life tables of the German Bureau of the Census (Statistisches Bundesamt), and SSW depends also on the joint survival probabilities of spouses through survivor pensions. We assume independence of survival of spouses to compute the joint probability. We also have to predict future contribution rates and pensions. In order to obtain consistent policy simulations, they are simulated using the macro-economic pension model underlying Börsch-Supan (1995). This internal consistency is important. Assume a policy proposal that reduces the replacement rate by 20 percent. This immediately lowers the contribution rates by 20 percent if the system is PAYG and financed through contributions. The effect on SSW is ambiguous and varies by cohort. Table 5.4 shows the average SSW in our sample and its change for each individual—the accrual of social security when retirement is postponed by one year. Note that the averages in the right-side panel are not the first differences of the average SSW in the left-side panel since the aggregate figures relate to different individuals in our unbalanced panel. 5.4.6 Specification of Incentive Variables We computed five different incentives measures.

• ACCRUAL: the accrual of SSW if retirement is postponed by one year • ACCRUALRATE: the accrual divided by the level of SSW

Micro-Modeling of Retirement Decisions in Germany Table 5.4

Social Security Wealth and Its Accrual SSW

Age 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 Total

305

Accrual

Mean

SD

N

Mean

SD

N

167.209 169.280 166.391 168.211 171.161 170.002 172.426 175.892 182.616 191.603 194.370 194.654 197.866 197.180 196.372 196.469 175.948

82.603 80.628 80.242 80.550 81.672 81.828 80.792 81.324 77.848 75.784 76.811 76.697 76.820 76.052 77.633 78.187 80.691

913 948 930 885 812 736 656 572 507 434 375 325 275 225 182 144 8919

–10.084 –10.494 –8.327 –8.730 –8.991 –9.360 –8.565 –8.929 –9.064 –10.997 –11.643 –12.149 –12.855 –13.374 –13.869 –14.424 –9.908

5.083 5.175 4.356 4.651 4.612 4.713 4.565 4.690 4.780 4.302 4.512 4.652 4.846 5.038 5.330 5.637 4.995

913 948 930 885 812 736 656 572 507 434 375 325 275 225 182 144 8919

Note: All figures in € 1995 (€1 has a purchasing power of about US$1.00) SD = standard deviation; N = number of observations.

• TAXRATE: the accrual divided by the (potential) gross earnings during the year of postponement

• PEAKVAL: the maximum of future SSW over all possible retirement ages minus the SSW for immediate retirement

• OPTVAL: the option value of postponing retirement by one year The pension-wealth accrual function, a function of the retirement age R, is the change in SSW when retirement is postponed from age R – 1 to age R. We have seen this first incentive variable already in table 5.4. We convert this variable into a rate by defining SSWS (R)  SSWS (R  1) ACCRS (R)   , SSWS (R  1) which is displayed in table 5.5. The lack of actuarial fairness of the German public pension system creates a negative accrual of pension wealth between 5 and 8 percent during the early retirement window when retirement is postponed by one year. The average loss in our sample is about DM 10,000 (roughly US$5,000 at purchasing power parity). A negative accrual can be interpreted as a tax on further labor force participation. We therefore compute as an implicit tax rate the ratio of the (negative) SSW accrual to the gross wage (YLAB) that workers would earn if they postponed retirement to age R.

306 Table 5.5

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni Accrual Rates and Implicit Tax Rates Accrual Rate

Age 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 Total

Implicit Tax Rate

Mean (%)

SD (%)

N

Mean (%)

SD (%)

N

–7.7 –8.3 –6.9 –7.7 –7.2 –8.2 –8.0 –7.6 –5.0 –5.8 –6.0 –6.3 –6.5 –6.8 –7.1 –7.4 –7.2

15.1 16.6 14.4 17.3 14.3 18.2 20.3 20.2 1.7 0.4 0.4 0.4 0.5 0.5 0.5 0.5 14.3

885 927 904 862 788 711 637 553 493 432 374 325 275 225 182 144 8717

34.4 35.7 28.6 29.6 30.4 31.9 29.4 30.6 31.8 37.9 40.1 42.2 44.5 47.0 49.6 51.8 34.1

13.5 13.0 11.6 11.8 11.4 11.6 11.4 11.0 11.7 6.8 7.0 6.6 6.7 7.0 7.5 7.1 12.4

886 923 908 866 798 726 648 568 502 431 374 325 275 225 182 143 8780

Note: SD = standard deviation; N = number of observations.

SSWS (R)  SSWS (R  1) TAXRS (R)    YLABR This implicit tax rate can be rewritten as the product of two terms. The first term represents the effect of postponing retirement through mortality, discounting, and the adjustment of benefits to retirement age, while the second term is the net replacement rate. YRETR  ET YLABN R TAXRS (R)  [aR

(  1)  1]
REPLR If benefits are actuarially fair in a financial sense,   1  1/aR
, and the resulting tax rate is zero. This is not the case in Germany (see table 5.5). It shows that the early retirement incentives in Germany have been strong. The tax rate at age sixty was about 30 percent, and increased with the retirement age to exceed 40 percent at age sixty-five. These numbers refer to the pre-1992 legislation applicable to our sample. Today’s implicit tax rates are about half of those in table 5.5 (see Börsch-Supan and Schnabel 1999). These first three measures of one-year accrual only account for the immediate benefit to working an additional year. But an additional year of work also sustains the option of retiring at an even later date. The value of this choice can be important if there are large nonlinearities in the accrual

Micro-Modeling of Retirement Decisions in Germany Table 5.6

Peak Value and Option Value Peak Value

Age 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 Total

307

Option Value

Mean

SD

N

Mean

SD

N

–10.084 –10.493 –8.326 –8.728 –8.986 –9.320 –8.531 –8.904 –9.064 –10.997 –11.643 –12.149 –12.855 –13.374 –13.869 –14.424 –9.900

5.085 5.176 4.357 4.652 4.622 4.764 4.612 4.717 4.780 4.302 4.512 4.652 4.846 5.038 5.330 5.637 5.007

913 948 930 885 812 736 656 572 507 434 375 325 275 225 182 144 8919

66.414 61.179 59.015 55.770 49.709 42.628 33.888 25.936 15.646 5.900 2.806 0.428 –1.341 –3.188 –4.984 –6.642 38.167

111.233 101.163 93.312 118.974 115.357 98.487 51.254 44.513 33.281 16.255 12.175 7.975 6.873 5.724 4.855 4.029 87.540

913 948 930 885 812 736 656 572 507 434 375 325 275 225 182 144 8919

Note: See table 5.4.

profile. For example, if there is a small negative accrual at age fifty-nine, but a large positive accrual at age sixty, it would be misleading to say that the system induces retirement at age fifty-nine; the disincentive to work at that age is dominated by incentives to work at age sixty. One way of capturing this possibility is to use the “peak value” calculation suggested by Coile and Gruber (1999). Rather than taking the difference between SSW today and next year, peak value takes the difference between SSW today and in the year in which the expected value of SSW is maximized: PEAKVALS (R)  SSWS (R)  max [SSWS(T )]. T R

This measure therefore captures the tradeoff between retiring today and working until a year with a much higher SSW. In years beyond the year in which SSW peaks, this calculation collapses to the simple one-year accrual variable. In fact, PEAKVAL turns out to be virtually identical to ACCRUAL since pension accrual is negative in most cases for the whole sequence of retirement ages (see the averages in table 5.6). All these measures include the financial aspects of the retirement decision only. Alternatively, one might consider the consumption utility of net earnings and pension benefits and also account for the utility aspects of the labor-for-leisure trade-off. To this end, we employ as the fifth and final incentive variable the option value to postpone retirement (Stock and Wise

308

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

Fig. 5.4

Grid search for three estimation variants

Source: Authors’ estimates based on GSOEP panel of males (available at http://www. diw-berlin.de/gsoep). See text for explanation of legend.

1990). This value expresses for each retirement age the trade-off between retiring now (resulting in a stream of utility that depends on this retirement age) and keeping all options open for some later retirement date (with associated streams of utility for all possible later retirement ages). Let Vt(R) denote the expected discounted future utility at age t if the worker retires at age R, specified as follows: R1



st

sR

Vt (R)  ∑ u(YLABNET )
as
 st  ∑ u(YRETs (R))
as
 st. s YLAB  after-tax labor income at age s, s  t . . . R – 1 YRETs (R)  pension income at age s, s R R  retirement age

 marginal utility of leisure (to be estimated) a  probability to survive at least until age s   discount factor  1/(1 r) NET s

Utility from consumption is represented by an isoelastic utility function in after-tax income, u(Y )  Y . (Remember that pension income in Germany is effectively untaxed.) To capture utility from leisure, utility during retirement is weighted by 1, where 1/ is the marginal disutility of work. We employed a grid search for the parameter , applied to three specifications (see figure 5.4). The parameter gets smaller as more covariates are used: It is larger than 4 if only a second-order age polynomial is included (plus option value), 2.8 if initial SSW is added, and 2.5 if a large set of regressors including a full set of age dummies is added (see table 5.7).

Micro-Modeling of Retirement Decisions in Germany Table 5.7

309

Definitions of Other Explanatory Variables

Age Married Health College Medium skilled Homeowner No assets Financial assets Exp Yhat Former self-employed Former civil servant Kids

Age of person Marital status: 1 = married, 0 = not married 0 = poor, . . . , 10 = excellent 1 = college degree, 0 = else 1 = medium skilled (only vocational training or high school) 1 = homeowner 1 = no wealth 1 = owner of financial assets Work experience Estimated labor income Self-employed before retirement Civil servant before retirement Children in household

The option value for a specific age is defined as the difference between the maximum attainable consumption utility of the worker postpones retirement to some later year minus the utility of consumption that the worker can afford if the worker would retire now. Let R ∗(t) denote the optimal retirement age if the worker postpones retirement past age t, that is, max(Vt (r)) for r t. With this notation, the option value is G(t)  Vt (R ∗(t))  Vt (t). Since a worker is likely to retire as soon as the utility of the option to postpone retirement becomes smaller than the utility of retiring now, retirement probabilities should depend negatively on the option value. The option value captures the economic incentives created by the pension system and the labor market because the retirement income YRETs(R) depends on retirement age according to the adjustment factors and on previous labor income by the benefit rules summarized in sections 5.2 and 5.3. The option value is also closely linked to the pension accrual. This is most easily seen in a simple two-period comparison and for  equals 1. In this crude approximation, a worker of age R in the first period will retire early if


W(R) YLABNET 
W(R  1), where W(t) denotes the present discounted value of pension benefits when retiring at age t. Using the definition of TAXR(R), it follows that a worker will retire in the first period if TAXR(R) 1/ . Hence, according to this crude approximation, the tax rates well above 50 percent exerted by the current public pension system in Germany will lead to early retirement. We compute the option value for every person in our sample, using the applicable pension regulations and the imputed earning histories. The parameters chosen are a discount rate  of 3 percent, a curvature parameter

310

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

 of 1.0, and a relative utility parameter of 2.8. Additional private pension income is ignored because it represents only a very small proportion of retirement income as described before. Table 5.6 shows the sample averages. 5.5 Regression Results The variable to be explained is old-age labor force status. Because Germany has very few part-time employees, we model only two states—fully in the labor force and fully retired—unlike the competing-risk analysis of Sueyoshi (1989). We use definition I for retired, based on the self-assessed labor force status (see section 5.4.3). In each of the following regressions, our main explanatory variable is one of the four incentive variables described in the previous section: accrual rate, implicit tax rate, peak value (which is essentially identical to the accrual of SSW), and option value. The other explanatory variables are the usual suspects: an array of socioeconomic variables, such as gender, marital status, wealth (indicator variables of several financial and real-wealth categories), and a self-assessed health measure ranging from 0 for poor to 10 for excellent health. We do not use the legal disability status as a measure of health since this is endogenous to the retirement decision. The desire for early retirement may prompt workers to seek disability status, and frequently the employer helps in this process to alleviate restructuring. Until recently, disability status was granted for labor market reasons without a link to health. We link the explanatory variables to the dependent variable by a binary probit model. This does some injustice to the panel nature of our data and probably underestimates the true effect (see Börsch-Supan 2001b, who experiments with several specifications of panel probit models with parametrized correlation patterns over time). This more complicated models can be interpreted as partly nonparametric hazard models for multiple spell data, permitting unobserved heterogeneity and state dependence without imposing a functional form on the duration in a given state, while the simple probit model ignores these temporal effects.22 We conducted several randomeffect estimates that correct for some of the intertemporal correlations. The effects of the incentive variables were slightly strengthened, but the results did not change significantly. Note that our estimation sample includes repeated observations of the same person only while this person is employed. Once the person retires, we assume that this is an absorbing state and include only the first observation in retirement. Hence, our dependent variable is in fact the probability to retire, given that the sample person has 22. Flexible-hazard-rate models of retirement have been estimated by Sueyoshi (1989) and Meghir and Whitehouse (1997), and parametric-hazard-rate models for German data have been estimated by Schmidt (1995) and Börsch-Supan and Schmidt (1996).

Micro-Modeling of Retirement Decisions in Germany

311

worked during the year before, pt  Prob(retired in t | worked in t – 1). We then compute the survivor function S(t) conditional on working until the beginning of our window period (age fifty-three) as the product of (1 – pt ) from age fifty-three to t. The probability of choosing a retirement age a is then pa
S(a) and the expected retirement age is Σpa
S(a)
a. Inserting the option value in this type of a regression model is a practical estimation procedure that can be interpreted as a flexible discrete-time duration model explaining the timing of retirement entry. It ignores, however, the structure of the dynamic optimization that underlies the workers decision regarding when to retire.23 Nevertheless, previous experimentation has shown that this pragmatic approach generates robust estimates of the average effects of the incentive variables on retirement, although it is likely to fail the individual variation as precisely as the true dynamic optimization model.24 Identification of the incentive variables is possible only if we have meaningful variation in these variables. Sources of variation are

• The level of SSW reached at the earliest retirement age, mainly generated by variation in labor force histories;

• The upper threshold for the social security contributions, mainly generated by their changes over time and the different earnings levels; Differences in the pension rules between single workers and couples; Widely varying age differences between husband and spouse; Restricted eligibility of self-employed; Restricted eligibility of women with less than fifteen years of service; Differences in the pension formula between private-sector employees and civil servants; • Differences in the ratio between contribution rates and pension benefits across cohorts (younger cohorts have a substantially lower internal rate of return); and • Several minor rule changes during our the sample period.

• • • • •

We estimated twenty-four different models: We use four different incentive variables as our main regressors (accrual rate, tax rate, option value, and peak value; see section 5.4.6). For each of these incentive variables, we run probit regressions with three age specifications (linear, quadratic, and a full set of age dummies) as well as with and without including SSW. We pool public and private workers, but have separate regressions for males and females. We first summarize our main results separately for men and women. Tables 5.8 and 5.10 report the goodness of fit, and tables 5.9 and 5.11 the impact of the incentive variables, measured as the change in the probability of 23. The full underlying dynamic programming model has been estimated by Rust and Phelan (1997). 24. See Lumsdaine, Stock, and Wise (1992).

Table 5.8

Goodness of Fit (males)

Without SSW With SSW

Without SSW With SSW

Accrual Rate

Tax Rate

Age

Age

Linear

Quadratic

Dummies

Linear

Quadratic

Dummies

–1530.1362 18.2 –1528.6764 18.3

–1524.0716 18.6 –1522.7746 18.6

–1492.562 20.2 –1491.2561 20.3

–1532.5654 18.1 –1532.5653 18.1

–1529.0165 18.3 –1529.0016 18.3

–1481.4323 20.8 –1481.3642 20.8

Option Value

Peak Value

Age

Age

Linear

Quadratic

Dummies

Linear

Quadratic

Dummies

–1527.4082 18.4 –1522.938 18.6

–1525.2906 18.5 –1521.7785 18.7

–1490.2695 20.3 –1485.3395 20.6

–1536.2424 17.9 –1536.2335 17.9

–1533.1139 18.1 –1533.0664 18.1

–1488.696 20.4 –1488.6922 20.4

Source: GSOEP, working sample of men, 1984–1997 (available at http://www.diw-berlin.de/gsoep). Note: Percentages in italics are log-likelihood values and pseudo R 2.

Table 5.9

Without SSW With SSW

Without SSW With SSW

Marginal Effect of Incentive Variables (males) Accrual Rate

Tax Rate

Age

Age

Linear

Quadratic

Dummies

Linear

Quadratic

Dummies

0.0001396 3.5 0.0001509 3.9

0.0001381 3.6 0.0001481 3.9

0.0001356 3.6 0.0001460 3.9

–0.08456341 –4.7 –0.08454737 –4.6

–0.07667369 –4.4 –0.07632455 –4.3

–0.17190381 –7.2 –0.17314054 –7.1

Option Value

Peak Value

Age

Age

Linear

Quadratic

Dummies

Linear

Quadratic

Dummies

–0.00023237 –5.5 –0.00030332 –6.1

–0.00020934 –5.0 –0.00027806 –5.6

–0.00024276 –5.5 –0.0003286 –6.2

–0.0012644 –3.8 –0.00126031 –3.7

–0.00107072 –3.2 –0.00105993 –3.2

–0.00292954 –5.7 –0.00293449 –5.7

Source: GSOEP, working sample of men, 1984–1997 (available at http://www.diw-berlin.de/gsoep). Note: Numbers in italics are ∂P/∂x and t-statistics.

Table 5.10

Without SSW With SSW

Without SSW With SSW

Goodness of Fit (females) Accrual Rate

Tax Rate

Age

Age

Linear

Quadratic

Dummies

Linear

Quadratic

Dummies

–691.58206 18.0 –691.28715 18.1

–684.4851 18.9 –684.2648 18.9

–632.7585 24.9 –632.60559 25.0

–687.97887 18.4 –687.95991 18.4

–681.87265 19.2 –681.86189 19.2

–631.03328 25.1 –631.01395 25.1

Option Value

Peak Value

Age

Age

Linear

Quadratic

Dummies

Linear

Quadratic

Dummies

–685.97528 18.7 –684.17131 18.9

–679.93327 19.4 –678.30671 19.6

–627.72675 25.5 –625.19017 25.8

–678.58724 19.6 –677.54142 19.7

–672.18414 20.3 –671.08436 20.4

–623.80517 26.0 –621.56825 26.3

Source: GSOEP, working sample of men, 1984–1997 (available at http://www.diw-berlin.de/gsoep). Note: Percentages in italics are log-likelihood values and pseudo R 2.

Table 5.11

Without SSW With SSW

Without SSW With SSW

Marginal Effect of Incentive Variables (females) Accrual Rate

Tax Rate

Age

Age

Linear

Quadratic

Dummies

Linear

Quadratic

Dummies

0.0000087 0.7 0.0000074 0.6

0.0000115 0.8 0.0000102 0.7

0.0000046 0.3 0.0000037 0.2

–0.01152618 –2.8 –0.01180793 –2.7

–0.01250953 –2.4 –0.01278009 –2.3

–0.01550037 –1.9 –0.01608508 –1.8

Option Value

Peak Value

Age

Age

Linear

Quadratic

Dummies

Linear

Quadratic

Dummies

–0.00005129 –3.4 –0.00005499 –3.7

–0.00006159 –3.1 –0.00006957 –3.4

–0.00010106 –3.1 –0.00013015 –3.7

–0.00133364 –5.1 –0.00155703 –5.2

–0.00159996 –5.0 –0.00189241 –5.1

–0.00270272 –4.2 –0.00384073 –4.6

Source: GSOEP, working sample of men, 1984–1997 (available at http://www.diw-berlin.de/gsoep). Note: Numbers in italics are ∂P/∂x and t-statistics.

314

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

Table 5.12

OPTVAL FSSW AGEDUM55* AGEDUM56* AGEDUM57* AGEDUM58* AGEDUM59* AGEDUM60* AGEDUM61* AGEDUM62* AGEDUM63* AGEDUM64* AGEDUM65* AGEDUM66* AGEDUM67* AGEDUM68* AGEDUM71* HEALTH MARRIED* UNI* SKILL VEIGEN* VNULL* VWP* VPDAUER VPDAUER2 YHAT YHAT2 ds* db* KIDS*

Probit Estimates for Male Subsample—Incentive Variable OV dF /dx

SE

z

P z

x-bar

–.003053 –.0001714 –.0035989 .0171014 .0283252 .03864 .0422839 .1134217 .1363814 .112833 .1994108 .4444502 .3689588 .7122481 .4784933 .3776805 .6680958 –.0101493 –.0076327 –.027545 .0029414 .0062658 –.0050004 .0156084 .0001696 –3.57e-06 .0012985 2.80e-06 –.022587 .0464132 –.0075237

.0000485 .0000361 .0146005 .0172827 .0191398 .0213393 .0230766 .0357241 .0140872 .0408826 .0552118 .0744151 .0861118 .084577 .2128542 .2945118 .3101597 .0009479 .0119856 .007122 .0058029 .0050534 .0078429 .0067196 .001485 .0000249 .0003284 5.51e-07 .0057826 .0141686 .0048206

–6.07 –4.67 –0.24 1.09 1.72 2.20 2.26 4.60 4.97 4.05 5.69 8.93 6.71 8.69 3.34 2.03 2.53 –11.65 –0.68 –2.96 0.51 1.24 –0.61 2.51 0.11 –0.14 3.92 4.95 –3.13 4.14 –1.55

0.000 0.000 0.810 0.275 0.086 0.028 0.024 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.042 0.012 0.000 0.498 0.003 0.612 0.216 0.541 0.012 0.909 0.886 0.000 0.000 0.002 0.000 0.122

249.725 316.764 .137001 .133765 .123902 .111111 .096933 .081368 .062567 .045307 .033595 .022346 .01125 .006164 .001079 .00616 .000308 9.07096 .95161 .128833 .869934 .53059 .087225 .229619 38.3005 1521.28 59.9149 4499.21 .107104 .093389 .419171

95% C.I. –.0004 –.000242 –.032215 –.016772 –.009188 –.003184 –.002945 .043404 .055852 .032704 .091198 .298599 .200183 .540834 .061307 –.199552 .060194 –.012007 –.031124 –.041504 –.008432 –.003639 –.020372 .002438 –.002741 –.000052 .000655 1.7e-06 –.033921 .018643 –.016972

–.00021 –.000101 .025018 .050975 .065838 .080464 .087513 .18344 .216911 .192961 .307624 .590301 .537735 .883662 .89568 .954913 1.276 –.008292 .015859 –.013586 .014315 .01617 .010371 .028779 .00308 .000045 .001942 3.9e-06 –.011253 .074183 .001925

Obs. P .0826013 Pred. P .0412237 (at x-bar) No. of obs. = 6,489 LR 2 (31) = 870.62 Prob 2 = 0.0000 Pseudo R 2 = 0.2353 Log-likelihood = –1414.5558 Source: GSOEP, working sample of men, 1984–1997 (available at http://www.diw-berlin.de/gsoep). Note: An asterisk denotes dummy variables.

being retired when the incentive variable is changed infinitesimally. Tables 5.12 and 5.13 show full regression results for our favorite specification (option value with SSW and with a full set of age dummies). The other specifications produce very similar results in terms of significance and signs. Using tax rate and peak value yield significantly better fits than accrual rate and option value in almost all specifications. There is little difference

Micro-Modeling of Retirement Decisions in Germany Table 5.13

OPTVAL FSSW AGEDUM55* AGEDUM56* AGEDUM57* AGEDUM58* AGEDUM59* AGEDUM60* AGEDUM61* AGEDUM62* AGEDUM63* AGEDUM64* AGEDUM65* AGEDUM66* AGEDUM67* AGEDUM69* AGEDUM70* AGEDUM71* HEALTH MARRIED* UNI* SKILL VEIGEN* VNULL* VWP* VPDAUER VPDAUER2 YHAT YHAT2 ds* db* KIDS*

315

Probit Estimates for Female Subsample—Incentive Variable OV dF /dx

SE

z

P z

x-bar

–.0001008 –.0000474 –.000499 .0009751 –.0006681 .0002302 .0042321 .0606805 .1375386 .0385825 .0209055 .0712695 .2828068 .4289469 .0305762 –.0046263 .010992 .017278 –.0018635 –.028244 –.0063273 –.005533 –.0005057 –.0016109 .0013406 .0001974 –3.72e-06 .0009701 –2.41e-06 –.006747 .0410058 .0007423

.0000657 .0000351 .0044412 .0049257 .0043968 .0047638 .0073449 .0448888 .0839864 .0360184 .0247693 .0611771 .1450685 .2050834 .0538137 .0085027 .0402127 .0628126 .001441 .0195633 .0056794 .004353 .0019191 .0028384 .0026261 .0003477 6.74e-06 .0005632 6.26e-07 .0053123 .0534307 .0021086

–3.54 –2.13 –0.11 0.21 –0.15 0.05 0.78 4.97 6.78 2.92 1.79 3.48 6.64 5.55 1.08 –0.40 0.40 0.44 –4.52 –4.46 –1.53 –2.82 –0.27 –0.58 0.56 0.67 –0.63 3.97 –2.02 –2.66 1.63 0.36

0.000 0.033 0.912 0.835 0.882 0.961 0.434 0.000 0.000 0.003 0.074 0.001 0.000 0.000 0.279 0.691 0.690 0.662 0.000 0.000 0.126 0.005 0.786 0.561 0.575 0.504 0.529 0.000 0.043 0.008 0.103 0.717

157.169 118.39 .152008 .137986 .12747 .114404 .100064 .079987 .051307 .028043 .02167 .015296 .011791 .004461 .002231 .001912 .000956 .000637 9.42065 .950605 .035692 .476099 .515934 .108987 .202996 20.1118 558.3 28.7974 6598 .089229 .007967 .345124

95% C.I. –.000229 –.000116 –.009204 –.008679 –.009286 –.009107 –.010164 –.0273 –.027072 –.032012 –.027641 –.048636 –.001522 .026991 –.074897 –.021291 –.067823 –.105832 –.004688 –.066587 –.017459 –.014065 –.004267 –.007174 –.003806 –.000484 –.000017 –.000134 –3.6e-06 –.016887 –.063716 –.00339

.000028 .000021 .008206 .010629 .00795 .009567 .018628 .148661 .302149 .109177 .069241 .191174 .567136 .830903 .136049 .012039 .089808 .140388 .000961 .010099 .004804 .002999 .003256 .003952 .006488 .000879 9.5e-06 .002074 –1.2e-06 .003937 .145728 .004875

Obs. P .0761632 Pred. P .0080778 (at x-bar) No. of obs. = 3,138 LR 2 (31) = 462.90 Prob 2 = 0.0000 Pseudo R 2 = 0.2739 Log-likelihood = –613.60367 Source: GSOEP, working sample of men, 1984–1997 (available at http://www.diw-berlin.de/gsoep). Note: An asterisk denotes dummy variables.

between including SSW or not, although introducing age dummies makes a large difference. Judging from the goodness of fit, the regression with age dummies but without SSW, is our favorite specification. The pseudo R2 is just about 20 percent, a satisfactory but not excellent value. All incentive variables have the correct sign and are highly significant.

316

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

They are very robust across all the different specifications, including inclusion of other covariates, sample selection, and definition of retirement (not shown in table). Including age dummies yields larger marginal effects and better precision, while including SSW has a very small weakening effect. The estimation sample also includes civil servants. We have programmed the incentive variables for civil servants using the pension rules for civil servants, which should lead to stronger incentives for early retirement. However, estimates for a sample of civil servants only are disappointing. The most probable reason is that we do not capture the incentives created by promotion possibilities, the main reason for civil servants to retire later than measured by our incentive variables. Turning to the female sample, results are much weaker than for males. While the overall fit is comparable and sometimes even better, the incentive variables are very weak. Only option value and peak value are significant, with one incorrect sign for peak value in the specification with SSW and linear age. Table 5.12 shows the full regression results for our favorite specification in the full sample. The incentive variable (here, option value) is highly significant as pointed out before. The set of age dummies is also highly significant and elevates the probabilities to retire after ages sixty, sixty-three, and sixtyfive, the earliest retirement ages under the various pathways (see table 5.1). There clearly is an independent effect of age and the incentive variable on retirement. Self-reported health is also highly significant: Healthier workers retire substantially later than those males who report poor health. Married males do not have a different retirement behavior than single males. However, if there is (still) a child in the household retirement is more likely to be deferred. The effect of a college degree on retirement age is very strong and is present although we have an income measure (yhat and yhat2) as an additional control. The wealth variables indicate that there is a wealth effect, also weak and barely significant: Persons with higher wealth (indicated by homeownership and financial assets) can afford an earlier retirement. Also, higher labor income weakens labor force attachment. Note that the higher opportunity costs of retirement have already been accounted for in the option value variable, and hence, this income effect is over and above this plus the wealth effect. Two dummy variables indicate the former labor force status. These variables take the value of 1 if the person is actually or used to be a self-employed or a civil servant. The model indicates that the self-employed tend to work longer, while civil servants retire earlier. Both result confirm our expectations. The peaks of the age dummies are now much more pronounced at age sixty and sixty-five, in accordance with the different rules for women. Most socioeconomic variables have similar (but weaker) effects compared with the male sample. Different, however, is the effect of being married: Married women retire later, probably because they have raised children and there-

Micro-Modeling of Retirement Decisions in Germany

317

fore have an interrupted earnings record such that they are not yet eligible for retirement at age sixty. 5.6 Simulation Results We now apply the estimated coefficients to several simulation experiments. We first simulate reforms that are close to what happened in Germany in the 1992 reform and what the next reform step may strengthen: shifting the retirement age up by making the system more actuarially fair. Second, we simulate several reforms not specific to Germany and unlikely to happen, but which are used to compare the retirement incentive effects across the countries represented in this volume. 5.6.1 Reform Options Specific to Germany The first country-specific experiment shows what is likely to happen when the 1992 reform is fully phased in. The experiment applies the adjustment factors for early retirement that have been introduced by the pension reform 1992 (3.6 percent per year of early retirement) and compares it to the previous situation without any explicit adjustments. The 1992 adjustment factors have been phased in after our sample period and will take full force from the year 2004 onward. They are not actuarial fair, and they are not effective before age sixty because they are overruled by the special earnings-point credits given under disability insurance. The second country-specific experiment goes one step further and introduces a geometric adjustment of 6 percent per year that comes closer to a actuarially fair adjustment. The experiment can therefore be thought of as a preview of a potential pension reform after the 2002 elections. It is applied to all ages in the window period (ages fifty-four to sixty-nine), anchored at the pivotal retirement age of sixty-five. For each policy scenario we use the estimated parameter values in order to compute the probability to retire at age x given that the worker has worked until age fifty-three. We first display the baseline probabilities (i.e., predicted under the pension rules of 1972 valid in our sample period). We then predicted probabilities under the hypothetical new rules (see figures 5.5 and 5.6 based on a specification with age and age squared, rather than linear age or a set of age dummies). The figures show the shift to the right of the distribution, resulting in an increase of the average retirement age. This resulting increase in retirement age is displayed in table 5.14. It amounts to eight months for the 1992 reform, and seventeen months for a system that is almost actuarially fair. Given that the average retirement age is about sixty years in 1999 for German males, the adjustment factor of 6 percent would imply an increase of the retirement age to sixty-one-andone-half years for males. The impact of such a reform on the budget of the PAYG system would be considerable. Given that the average duration of

318

Fig. 5.5

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

Baseline and predicted distribution of retirement ages (1992 reform)

Source: GSOEP, working sample of males, 1984–1997 (available at http://www.diw-berlin.de/ gsoep).

Fig. 5.6

Baseline and predicted distribution of retirement ages (fair system)

Source: GSOEP, working sample of males, 1984–1997 (available at http://www.diw-berlin.de/ gsoep).

pension receipts was sixteen years prior to the reform, expenditure would decrease by roughly 10 percent through this effect. A second effect works through the extended working life, which leads to higher contributions. Two additional years, relative to forty service years, increase the contributions to the PAYG system by 5 percent—provided that deferred take-up of

Micro-Modeling of Retirement Decisions in Germany Table 5.14

319

Effects of Policy Reforms on Expected Retirement Age With SSW

Without SSW

Age

Age

Linear

Quadratic

Dummies

Linear

Quadratic

Dummies

Accrual rate Tax rate Option value Peak value

–0.10 –0.12 1.90 2.31

Simulation A (1992 Reform), Men –0.10 –0.09 0.41 –0.12 –0.11 0.41 1.74 1.55 0.39 2.16 1.86 0.39

0.34 0.34 0.32 0.31

0.72 0.73 0.73 0.73

Accrual rate Tax rate Option value Peak value

Simulation B (6% Geometric Adjustment Factors), Men –0.10 –0.10 –0.09 0.91 0.79 –0.12 –0.12 –0.11 0.91 0.79 2.63 2.45 2.12 1.11 0.93 3.06 2.91 2.39 1.11 0.93

1.57 1.58 1.85 1.85

Accrual rate Tax rate Option value Peak value

–0.02 –0.02 1.01 1.42

Simulation A (1992 Reform), Women –0.02 –0.01 0.12 –0.02 0.00 0.12 0.89 0.77 0.42 1.28 1.18 0.46

0.08 0.08 0.26 0.27

0.08 0.09 0.40 0.52

Accrual rate Tax rate Option value Peak value

Simulation B (6% Geometric Adjustment Factors), Women –0.02 –0.02 –0.01 0.33 0.25 –0.02 –0.02 0.00 0.34 0.26 1.48 1.32 1.14 1.32 1.05 1.98 1.80 1.62 1.41 1.10

0.19 0.20 1.13 1.36

Source: GSOEP, working sample of men, 1984–1997 (available at http://www.diw-berlin.de/ gsoep).

pensions implies additional employment. Moreover, there is a third budgetary effect (compared to the no-reform case) since pension benefits are now lower for all who retire early. This would save another 18 percent. 5.6.2 Simulations for Cross-National Comparisons This second set of simulations serves as a vehicle for an extensive crossnational comparison of the effects that the early retirement incentives exert on retirement behavior. We use two hypothetical reform scenarios (the “three-year-shift reform” and the “common reform,” later explained in more detail) and apply them systematically to several variants of our estimated models of retirement. These variants include the option value and the peak value model, each of which is estimated using a linear and a dummy-variable age specification. In the latter case and in combination with the three-year-shift reform, we introduce yet another two variants: one for keeping the dummy variables at their original ages and one for shifting them along with the shift in the incentive variables. These latter variants are designed to bracket possible behavioral effects that are embedded in the

320

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

age dummies; for a particular example, the habitual effects associated with age sixty-five as a psychological anchor for retirement decisions. The three-year-shift reform increases the ages of early and normal retirement by three years (and the corresponding adjustment factors, if applicable) from the current age in the countries represented by this study. The common reform changes all national systems to a common system with an early retirement age of sixty years, a normal retirement age of sixty-five years, a 60 percent replacement rate at age sixty-five, and a 6 percent per year actuarial adjustment, pivoted at age sixty-five. In the following set of figures, we show all our results both in terms of hazard rates (left-side panels) and the cumulative distribution function (inverse survival function, right-side panels). For convenience, the hazard rates are also tabulated in the appendix. We summarize our results in table 5.15 which displays the expected average retirement ages for all simulations. Figure 5.7 shows the fit of the option versus the peak value model used

Table 5.15

Expected Retirement Age Men

Women

61.77

61.89

Sample Sample frequencies

Base Simulation Option value model Linear age Dummies Peak value model Linear age Dummies

62.01 61.79

62.02 61.89

62.01 61.79

62.02 61.90

Three-Year-Shift Simulation Option value model Linear age 63.52 Dummies fixed 63.55 Dummies shifted 65.52 Peak value model Linear age 62.65 Dummies fixed 62.46 Dummies shifted 65.04 Common Reform Simulation Option value model Linear age 64.31 Dummies fixed 64.17 Dummies changed 63.55 Peak value model Linear age 63.56 Dummies fixed 63.30 Dummies changed 62.46

64.50 64.21 66.23 62.34 62.55 65.16

62.64 62.60 64.21 62.32 62.51 62.55

Note: Expected value is taken over distribution truncated at age 69.

Fig. 5.7

Fit of base case

Fig. 5.7

(cont.)

Micro-Modeling of Retirement Decisions in Germany

323

in these simulations. With a linear age specification, the peak value model slightly underestimates retirement at age sixty-five, but in the other ages the two models are very close. They deliver, of course, identical simulations with a full set of dummy variables for each age since in this case the model is fully saturated. Table 5.15 shows the variation among the four models in terms of expected retirement age and how closely the match the actual retirement age. Figure 5.8 summarizes all simulations. Focusing on the cumulative retirement function, the three-year-shift reform is much more incisive for German men than the common reform, since the common reform keeps the current early retirement age as well as the normal retirement age unchanged. The common reform lowers the replacement rate from currently around 70 percent by 10 percentage points. This actually has very little effects on retirement age. The shift seen in figure 5.8 is mainly an effect of changing the actuarial adjustments and therefore closely corresponds to what we have seen in simulation B in the previous subsection. The effects are more complicated for women. Women have an earlier retirement age (sixty rather than sixty-three), so that the common reform has an even larger effect on women than on men. However, the early retirement age applies only to women with thirty-five years of service. For women with an interrupted work history, the reform has little effect at all because they are restricted to enter retirement at age sixty-five anyway. Figure 5.8 also shows that the option value model generally predicts larger effects than the peak value model. This difference is isolated in figure 5.9. As seen by the hazard rates, the option value model has a more subtle pattern of peaks than the peak value model, which does not really capture the in-between-peak at age sixty-three. Another eye-catching difference in figure 5.8 stems from the impact of shifting the dummies as an approximation of the effect, by which all agespecific behavior is shifted by three years. Figure 5.10 isolated these differences for the three-year-shift reform and shows that this makes a huge difference—by about two years concerning the expected retirement age. While these two years are probably an exaggeration of the long-run impact of a later retirement age, the short-run impact measured by keeping the dummies in place is probably an underestimate. We finish this paper by summarizing that a reform policy of changing the actuarial adjustments, the early retirement ages, or both can indeed shift the retirement age quite substantially. Considering the overall length of retirement (which is currently about seventeen years in Germany), the orders of magnitude—about two years for a set of feasible reform options—is quite significant.

Fig. 5.8

Base case (BC), common reform (CR), and three-year-shift (
3)

Fig. 5.8

(cont.)

Fig. 5.8

(cont.) Base case (BC), common reform (CR), and three-year-shift (
3)

Fig. 5.8

(cont.)

Fig. 5.8

(cont.) Base case (BC), common reform (CR), and three-year-shift (
3)

Fig. 5.8

(cont.)

Fig. 5.9

Option value versus peak value: Simulations

Fig. 5.9

(cont.)

Fig. 5.9

(cont.) Option value versus peak value: Simulations

Fig. 5.9

(cont.)

Fig. 5.9

(cont.) Option value versus peak value: Simulations

Fig. 5.9

(cont.)

Fig. 5.10

Simulations with fixed versus shifted dummies

Fig. 5.10

(cont.)

338

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

Appendix Table 5A.1

Sample Hazard Rates Men

Age 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75

Women

N

Mean

SD

95% Lower Bound

861 889 868 804 721 629 528 406 294 218 145 73 40 7 4 4 3 2 1 2 1 0

0.013 0.015 0.026 0.041 0.060 0.073 0.155 0.185 0.146 0.225 0.428 0.315 0.700 0.429 0.250 0.000 0.000 0.500 0.000 0.000 0.000 •

0.004 0.004 0.005 0.007 0.009 0.010 0.016 0.019 0.021 0.028 0.041 0.055 0.073 0.202 0.250 0.000 0.000 0.500 • 0.000 • •

0.005 0.007 0.016 0.027 0.042 0.053 0.124 0.147 0.106 0.169 0.346 0.206 0.552 –0.066 –0.546 0.000 0.000 –5.853 • 0.000 • •

95% Upper Bound

N

Mean

0.020 0.023 0.037 0.055 0.077 0.094 0.186 0.223 0.187 0.281 0.509 0.424 0.848 0.923 1.046 0.000 0.000 6.853 • 0.000 • •

470 477 433 400 359 314 251 161 88 68 48 37 14 7 5 6 3 2 0 0 0 0

0.017 0.019 0.025 0.025 0.033 0.051 0.219 0.348 0.159 0.103 0.188 0.486 0.643 0.286 0.000 0.167 0.333 0.500 • • • •

SD

95% Lower Bound

95% Upper Bound

0.006 0.006 0.008 0.008 0.009 0.012 0.026 0.038 0.039 0.037 0.057 0.083 0.133 0.184 0.000 0.167 0.333 0.500 • • • •

0.005 0.007 0.011 0.010 0.015 0.026 0.168 0.273 0.081 0.029 0.073 0.318 0.356 –0.166 0.000 –0.262 –1.101 –5.853 • • • •

0.029 0.031 0.040 0.040 0.052 0.075 0.271 0.422 0.237 0.177 0.302 0.655 0.930 0.737 0.000 0.595 1.768 6.853 • • • •

Note: N = number of observations; SD = standard deviation. Dots indicate observations.

Table 5A.2

Base Case: Empirical and Simulated Hazard Rates (men) Option Value Model

Age 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72

Peak Value Model

Sample

Linear

Dummies

Linear

Dummies

0.000 0.013 0.015 0.026 0.041 0.060 0.073 0.155 0.185 0.146 0.225 0.428 0.315 0.700 0.429 0.250 0.000 0.000 0.500 0.000

0.000 0.007 0.013 0.023 0.039 0.064 0.100 0.141 0.182 0.211 0.258 0.289 0.310 0.412 0.424 0.379 0.507 0.472 0.550 0.671

0.000 0.012 0.014 0.027 0.041 0.059 0.076 0.156 0.181 0.145 0.225 0.427 0.316 0.698 0.425 0.239 0.000 0.000 0.500 0.000

0.000 0.009 0.016 0.024 0.039 0.059 0.088 0.136 0.181 0.218 0.285 0.323 0.280 0.356 0.437 0.447 0.537 0.540 0.623 0.717

0.000 0.012 0.014 0.026 0.041 0.059 0.075 0.156 0.182 0.145 0.255 0.426 0.326 0.692 0.426 0.234 0.000 0.000 0.500 0.000

Table 5A.3

Three-Year-Shift Reform: Simulated Hazard Rates (men) Option Value Model

Peak Value Model

Age

Linear

Dummies

Shifted

Linear

Dummies

Shifted

53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72

0.000 0.003 0.005 0.008 0.013 0.023 0.036 0.050 0.077 0.097 0.142 0.179 0.222 0.274 0.265 0.252 0.387 0.364 0.493 0.617

0.000 0.004 0.004 0.009 0.012 0.018 0.022 0.046 0.065 0.052 0.106 0.265 0.211 0.517 0.240 0.131 0.000 0.000 0.432 0.000

0.000 0.004 0.005 0.006 0.006 0.008 0.015 0.017 0.025 0.025 0.065 0.074 0.065 0.112 0.224 0.129 0.482 0.196 0.179 0.000

0.000 0.010 0.016 0.024 0.039 0.059 0.087 0.107 0.146 0.179 0.237 0.284 0.328 0.400 0.351 0.363 0.465 0.455 0.559 0.649

0.000 0.013 0.014 0.028 0.042 0.059 0.073 0.080 0.096 0.072 0.120 0.304 0.475 0.805 0.203 0.092 0.000 0.000 0.316 0.000

0.000 0.013 0.014 0.012 0.014 0.015 0.035 0.034 0.049 0.051 0.088 0.094 0.075 0.117 0.187 0.313 0.674 0.149 0.077 0.000

Table 5A.4

Common Reform: Simulated Hazard Rates (men) Option Value Model

Peak Value Model

Age

Linear

Dummies

Shifted

Linear

Dummies

Shifted

53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72

0.000 0.006 0.010 0.016 0.026 0.041 0.062 0.086 0.137 0.187 0.269 0.341 0.401 0.456 0.398 0.350 0.498 0.539 0.602 0.724

0.000 0.009 0.010 0.018 0.026 0.035 0.042 0.088 0.130 0.123 0.238 0.497 0.426 0.741 0.394 0.209 0.000 0.000 0.564 0.000

0.000 0.004 0.004 0.009 0.012 0.018 0.022 0.046 0.065 0.052 0.106 0.265 0.211 0.517 0.240 0.131 0.000 0.000 0.432 0.000

0.000 0.010 0.017 0.026 0.040 0.060 0.085 0.099 0.139 0.175 0.241 0.293 0.353 0.421 0.404 0.410 0.525 0.626 0.689 0.781

0.000 0.014 0.015 0.034 0.047 0.064 0.071 0.068 0.088 0.069 0.127 0.333 0.524 0.847 0.331 0.147 0.000 0.000 0.690 0.000

0.000 0.013 0.014 0.028 0.042 0.059 0.073 0.080 0.096 0.072 0.120 0.304 0.475 0.805 0.203 0.092 0.000 0.000 0.316 0.000

Table 5A.5

Base Case: Empirical and Simulated Hazard Rates (women) Option Value Model

Age 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72

Peak Value Model

Sample

Linear

Dummies

Linear

Dummies

0.000 0.017 0.019 0.025 0.025 0.033 0.051 0.219 0.348 0.159 0.103 0.188 0.486 0.643 0.286 0.000 0.167 0.333 0.500 0.000

0.000 0.011 0.018 0.028 0.043 0.067 0.095 0.137 0.16 0.200 0.229 0.257 0.324 0.377 0.523 0.634 0.681 0.822 0.905 0.000

0.000 0.016 0.018 0.024 0.024 0.033 0.052 0.221 0.346 0.160 0.105 0.185 0.488 0.644 0.303 0.000 0.180 0.337 0.500 0.000

0.000 0.013 0.021 0.026 0.040 0.059 0.079 0.150 0.179 0.215 0.250 0.280 0.256 0.262 0.561 0.640 0.704 0.873 0.942 0.000

0.000 0.016 0.018 0.024 0.024 0.032 0.051 0.222 0.346 0.162 0.107 0.186 0.486 0.645 0.313 0.000 0.167 0.337 0.500 0.000

Table 5A.6

Three-Year-Shift Reform: Simulated Hazard Rates (women) Option Value Model

Peak Value Model

Age

Linear

Dummies

Shifted

Linear

Dummies

Shifted

53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72

0.000 0.006 0.009 0.015 0.022 0.034 0.049 0.072 0.098 0.135 0.167 0.204 0.283 0.310 0.401 0.488 0.580 0.762 0.870 0.000

0.000 0.007 0.007 0.010 0.009 0.011 0.017 0.099 0.198 0.082 0.05 0.123 0.421 0.544 0.134 0.000 0.071 0.217 0.391 0.000

0.000 0.007 0.008 0.009 0.009 0.010 0.013 0.013 0.015 0.028 0.121 0.221 0.101 0.058 0.220 0.622 0.781 0.312 0.000 0.000

0.000 0.013 0.021 0.028 0.041 0.060 0.081 0.093 0.120 0.155 0.187 0.225 0.272 0.285 0.346 0.410 0.511 0.683 0.814 0.000

0.000 0.017 0.019 0.026 0.025 0.033 0.052 0.111 0.212 0.083 0.050 0.116 0.518 0.691 0.057 0.000 0.023 0.056 0.136 0.000

0.000 0.017 0.019 0.016 0.017 0.018 0.029 0.019 0.024 0.044 0.110 0.205 0.067 0.030 0.111 0.541 0.794 0.126 0.000 0.000

Table 5A.7

Common Reform: Simulated Hazard Rates (women) Option Value Model

Peak Value Model

Age

Linear

Dummies

Shifted

Linear

Dummies

Shifted

53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72

0.000 0.003 0.005 0.009 0.013 0.021 0.034 0.053 0.074 0.108 0.138 0.175 0.259 0.268 0.346 0.415 0.516 0.688 0.818 0.000

0.000 0.004 0.004 0.006 0.004 0.005 0.010 0.067 0.145 0.058 0.037 0.096 0.380 0.471 0.090 0.000 0.041 0.126 0.277 0.000

0.000 0.007 0.007 0.010 0.009 0.011 0.017 0.099 0.198 0.082 0.054 0.123 0.420 0.544 0.134 0.000 0.071 0.217 0.391 0.000

0.000 0.007 0.010 0.015 0.022 0.032 0.048 0.081 0.104 0.135 0.161 0.193 0.271 0.277 0.325 0.375 0.467 0.592 0.717 0.000

0.000 0.008 0.007 0.012 0.010 0.012 0.022 0.092 0.177 0.065 0.036 0.086 0.515 0.675 0.048 0.000 0.015 0.024 0.062 0.000

0.000 0.017 0.019 0.026 0.025 0.033 0.052 0.111 0.212 0.083 0.050 0.116 0.518 0.691 0.057 0.000 0.023 0.056 0.136 0.000

342

A. Börsch-Supan, R. Schnabel, S. Kohnz, and G. Mastrobuoni

References Birg, H., and A. Börsch-Supan. 1999. Für eine neue Aufgabenteilung zwischen gesetziclicher und privater Altersversorgung (A new division of labor between public and private pensions). Berlin: GDV. Börsch-Supan, A. 1992. Population aging, social security design, and early retirement. Journal of Institutional and Theoretical Economics 148:533–57. ———. 1995. Regional development, capital flows, and trade policies in an aging Europe. In Neue entwicklungen in der regionalökonomik (New developments in regional economics), ed. B. Gahlen, H. Hesse, and H. J. Ramser, 237–63. ———. 1998. Germany: A social security system on the verge of collapse. In Redesigning social security, ed. H. Siebert, 129–59. Tübingen, Germany: J. C. B. Mohr. ———. 2000a. A model under siege: A case study of the German retirement insurance system. The Economic Journal 110 (461): F24–45. ———. 2000b. Data and research on retirement in Germany. Report for the National Academy of Sciences, Washington, D.C. University of Mannheim, Department of Economics. Mimeograph. ———. 2000c. Incentive effects of social security on labour force participation: Evidence in Germany and across Europe. Journal of Public Economics 78:25–49. ———. 2001a. Blaupause für eine nachhaltige rentenreform (Blueprint for a sustainable pension reform). University of Mannheim, Department of Economics. Mimeograph. ———. 2001b. Incentive effects of social security under an uncertain disability option. In Themes in the economics of aging, ed. D. A. Wise, 281–310. Chicago: University of Chicago Press. Börsch-Supan, A., and P. Schmidt. 1996. Early retirement in East and West Germany. In Employment policy in the transition to free enterprise: German integration and its lessons for Europe, London, ed. R. Riphahn, D. Snower, and K. Zimmermann, 83–102. Heidelberg, Germany: Springer. ———. 1998. Social security and declining labor force participation in Germany. American Economic Review 88 (2): 173–78. ———. 1999. Social security and retirement in Germany. In International social security comparisons, ed. J. Gruber and D. A. Wise, 135–80. Chicago: University of Chicago Press. Bundesministerium für Arbeit und Sozialordnung (BMA). 1997. Statistisches taschenbuch (Statistical handbook). Bonn, Germany: Bundespresseamt. Burkhauser, R. 1991. An introduction to the German socio-economic panel for English speaking researchers. Syracuse University, Department of Economics. Mimeograph. Casmir, B. 1989. Staatliche rentenversicherungssysteme im internationalen Vergleich (Public pensions: An international comparison). Frankfurt, Germany: Lang. Coile, Courtney. 1999. Retirement incentives and couples’ retirement decisions, Ph.D. diss. MIT, Department of Economics. Coile, Courtney, and Jonathan Gruber. 1999. Social security and retirement. MIT, Department of Economics. Mimeograph. Gruber, J., and D. A. Wise, eds. 1999. Social security and retirement around the world. Chicago: University of Chicago Press. Jacobs, K., M. Kohli, and K. Rein. 1990. Germany: The diversity of pathways. In Time for retirement: Comparative studies of early exit from the labor force, ed. M. Kohli, M. Rein, A. M. Guillemard, and H. van Gunsteren, 181–221. Cambridge: Cambridge University Press.

Micro-Modeling of Retirement Decisions in Germany

343

Lumsdaine, R. L., J. H. Stock, and D. A. Wise. 1992. Three models of retirement: Computational complexity versus predictive validity. In Topics in the economics of aging, ed. D. A. Wise, 16–60. Chicago: University of Chicago Press. Meghir, C., and E. Whitehouse. 1997. Labour market transitions and retirement of men in the UK. Journal of Econometrics 79:327–54. Riphahn, Regina T. 1995. Disability retirement among German men in the 1980s. Ludwig Maximilians Universität. Münchner Wirtschaftswissenschaftliche Beiträge (Munich Economics Discussion Papers), no. 95-20. München, Germany. Riphahn, Regina T., and P. Schmidt. 1995. Determinanten des ruhestandes: Lockt der ruhestand oder drängt der arbeitsmarkt (Determinants of retirement: Is leisure pulling or the labor market pushing)? Center for European Economic Research (ZEW) Discussion Paper no. 95-10. Mannheim, Germany. Rust, J. 1990. Behavior of male workers at the end of the life cycle: An empirical analysis of states and controls. In Issues in the economics of aging, ed. D. A. Wise, 317–79. Chicago: University of Chicago Press. Rust, J., and C. Phelan. 1997. How Social Security and Medicare affect retirement behavior in a world of incomplete markets. Econometrica 65 (4): 781–831. Schmähl, W. 1991. Alterssicherung in der DDR und ihre umgestaltung im zuge des deutschen einigungsprozesses—einige verteilungspolitische aspekte (Pensions in the GDR and their changes during reunification—some distributional aspects). In Sozialpolitik im vereinten Deutschland (Social policy in a unified Germany), ed. G. Kleinhenz, 49–95. Berlin: Duncker and Humblot. Schmidt, P. 1995. Die wahl des rentenalters—Theoretische und empirische analyse des rentenzugangsverhaltens in West- und Ostdeutschland (The choice of retirement age—theoretical and empirical analyses of retirement behavior in West and East Germany). Frankfurt, Germany: Lang. Schnabel, R. 1998. Rates of return of the German pay-as-you-go pension system. Finanzarchiv N.F. 55 (3): 374–99. ———. 1999. Opting out of social security: Incentives and participation in the German public pension system. SFB504-Discussion Paper no. 99-42. University of Mannheim, Department of Economics. Siddiqui, S. 1997. The pension incentive to retire: Empirical evidence for West Germany. Journal of Population Economics 10 (4): 463–86. Stock, J. H., and D. A. Wise. 1990. The pension inducement to retire: An option value analysis. In Issues in the economics of aging, ed. D. A. Wise, 205–30. Chicago: University of Chicago Press. Sueyoshi, G. T. 1989. Social security and the determinants of full and partial retirement: A competing risk analysis. NBER Working Paper no. 3113. Cambridge, Mass.: National Bureau of Economic Research. Verband Deutscher Rentenversicherungsträger (VDR). 1999. Die rentenversicherung in zeitreihen (German public pensions in time series). Frankfurt am Main, Germany: VDR.

6 Micro-Modeling of Retirement Behavior in Italy Agar Brugiavini and Franco Peracchi

6.1 Introduction This paper presents an empirical analysis of the retirement decisions of Italian workers. We emphasize the role played by dynamic incentives that are built in the social security system, which encourage (or discourage) retirement. The basic idea is that, at any given age and based on the available information, workers compare the expected present value of two alternatives, retiring today and working one more year, and then choose the one that is best. A key role in this kind of comparison is played by social security wealth for which the level and the changes on a year-to-year basis and over the worker’s residual life reflect the institutional features of the social security system. The various incentive measures that we consider differ in the precise weight given to the social security wealth that workers accrue as they continue to work. These incentive measures are relevant in explaining retirement decisions as there are substantial gains or losses from retiring at particular ages, according to the prevailing legislation. Notice, however, that our model does not pretend to be a structural representation of the reAgar Brugiavini is associate professor of economics at the University of Venice, Ca’Foscari, Italy, and research associate at the Institute for Fiscal Studies, London, United Kingdom. Franco Peracchi is professor of econometrics at the University of Rome, Tor Vergata, Italy. We thank Franco Mariuzzo and Ronni Pavan for excellent research assistance. We also thank the Bank of Italy, the Instituto Nazionale della Previdenza Sociale (INPS) administration and the Instituto Nazionale di Statistica (ISTAT) for providing us with data. Financial support from the European Union (EU)-sponsored Training and Mobility of Researchers Programme (TMR) network on Saving and Pensions, and the European Science Foundation and the Ministero dell’Università e della Ricera Scientifica e Tecnologica (MIUR) are gratefully acknowledged. The usual disclaimer applies.

345

346

Agar Brugiavini and Franco Peracchi

tirement process, as a worker’s decision is modeled here following a simple reduced-form approach in which the incentive measures enter as predictors of a worker’s binary choice in addition to standard variables, such as sex, age, and other background variables. We use a longitudinal sample of Italian workers drawn from the social security archive containing the information on private-sector employees. We have twenty years of data available and can follow workers from the moment they start an employment spell in the private sector until they permanently leave the archive. Hence, we can model age-earnings profiles for all workers in the sample. This enables us to project their earnings forward and compute social security wealth in order to carry out the empirical analysis described above. Results from the estimated binary-choice model are then used to predict retirement probabilities under alternative policy arrangements. A basic feature of the simulated policies is to make retirement (particularly early retirement) more costly or to make eligibility requirements more stringent. Because changes in policy imply changes in the incentive measures, we expect to observe an impact on retirement probabilities. Our results suggest that incentives are important in explaining retirement decisions, although the interaction of age and dynamic incentives is rather complex. In some of our policy simulations, we observe a shift in the age hazard into retirement toward older ages and also (although less frequently) an increase of the mean retirement age in response to the policy change. The remainder of the paper is organized as follows. Section 6.2 reviews the recent situation of the Italian social security system in terms of expenditure trends and financial viability. It also describes recent institutional changes and labor market trends. Section 6.3 provides a description of the available data and discusses the methodology and the results of estimating age-earnings profiles. Section 6.4 looks at the definitions of the incentive variables and highlights some of the methodological issues involved in computing these measures in our sample. Finally, section 6.5 presents the results of the econometric exercise, as well as the results of the simulations. 6.2 Recent Situation and Institutional Details In this section, we briefly describe the current situation of the Italian social security system. After summarizing the institutional details of the system before the beginning of the reform process in 1992, we look at the steady state characteristics of the system introduced with the 1995 reform. For the scope of this study, the transitional arrangements are also important. Finally, we discuss recent trends in retirement patterns and review the results of the available empirical literature.

Micro-Modeling of Retirement Behavior in Italy

347

6.2.1 Recent Expenditure Trends Figure 6.1 shows the historical trends in the number of pensions (top-left panel), average pension amounts (top-right panel), total pension expenditure (bottom-left panel) and the ratio of pension expenditure to gross domestic product (GDP; bottom-right panel) as measured by the National Institute of Statistics (ISTAT 1999). We only consider expenditure on oldage, disability, and survivor pensions, thus excluding noncontributive pensions. In the twenty-five years from 1975 to 1999, the number of pensions outstanding grew by more than 40 percent, from 12.4 to 17.8 million, and the average pension amount more than doubled in real terms, from L7.1 to L16.1 million at 1998 prices. Therefore, total expenditures increased by more than three times and the ratio of pension expenditure to GDP rose by more than 5 percentage points, from 8.4 to 13.6 percent. No other country in the EU experienced such a dramatic growth. The increase in the number of pensions outstanding reflects both the progressive aging of the Italian population and the steady reduction of the average retirement age, largely due to people taking advantage of the possibility of retiring with a seniority pension ( pensione di anzianita’) after

Fig. 6.1 Number of pensions, pension expenditure, average pension, and expenditure-GDP ratio, 1975–98: Old-age, disability, and survivor pensions Source: ISTAT (various years).

348

Agar Brugiavini and Franco Peracchi

thirty-five years of contributions to the system (or even less for publicsector employees) without any actuarial reduction. On the other hand, the rise of average pensions is due to the sharp increase of lifetime earnings of successive cohorts of workers, a number of legislated changes that made the system progressively more generous, and the fact that, until 1992, outstanding pensions were linked to productivity growth. Notice that, after increasing very quickly during the 1970s, pension expenditure slowed down during the first half of the 1980s. Expenditure resumed growing very rapidly from the mid-1980s to the early 1990s when two major reforms, in 1992 and 1995 (known respectively as the “Amato reform” and the “Dini reform” from the names of the prime ministers then in charge), apparently succeeded in stabilizing the ratio of pension expenditure to GDP. The next subsections discuss in more detail the rules prevailing before 1993 and the reforms of the 1990s. 6.2.2 The Rules Prevailing Before 1993 The Italian social security system is based on a variety of institutions administering public pension programs. About two-thirds of the workforce is insured with the National Institute of Social Security (Istituto Nazionale della Previdenza Sociale, or INPS).1 This is responsible for a number of separate funds, of which the most important covers the private-sector employees (Fondo Pensioni Lavoratori Dipendenti, or FPLD). In this section, we describe the main rules prevailing before 1993. In 1992, a major reform was introduced, followed by another major reform in 1995, and further changes in 1997. These reforms are discussed in section 6.2.3. We focus on the FPLD fund because our empirical exercise uses data on private-sector employees. This description is incomplete, however, as important differences exist between the rules for private- and public-sector employees. Furthermore, the self-employed are characterized by a separate fund and enjoy a particularly favorable treatment in terms of both contributions and benefit calculation. Payroll Social Security Taxes The resources to the system come mainly from the employers’ and employees’ contributions. Outlays exceed revenue, however, and the resulting deficit is financed by the central government, which has come under increasing pressure to pay for pensions. The payroll tax is unevenly shared between the employer and the employee. For the FPLD, the total payroll tax was 24.51 percent of gross earnings until 1992, of which 7.15 percent was levied on the employee. This grew to 27.17 percent in 1992 (of which the worker paid 8.34 percent and 1. It covers the vast majority of the private-sector employees and the self-employed. Publicsector employees are covered by a completely separate administration Instituto Nazionale della Prividenza per i Dependenti delle Amministrazioni Pubbliche (INPDAP).

Micro-Modeling of Retirement Behavior in Italy

349

33 percent in 1998 (of which the worker pays 8.89 percent). Social security taxes for public-sector employees have been lower in the past but are now in line with those in the private sector.2 Employees contribute a further 7.41 percent to a severance-pay fund known as Trattamento di Fine Rapporto (TFR). These contributions are retained by the employer and build up in a fund that offers a legislated rate of return (1.5 percent plus 75 percent of the annual inflation rate) and provides a lump-sum benefit when the employee leaves the firm. Eligibility Before 1993, eligibility requirements were met when a male worker reached age sixty (for a female worker, age fifty-five) after contributing for at least fifteen years. However, the existence of an early retirement option made the age requirement largely irrelevant for private-sector workers with uninterrupted work careers because they could claim early retirement benefits at any age after completing thirty-five years of contributions.3 A feature that is important when discussing labor supply incentives provided by the system is the retirement earnings test. Italian workers can draw a pension and earn income at the same time, but there are earnings cutoffs that tend to discourage this choice. The cutoffs changed over time and have been affected by the reforms. Prior to 1993, private-sector employees could receive an old-age pension along with earnings only if earnings did not exceed the minimum pension benefit. Early retirement benefits could not be received along with earnings.4 Eligibility criteria and the rules for benefit calculations were responsible for the highly redistributive nature of the pension program. Most importantly, they affected retirement decisions and the choice between dependent employment and self-employment in a nontrivial fashion. In particular, there was a clear incentive to early retirement since no actuarial penalty applied to early retirees. For example, a private-sector employee who started working at age sixteen could retire at age fifty-one with a full pension. This helps to explain why exit from the labor force increased significantly over time for the age group fifty to fifty-nine. Benefit Computation Pensionable earnings for private sector employees (covered by the FPLD) were computed by averaging the earnings of the last five years 2. While 33 percent is the level of the payroll tax used in the benefit-calculation formula, the actual amount paid is 32.7 percent. In contrast to employees, social security taxes for the selfemployed remain considerably lower; the tax rate is notionally set at 20 percent, but the effective payroll tax rate is 15 percent. 3. For a male public-sector employee, only twenty years of contributions were required (fifteen years for a married woman). 4. Note that the self-employed could claim early retirement without restrictions on earnings. Also, cutoff points on earnings did not apply to this group of workers when claiming oldage benefits. Early retirement benefits are reduced for the working self-employed only in 1998.

350

Agar Brugiavini and Franco Peracchi

before retirement.5 Earnings were taken before tax and converted into real amounts using a consumer price index. Pension benefits were then obtained as the product of three terms: pensionable earnings, the number of years of contribution (up to a maximum of forty), and a 2 percent factor per year of contribution (referred to as the “rate of return”). Hence, a worker could get 80 percent of his pensionable earnings at most. If retirement was postponed, additional years of work beyond age forty did not count for benefit computation, although they were included in pensionable earnings as they replaced earnings of earlier years. As discussed later in section 6.4, this has important implications for the age profiles of social security wealth. The system was highly progressive (and it still is) because of both earnings caps (i.e., earnings entering the benefit computation were capped) and a minimum benefit level. Between 1969 and 1988, only earnings up to a legislated limit would contribute to pensionable earnings, while after 1988 different rates of return would be applied to different pensionable-earnings brackets.6 Contributions were not subject to income taxes, but social security benefits were taxed at the normal income tax rates.7 For all funds, benefits were indexed both to consumer price inflation and real earnings growth.8 The former was measured by the consumer price index but was implemented in a slightly staggered fashion (e.g., if the pension amount is more than three times the minimum benefit, then indexing is based only on 75 percent of the price change). The measure of earnings growth took into account changes in real wages in both the private- and the public-sector. Minimum Benefit This provision of the Italian system is important for at least two reasons. First, the number of retirees involved is nonnegligible (see table 6.4). Second, the minimum benefit is often used as a benchmark for income transfers from other assistance programs. In practice, if the benefit formula results in a benefit level below a legislated threshold, then the benefit itself is 5. It is worth recalling that public-sector employees had their benefit level based on final salary rather than average earnings of the last five years of work. 6. For example, in 1985 pensionable earnings in excess of €16,527 (1.6 times the average earnings of that year) would not be included in benefit calculations. After 1988, the constraint became less stringent, but a lower rate of return was applied to pensionable earnings in excess of the legislated limit. In 1995, a 2 percent rate applied to the first €29,438 (again, 1.6 times the average earnings), a 1.5 percent rate to pensionable earnings between €29,438 and €39,250 (2.2 times the average earnings), and a 1.25 percent rate to pensionable earnings between €39,250 and €49,063 (2.7 times the average earnings). Finally, a 1 percent rate of return applied to the top earnings bracket. The generosity of the system toward low-income workers was further increased by the minimum benefit provision, that is, a floor on the benefit level. 7. The principle of taxing pension benefits, but not taxing contributions, remains valid after the reforms. 8. These growth rates were combined in a specific index computed by the INPS administration (coefficiente di perequazione). The frequencies were every four months between 1983 and 1986, every six months between 1986 and 1992, and annually after the 1992 reform.

Micro-Modeling of Retirement Behavior in Italy

351

increased up to the threshold. Until 1983, this provision could be applied to more than one pension for the same retiree. Now, it only applies to one pension, leaving the other benefits at their original level. Minimum benefits are means tested. Until 1992, the test was based only on the claimant’s income, excluding the income of the spouse. For example, in 1985, the means test had a cutoff at twice the minimum level (roughly €4,130 of that year, corresponding to 17 percent of mean household income of the same year). A limit on individual income still applies to singles. For married couples, what matters now is instead the sum income of both spouses, which must be less than four times the minimum level (approximately €4,130 in 1995, corresponding to 18 percent of mean household income). 6.2.3 The Reforms of the 1990s Some of the issues emerging from the preceding description of the pre1993 system have been tackled by the recent reforms. A first reform (the Amato reform) was passed by Parliament in 1992 and took effect in 1993. It raised the normal retirement age and the minimum number of years of contribution by five years, lengthened the reference period for calculating pensionable earning, restricted the special eligibility conditions applying to public-sector employees, eliminated pension indexation to real wage growth by indexing pensions only to price inflation, and increased social security contributions. The 1992 reform was the first signal of a coherent attempt at redesigning the social security system and reducing pension expenditures. However, it left the rules governing early retirement almost untouched and, overall, it did not produce the much-needed short-term savings in the social security budget.9 This partly justified the need of a second reform in 1995 (the Dini reform), which totally changed the basic rules for granting benefits to future retirees and tried to harmonize the actuarial rates of return for early and late retirees. The 1995 reform goes under the heading of “virtual” funding since public pensions are still financed through a PAYG scheme, but each worker holds a claim based on a fund that remains only virtual. The new system is therefore a notionally defined-contribution system that is similar to the one recently adopted by Sweden. Crucial elements of the reform are: (a) how the accrued value of the virtual fund is computed, (b) how it is then converted into an annuity at the time of retirement, and (c) the indexation rule adopted for the pensions outstanding. Lifetime tax payments are capitalized at an annual rate equal to a (five-year) moving average of past GDP-growth rates. The level of benefits (the annuity) no longer depends on final earnings, as in the previous system, but is instead proportional to the value of accrued social security tax payments. The proportionality factor increases more than 9. Section 6.2.4 analyzes some of the effects of the 1992 reform and of the 1995 reform.

352

Agar Brugiavini and Franco Peracchi

proportionally with the retirement age up to age sixty-five and then flattens out. Its current age profile implies an actuarial reduction of 23 percent in pension benefits for retirement at age fifty-seven, relative to retirement at age sixty-five, and has been legislated taking into account two key elements: the average residual life expectancy at retirement based on the 1990 life tables and a fixed real rate of return of 1.5 percent that reflects long-run forecasts of annual GDP growth. Finally, outstanding pensions are going to be indexed to price inflation only and not to real wage growth. The 1995 reform changed eligibility rules by allowing people to retire at any age after reaching age fifty-seven provided that they meet two conditions: (a) they must have contributed to the system for at least five years, and (b) the value of their pension must be at least 1.2 times the noncontributive pension that is paid to those aged sixty-five and older who have no other income source (“social” pension). Besides changing the benefit formula and the eligibility rules, the 1995 reform also took a number of steps in the direction of unifying the rules of the many schemes in which the Italian social security system is organized. Once phased in, the 1995 reform will imply a more transparent and actuarially fair pension system. The reason is twofold. First, benefits are more clearly linked to contributions than it was the case with the previous final-salary type formulae, thus reducing negative incentive effects on labor supply. Second, the whole workforce is now covered by essentially the same system, thus reducing incentive effects in favor of certain types of employment (public-sector employees and the self-employed). The system is not completely neutral, however, and a number of provisions still exist that tend to favor the self-employed.10 Table 6.1 summarizes the key features of the three regimes: the one prevailing before the 1992 reform (denoted as pre-1993 regime), the one prevailing at the steady state after the 1992 reform, and the ones after the 1995 reform. In practice, both the 1992 and the 1995 reforms are characterized by a very long transitional period. For workers with less than fifteen years of contributions at the end of 1992, the provisions for the transitional period establish eligibility and benefit-computation criteria on a pro-rata basis. This method allows the rules of the old regime to hold for the fraction of years in employment under that regime, while the remaining fraction is regulated by the new rules. For these workers, eligibility and social security benefits are therefore computed taking into account three different systems of legislation. For workers with at least fifteen years of contribution at the end of 1992, the rules of the pre-1992 regime apply with only small changes. 10. For example, the self-employed still enjoy lower contribution rates (see also note 2). Furthermore, not all funds are affected by the reform; for example, a number of special funds for professional self-employed (such as lawyers, architects and building engineers, and accountants) still work according to their own rules, although attempts to harmonize these rules with those of the employees fund are taking place.

Micro-Modeling of Retirement Behavior in Italy Table 6.1

353

Key Features of the Pre-1993 Regime and the 1992 and 1995 Reforms (at the steady-state) Pre-1993 Regime

Normal retirement age

60 (men) 55 (women)

Transitional period

1992 Reform

1995 Reform

65 (men) 60 (women)

Any age after 56 (both men and women)

Until about 2032

Until about 2035

Pensionable earnings

Average of last 5 years real earnings (converted to real values through price index)

Career average earnings (converted to real values through price index + 1%)

Career contributions (capitalized using a 5-year moving average of GDP growth rate)

Pension benefit

2%
(pensionable earnings)
(t), where t is years of tax payments (40)

2%
(pensionable earnings)
(t), where t is years of tax payments (40)

Proportional to capitalized value of career contributions, the proportionality factor increasing with age at retirement (from .04720 at age 57 to .06136 at age 65)

Pension indexation

Cost of living plus real earnings growth

Cost of living

Cost of living

Pension to survivor

60% to spouse, 20% to each child; 40% to each child (if no spouse)

Same

Same

Years of contributions for eligibility

15

20

5

Early retirement provision

Any age if contributed to SS for 35 years or more, no actuarial adjustment

Any age if contributed to SS for 35 years or more, no actuarial adjustment

No early retirement provision

Total payroll tax

24.5% of gross earnings

27.17% of gross earnings

32.7% of gross earnings

Therefore, people will retire under the pre-1993 system until about the year 2015. During the following fifteen to twenty years, an increasing fraction of a retiree’s pension will be computed on the basis of the new system. It will only be around 2030 that a significant number of workers will start retiring fully under the 1995 rules. In our paper, the only relevant transitional phase is the one introduced by the 1992 reform. According to these rules, the normal retirement age for a private-sector employee gradually increases to reach sixty-five for men and sixty for women in year 2000. It should be mentioned that between 1992 and 1997 there have been spells (typically lasting between six months and one year) in which many employees were not allowed to take early

354 Table 6.2

Agar Brugiavini and Franco Peracchi Rules for Benefit Computation Prevailing During the Transitional Period After the 1992 Reform

Case

Pensionable Earnings Computation a

A1: Private sector—Senior (years before 1993) A2: Private sector—Senior (years after 1993) B1: Private sector—Junior (years before 1993) B2: Private sector—Junior (years after 1993)

C1: Public sector—Senior (years before 1993) C2: Public sector—Senior (years after 1993) D1: Public sector—Junior (years before 1993) D2: Public sector—Junior (years after 1993)

Average of last 5 years’ real earnings Average of last 6.5 years’ earnings, to gradually become last 10 years’ earnings after 2002 Same as A1 Average of last 8 years’ earnings, to gradually become last 15 years’ earnings after 2002 and further increasing thereafter. Last month’s real earnings Average of last 1.5 years’ earnings, to gradually become last 10 years’ earnings after 2012 Same as C1 Average of last 3 years’ earnings, to gradually become last 10 years’ earnings after 2012 and further increasing thereafter

Notes: Table 6.2 focuses on employees, but similar rules apply to the self-employed. Consider, for example, a case A worker with at least 15 years of seniority at the end of 1992. His pension depends on a weighted average of pensionable earnings computed under case A1 and pensionable earnings computed under case A2, with weights determined by the number of years in the system before and after 1993 respectively. a Seniors  At least 15 years in the system at the end of 1992. “Years before 1993” are the years of valid tax payments to the social security administration completed before 1993.

retirement.11 During the transitional phase, the benefit calculation distinguishes workers depending on their seniority and deals differently with contributions paid before and after the reform as detailed in table 6.2. It is clear that the transitional rules mainly affect younger workers. 6.2.4 Future Prospects So far, the effects of the 1995 reform on pension expenditure have been only minor. After 1995, pension expenditure resumed growing at rates that, although lower than in the past, have nevertheless been higher than GDP growth rates (figure 6.1). Furthermore, the slowdown of expenditure growth is largely the result of decisions taken in 1992, namely the switch from the double indexing (to price inflation and productivity growth) of pensions to price-inflation indexing; the introduction of limitations to early retirement; and the gradual increase of the normal retirement age. However, as we discuss below, the changes in eligibility for old-age and early retirement benefits brought about by the reforms have not been particularly effective: the elimination of the double indexation has been by far the most important change. 11. The 1995 reform has gradually removed these constraints. In the transitional phase starting in 1996, public-sector employees who claimed early retirement benefits suffered minor reductions on the basis of actuarial adjustment factors.

Micro-Modeling of Retirement Behavior in Italy

355

Fig. 6.2 Recent trends in labor force participation rates and employment rates, October 1992–April 2000 (index October 1992
1,000) Source: Authors’ calculations based on the micro-files of the Labor Force Survey 1992–2000.

In order to evaluate the future prospects of the Italian public pension system, we need to start from the available demographic projections. Indeed, the ones produced by ISTAT for the period 1996–2050 have attracted considerable attention in the public policy debate. They portray a rather worrisome picture as the elderly dependency ratio—that is, the ratio between the elderly population (people aged sixty and older) and the working age population (people aged twenty to fifty-nine)—is expected to grow from 40 percent in 1996 to 83 percent in 2050. These projections have been criticized for being too optimistic because they rule out further declines in mortality. On the other hand, they appear to underestimate migration, which could play an offsetting role. Besides demographics, the other crucial variable for judging the future viability of the social security system is labor market trends. Past trends in labor force participation, described in Brugiavini (1999), show a progressive detachment of older workers from the labor force. Here we look at more recent evidence on the labor market behavior of older workers. Figure 6.2 shows the time series of labor force participation rates (top panels) and employment rates (bottom panels) by age group for the period

356

Agar Brugiavini and Franco Peracchi

Table 6.3

Current Retirement Eligibility Rules INPS (Private Sector)

INPDAP (Public Sector)

Self-Employed

Year

Age and Years of Contribution

Only Years of Contributions

Age and Years of Contribution

Only Years of Contribution

Age and Years of Contribution

Only Years of Contribution

1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008

54 and 35 55 and 35 55 and 35 56 and 35 57 and 35 57 and 35 57 and 35 57 and 35 57 and 35 57 and 35 57 and 35

36 37 37 37 37 37 38 38 39 39 40

53 and 35 53 and 35 54 and 35 55 and 35 55 and 35 55 and 35 57 and 35 57 and 35 57 and 35 57 and 35 57 and 35

36 37 37 37 37 37 38 38 39 39 40

57 and 35 57 and 35 57 and 35 58 and 35 58 and 35 58 and 35 58 and 35 58 and 35 58 and 35 58 and 35 58 and 35

40 40 40 40 40 40 40 40 40 40 40

Source: INPS, Information on Early Retirement (available at http:www.inps.it) and Ministero del Lavoro, Social Security (available at http://www.miniwelfare.it). Note: Rules prevailing after 1998 according to the Law 449/1997. These rules apply to white-collar employees, they differ only slightly for blue-collar employees.

from October 1992 to April 2000 and separately for men and women. The data have been computed from the micro-files of the quarterly labor force survey and have been normalized by taking indexes with an October 1992 base equal to 100. For men, a relatively stable labor force participation rate (and employment rate) is observed after 1998 in all age groups. This occurs after a period of marked decline in participation. It is interesting to note that, after 1996, labor force participation is slowly starting to increase for the age group fifty to fifty-four, while it remains at a low level for the other two age groups. Women aged fifty to fifty-four and fifty-five to fifty-nine also show increasing participation after the beginning of 1995, whereas men and women aged sixty to sixty-four show a very similar negative trend. The interpretation of these findings is that the recent reforms have affected retirement behavior in two ways. Initially, reforms had mainly an announcement effect and workers—particularly younger retirees— claimed retirement as soon as they could in order to avoid losing the option of leaving the labor market in subsequent years. After 1997, however, the eligibility rules built into the new system start to become binding (see table 6.3), and workers, particularly the younger cohorts of potential retirees, were forced to delay retirement.12 Overall, there is some evidence that the recent reforms of the Italian social security system have begun to 12. See also Franco (2000) for similar arguments in an appraisal of the recent Italian reforms.

Micro-Modeling of Retirement Behavior in Italy

357

have an impact on the retirement behavior of older workers, although this is unlikely to revert the pension expenditure trends of the recent past or to counteract the effects of the aging process of the Italian population. 6.2.5 Literature Review To our knowledge, only three studies are currently available in Italy that try to explain the individual decisions to retire from the labor force, particularly the choices of retirement age and the specific pathway of exit. The problem, however, is of great importance. Take, for example, the dramatic decline in labor force participation of men aged fifty to fifty-nine following the start of the lengthy process of social security reforms in 1992. This decline has been missed entirely by the reduced-form model used by the actuaries of the main social security funds, which are based on simple extrapolation of the trends prevailing up to the early 1990s (see, e.g., INPS 1995). The first study, by Miniaci (1998), analyses the effects of socioeconomic and demographic characteristics on the choice of labor force status using the 1995 Survey of Household Income and Wealth (SHIW) carried out by the Bank of Italy. The sample consists of heads of households and their partners who were in the labor force at the age of forty-five (if women) or fifty (if men). The retirement age is obtained from the retrospective questions asked in the survey about the time when pension benefits were first received. The basic model is a multinomial logit specification with three states: working, receiving an old-age or seniority pension, and receiving other pensions (invalidity, survivor, or social pensions). The paper also estimates a Cox proportional-hazard model for the duration of stay in employment and separate pathways of exit (old-age or seniority pension, invalidity, or other), with and without the inclusion of the predicted value of the replacement rate among the covariates. The main findings show the presence of strong cohort effects towards earlier retirement, later retirement ages among the better-educated workers, lower replacement rates and later retirement ages among the self-employed, no evidence that public employees retire earlier, no evidence of a differential role of invalidity and other pensions between the north and the south, and little evidence of a replacement-rate effect on the expected retirement age. The second and third study, by Spataro (2000) and Colombino (2000), respectively, take a more structural approach. The paper by Spataro estimates an option value model using a subset of the panel component of the SHIW for the years 1991 and 1993 (namely, men aged forty-five to sixtyfive who are full-time employees at the end of 1990). The estimation method is a pseudo–maximum likelihood based on the normality assumption. He compares the empirical results with reduced-form probit specifications and the original results in Stock and Wise (1990). In particular he shows that, relative to U.S. workers, Italian workers value their leisure

358

Agar Brugiavini and Franco Peracchi

more highly, are more risk averse, and have a lower intertemporal discount rate. He also shows that the option value model is unable to capture the peak in the retirement hazard at age sixty. Finally, Colombino (2000) extends the probit model estimated by Zweimüller, Winter-Ebmer, and Falkinger (1996) for Austria to provide a structural interpretation in terms of utility comparisons between two mutually exclusive states (employed and retired). This interpretation is based on specific assumptions about the time profiles of the instantaneous utility of being employed and retired. The model is estimated using the 1993 SHIW for those aged forty and older. The empirical implementation of the model relies crucially on the imputation of potential earnings for those who are currently retired and on potential pension benefits for those who are currently employed. One of the main findings is that women are more sensitive than men to the rules and incentives of the system. The paper also contains an extension to the joint decision of a couple. This extension leads to a bivariate probit model whose parameters are subject to constraints and admit a structural interpretation. The model is used to simulate the effects of various policies (cut of pension benefits, elimination of seniority pensions, and complete phasing in of the 1992 and 1995 reforms), distinguishing between the effects due to the changes in the way pension benefits are computed, the changes in the criteria for pension eligibility, and the behavioral response to both changes. The main finding here is that behavioral effects are small, but not negligible. 6.3 Data Description and Earnings Model 6.3.1 The Data Unlike the papers reviewed in section 6.2.5, all based on the Bank of Italy’s SHIW, we use a random sample of administrative records from one of the INPS archives.13 The sample is drawn from the so-called INPS Workers Archive (Archive O1M), which contains records on all privatesector employees insured with INPS. The information on each employee is filled in by the employer on a standard form containing a small number of entries. We have a random sample of these employees in the form of a panel 13. This is a subsample (one out of one hundred) workers born either on 1 March or on 1 October of any possible year contained in the archive. We carried out a parallel analysis of retirement decisions using the Bank of Italy’s SHIW data. The SHIW is a cross-sectional sample and contains a full set of demographic variables, relative to the household and its members, plus labor market variables. However, the panel component of SHIW is characterized by a very small sample size and very short time span: very few transitions into retirement are actually observed. Also questions in the SHIW survey are retrospective, and the survey is carried out, for the relevant period, every two years. Therefore, we decided to restrict the attention to the INPS sample.

Micro-Modeling of Retirement Behavior in Italy

359

covering a period of about twenty years from 1973 to 1994. The sample contains 10,000 workers entering the archive at any time during the period considered. Employment spells can last any number of years, and individuals can leave the sample and enter again in any subsequent year. The panel is therefore highly unbalanced. The main advantages of using these data are that they span a fairly long time period and contain information on gross earnings (as opposed to net earnings, as in the SHIW), which form the basis for the calculation of social security benefits. However, there are several shortcomings. 1. The data set only covers private-sector employees, leaving out publicsector employees and the self-employed. Even for private-sector employees, however, coverage is not full, and a small fraction of them is not included. 2. The reason for a worker leaving the archive is not known: In addition to retiring, workers could die, become self-employed or public-sector employees, or simply stop working. 3. Important covariates (e.g., education level, spousal information, and other family background variables) are missing, and hence, we have very few demographic controls available: we do not know about marital status and cannot say much about differential mortality. 4. There is no information on receipt of disability or other types of benefits. The initial sample selection, carried out in order to estimate suitable earnings histories, is as follows. We focus on workers between eighteen and seventy years of age. We drop observations for which one important indicator (such as age) is missing and individuals who work less than twentysix days a year. We also exclude from the analysis workers belonging to special INPS funds (nursery school teachers, local authorities, employees, and so forth).14 In order to estimate earnings profiles and eventually measure social security wealth, we further limit the sample by including only workers who are present in the sample for an uninterrupted period of at least five years (workers often appear for one year and then disappear from the sample for a long spell). The five-year minimum requirement is motivated by the fact that it corresponds to the minimum contributive period under the 1995 reform. We only keep workers who do not have substantial gaps (more than ten years missing) in their records. This is because we cannot say whether in that time span they were engaged in other labor market or nonlabor market activities (such as maternity leaves or undertaking fur14. We could include these observations to add variability across funds, but these workers represent only a small number (less than 100 observations) and tend to exhibit many gaps in their careers.

360

Agar Brugiavini and Franco Peracchi

Table 6.4

Fraction of Minimum Benefit Awards Over Total Awards, by Sex and Type of Benefits (end of 1999) Old-Age and Early Retirement

Disability

Survivors

Age

Male

Female

Male

Female

Male

Female

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75

0.000 0.000 0.004 0.007 0.011 0.011 0.012 0.012 0.016 0.018 0.019 0.021 0.024 0.028 0.050 0.080 0.092 0.100 0.111 0.121 0.101 0.103 0.116 0.128 0.133 0.136

0.000 0.000 0.008 0.020 0.034 0.044 0.056 0.063 0.086 0.252 0.328 0.408 0.433 0.453 0.464 0.382 0.382 0.390 0.402 0.406 0.400 0.385 0.380 0.374 0.372 0.327

0.153 0.169 0.175 0.174 0.182 0.191 0.188 0.191 0.198 0.203 0.207 0.215 0.220 0.225 0.263 0.338 0.387 0.389 0.402 0.400 0.410 0.418 0.433 0.439 0.448 0.535

0.250 0.310 0.308 0.325 0.328 0.341 0.345 0.345 0.359 0.417 0.520 0.565 0.589 0.602 0.610 0.616 0.622 0.620 0.629 0.630 0.639 0.647 0.655 0.659 0.661 0.684

0.390 0.326 0.326 0.298 0.309 0.293 0.290 0.283 0.268 0.263 0.265 0.263 0.265 0.283 0.265 0.250 0.251 0.254 0.255 0.246 0.268 0.260 0.258 0.253 0.256 0.212

0.325 0.308 0.306 0.297 0.304 0.309 0.300 0.300 0.301 0.284 0.259 0.254 0.247 0.247 0.242 0.259 0.256 0.245 0.238 0.231 0.227 0.226 0.219 0.214 0.208 0.193

Note: Authors’ calculations based on the INPS Workers Archive and Pension Archive of outstanding pension benefits. This contains the universe of public pension benefits.

ther education). The choice of a ten-year interval is arbitrary and is based on a preliminary inspection of the data.15 15. It should be noted that, in order to gain variability in social security benefits, we did experiment with a larger sample including almost all workers, regardless of the existence of gaps in their careers. However, this did not add valuable information since the majority of workers with substantial spells out of the private sector would end up qualifying for minimum benefits (the level of which is fixed by legislation each year) or for an old-age income guarantee (pensione sociale). Hence, there would be very little correlation between earnings histories and pension benefits for these individuals, and the effects of potential reforms in changing the incentives to retire would be negligible (these workers would basically qualify for the minimum benefit under all regimes). Therefore, these cases would end up blurring the results rather than adding variability to be exploited. Finally, our choice of the ten-year threshold and the requirement of a five-year minimum presence in the archive gives us an estimated sample percentage of minimum benefit recipients that is not too far from what we observed in the universe of pension awards as recorded by the INPS Administration (see table 6.4).

Micro-Modeling of Retirement Behavior in Italy Table 6.5

361

Distribution, by Labor Force Status in the SHIW

Year

Private-Sector Employee

Public-Sector Employee

Self-Employed

Pensioner

Other

1978 1979 1980 1981 1982 1983 1984 1986 1987 1989 1991 1993 1995 Total

44.41 41.23 39.42 43.56 42.44 39.64 36.85 37.15 45.79 40.37 35.02 26.72 26.64 37.00

12.96 15.07 15.22 14.21 15.29 15.21 16.56 15.46 6.95 10.50 13.16 16.50 15.36 13.69

15.42 15.45 15.09 14.55 17.42 15.01 15.17 14.61 7.27 16.63 13.23 12.70 13.54 13.80

25.76 26.36 28.44 26.24 23.53 28.76 29.37 31.34 28.87 30.89 34.69 37.48 37.21 31.39

1.46 1.88 1.83 1.44 1.31 1.38 2.05 1.45 11.12 1.61 3.91 6.60 7.25 4.12

Notes: Private-sector employee  active workers in private sector; public-sector employee  active worker in public sector; self-employed  a wide category of active workers working as entrepreneurs, professionals, and self-employed proper; pensioner  self-reported pensioner.

In order to check the quality of our data, we carried out an extensive comparison between the basic variables in the INPS sample and those in the Bank of Italy’s SHIW, using the whole sequence of SHIW cross sections for all available years between 1978 and 1995. Here we only report some of our results. We look at changes in the composition of the workforce. Table 6.5 measures the relative importance of private-sector employees on total employment and shows a clear trend towards a reduced importance of employment in the private sector and a corresponding increase in the importance of self-employment and employment in the public sector. Hence, a drawback of the INPS data is that an increasingly important section of the workforce is not covered. Turning to earnings of private-sector employees, notice that the SHIW only collects information on net earnings while the INPS data provide gross earnings figures (as earnings are recorded by the social security administration before any tax due). To carry out the comparison, the SHIW data were “grossed-up” using a procedure based on the information available at both the individual and the household level. There exist some differences in average gross earnings between the two samples: a more careful comparison was carried out by controlling for age and cohort effects. In particular, we estimated a simple model of gross earnings as a function of age and birth cohort, with synthetic cohorts defined according to the year of birth of each worker. This analysis confirmed a substantial agreement

362

Agar Brugiavini and Franco Peracchi

in the earnings data for the two samples, both the grossing-up procedure adopted for SHIW and the different sampling frame could partly explain the remaining differences. 6.3.2 Earnings Projection The information needed in order to model age-earnings profiles in the INPS sample consists of age, gender, occupation, sector of employment, and region of working activity.16 The specification of a model for the age-earnings profile represents an essential step in the estimation of social security wealth at the individual level. This is especially important in Italy, as the process of social security reform involves moving from a final-salary type of benefit formula (pre1993 system) to a lifetime-earnings formula (1992 reform) and to a formula based on the value of lifetime contributions (1995 reform).17 We subsequently describe two additional hypothetical reforms that also involve extending the benefit calculation period. It turns out that results are very sensitive to the way that earnings projections, backward projections in particular, are carried out. For example, what may seem a negligible overestimation of real earnings in the early years can have marked effects on benefit calculations in the 1995 reform, where the whole earnings history matters and revaluation of past earnings is based on a five-year moving average of past GDP-growth rates. Three alternative earnings-modeling strategies were considered. In the first strategy, individual real age-earnings profiles are assumed to be completely flat after the last year of observed earnings. This corresponds to the assumption that, at the individual level, the real-earnings process is a random walk with no drift. In practice, the “jump-off” point for the earnings projections is taken to be the average of the last three years of observed earnings. This jump-off point pins down the level of the age-earnings profile for each individual.18 Note that this might seem to underestimate future earnings growth, particularly for younger cohorts, but since our “sample at risk” (as defined later) consists mainly of older cohorts, the problem may not be too severe.19 Furthermore, for ages above fifty, earnings are lower on average and very noisy, possibly because of part-time work or the coexistence of early retirement benefits and working activities. When going back16. This is actually the region where the firm is located. Hence a comparison with the SHIW and national accounts data reveals that there seems to be a higher number of workers located in the northwest, where many large firms have their headquarters. 17. In this and the following sections, we only describe results for the 1995 reform; results for the other cases are available upon request from the authors. 18. When going backward, the jump-off point corresponds to the average of the first three observations available for each individual. 19. The cohorts at risk are defined according to year of birth: For the oldest cohort, these are birth years between 1918 and 1926, for the next cohort 1927–1936, and for the youngest cohort 1937–1944.

Micro-Modeling of Retirement Behavior in Italy

363

ward, using a flat earnings profile would grossly overestimate the level of earnings at earlier ages and grossly underestimate real earnings growth. To avoid this problem, individual earnings were assumed to grow at the annual growth rate of aggregate earnings, for the years when this information is available, and at a constant real rate of 1.5 percent otherwise.20 In the second strategy, real earnings are projected into the future using group-specific growth rates obtained from the available sample, with groups defined by sex, age, birth cohort, and occupation. We did not really pursue this route, as some major drawbacks soon became apparent. Because group-specific growth rates are obtained using the sample information, we face the problem that individual-specific growth rates may differ dramatically from those estimated for the relevant group, but at the same time, group-specific growth rates may differ substantially from the macroeconomic growth rate. In addition, for younger cohorts we would often get very high growth rates going backward and very low growth rates going forward. Finally, notice that our first data point is in 1973 although we need to go back to the 1930s for some of our workers in order to complete their working history. Hence, we would be forced to use a hybrid procedure, which makes use of aggregate growth rates when projecting backwards into the distant past, while using group-specific growth rates for the recent past. In the third strategy, real earning projections are obtained using a first order autoregressive process (AR[1]) specification. This estimation procedure did not give satisfactory results in terms of forecasting future earnings. For all these reasons, we opted for the first solution, which seems to be the safest. Notice that in projecting earnings forward, individuals are assumed to form expectations by using the model—in other words, for each age, we only use actual earnings up to that age and project earnings from that age forward according to the forecasting model. 6.3.3 Transitions into Retirement The INPS sample contains no information on the reasons for leaving the archive. Thus, in order to use these data, we have to make the strong assumption that every exit from the archive is due to retirement. In fact, rather than retiring, a worker could have died or moved from private-sector employment to public-sector employment or to self-employment. Our identifying assumption is that, for the range of ages that we consider (from age fifty to sixty-five), exit from the INPS archive is due to retirement and not to other reasons. This assumption is backed up by what we observe in the SHIW sample, which has available the full set of information concerning the occupational 20. Aggregate earnings are equal to the earnings series put together by Rossi, Sorgato, and Toniolo (1993) for the years before 1970 and to national account statistics for subsequent years up to 1999.

364

Agar Brugiavini and Franco Peracchi

Table 6.6

Age 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

Transitions Observed in the SHIW Panel Sample, by Age (%) From Work to Retirement

From Work to Disability

From Disability to Retirement

6.11 11.79 10.12 9.03 2.22 15.03 17.46 20.16 32.65 22.07 18.51 15.78 36.66 40.74 76.47

0.27 0.00 0.77 0.28 0.56 1.12 1.04 0.25 0.53 0.00 0.28 0.50 0.00 0.00 0.22

0.27 2.85 0.00 0.00 0.85 0.56 0.78 1.03 1.59 1.91 1.41 2.79 2.08 1.51 1.83

status in each year.21 Table 6.6 shows the transitions observed in the panel component of the SHIW sample for the years 1989, 1991, 1993, and 1995. Each entry of the table gives the proportion of transitions as a percentage of the observations in that age group starting from the sample of workers in 1989. First, it is apparent that, after age fifty-four, the transitions into retirement are substantial. Furthermore, the overwhelming majority of transitions are from work to retirement, very few cases of transitions from work to disability are observed and only some cases from disability to retirement.22 We do not report transitions from retirement back into the labor force as these are basically nonexistent, and we can safely assume that retirement is an absorbing state. This evidence suggests that, for Italian workers, the only relevant alternative escape route from the labor force is via disability. Many other “softlanding” or bridging plans exist, but they would all fall in the category of preretirement or early retirement, and in our data, they would effectively correspond to retirement.23 As far as disability benefits are concerned, after the changes legislated in 1984, their importance as an escape route has greatly diminished.24 This is shown quite clearly in figure 6.3, which pre21. In the SHIW sample, different definitions of pensioner are available based on selfreported occupational status, on earnings, and on benefits receipts. However, no marked difference in the distribution of retired people by age emerged from adopting different definitions. 22. A disability pension is automatically converted into an old-age pension when the recipient reaches the age of eligibility. 23. For example, in the SHIW sample we do know about disability insurance, but early retirement and preretirement are recorded as retirement. 24. The fall in disability benefits out of total benefits occurring after 1984 is documented in Brugiavini (1999).

Micro-Modeling of Retirement Behavior in Italy

365

Fig. 6.3 Flow of new pensions in 1997 and stock of existing pensions at the end of 1997 (in thousands) by sex and age of the beneficiary and type of pension Note: Figures are based on the INPS Archive of Outstanding Pension Benefits. This contains the universe of public pension benefits.

sents the number of new pension awards in 1997 (flow data) and the number of outstanding pensions at the end of 1997 (stock data). The data, kindly provided by ISTAT, are tabulations based on the INPS Pensions Archive and are broken down by sex and age of the beneficiary and by type of pension (old-age and early retirement, survivors, and disability).25 In the age range that we consider—fifty to seventy—the number of disability pensions is clearly negligible relative to old-age pensions, especially in the case of new awards (flow data). Disability becomes important relative to old-age pensions only if we consider the stock of pensions at ages above sixty-five, which reflects the very generous policy followed until the mid1980s, a period that we do not model in this paper. Finally, figure 6.4 presents two descriptions of the retirement hazard based on the INPS sample. The first is a nonparametric estimate based on the raw data, which shows sharp peaks at age sixty and age sixty-five and 25. The INPS pensions archive contains all outstanding pensions paid out by the INPS fund and by the other public funds, including the public-sector employees fund and the local authorities employees fund. This archive contains the universe of all public pension benefits.

366

Agar Brugiavini and Franco Peracchi

A

B

Fig. 6.4

Hazard functions: A, Men; B, Women

Note: Each picture shows two hazard functions. The nonparametric hazard is the one obtained from the raw hazard by simply relating the flow into retirement at each age to the stock of workers of that age. The baseline Cox hazard is obtained by using Cox proportional-hazard model with age as the only explanatory variable.

substantial exit already at age fifty. The second is a semiparametric estimate of the baseline hazard in a proportional hazard model with a number of basic demographic variables introduced as covariates that proportionally shift the baseline hazard. For men there is an important spike at age sixty, but a lot of action also at other ages, whereas for women there are sev-

Micro-Modeling of Retirement Behavior in Italy

367

eral important ages at which the conditional probability of leaving the labor force peaks. 6.4 Social Security Wealth and Incentive Measures Section 6.4.1 reminds the reader of the definition of social security wealth and the related incentive measures used in the econometric application. Section 6.4.2 lists our basic assumptions in the calculation of social security wealth, whereas section 6.4.3 describes the social security regimes modeled in the simulations. 6.4.1 Definition of Social Security Wealth and Incentive Measures For a worker of age a, we define social security wealth (SSW) in case of retirement at age h  a as the expected present value of future pension benefits S

∑

SSWh 

s Bs (h),

sh1

where S is the age of certain death, s  s–a s is a discount factor that depends on the rate of time discount  and the survival probability s at age s conditional on being alive at age h, and B(h) is the pension benefit expected at age s  h  1 in case of retirement at age h. Pension benefits are net of income taxes. Given the SSW, we define three incentive measures for a worker of age a. First, social security accrual (SSA) is the difference in SSW due to postponing retirement from age a to age a  1, S

SSA a  SSWa1 SSWa 

∑

s [Bs (a  1) Bs (a)] a1 Ba1 (a).

sa2

The SSA is negative if the expected present value of pension benefits foregone by postponing retirement by one year is greater than ΣSsa2 s [Bs (a  1) – Bs (a)], the expected present value of the increment in the flow of pension benefits. The rescaled negative accrual a  –SSAa /Wa1 , where Wa1 are expected net earnings at age a  1 based on the information available up to age a, is called the implicit tax or subsidy of postponing retirement from age a to age a  1. Second, peak value is PVa  maxh(SSWh – SSWa ), h  a  1, . . . , R, where R is the mandatory retirement age (the latter does not exist in Italy, but given the retirement evidence, we find it reasonable to put R equal to 70). Thus, the peak value is the maximum difference in SSW between retiring at future ages and retiring at the current age. Third, option value is OVa  maxh (Vh – Va ), h  a  1, . . . , R, where S

Va 

∑

sa1

s [kBs (h)] 

368

Agar Brugiavini and Franco Peracchi

is the intertemporal expected utility of retiring at age a, while h

Vh 

S

∑

sW s 

sa1

∑

s [kBs (h)] 

sh1

is the intertemporal utility of retiring at age h a. Thus, the option value is the maximum utility difference between retiring at future ages and retiring at age a. We parameterize the model by  equals 1 and k equals 1.25. Under these assumptions, Va  1.25 SSWa and h

Vh 

∑

sW s  1.25SSWh .

sa1

If expected earnings are constant at Wa (as assumed in our earnings model), then h

Vh – Va  Wa

∑

s  1.25(SSWh – SSWa ).

sa1

That is, the peak value and the option value are proportional to each other except for the effect due to the term Σhsa1 s . These three incentive measures are consistent with the view that, in deciding whether or not to retire, a worker compares the expected gain from each of the two alternatives. Note that, in computing the incentive measures, we used the assumption that workers revise their expectations at each age. Hence for each given age, a worker projects the path of observed earnings according to the model and then computes their SSW and the incentive variables, taking into account the information currently available. This requires recomputing SSW and the corresponding incentive measures for each year until retirement. Reforms are assumed to come as a surprise to workers. As is clear from the above formula, the social security accrual depends crucially on the expected present value s [Bs (a  1) – Bs (a)] of the increment in pension benefits at age s resulting from postponing retirement by one year. Let t denote the number of years of contributions for a worker of age a, and assume that pension benefits remain constant in real terms. Under the pre-1993 regime (but also during the transitional period that we assume as the baseline regime), if t 40 then Bs (a  1) Bs (a)  0.02t[W
(a  1) W
(a)]  0.02W
(a  1), where W
(a) denotes pensionable earnings in case of retirement at age a; whereas if t  40, then Bs (a  1) Bs (a)  0.8[W
(a  1) W
(a)]. In the special case when W
(a  1)  W
(a)  W
, we obtain Bs (a  1) Bs (a) 


, t 40 0.02W 0, otherwise.

Micro-Modeling of Retirement Behavior in Italy

369

Under the 1995 reform, we instead have Bs (a  1) Bs (a)  (a  1)M(a  1) (a) M(a), where  (a) represents the legislated conversion factor used to transform the worker’s “notional account” into a pension annuity, and M (a) denotes the value of such a notional fund in case of retirement at age a. If the fiveyear moving average of GDP growth rates (used to capitalize past earnings) is positive, then the difference [Bs (a  1) – Bs (a)] is always positive because M (a  1) M (a) and  (a  1)   (a), (with equality for a 65). Note that, for many ages and under most regimes, SSW is a monotonically decreasing function of age. In all these cases, the maximum value of SSW over current and future years is attained at the current age. Therefore, once the eligibility criteria are met, the age profile of the peak value looks very similar to the age profile of the accrual. This explains why the pictures portraying the age profile of the peak value and the accrual look often similar (see section 6.4.2). We do have cases, however, of local maxima (e.g., under the 1995 reform). 6.4.2 Basic Assumptions and Calculation of Social Security Wealth In the actual calculation of SSW, we assume a real discount factor of 1.5 percent ( equals 0.985). Benefits are defined in real terms and the indexation rules prevailing under each legislation are implemented (e.g., before the 1992 reform, we apply indexation to both price inflation and real wages). We also assume that real earnings growth after 1994 (the last year of the INPS sample) is constant at 1.5 percent. We carry out calculations as follows. 1. We estimate SSW for men and women separately. 2. Unlike most other countries in this project, we assume that workers are single. In fact, from the data, we are unable to tell whether or not a worker has a spouse. In the Italian legislation, the only major difference between a single worker and a married worker is eligibility to survivors’ pension (there is no dependent-spouse benefit26). We did not attempt an imputation procedure to assign workers a spouse.27 3. Disability benefits have not been taken into account because multiple exit routes are not relevant in the Italian case. We experimented with 26. There is a difference in the rebates on income tax and in the calculation of minimum benefit, particular in the way in which means testing is carried out. 27. We could readily generalize our calculations to include pensions to a surviving spouse. For example, we could randomly assign to men a wife who is three years younger. However, this imputation would not produce extra variability in SSW, because pension wealth of survivors depends on the characteristics of the deceased and on the survivor’s age. Hence, in order to have an impact on the variability of the profile of SSW by age, we should also assign randomly the age of the imputed wife, which may produce a considerable amount of noise for a little gain.

370

Agar Brugiavini and Franco Peracchi

adding this alternative exit route in an ad hoc fashion by using the observed disability probabilities, but this had no effect on the main findings. 4. We do not account for the lump-sum benefit represented by the TFR, since the paper by Brugiavini (1999) shows that this lump sum benefit does not alter dynamic incentives. 5. We assume variation in mortality only by sex and age. 6. We look at all individuals considered at risk (i.e., aged fifty to seventy). Note that, in our sample, the first worker who becomes eligible for retirement under the baseline regime is aged fifty, and the econometric application makes use of the full range of ages between fifty and seventy. However, since there are very few exits until age fifty-two and very few individuals working after age sixty-five, in some tabulations (e.g., table 6A.1 in the appendix) we present results starting at age fifty-three and group in a single age interval all workers aged sixty-five and over. In estimating the model, we also had to deal with the fact that the actual age of entry into the labor market is not always known. We used the information on the initial occupational level to get a reasonable proxy for educational attainments. This was then used to impute an initial age for the worker’s contributive history. Eligibility rules and benefit computation rules prevailing under each regime are rather complex (see section 6.2), and some shortcuts were made. Finally, we computed SSW net of income tax, by subtracting from gross pension benefits income taxes as due. 6.4.3 Social Security Regimes Modeled in the Simulation We estimate SSW and incentive variables under five alternative regimes:28 1. The rules prevailing before 1993; 2. The rules actually prevailing during the transitional phase that starts in 1993, and this case represents our baseline; 3. The rules prevailing in the steady state, once the 1995 reform is fully phased in; 4. Policy simulation 1 in which, starting from the current system (the baseline case), we raise the normal retirement age by three years while holding constant all other features; and 5. Policy simulation 2, which entails a different pension program altogether that features an early retirement age of sixty and a normal retirement age of sixty-five. It provides a retiree with a benefit replacing 60 per28. We actually also simulated the rules prevailing after 1992 in the steady state once the Amato reform is fully phased in. But, since this case is only of theoretical interest and the 1992 reform will never reach its steady state values as subsequent changes took place in 1995 and 1997, we did not report results for this case.

Micro-Modeling of Retirement Behavior in Italy

371

cent of their projected earnings when they turn sixty-five. It applies an actuarial reduction of 6 percent per year for early claiming and an actuarial increase of 6 percent per year for later claiming. It essentially makes early retirement costly and introduces age neutrality in retirement choices. To simplify the presentation, we only report the results obtained for the baseline, the 1995 reform, and the two policy simulations. Table 6A.1 of the appendix presents descriptive statistics (mean, standard deviation, and selected percentiles) for the three incentive measures. In the baseline case, eligibility is reached very early and SSW tends to decrease monotonically after the eligibility age. By comparing the two initial panels of table 6A.1, one has a first impression of the effects of the 1995 reform: under the post-1995 regime, SSW becomes positive only late in life, and median SSW is lower. We also report descriptive statistics for the implicit tax on work, an incentive measure that was used extensively in Gruber and Wise (1999). On average, the estimated implicit tax in our sample is fairly close to the one obtained for a reference individual in the Gruber and Wise study.29 In particular, taking the comparable case of a single male worker under the pre-1993 legislation, the average implicit tax of the sample is in line with the one computed in Brugiavini (1999), but it differs substantially at the two crucial ages of sixty and sixty-five. Figures 6.5 through to 6.8 complement this information by presenting nonparametric estimates of the density for each incentive measure and for SSW. This is done first under different policy regimes separately for men and women (figures 6.5 and 6.6) and then at selected ages for men only (figures 6.7 to 6.8). These estimates have been computed by the kernel method using only the observations with a positive value of the SSW. The effects of the reform on the density of the relevant variables produced by the policy simulation 1 (three-year delay) are negligible, and we do not report them in these graphs. The 1995 reform and the policy simulation 2 (actuarial adjustment) have a more marked impact. For men, both reforms imply a shift to the left of the distribution of SSW (see figure 6.5) and, hence, a lower mean and no substantial change in dispersion. However accrual and peak values show a substantial change in the level of concentration around the mean as a result of the reforms, while the option value exhibits an interesting bimodal shape. For women, as a result of the reforms, there are more important changes in the location and dispersion of all variables, including SSW. For men, we also look at the distribution of SSW and all incentive measures for a few selected ages. These distributions are much more sensitive to age under the baseline than under the reforms. This confirms the point that the pre-1993 legislation (essentially the baseline) was highly non-neutral with respect to age. 29. In the Gruber and Wise study (1999; see Brugiavini 1999 for Italy) the reference individual is basically a worker characterized by the median earnings of his cohort.

372

Agar Brugiavini and Franco Peracchi

Fig. 6.5 Kernel-estimated density of social security wealth and incentives (SSW in 1998 euro  104), men Note: We do not show the results for the plus-three-years reform as these almost entirely overlap with the baseline densities.

6.5 Modeling Retirement Choices In this section we present the results of modeling exit into retirement using probit models that include, in addition to a standard set of covariates (such as age, occupation, and sector), the incentive measures discussed in the previous section. Table 6.7 provides mean values and standard deviations of the relevant variables. 6.5.1 Probability of Retirement We present two tables of probit estimates, one for men (table 6.8) and one for women (table 6.9). In either case, the response variable is a binary indicator, representing exit from the INPS sample between the year t and the year t  1. As discussed in section 6.4, we assume that exit from the INPS sample corresponds to retirement. The population at risk consists of workers aged between fifty and seventy in any of the relevant years. The sample used for estimation includes all pairs of years from 1980 to 1981

Micro-Modeling of Retirement Behavior in Italy

373

Fig. 6.6 Kernel-estimated density of social security wealth and incentives (SSW in 1998 euro  104):Women Note: We do not show the results for the plus-three-years reform as these almost entirely overlap with the baseline densities.

through 1993 to 1994. Because in some cases we have to model earnings profiles going back fifty years, given the existing limitations on aggregate wage data, we restrict the analysis to individuals at risk after 1980. In this way, our oldest worker is aged seventy in 1980, and we only need to backcast earnings to the year 1930.30 For each incentive measure, two basic specifications are considered for a total of six estimated models. The first column of each table corresponds to the use of the accrual as the incentive measure and shows the results obtained for a general specification that includes, in addition to accrual and SSW, a set of sectoral and regional indicators, a linear age term, and a set of earnings measures relevant for the retirement choice (i.e., a quadratic polynomial in expected earnings and a quadratic polynomial in pensionable earnings). It should be noted that, for all ages, expected earnings in the 30. Retirement is not mandatory. Given that we assume an individual at risk up to age seventy and given that we cannot exclude that they started working at age twenty, we cannot rule out the possibility that this individual worked for fifty years.

374

Agar Brugiavini and Franco Peracchi

Fig. 6.7 Kernel-estimated densities, social security wealth, and incentives for selected ages: Men—baseline

next period are computed based on the projection model and the information available to the worker at that age. We refer to this as model M1. The second column shows the results for a specification in which the effect of age is modeled through a set of age dummies, rather than a linear age term, while the rest of the specification is unchanged; we refer to this as model M2. The same pattern is repeated for the other columns: columns (3) and (4) correspond to the use of the peak value with, respectively, a linear age trend (model M1) and a set of age dummies (model M2), while columns (5) and (6) correspond to the use of the option value. Since coefficients from a probit analysis do not have an immediate interpretation, we also provide the probability effect of the relevant variables. These are shown underneath the coefficient estimated for SSW and for each incentive variable. Since the scale of each of these variables differs, we measure the probability effect (for the reference individual) by increasing the variable of interest from the mean level to the level of the mean plus one standard deviation, holding the other variables constant. In our baseline specification, SSW and all incentive measures are computed according to the transitional rules introduced after 1992. This represents the relevant regime for the workers in our sample. Overall, using

Fig. 6.8 Kernel-estimated densities, social security wealth and incentives for selected ages: A, Men—common reform; B, Men—1995 reform

A

Fig. 6.8 (cont.) Kernel-estimated densities, social security wealth and incentives for selected ages: A, Men—common reform; B, Men—1995 reform

B

Table 6.7

Summary Statistics of Variables Used in the Probit Analysis Women

Variable Year Age No. of jobs held Industry dummy Agriculture dummy Building dummy Trade dummy Transport dummy Financial dummy Other sector dummies Earnings Years of contribution SSW Baseline 1995-ref S1 S2 Option value Base Ref S1 S2 Peak value Baseline 1995-ref S1 S2 Accrual Baseline 1995-ref S1 S2 Retirement dummy Northwest Northeast Center South Islands Blue-collar White-collar Manager Other levels

Men

Mean

SD

Mean

SD

87.74 53.06 0.58 0.47 0 0.00 0.14 0.02 0.03 0.30 13,736.26 32.44

4.07 3.51 1.11 0.49 0 0.08 0.35 0.15 0.17 0.46 52.88 3.93

87.14 53.80 0.91 0.50 0.01 0.11 0.08 0.09 0.04 0.14 18,953.01 33.12

4.17 3.62 1.58 0.49 0.10 0.31 0.27 0.29 0.20 0.35 14,495.15 4.43

0.52 0.17 0.22 0.07

0.79 0.45 0.56 0.28

0.53 0.27 0.25 0.07

0.73 0.56 0.55 0.29

236.93 312.23 290.92 306.90

179.69 176.62 173.39 162.65

269.36 343.64 313.94 335.34

240.63 260.39 239.15 232.58

101.84 118.09 137.40 113.16

86.03 63.71 72.49 43.15

87.60 106.36 115.64 98.40

82.20 78.42 74.77 48.71

–2.00 0.27 –0.72 –0.01 0.12 0.45 0.21 0.27 0.04 0.01 0.62 0.32 0.00 0.04

7.27 1.36 3.87 1.75 0.33 0.49 0.41 0.44 0.19 0.12 0.48 0.46 0.07 0.20

–2.02 0.34 –1.11 –0.15 0.11 0.38 0.22 0.21 0.11 0.05 0.73 0.23 0.03 0.00

6.32 2.51 4.32 2.85 0.31 0.48 0.41 0.41 0.32 0.22 0.44 0.42 0.17 0.04

Note: SD = standard deviation. Earnings are in 1998 euros; SSW is in 1998 euro
(106). SSW and incentive variables are distinguished according to the baseline case, the 1995 reform, and the two hypothetical reforms.

–0.023** (0.004) –0.244 –0.128** (0.048) –0.161 –0.150** (0.006) 0.105** (0.008) 0.117** (0.008) –0.070** (0.011) 0.093** (0.008)

M1 (1)

M2 (2)

1.34** 1.33** 1.43** 1.53** 1.49** 1.75** 1.48** 1.61**

(0.259) (0.259) (0.258) (0.258) (0.262) (0.261) (0.266) (0.267)

–0.024** (0.004) –0.079 –0.071** (0.053) –0.022 –0.147** (0.006) 0.102** (0.008) 0.114** (0.009) –0.068** (0.011)

Accrual

0.005** (0.013) 0.075 0.006** (0.125) 0.088 –0.161** (0.007) 0.109** (0.008) 0.131** (0.008) –0.075** (0.011) 0.101** (0.009)

M1 (3)

M2 (4)

1.35** 1.35** 1.42** 1.56** 1.54** 1.79** 1.54** 1.68**

(0.259) (0.259) (0.258) (0.258) (0.262) (0.261) (0.266) (0.267)

–0.001** (0.013) –0.314 –0.095** (0.130) –0.080 –0.159** (0.007) 0.106** (0.008) 0.132** (0.009) –0.073** (0.011)

Peak Value

Male Retirement Probit Analysis (INPS sample 1980–93, population at risk aged 50–70)

Earnings (T + 1) Earnings (T + 1) 2 Pensionable earnings Pensionable earnings 2 Age 50 51 52 53 54 55 56 57

SSW

Incentive

Table 6.8

0.008** (0.0005) 0.504 0.091** (0.096) 0.141 –0.176** (0.008) 0.105** (0.008) 0.127** (0.009) –0.070** (0.012) 0.104** (0.009)

M1 (5)

M2 (6)

1.34** 1.33** 1.43** 1.53** 1.49** 1.75** 1.48** 1.61**

(0.259) (0.259) (0.258) (0.258) (0.262) (0.261) (0.266) (0.267)

–0.005 (0.0006) –0.185 0.088** (0.107) 0.028 –0.168** (0.009) 0.104** (0.009) 0.127** (0.009) –0.070** (0.012)

Option Value

7,446 0.322 –1765

–1.353* (0.094)

(0.268) (0.269) (0.267) (0.296) (0.135) (0.340) (0.356) (0.435) (0.516) (0.263)

7,446 0.342 –1716

1.70** 1.81** 2.51** 1.88** 1.99** 2.10** 2.55** 1.95** 2.25** –2.451 7,446 0.316 –1782

–1.435* (0.129)

(0.268) (0.269) (0.267) (0.296) (0.315) (0.340) (0.356) (0.435) (0.516) (0.138)

7,446 0.335 –1734

1.76** 1.86** 2.59** 1.88** 1.99** 2.10** 2.55** 1.95** 2.25** –1.345 7,446 0.316 –1781

–1.497* (0.105)

(0.268) (0.269) (0.267) (0.296) (0.315) (0.340) (0.356) (0.435) (0.516) (0.114) 7,446 0.334 –1734

1.70** 1.81** 2.51** 1.88** 1.99** 2.10** 2.55** 1.95** 2.25** –2.518*

Notes: N = number of observations. Dependent variable is a dummy for being retired. Standard errors in parentheses. The number under the coefficient indicates the probability effect measured by the percentage change in the probability. The difference in probabilities is obtained by first generating a projected probability for the reference individual (mean value of continuous variables and 0 for the dummy variables) at the mean and then generating a projected probability at the mean plus 1 standard deviation of the relevant variable (other things equal). Hence the change in the independent variable depends on the scale of the variable itself. For example, in the case of SSW, we start from a mean of €538 thousand and reach a value of €1,271 thousand (mean plus 1 standard deviation). The reference category for the age dummies is age 49. All specifications include region, sector of employment, and number of jobs held, and because of sample size, the specifications of dummy variables for women change slightly from the one used for men (see table 6.10). **Significant at the 5 percent level.

N Pseudo R 2 Log-likelihood

58 59 60 61 62 63 64 65 66+ Constant

–0.020** (0.006) –0.240 –0.014** (0.094) –0.015 –0.240** (0.025) 0.260** (0.089) 0.199** (0.034) –0.219** (0.115) 0.062** (0.018)

M1 (1)

M2 (2)

0.97** 1.09** 1.00** 0.97**

(0.266) (0.268) (0.273) (0.282)

–0.021** (0.007) –0.056 0.143** (0.159) 0.024 –0.241** (0.025) 0.288** (0.088) 0.187** (0.034) –0.260** (0.115)

Accrual

–0.004** (0.002) –0.522 0.321** (0.248) –0.345 –0.241** (0.028) 0.282** (0.092) 0.229** (0.034) –0.292** (0.118) 0.059** (0.019)

M1 (3)

M2 (4)

1.04** 1.21** 1.14** 1.16**

(0.282) (0.185) (0.291) (0.303)

–0.007** (0.002) –0.907 –0.476** (0.291) –0.094 –0.229** (0.028) 0.282** (0.091) 0.238** (0.035) –0.315** (0.117)

Peak Value

Female Retirement Probit Analysis (INPS sample 1980–93, population at risk aged 50–70)

Earnings (T + 1) Earnings (T + 1)2 Pensionable earnings Pensionable earnings2 Age 50 51 52 53

SSW

Incentive

Table 6.9

–0.007** (0.001) –0.231 0.030** (0.251) –0.002 –0.249** (0.034) 0.299** (0.090) 0.216** (0.035) –0.272** (0.118) 0.061** (0.021)

M1 (5)

M2 (6)

1.05** 1.18** 1.07** 1.03**

(0.288) (0.289) (0.292) (0.301)

–0.004** (0.002) –0.685 –0.359** (0.310) –0.064 –0.213** (0.035) 0.327** (0.089) 0.232** (0.035) –0.318** (0.118)

Option Value

Note: See table 6.9.

N Pseudo R 2 Log-likelihood

54 55 56 57 58 59 60 61 62 63 64 65 66+ Constant

1770 0.319 –458

–1.069** (0.229)

(0.332) (0.338) (0.348) (0.369) (0.406) (0.431) (0.383) (0.441) (0.565) (0.531) (0.674) (0.680) (0.680) (0.201)

1770 0.343 –443

0.57** 1.10** 1.28** 1.02** 0.59** 0.50** 1.33** 1.10** 0.69** 1.10** 0.92** 1.15** 1.15** –1.750* 1770 0.314 –461

–0.795** (0.275)

(0.345) (0.345) (0.355) (0.373) (0.402) (0.430) (0.383) (0.442) (0.558) (0.529) (0.665) (0.654) (0.654) (0.283) 1770 0.342 –443

0.71** 1.21** 1.34** 1.09** 0.67** 0.57** 1.43** 1.08** 0.73** 1.06** 0.91** 1.10** 1.10** –1.363* 1770 0.313 –463

–1.032** (0.246)

(0.345) (0.350) (0.357) (0.374) (0.374) (0.428) (0.382) (0.440) (0.556) (0.529) (0.659) (0.656) (0.656) (0.271) 1770 0.339 –445

0.71** 1.27** 1.37** 1.13** 0.70** 0.58** 1.44** 1.06** 0.71** 1.03** 0.87** 1.05** 1.05** –1.556*

382

Agar Brugiavini and Franco Peracchi

instead the pre-1993 rules leads to negligible differences in terms of SSW and eligibility for our sample of workers. This is because, as already mentioned, the rights of workers near retirement were changed only marginally by the reforms of the 1990s. All the specifications are satisfactory in terms of explaining variability in the data, as indicated by pseudo R 2 values ranging between 32 and 34 percent. The use of age dummies increases the fit relative to the model with a linear age term, but only marginally. This suggests that age is an important determinant of retirement decisions but, despite the presence of impressive spikes in the hazard, we get only marginal gains by making use of a fully parameterized model. Hence, these spikes may be less important than they first appear in explaining the age-retirement process, as most of the action comes from the exits taking place between age fifty and age sixty. For men, the accrual is the only incentive variable that is statistically significant and has the expected negative sign. For women, instead, the incentive variables have the expected sign and are significant, at least for the specifications with age dummies. For the estimates based on the accrual, we also find a nonnegligible probability effect of the incentive variable (a one-percent increase in the accrual decreases the retirement probability by 24 percent). In some cases, SSW has a negative effect on retirement. This is somewhat surprising as it suggests that workers with higher levels of SSW tend to postpone retirement (i.e., have a taste for work) even after controlling for the type of job and the occupational sector. However, the negative coefficient is hardly ever significant. A better grasp of the importance of the age effects can be gained by looking at figure 6.9. In the top part, we present results for the general model with age dummies (M2). We plot the raw hazard computed in the estimation sample versus the projected hazards both for men (left panel) and women (right panel). The projected hazards are obtained, on the basis of the general model M2 and for the baseline case, by setting the incentive variables and all continuous variables to their mean value and setting to zero all dummies, except the age dummies. The different lines drawn for the projected hazards on the same graph correspond to the different incentive measures. Figure 6.10 compares the hazard function and the cumulative distribution function (CDF) of the raw data with those implied by our estimates of model M2 (the baseline). The raw hazard and the raw CDF have a number of interesting features. In particular, while the hazard shows significant spikes at age fifty-five and sixty for men (more spikes for women), it is clear from the CDF that half of the sample has already retired by age fifty-seven for men and by age fifty-five for women. The results obtained for the two models described above (models M1 and M2) suggest that the linear age term does not capture the important spikes in the data, but the use of a full set of age dummies provides an age profile for the hazard that is fairly close to the raw hazard, although at the cost of saturating the model.

Micro-Modeling of Retirement Behavior in Italy

Fig. 6.9

383

Age effects in model M2 and model M3

Note: The left panels are for men and the right panels for women. The continuous line is the raw hazard. The estimated hazards are based (a) on the complete model M2 for each of the incentive variable (top panels) and (b) on the parsimonious representation based on a cubic in age plus the relevant dummies M3 (bottom panels). We convert in probability space the age effects for a representative worker. This worker is characterized by the mean values of continuous variables and zero for all the dummy variables (apart from age and age dummies). Results are based on estimation sample.

In order to separate the effect of age on retirement due to preferences from the age effect related to incentives and to provide a parsimonious representation of the age effects, we carry out an additional probit analysis in which we replace the linear age term by a combination of a cubic function of age-plus-three dummies at ages fifty-five, sixty, and sixty-five.31 This model, which we call model M3, tries to capture the fact that the raw hazard tends to increase smoothly with age, except for the presence of three important spikes at ages fifty-five, sixty, and sixty-five. We interpret the estimated cubic-age trend as the “pure” effect of preferences on retirement. The results of the probit analysis are not presented for the sake of brevity. We simply note that the main results obtained for model M2 remain valid for model M3 and that the age effects are significant overall. We use this specification M3 in two ways: (a) directly, to make projections on the basis 31. For women, only two age dummies at fifty-five and sixty are added.

384

Agar Brugiavini and Franco Peracchi

Fig. 6.10

Hazard and CDF: Raw data versus baseline estimates in model M2

Note: The top panels present the hazard functions while the bottom panels present the corresponding CDF. The results for the baseline are obtained from the general model M2. These graphs are based on the estimation sample (7,446 observations for men and 1,770 observations for women) where individuals drop from the sample at the age of actual retirement. The picture looks much different for the simulation sample, where individuals are assumed to be at risk up to a maximum age of seventy.

of an “intermediate” specification in simulating retirement decisions, and (b) indirectly, to impose the estimated age effect on the results of specification M2 in order to “purge” the effect of age-preferences from the general model M2.32 Results are described in figure 6.9 (bottom panels), where the age effects implied by model M3, along with the raw hazard, are graphed 32. By intermediate specification we mean one that models age in a satisfactory way without saturating the model. In view of the 1995 reform, which tries to implement an actuarially fair system, it would be reasonable to remove the three dummies and to allow the cubic specification in age to capture taste for retirement. However, in our sample, which does not cover the postreform years, we have a basic identification problem on the interpretation of the spikes occurring at particular ages, as these may be mostly due to legislation but could also emerge as a result of habits and peer effects in retirement choices.

Micro-Modeling of Retirement Behavior in Italy

385

against age for both men and women. This specification fits closely the age pattern observed in the data, particularly for men. 6.5.2 Simulating Retirement Choices We carry out simulations separately for men and women. This is done on a simulation sample in which earnings are projected forward as well as backward to cover the necessary time span, regardless of actual entry into the sample and of possible gaps in earnings. In all cases, we maintain the assumption that policy changes are not anticipated. We perform six types of simulations based on three policy variants to be contrasted with the baseline. The first policy variant envisages a forward shift of three years in the retirement ages. The second policy variant (referred to as the “common reform”) consists of an actuarial adjustment of benefits of 6 percent per year and is designed to make early retirement costly to individuals. The last variant is the actual 1995 reform of the Italian social security system. Simulation S1 (Model M1) Starting from the model, which only includes a linear age trend (model M1), we project the estimated hazard on the simulation sample by changing the SSW and the incentive variables according to the chosen reform. Results for men are shown in figure 6.11 (accrual and peak value) and the left panels of figure 6.12 (option value). The effect of a policy change is significant only in the case of the accrual. The largest effect is under the common reform. Results are not much different for women (left panels of figure 6.14). Simulation S2 (Model M2) Starting from the model with age dummies (model M2), we project the estimated hazard by changing the incentive variables according to the chosen reform while leaving all the other variables (including the age dummies) unchanged. Since the model is saturated and the age dummies are very important in explaining variability in retirement probabilities, this case leaves little room for the effects of policies. Results are shown for men in figure 6.12 (right panels) and figure 6.13. The only case where the effect of a policy change is nonnegligible is for the accrual (figure 6.12 M2-S2 for men and figure 6.14 for women). Simulation S3 (Model M2) Starting from the model with age dummies (model M2), we project the estimated hazard by changing all the relevant variables (incentive variables and age dummies) according to the policy change. In this experiment we want to measure the impact of policies going through the incentives and through a direct effect of age, and hence it is important to net out the effect that age has on retirement due to individual preferences. To this end, in

386

Agar Brugiavini and Franco Peracchi

Fig. 6.11

Simulation M1-S1 for accrual and peak value, men

Note: The top panels show the hazards and the bottom panels the corresponding CDF. Results are obtained by using the coefficients from the baseline M1 estimated on the estimation sample of 7,446 men for each year and projected both on the baseline data and on the policy simulation data.

accordance with the policy regime, the age dummies are adapted ex post facto. For the policy entailing a three-year shift in retirement age, this feature is modeled in the incentive variables and subsequently imposed on the age dummies (effectively shifting the hazard to the right by three years). In the case where the reform entails an actuarial adjustment by age, agedummy coefficients are adjusted by assuming that the underlying hazard should be smoother than the observed one. In order to impose this behavior, we make use of the information obtained in model M3 and adjust the age dummies according to a cubic function in age. The final effect of using model M3 is to preserve the important breaking points while making the rest of the simulated hazard smoother. Results for men are presented in figures 6.15 (accrual and peak value)

Micro-Modeling of Retirement Behavior in Italy

Fig. 6.12

387

Simulations M1-S1 for option value and M2-S2 for accrual, men

Note: The top panels show the hazards and the bottom panels the corresponding CDF. Results are obtained by using the coefficients from the baseline M1 estimated on the estimation sample of 7,446 men for each year and projected both on the baseline data and on the policy simulation data.

6.16 (option value) and for women in figure 6.17 (accrual only). The effects of policies are more marked than in the previous cases. In particular, both policies imply a tendency to delay retirement, as it is clearly documented by the CDF. The common reform has a stronger impact than the three-year shift. It is also interesting to note that under the common reform, the expost-facto adjustment of the age dummies smoothes out most spikes, leaving only one important spike at age sixty. Simulation S4 (Model M2) In this simulation, we consider the effects of the 1995 Dini reform described in section 6.2. This reform is in many respects similar to the common reform, as it computes benefits at each age according to an actuarial

388

Agar Brugiavini and Franco Peracchi

Fig. 6.13

Simulations M2-S2 for peak value and option value, men

Note: The top panels show the hazards and the bottom panels the corresponding CDF. Results are obtained by using the coefficients from the baseline M1 estimated on the estimation sample of 7,446 men for each year and projected both on the baseline data and on the policy simulation data.

adjustment factor and restricts retirement ages to a window. For this simulation, we contrast the estimated hazard with the baseline by looking at models M2 and M3 only. The results presented in figure 6.16 (right panels) suggest that the incentive measures do not have a strong impact in this case. Simulation S5 (Model M3) We look at the results obtained for the intermediate model M3 by contrasting the baseline and the policy changes. The assumption here is that the age polynomial should capture only preferences. Results for men are shown in figures 6.18 and 6.19. In particular, figure 6.18 and the left panel of figure 6.19 present the two hypothetical reforms, while the right panels

Micro-Modeling of Retirement Behavior in Italy

Fig. 6.14

389

Simulations M1-S1 and M2-S2 for accrual, women

Note: The top panels show the hazards and the bottom panels the corresponding CDF. Results are obtained by using the coefficients from the baseline M1 and then M2 estimated on the estimation sample of 1,770 women for each year and projected both on the baseline data and on the policy simulation data.

of figure 6.19 are for the actual 1995 reform. When incentives are measured by the accrual, there is a nonnegligible effect towards delayed retirement for both the common reform and the 1995 reform. Finally, in figure 6.20 we carry out an experiment (only by making use of the accrual) that simulates the relevant policy changes on the basis of model M3. This is done by applying the incentive variables of the threeyear adjustment and the common reform on the basis of coefficients estimated under specification M3. At the same time, we shift forward the three important spikes (age fifty-five, sixty, and sixty-five) by three years, while leaving the cubic polynomial in age in its original form. Figure 6.20 presents the results, it suggests that we obtain significant changes in the hazard

390

Agar Brugiavini and Franco Peracchi

Fig. 6.15

Simulations M2-S3 for accrual and peak value, men

Note: The top panels show the hazards and the bottom panels the corresponding CDF. Results are obtained by using the coefficients from the baseline M1 estimated on the estimation sample of 7,446 men for each year and projected both on the baseline data and on the policy simulation data.

not only for the common reform, but also an impact on the CDF, leading to a nonnegligible increase in mean retirement age (see table 6.10). 6.6 Conclusions This paper analyses retirement behavior of Italian workers by first estimating probit models and then making use of the econometric model to simulate exits from the labor force. Results are mixed. The probit analysis provides, overall, a good fit for the estimated retirement hazards and the correct sign for the incentive variables (i.e., when the dynamic incentives increase workers tend to delay retirement). When comparing the models with actual behavior, one sees that,

Micro-Modeling of Retirement Behavior in Italy

391

Fig. 6.16 Simulations M2-S3 for option value and simulation M2-S4 based on the legislation of the 1995 reform, men Note: The top panels show the hazards and the bottom panels the corresponding CDF. Results are obtained by using the coefficients from the baseline M2 estimated on the estimation sample of 7,446 men for each year and projected both on the baseline data and on the policy simulation data. The right panels refer to the reform enacted in Italy in 1995 known as the Dini reform.

despite the adequate fit, all models tend to underestimate actual retirement as measured by mean retirement age (see table 6.10 for men). However, in this paper, the important comparison is between the baseline scenario and the simulated reforms. The reforms are implemented in two steps: first by allowing only for a change in the incentives (SSW as well as dynamic incentives) and then by looking at the full impact of the reforms through eligibility. However, while the effects of the reforms, as captured by the incentive variables, are clearly seen on the hazard, these are not of a significant magnitude. Of the incentive variables, the most effective is the accrual, and in some cases (e.g., model M2, simulation S3), we can see a substantial change in the hazard caused by the policy change. In particular,

Fig. 6.17

Simulations M2-M3 for accrual, women

Note: The top panels show the hazards and the bottom panels the corresponding CDF. Results are obtained by using the coefficients from the baseline M1 and then M2 estimated on the estimation sample of 1,770 women for each year and projected both on the baseline data and on the policy simulation data.

Fig. 6.18

Simulation M3 based on a flexible specification of the age effects, men

Notes: Results for the simulated hazards are obtained by using the coefficients from the baseline M3 estimated on the estimation sample of 7,664 men for each year. The left panel makes use of the accrual value and the right panels of the peak value for the baseline and for the policy changes.

Micro-Modeling of Retirement Behavior in Italy

Fig. 6.19

393

Simulation M3 based on a flexible specification of the age effects, men

Note: Results for the simulated hazards are obtained by using the coefficients from the baseline M3 estimated on the sample of 7,664 men for each year. The left panel makes use of the option value for the standard reforms and the right panels make use of the accrual value for the 1995 reform.

both the hypothetical common reform (based on an actuarially-fair scheme) and the actual 1995 reform show a clear move toward an ageneutral system as opposed to the baseline scenario. However, this is not always sufficient to produce significant changes in mean retirement age and in the unconditional retirement probabilities as described by the CDF.

Fig. 6.20

Simulation M3 based on flexible specification of the age effects

Note: The left panels are for men and the right panels for women. Results for the simulated hazards are obtained by using the coefficients from the baseline M3. In this model, we shift the three important dummies by three years in order to impose an age effect due to the policy change. Table 6.10

M1-S1-ACC M1-S1-Peak V M1-S1-OV M2-S2-ACC M2-S2-Peak V M2-S2-OV M2-S3-ACC M2-S3-Peak V M2-S3-OV M3-ACC M3-S5-ACC

Estimated Mean Retirement Age of Men (estimation sample: N
7,664) Baseline

Three-Year Reform

Common Reform

1995 Reform

55.23 55.26 55.28 55.97 55.92 55.93 58.35 58.40 58.43 55.71

55.23 55.19 55.18 56.01 55.93 55.97 58.45 58.42 58.45 55.70 55.84

55.24 55.25 55.16 55.81 55.80 55.94 58.41 58.32 58.66 55.74 55.88

55.43 55.19 54.95 55.71 55.92 55.81 58.58 58.43 58.38 55.90

Notes: Sample mean retirement age = 56.63. ACC  accrual value; OV  option value. Unconditional mean retirement age is based on the CDF corresponding to each case (for example, M2-S3-OV gives the mean retirement age corresponding to the simulated retirement probabilities based on model M2, policy S3, when the incentive variable is the option value). For model M3, case S5, ACC, mean retirement ages for the baseline and the 1995 reform are not shown as the eligibility rules of these cases are not coherent with the policy change under investigation. Blank cells indicate that no calculation was done in that case.

Micro-Modeling of Retirement Behavior in Italy

395

Appendix Table 6A.1

Summaries of the Distribution of Social Security Wealth and Incentive Measures for Men, by Age Accrual

Age

Median SSW

Median

10%

Observations

From Volume

53 54 55 56 57 58 59 60 61 62 63 64 65

0 112.48 129.96 135.52 142.30 137.97 134.32 129.30 128.73 126.19 124.46 120.36 101.24

0 0 –1.00 –1.25 –2.61 –2.66 –3.69 –9.76 –4.42 –4.87 –4.96 –4.49 –4.91

0 –4.60 –7.68 –7.69 –16.78 –16.19 –13.40 –21.44 –23.55 –23.23 –23.037 –20.233 –20.94

Baseline (Transition) Case 0 1.20 .051 6.40 0 6.50 .082 7.04 .867 8.48 .666 8.48 0 7.10 –1.86 11.075 –.111 13.54 0.983 11.82 –1.00 10.80 1.63 17.15 –1.99 13.32

0 0 .102 .124 .287 .298 .403 1.634 .528 .591 .628 .487 0.923

712 641 600 516 466 405 333 279 110 69 46 32 39

.282 .301 .326 .356 .378 .623 .632 .633 .638 .648 .651

53 54 55 56 57 58 59 60 61 62 63 64 65

0 0 0 0 116.62 115.51 114.48 112.90 117.25 114.98 111.96 117.41 95.36

0 0 0 0 0.33 1.42 –.59 –1.23 –.775 .084 .065 .132 –2.18

0 0 0 0 –1.065 –1.195 –2.00 –2.59 –2.69 –1.08 –.93 –.94 –3.14

Post-1995 (DINI) Case 0 0 0 0 0 0 0 0 4.43 3.41 7.23 5.19 2.90 4.06 2.24 4.62 6.64 6.60 13.94 7.13 24 8.29 15.49 8.19 1.09 1.56

0 0 0 0 –.034 –.111 .063 .154 .123 .009 –.006 –.015 .256

712 641 600 516 466 405 333 279 110 69 46 32 39

53 0 54 0 55 0 56 0 57 129.25 58 128.65 59 126.59 60 126.83 61 127.18 62 120.34 63 124.46 64 120.85 65 102.4 (continued )

0 0 0 0 –.504 –1.37 –2.72 –8.41 –4.27 –4.6 –4.18 –4.49 –4.91

Simulation (3-year Increment) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 –5.69 0 5.53 .057 –6.94 0 5.71 .136 –8.25 0 5.57 .338 –19.72 0 8.71 1.21 –21.99 0 10.45 .496 –23.23 0 8.36 .591 –21.29 –1.00 9.67 .496 –20.23 1.64 14.59 .500 –20.94 –1.99 13.32 .659

712 641 600 516 466 405 333 279 110 69 46 32 39

90%

SD

Median Tax/Subsidy

Table 6A.1

(continued) Accrual

Age

Median SSW

Median

53 54 55 56 57 58 59 60 61 62 63 64 65

0 0 0 0 0 0 0 83.88 94.64 100.14 105.87 105.12 89.84

0 0 0 0 0 0 0 –2.18 –.356 –.623 –.602 –1.259 –2.02

10%

90%

SD

Median Tax/Subsidy

Common Reform (Act. Adjustment) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 –11.30 5.69 6.94 .652 –18.93 7.44 11.51 .035 –10.53 6.18 13.27 .064 –8.97 24.11 12.52 .063 –11.06 3.79 13.80 .119 –13.71 6.80 10.41 .352

Peak Value Median

10%

53 54 55 56 57 58 59 60 61 62 63 64 65

147.90 3.27 –.659 –1.17 –2.58 –2.66 –3.56 –9.76 –4.42 –4.68 –4.96 –4.49 –5.24

69.35 –4.61 –7.15 –7.69 –15.44 –16.11 –13.40 –21.44 –23.55 –23.23 –23.04 –20.23 –23.47

53 54 55 56 57 58 59 60 61 62 63 64 65

123.39 121.76 120.33 119.27 5.72 4.53 –.485 –1.14 –.246 .425 .161 .132 –2.18

92.12 88.57 88.40 85.38 –.94 –1.17 –1.96 –2.56 –2.70 –1.08 –.938 –.94 –3.14

Observations

90%

From Volume

712 641 600 516 466 405 333 279 110 69 46 32 39 Option Value

Median

10%

90%

SD

Baseline (Transition) Case 187.01 54.86 301.17 174.19 83.58 143.20 176.58 85.00 123.77 178.90 81.57 112.70 64.90 47.49 99.78 41.13 44.33 89.48 14.46 45.34 79.33 –1.76 15.61 –0.825 2.49 18.97 53.27 1.83 14.49 51.99 –1.00 14.00 45.16 –2.85 18.72 45.20 –1.99 14.54 10.01

114.81 48.06 11.68 20.86 12.75 –1.29 –2.53 –22.6 –25.63 –20.24 –24.30 –6.4 –25.09

531.27 477.51 474.82 450.28 248.62 246.70 221.00 143.92 254.06 366.24 428.20 136.25 81.99

219.23 234.33 231.79 206.05 140.10 145.95 154.66 125.00 164.45 165.17 171.61 125.76 54.87

171.18 156.24 153.79 159.76 47.52 38.39 28.38 3.70 –1.91 3.04 6.90 16.25 –3.00

598.8 567.08 549.60 538.11 322.79 298.18 255.88 230.24 373.90 551.25 601.55 397.06 110.98

248.99 235.53 231.53 213.15 154.98 164.90 163.95 160.40 205.02 206.55 208.60 161.15 78.62

182.56 177.31 173.22 169.36 58.08 57.38 60.29 52.94 60.33 56.27 56.48 35.13 2.52

SD

1995 Reform Case 63.2 334.11 61.52 325.63 65.30 309.88 66.59 300.34 34.45 157.37 5.715 142.32 31.86 128.35 31.08 51.42 34.84 95.46 29.00 94.09 24.06 81.32 14.40 68.84 2.17 33.55

Micro-Modeling of Retirement Behavior in Italy Table 6A.1

397

(continued) Peak Value

Median

10%

90%

53 54 55 56 57 58 59 60 61 62 63 64 65

143.40 143.95 140.52 139.49 –.285 –1.24 –2.72 –8.42 –4.27 –4.60 –4.18 –4.06 –4.91

69.39 69.42 56.24 13.07 –5.69 –6.91 –8.25 –19.72 –21.99 –23.23 –21.29 –20.23 –20.94

179.69 179.13 176.99 176.49 167.40 169.64 168.42 56.06 35.56 192.40 –1.00 2.06 –1.99

53 54 55 56 57 58 59 60 61 62 63 64 65

102.58 103.13 102.07 102.87 101.34 100.84 101.02 –2.18 –.116 –.624 –.371 –.91 –2.02

62.73 62.69 62.69 62.73 62.69 62.69 62.73 –11.30 –18.93 –10.53 –8.95 –11.06 –13.71

135.40 133.10 133.14 135.32 133.16 134.74 131.10 15.22 19.35 61.64 85.52 31.29 6.80

Option Value SD

Median

10%

90%

SD

Simulation 1 Case 51.07 316.72 50.78 311.10 62.86 290.96 61.70 279.56 79.23 123.54 83.48 107.79 85.68 92.74 58.54 3.36 82.81 59.44 100.64 64.56 12.94 52.15 16.36 53.09 14.19 19.99

119.70 106.85 94.64 96.34 39.37 16.20 .328 –21.58 –25.22 –20.51 –10.81 –6.40 –20.61

545.38 516.38 496.41 491.54 450.03 425.61 396.20 205.77 346.32 553.95 462.73 157.65 88.68

219.99 208.64 211.96 187.54 191.51 204.80 210.26 174.40 243.86 270.40 176.74 139.04 56.24

Simulation 2 Case 42.61 315.54 42.12 309.91 43.55 295.40 37.43 288.26 35.60 272.98 40.62 260.22 45.08 248.52 25.95 17.88 39.72 103.99 42.41 96.24 44.06 82.96 29.24 74.46 12.80 25.13

119.56 102.55 105.73 111.34 102.65 98.26 89.50 –9.96 –21.86 –4.60 –1.88 –1.30 –15.89

544.01 514.72 494.77 490.25 455.71 440.76 398.04 202.95 332.89 566.63 653.44 283.30 105.05

228.55 216.92 212.44 183.38 167.42 181.99 189.26 161.45 216.70 224.92 233.20 171.98 75.21

Note: SD = standard deviation.

References Brugiavini, A. 1999. Social security and retirement in Italy. In Social security and retirement around the world, ed. J. Gruber and D. Wise, 181–237. Chicago: University of Chicago Press. Colombino, U. 2000. Un modello per la valutazione degli effetti degli incentivi individuali sulle decisioni di pensionamento (A model for evaluating the effects of individual incentives on retirement decisions). University of Turin, Department of Economics. Mimeograph. Franco, D. 2000. Italy: A never-ending pension reform. Paper presented at NBERKiel Institute Conference, Coping with the Pension Crisis—Where Does Europe Stand? 20–21 March, Berlin, Germany. Gruber, J., and D. Wise. 1999. Social security and retirement around the world. Chicago: University of Chicago Press. Instituto Nazionale della Previdenza Sociale (INPS). 1995. Il modello INPS. Una

398

Agar Brugiavini and Franco Peracchi

proiezione al 2030 del Fondo pensioni lavoratori dipendenti e delle Gestion i pensionistiche dei lavoratori autonomi (The INPS model. A projection to year 2030 of the pension funds of private sector employees and self-employed workers). Rome: INPS. Mimeograph. Instituto Nazionale di Statistica (ISTAT). Various years. I trattamenti pensionistici. Rome: ISTAT. Miniaci, R. 1998. Microeconometric analysis of the retirement decision in Italy. Organization for Economic Cooperation and Development (OECD) Economics Department Working Paper no. 205. Paris: OECD. Rossi, N., A. Sorgato, and G. Toniolo. 1993. I conti economici italiani: Una ricostruzione statistica, 1890–1990 (The Italian national accounts: A statistical reconstruction, 1890–1990). Rivista di Storia Economica 10 (1): 1–47. Spataro, L. 2000. Le scelte di pensionamento in Italia: Un’applicazione (ed estensione) del modello “option value” (Retirement choices in Italy: An application [and extension] of the option value model). Studi Economici 72:1–30. Stock, J. H., and D. A. Wise. 1990. Pensions, the option value of work, and retirement. Econometrica 58 (5): 1151–80. Zweimüller, J., R. Winter-Ebmer, and J. Falkinger. 1996. Retirement of spouses and social security reform. European Economic Review 40 (2): 449–72.

7 Social Security and Retirement in Japan: An Evaluation Using Micro-Data Takashi Oshio and Akiko Sato Oishi

7.1 Introduction The main purpose of this paper is to analyze empirically the impact of social security incentives on retirement decisions of older employees in Japan. It is important because the more elderly people stay in the labor market, the less the demographic pressures social security programs will have to struggle with. Estimation and simulation results in this paper will provide micro-economic foundations for the impact analysis that is crucial to discussions about pension reforms in Japan. Japan is now facing a very rapid population aging. The share of people aged sixty-five years or above of total population was 16.2 percent in 1998, roughly the Organization for Economic Cooperation and Development (OECD) average. Looking forward, however, the share of elderly people is expected to grow faster than in any other advanced country, reflecting a very low fertility rate, which dropped to 1.34 in 1999. Indeed, the National Institute of Population and Social Security Research (NIPSSR) revised down its population projections in 1997. In its new “middle” projection, the NIPSSR assumes that the fertility rate would return to only 1.61 by 2050—a much more conservative figure than the previously assumed 1.80. The NIPSSR also projects that the share of people aged sixty-five or above would grow to 27.4 percent in 2025 and 32.3 percent in 2050. Many analysts, however, argue that NIPSSR’s “pessimistic” scenario, which assumes that the fertility rate would remain as low as 1.38 even in 2050, seems to be Takashi Oshio is associate professor of economics at Tokyo Gakugei University. Akiko Sato Oishi is a senior researcher in the Department of Theoretical Social Security Research at the National Institute of Population and Social Security Research.

399

400

Takashi Oshio and Akiko Sato Oishi

more plausible. If this were the case, the pace of population aging would be more dramatic than is now widely anticipated. Rapid population is a big challenge to Japan’s long-term fiscal strategy. Social security expenditures, including public pensions, health care, and social welfare benefits, amounted to ¥69.4 trillion in 1997, equivalent to 17.8 percent of national income. Public pension benefits were ¥36.4 trillion, covering 52.4 percent of overall social security expenditures. It is likely that social security expenditures will grow substantially over the coming decades. The most recent official projection, released by the Ministry of Health and Welfare in 1998, expects social security expenditures to grow to a level of 33.5 percent of national income by 2025, assuming no change in the current social security programs. The public pension system is the major determinant of the long-term trend in social security expenditures and fiscal balances beyond 2000. As in other industrialized countries, public pension insolvency is now one of the most serious challenges that an aging society poses to the Japanese economy. The Ministry of Health and Welfare estimates unfunded liabilities to be about ¥490 trillion—almost equivalent to nominal gross domestic product (GDP)—at the end of fiscal year 1999. Also, policy simulations conducted by the Economic Planning Agency (Yashiro et al. 1997) project the public pension fund will be exhausted by 2040 if the current system is not changed. In addition, the newly introduced corporate-accounting system, which became effective as of April 2000, will likely reveal substantial underfunding in corporate pensions and probably also make their reform inevitable. It is important to understand retirement incentive effects in order to assess the economic impact of pension reforms. The labor force participation rate in Japan is much higher than in other advanced countries: 74.8 percent for men aged sixty to sixty-four and 40.1 percent for women aged sixty to sixty-four in 1998 according to the Labor Force Survey (Management and Coordination Agency 1999). However, increasing social security benefits have been reducing labor force participation over the past few decades, even allowing for cyclical swings.1 Moreover, various cross-sectional studies have found that the existing pension scheme tends to reduce the incentive to work for elderly people (see section 7.3). It is widely recognized that an earnings-tested pension program, called the Zaishoku pension, tends to discourage the elderly from working. Retirement incentive effects will also be potentially critical for Japan’s economic growth and the financial position of the public pension, since postwar baby-boomers will become eligible for public pension benefits in 1. The labor participation rate of people aged sixty to sixty-four was 81.5 percent for men and 39.1 percent for women in 1970. The rate for men declined to 71.1 percent in 1988 and then rose to 75.6 percent in 1993, reflecting the economic boom; since then, it has been on a downtrend.

Social Security and Retirement in Japan

401

the next few years. With the total labor force diminishing due to a very low fertility rate, Japan’s growth potential will depend much more on labor force participation from the elderly. In addition, the sensitivity to social security provisions of the elderly is likely to increase in the long run and reflects two factors. First, the shares of self-employed and agricultural workers who are less sensitive to social security programs are likely to keep declining, reflecting structural change in the Japanese economy. Second, more women will likely enter labor force and become eligible for employees’ pension benefits. The structure of this paper is as follows. Section 7.2 reviews the institutional background, laying out the retirement policy landscape in Japan and setting out the relevant sources of income support for the elderly. Section 7.3 provides the research background with a brief review of previous studies on this topic in Japan; section 7.4 sets out the data on which our estimation and simulation is based; and section 7.5 constructs earnings histories and projections from our data, and section 7.6 sets up incentive measures (benefit accrual, option value, and peak value). Section 7.7 estimates the impact of incentive measures on retirement, and section 7.8 summarizes estimation results. Section 7.9 conducts policy simulations based on the estimated models, and section 7.10 concludes the paper. 7.2 Institutional Background 7.2.1 The Retirement Policy Landscape This section describes the retirement policy scheme as it existed for the years used in our analysis. Japan’s public pensions operate a two-tier system: One pays flat-rate basic pension (Kiso Nenkin) benefits to all residents, including the self-employed and unpaid family workers; the other pays earnings-related benefits only for private and public employees.2 Employees thus receive two forms of pension benefits: basic pension benefits and earning-related benefits. This basic pension, which is mainly for nonemployees, has little effect on retirement decisions because its benefits are relatively small and subject to no earnings criterion. The eligibility age of full basic pension benefits is sixty-five years old, with no earnings test. It incorporates a flat-tax and flat-benefit structure, and it is organized on an individual unit basis.3 The principal program for private-sector employees is the Kosei Nenkin Hoken (KNH; Private Employees Pension), which covers about 85 percent of all employees. Government employees, private-school teachers, and em2. See Takayama (1998) for more detailed and comprehensive information about Japan’s pension system. Discussions in this section owe much to chapter two of his book. 3. The flat tax and benefits per month were ¥11,700 and ¥65,000, respectively, in 1995.

402

Takashi Oshio and Akiko Sato Oishi

ployees in agriculture, forestry, and fishing organizations are covered by special programs provided by Kyosai Kumiai (Public Employees’ Pension; mutual aid associations), but those programs have almost the same structure as the KNH. Thus, our analysis of public pensions in this paper mainly focuses on the KNH, and treats Kyosai Kumiai members as KNH members. In what follows, we provide brief descriptions of the KNH as well as other income-support programs for elderly employees—including Zaishoku Pension, unemployment insurance, and wage subsidy. Kosei Nenkin Hoken (KNH) Under the KNH scheme, an individual’s benefits are calculated according to the following steps. First, an individual’s monthly wage (excluding semi-annual bonus payments) is converted into standard monthly earnings and graded into one of thirty levels. Second, the career-average monthly earnings are calculated over their entire period of coverage (up to age sixtyfour) and adjusted by wage income growth and converted into the current earnings level. Finally, benefits are calculated as the career-average monthly earnings
the number of contribution years
0.0075 (the accrual rate). For instance, forty-year contributors will earn 30 percent of the career-average monthly earnings. In addition, benefits are inflation indexed every year in terms of consumer prices, and adjusted for net wages every five years. The normal eligibility age for full KNH benefits is currently sixty, with some exceptions,4 while it is scheduled to increase gradually to sixty-five from 2001. A male KNH recipient currently gets both the full basic pension and earnings-related benefits at age sixty.5 In addition, his dependent wife (full-time housewife) can get her basic pension benefit with no contribution when she becomes sixty-five. Thus, total benefits that a typical couple receives are two basic pension benefits (for the husband and wife) plus earnings-related benefits (for the husband) that, in total, replace slightly less than 70 percent of average monthly earnings for about 50 percent of average annual wages—including bonus payments—of currently active male workers. Between the ages of sixty and sixty-four, one can get partial pension benefits (Zaishoku Pension, see later discussion) with an earnings test if one chooses to keep working. Beyond sixty-five, one gets full pension benefits without any earnings test but also has the option of delaying the receipt of pension benefits, with some actuarial adjustment. In addition, survivors’ benefits are available, but our analysis neglects them for simplicity. Contributions are based on the employee’s monthly standard earnings and are shared equally by the employee and employer. The total contribu4. The eligibility age for sailors and miners is fifty-five. 5. The KNH recipients are currently eligible for full basic pension benefits (in addition to the earnings-related component) at age sixty, while the eligibility age of basic pension benefits is sixty-five years for non-KNH recipients. A husband gets some additional spousal benefits (Kakyu Nenkin) until his dependent wife becomes sixty-five years old.

Social Security and Retirement in Japan

403

tion rate for KNH and basic pension is currently 17.35 percent, meaning that an employee and employer contribute 8.675 percent each. A female employee pays premiums at the same contribution rate, while a dependent housewife does not need to contribute. Zaishoku Pension (Earnings-Tested Kosei Nenkin Hoken for Elderly Workers) The Zaishoku Pension, which is a part of the KNH scheme, is a partial and earnings-tested pension for employees. Upon reaching age sixty and until age sixty-four, a KNH recipient who keeps working can receive reduced KNH benefits subject to an earnings test. This scheme is roughly equivalent to the early retirement system in many other OECD countries. The formula of the Zaishoku Pension is summarized as follows. If an individual earns even a small wage, benefits are reduced by 20 percent. If earnings are above ¥220,000 per month, benefits are reduced by one yen for each additional two yen increment in wages (i.e., the marginal tax rate is 50 percent). If earnings are above ¥340,000, benefits are reduced by the same amount of additional wage earnings (i.e., the marginal tax rate is 100 percent). One of the key elements of the 1994 reform was to make the formula of the Zaishoku Pension more favorable to elderly. Also note that one has to pay KNH contributions as long as they keep working, although they can expect an increase in future pension benefits. Unemployment Insurance Unemployment insurance (UI) adds temporary income support to retired employees. In many cases, an individual who reaches age sixty leaves the firm where they have been working and then start to receive KNH benefits. At the same time, it is normal to apply for UI benefits when quitting one’s previous job, regardless of any desire to find a new job. The UI benefits for those ages sixty to sixty-four replace 50–80 percent of wage earnings at age sixty for 300 days at most. Thus, there are many cases in which the total replacement rate—adding KNH and UI benefits together—is effectively more than 100 percent of income at the first retirement age, probably reducing the incentive to work. Furthermore, many people tend to stay out of the labor force after receiving UI benefits, meaning that UI is now used in a way very different from its original conception (see Yashiro and Nikami 1996). However, under a new law effective as of April 1998, an individual cannot receive UI and KNH benefits at the same time: As long as one is receiving UI benefits, one has to postpone receipt of KNH benefits. Wage Subsidy for Elderly Workers Another income support that potentially interacts with public pension programs is the wage subsidy to elderly workers (henceforth referred to as “WS”). This program was introduced in 1994 as a part of the public employment insurance scheme to replace the aforementioned UI benefits,

404

Takashi Oshio and Akiko Sato Oishi

which are considered to work ineffectively for elderly workers. The WS equivalent to 25 percent of the current wage is provided to an employee— subject to a certain wage ceiling—on the condition that they are sixty to sixty-four years old and their wage earnings are less than 85 percent of their preretirement wage at age sixty. This WS program is independent from the public pension scheme, but its economic implications are similar to those of the Zaishoku Pension. Both programs are applicable to the same age group (ages sixty to sixty-four) and subject to certain earnings criteria. The WS can be treated as a negative premium in calculating social security incentives. The WS equivalent to 25 percent of wage earnings well exceeds the employee’s share of KNH contributions (8.675 percent). Thus, the combination of the WS and the pension premium would add to an individual’s net pension wealth, although it may not be enough to offset the negative effect from postponing receipt of pension benefits. Disability Pension The disability pension, unlike in some European countries, is not used as interim income support for elderly workers who are on the path to retirement. The disability pension is strictly for those who are physically unable to work. Benefits are calculated in almost in the same way as those of KNH, while additional benefits of 25 percent are given to those who are categorized as having more serious disabilities. The eligibility conditions for the disability pension are generally strict: Most disabilities must originate from injuries, which prevents disability pension benefits from being used as a source for financing earlier retirement in Japan. There are about 285,000 recipients of the disability pension, covering only three percent of total old-age pension recipients. Thus, the disability pension will be neglected in our social security incentive calculations. Employer-Provided Pension In addition to public pension benefits and other income support, employer-provided pension programs—the Kosei Nenkin Kikin (Employees’ Pension Fund) and tax-qualified plans—cover about two-thirds of KNH participants. Employees can choose lump-sum retirement benefits, annuities at retirement, or both. In the case of the KNK, about 40 percent of recipients choose lump-sum benefits, 55 percent choose annuities, and 5 percent choose both in 1996. Benefits of employer-provided pensions are paid on top of public pension benefits, reflecting additional contributions that have been paid in addition to KNH premiums until retirement. In the model case of KNK, benefits from employer-provided pensions are assumed to be equivalent to about 27 percent of KNH benefits for each couple. Lump-sum retirement benefits vary substantially by firm size and tenure, but they are at a level of ¥20–30 million in the case of average employees in large firms.

Social Security and Retirement in Japan

405

These employer-provided pension and lump-sum retirement benefits work differently from public pension benefits. Their payments are closely linked to mandatory retirement6 at age sixty, regardless of an employee’s working status after that age. In fact, these benefits add to the incomes of an individual—wages, pension, and others—after the mandatory retirement and increase the disincentive to work through income effects. However, their present discounted value added over a lifetime is basically unchanged even if they continue to work, and they are thus unlikely to affect the timing of retirement for people beyond the age of sixty. Rather, as pointed out by Seike (1993), firms tend to adjust the amount of lump-sum retirement benefits to make their employees retire earlier than the mandatory retirement age. This is because middle-aged workers tend to levy heavy labor costs on firms under Japan’s seniority system. Seike argues that the present discount value of retirement benefits peaks in the early fifties and falls thereafter in some industries. Our analysis of social security incentives does not explicitly include the incentive effect of employerprovided pension or lump-sum retirement benefits. 7.2.2 1999 Pension Reform Act The 1999 Pension Reform Act incorporated measures to lower contributions paid by future generations, making it inevitable that eligibility conditions and benefit systems would become less generous than scheduled in the 1994 Pension Reform Act. In particular, the act proposed (a) a 5 percent reduction in pension benefits, (b) an increase in the eligibility age to sixty-five from sixty, and (c) the abolition of wage indexation for pension benefits. If these proposals are implemented as scheduled, the final contribution rate for KNH will be pushed up to 25.2 percent, from the current 17.35 percent, in contrast to the previously scheduled 34.5 percent. At the same time, the government plans to introduce U.S. 401(k)style defined-contribution private pensions to supplement the public pension scheme. The combination of these proposals, however, is not expected to solve insolvency problems, and whether or not and when the fertility rate will turn around remains an open question. Further policy measures thus remain to be discussed ahead of the next round of pension reforms. 7.2.3 Labor Market Participation of the Elderly Figure 7.1 and table 7.1 provide a rough picture of labor market participation and benefit program participation for elderly people in 1996. The 6. “Mandatory retirement” in this paper means the program in which at a certain age (sixty years in most cases) an employee is forced to leave the firm where they have been working fulltime for many years. This does not necessarily mean that they must fully retire at that age and become a beneficiary of social security programs. On the contrary, a large proportion of those who have experienced mandatory retirement continue to work in a new firm or even at the same firm with a new status such as a part-time employee.

Labor force participation ratio of the Japanese elderly, 1996

Source: Ministry of Labor (MOL; 1996).

Fig. 7.1

Social Security and Retirement in Japan Table 7.1

407

Labor Market and Benefit Program Participation in 1996, by Age and Gender Men

Women

55–59

60–64

60–64

65–69

55–59

60–64

65–69

Working Executives Employed, full time Employed, part time Self-employed, etc. Not working Unknown Total

92.8 13.1 59.6 2.0 18.1 6.9 0.3 100.0

Labor Market Participation 70.1 54.2 59.7 41.9 10.2 6.7 3.1 2.0 29.5 14.2 23.7 9.6 6.8 8.1 11.3 7.5 23.6 25.1 21.6 22.8 29.6 45.6 40.1 57.7 0.3 0.2 0.2 0.3 100.0 100.0 100.0 100.0

29.1 1.7 3.8 4.7 18.9 70.8 0.1 100.0

75.5 7.9 40.8 6.9 19.9 24.2 0.3 100.0

55.8 6.0 19.4 7.2 23.2 43.9 0.3 100.0

41.2 4.1 8.8 6.3 21.9 58.7 0.2 100.0

KNHa (excl. Zaishoku) Basic pension onlya Zaishoku Pension Kyosai Kumiaia Wage subsidy Employer-provided pension UI benefits

1.9 1.0 0.0 1.5 0.0

Benefit Program Participation 39.8 60.7 6.0 31.3 7.4 20.9 1.6 15.4 5.2 0.0 0.3 1.7 11.1 11.0 0.9 4.3 3.5 0.0 0.0 0.9

36.1 46.4 0.0 5.8 0.0

4.0 1.4 0.2 1.2 0.0

35.4 11.5 3.4 7.6 2.2

47.9 34.1 0.0 8.3 0.0

1.6 0.3

0.5 1.2

5.6 3.0

5.1 0.5

0.6 1.2

9.4 4.4

65–69

8.9 0.7

55–59

Total

0.4 1.2

1.9 1.6

Source: Authors’ calculations from the Survey on Labor Market Participation of Older Persons (SLMPOP; MOL 1996). Notes: KNH = Kosei Nenkin Hoken (employees’ pension); Zaishoku Pension = Earnings-tested KNH; Kyosai-Kumiai (mutual aid associations): Special programs for national and local government employees, etc.; UI = unemployment insurance. a Includes survivors’ benefits.

labor market participation rate is relatively high in Japan among OECD countries, but it drops sharply at age sixty because most employees have mandatory retirement and start receiving public pension benefits, private pension benefits, or both at that age. Also, beyond the age of sixty there are limited chances that they can get a full-time job: After leaving the firm at sixty, most people move to the secondary labor market, become part-time employees with lower wage earnings, or both. The bottom part of table 7.1 summarizes benefit receipt measures, the pattern of which roughly corresponds to that of labor force withdrawal. Of people aged between sixty and sixty-four, 35.4 percent receive KNH benefits, 7.6 percent receive Kyosai-Kumiai benefits, and 11.5 percent receive Basic Pension benefits only. Recipients of Zaishoku Pension benefits and WS are not a majority in the group of sixty to sixty-four year olds. This suggests that the earnings test for the Zaishoku Pension does not work effectively, especially for part-time workers, and those who continue to work as self-employed after retiring from company jobs can receive full KNH benefits, while they do not need to pay the premium. Also the WS, which was

408

Takashi Oshio and Akiko Sato Oishi

introduced in 1995, seems not to have been widely recognized yet, despite its relatively strong incentives to work. Table 7.1 poses an important question, that is, should we define retirement as the status of somebody who has stopped working or that of somebody who has started to claim benefits? Our preferred definition of retirement in Japan is the former—the absence of wage earnings. The receipt of public pension benefits is not an effective criterion for retirement in Japan. This is because pension benefits are given unconditionally to most citizens aged sixty-five years and older, and also because a large portion of pension beneficiaries aged sixty to sixty-four remains in the labor market. Another question is how we should deal with self-employed workers. There are some cases where one may become self-employed after retiring from a firm. Indeed, the self-employed contribute significantly to the high labor-participation rate of the elderly in Japan. In what follows, we categorize the self-employed (who have been employees) as retired even if they continue to receive self-employed income, because their working behavior seems quite different from that of employees. That said, there is still a gray zone between working and retirement. The narrowest definition of working should be “being employed full time (and receiving no pension benefits).” However, many people prefer to work part time and receive reduced pension benefits to supplement their wage income. Another question is whether people who say that they cannot find a job should be categorized as retired. We treat them as retired if they receive any pension benefits. Table 7.2 summarizes the combination of working status and public pension benefits for three groups, those aged fifty-five to fifty-nine, sixty to sixty-four, and sixty-five to sixty-nine. The relationship between working status and benefit claiming is so complicated that a clearcut definition of retirement cannot be established. So, let us consider the following three tentative definitions of retirement. Definition I: those who are not employed—in this definition, executives are assumed to be employed Definition II: those who are categorized as retired according to the Definition I excluding the self-employed and family workers Definition III: those who are categorized as retired according to the Definition II, excluding job-seekers that are not receiving any public pension benefits Table 7.3 summarizes the share of the retired according to these three definitions, based on the matrix of table 7.2. The share of retired to the total population is quite sensitive to these definitions, since many working people receive pension benefits. For the group aged sixty to sixty-four, for example, 67.4 percent of the sample is defined as the retired according to the definition I, while the shares of the retired are much lower according to the definitions II and III (44.3 percent and 39.3 percent, respectively). In

3.0 21.7 10.3 20.7 55.7 13.0 22.6 35.5 91.2

0.1 2.1 1.0 1.0 4.2

1.9 2.7 4.6 8.8

3.7 1.5 5.2 95.5

0.9 0.7 1.6 4.5

14.9 25.3 40.1 100.0

3.1 23.7 11.4 21.7 59.9

4.6 2.2 6.9 100.0

13.1 59.8 2.0 18.2 93.1

Total

Note: Calculations exclude samples whose working status is not known. Source: Authors’ calculations from the SLMPOP (MOL 1996).

Working Executives Employed, full time Employed, part time Self-employed, etc. Subtotal Not working Unable to find a job Not willing to work Subtotal Total

12.8 58.2 1.7 17.5 90.3

Not Receiving

0.3 1.6 0.3 0.7 2.8

Receiving Benefits

55–59

12.2 20.1 32.3 52.8

0.6 5.1 3.8 10.9 20.5

Females

15.9 9.5 25.4 63.5

2.9 15.7 6.0 13.6 38.1

Males

Receiving Benefits

Working Status and Public Pension Benefits, by Age and Gender

Working Executives Employed, full time Employed, part time Self-employed, etc. Subtotal Not working Unable to find a job Not willing to work Subtotal Total

Table 7.2

6.8 18.7 25.6 47.2

1.4 4.6 3.7 12.0 21.6

3.0 1.2 4.2 36.5

7.3 14.0 0.9 10.1 32.3

Not Receiving

60–64

19.0 38.9 57.9 100.0

2.0 9.7 7.5 22.9 42.1

18.9 10.7 29.6 100.0

10.2 29.6 6.9 23.7 70.4

Total

13.9 49.1 63.0 88.7

1.5 3.0 4.1 17.0 25.7

16.6 27.0 43.6 93.0

5.9 12.8 7.8 22.9 49.4

Receiving Benefits

2.0 5.8 7.8 11.3

0.2 0.8 0.6 1.9 3.5

1.1 1.0 2.1 7.0

0.3 1.5 0.4 2.2 4.9

Not Receiving

65–69

15.9 54.9 70.8 100.0

1.7 3.8 4.7 19.0 29.2

17.7 28.0 45.7 100.0

6.7 14.3 8.2 25.2 54.3

Total

410 Table 7.3

Males Definition I Definition II Definition III Females Definition I Definition II Definition III Total Definition I Definition II Definition III

a

Takashi Oshio and Akiko Sato Oishi The Share of Retired, by Different Definitions

Not employeda Def. I (excl. self-employed, etc.) Def. II (excl. job-seekers not receiving public pension benefits) Not employeda Def. I (excl. self-employed, etc.) Def. II (excl. job-seekers not receiving public pension benefits) Not employeda Def. I (excl. self-employed, etc.) Def. II (excl. job-seekers not receiving public pension benefits)

55–59

60–64

65–69

Total

100.0 25.4 7.2 3.5

100.0 53.5 29.9 26.9

100.0 70.9 45.8 44.7

100.0 49.2 27.0 24.4

100.0 61.9 40.3 27.4

100.0 80.9 58.1 51.2

100.0 89.9 70.9 69.0

100.0 77.0 55.8 48.4

100.0 44.5 24.5 16.0

100.0 67.4 44.3 39.3

100.0 80.8 58.9 57.3

100.0 63.6 41.9 36.8

Be reminded that, in Definition 1, executives are regarded as employees.

addition, the difference between definitions I and II indicates the importance of the self-employed in assessing the labor force participation of the elderly in Japan. It should also be remembered that Definition I might overstate the actual number of the retired for the younger group, because income from self-employment is likely to be their major source of income. 7.2.4 Pathways to Retirement For those who have been employed until age sixty, retiree categories are divided into two groups; those receiving public retirement benefits (KNH, other public pension benefits or both) and those receiving public and employer-provided retirement benefits. Some become self-employed but are covered by publicly provided benefit programs, employer-provided benefit programs, or both (with no earnings test). There are a variety of pathways to retirement due to a multiplicity of social security incentives. With a lack of longitudinal data, however, it is nearly impossible to trace all paths that are taken by individuals. So we roughly estimate major options and their probabilities based on the crosssectional data in the survey (see in section 7.4) and using simple assumptions. We assume an individual had been employed (with no public pension benefits or other public assistance program) until the age of sixty. Then, we trace major paths that they are likely to take through public assistance programs over the following ten years. For simplicity we divide the period after the age of sixty into two stages, one at sixty to sixty-four and the other at sixty-five to sixty-nine. We then estimate the paths that are taken by an individual who retires in the second stage, based on the observed probabilities of receiving each public benefit

Social Security and Retirement in Japan

411

in the cross-sectional data. With the lack of data, however, we cannot know how those who are still working in the second stage will behave beyond age seventy. We assume that they will follow the same pattern observed for actual retirees. Thus, it should be remembered that our estimation does not provide a full picture of retirement behavior in Japan. The procedure of our estimation is summarized as follows. We start with the second stage, ages sixty-five to sixty-nine, in which a retiree has two major options: to receive public pension benefits (referred to as SS hereafter) with or without employer-provided pension benefits (referred to as Pension hereafter); SS here includes not only KNH but also kyosai benefits, the latter of which are those paid to employees in the public sector and other special occupation groups. In addition, there are two possible forms of retirement: staying out of the labor force or becoming selfemployed—a retiree receives SS, Pension, or both benefits in either case. There is a very small group of nonbeneficiaries in each type, probably due to eligibility problems. For the first stage, ages sixty to sixty-four, some people will have already retired (or become self-employed) while others will have continued working. We roughly estimate the working or retirement status of an individual who is retired in the second stage as follows: If one had already been retired in the first stage, they must have started to get SS benefits (with or without Pension benefits); that is, they must have chosen the option of going directly to SS. If one remained employed, they could choose among the following five options. 1. To be employed with no public assistance (that is, going directly to SS) 2. To be employed with SS benefits 3. To be employed with SS and Pension benefits 4. To be employed with SS benefits and WS 5. To be employed with SS and Pension benefits and WS It should be noted here that the SS in options (2) to (4) include not only Zaishoku Pension, but also (full) KNK benefits, while the choice of being employed with WS only is neglected because of its minority. Full KNH benefits cannot be received when working, and earnings tests should be applied. However, there seem to be many cases in which people do not distinguish Zaishoku and KNH benefits or where earnings tests do not work effectively, especially for part-time jobs. The cases in which one is employed and obtains Pension benefits only, or is employed with Pension benefits and WS, can be ruled out as firms that provide Pension benefits are usually covered by the KNH program. In addition, as indicated in table 7.1, a significant proportion of retirees receive UI benefits at the first year of retirement. Hence, each option mentioned above has a pair of suboptions: to receive or not to receive UI at the

412

Takashi Oshio and Akiko Sato Oishi

first retirement age, meaning that there are ten (five times two) pathways to retirement in total. Combined with the probabilities of retiree categories in the second stage, we can estimate joint probabilities of pathways to retirement and retiree categories. If one retires in the first stage, they are assumed to remain retired in the second stage (although their retiree category may change). However, some combinations of the first-stage and secondstage options can be ruled out a priori: For instance, if an individual is employed without employer-provided Pension benefits in the first stage, they are unlikely to get them in the second stage. This is because Pension benefits, if applicable, must be paid at the mandatory age of sixty in most cases. Table 7.4 illustrates the major categories of retirees and their respective pathways to retirement, which are estimated from our cross-sectional data. As shown in this table, the most common pathway to retirement for Japanese employees is going to the SS program and receiving UI benefits (at the first retirement age). About one-third of retirees take this route. Adding the case of receiving no UI benefits—which is slightly less common than the direct path to SS—as well the path to a public pension program with no public or private income assistance until retirement raises the proportion to about 64 percent. Becoming self-employed after having been employed is a minority choice, covering only five percent of total retirees. Table 7.4 also reveals that about 26 percent of people choose to keep working while receiving SS benefits—which are earnings-tested Zaishoku or even full KNH benefits—before retirement. However, low probabilities for receiving WS confirm that this newly introduced plan has not been widely used so far. Another finding (not shown in the table) is that women depend less on employer-provided benefits than men, probably because women’s tenure as a full-time employee is generally shorter than men’s. 7.3 Research Background There have been many empirical analyses of retirement incentives for the elderly in Japan, in spite of the limited availability of longitudinal data. Most of these draw on cross-sectional data from the National Survey on Family Income and Expenditure, the Survey on Labor Market Participation of Older Persons (Ministry of Labor [MOL] 1996), or both. These analyses can be divided into the following two groups. The first group, which includes Takayama et al. (1990b), Seike (1993), Abe (1998), Ogawa (1998), and Iwamoto (2000), has estimated how social security benefits raise the probability of retirement. They have all found that, after excluding the sample selection bias, social security benefits create significant retirement incentives for the elderly. Recent research by Abe, Ogawa, and Iwamoto focuses on the negative impact of Zaishoku Pension benefits on labor supply. Each finds that the 1989 Pension Reform Act—which aimed to reduce the marginal tax rate of Zaishoku Pension

Table 7.4

Retiree Categories and Pathways to Retirement Retiree Category Retirees (excl. Self-Employed)

Pathway Total Directly to SS UI to SS SS (wk) to SS SS and pension (wk) to SS SS and WS (wk) to SS SS, pension, and WS (wk) to SS UI and SS (wk) to SS UI, SS, and pension (wk) to SS UI, SS, and WS (wk) to SS UI, SS, pension, and WS (wk) to SS Total Men Directly to SS UI to SS SS (wk) to SS SS and pension (wk) to SS SS and WS (wk) to SS SS, pension, and WS (wk) to SS UI and SS (wk) to SS UI, SS, and pension (wk) to SS UI, SS, and WS (wk) to SS UI, SS, pension, and WS (wk) to SS Total Women Directly to SS UI to SS SS (wk) to SS SS and pension (wk) to SS SS and WS (wk) to SS SS, pension, and WS (wk) to SS UI and SS (wk) to SS UI, SS, and pension (wk) to SS UI, SS, and WS (wk) to SS UI, SS, pension, and WS (wk) to SS Total

Self-Employed

SS Only

SS and Pension

No SS

SS Only

SS and Pension

No SS

30.37 33.58 8.35

3.11 3.44

0.22 0.25

1.47 0.92 1.11

0.01 0.01

0.15 0.09

2.03

0.18

0.84

0.11 0.20

0.02

9.24

0.69 2.24

0.12

0.93

83.31 25.00 32.33 7.20

0.07 0.23 11.26

0.47

4.37

3.68 4.76

0.13 0.17

2.73 1.95 1.85

2.38 0.28

0.20 0.14

0.04 1.32

3.08

0.27

1.09

0.15 0.36 14.54

0.31

8.22

1.93 1.67

0.40 0.34

0.12 0.06 0.11

0.75

0.03 0.80

0.34

0.00 0.00

0.02 0.01

0.01

0.59

0.01 0.04

0.00

8.59

0.05 0.65

0.00

0.51

93.79

0.04 0.03

0.22

9.32

39.82 34.32 9.97

0.24

0.38

0.85

75.79

0.01 0.35

0.00 0.04 5.08

0.74

0.35

0.00 0.01

0.03

Total

35.34 38.28 9.46 2.21 0.96 0.22 9.93 2.36 1.00 0.24 100.0 31.78 39.39 9.05 2.76 1.06 0.32 10.64 3.35 1.25 0.39 100.00 42.29 36.40 10.08 0.76 0.59 0.04 8.64 0.65 0.51 0.04 100.00

Notes: Wk  paid during working; UI  unemployment insurance; pension  employer-provided retirement benefits, which are mostly annuities and do not include lump-sum benefits.

414

Takashi Oshio and Akiko Sato Oishi

benefits—failed to significantly encourage the elderly to work. These analyses, however, treat social security benefits only on a flow basis, without a dynamic framework that considers how additional work changes benefits and, correspondingly, their wealth. The second group, which includes Takayama et al. (1990a), Seike (1991), Oshio (1997), and Yashiro and Oshio (1999), has been interested in the magnitude of social security wealth. Takayama et al. discuss how the public pension scheme affects the distribution of human capital through social security wealth. Seike finds that social security accrual turns negative at age sixty, consistent with a sharp drop in labor force participation at that age. Oshio estimates how the 1994 Pension Reform Act affects social security wealth and its accrual pattern. These analyses, however, do not empirically predict how retirement incentives based on social security wealth affect the retirement decisions of the elderly. One purpose of this paper is to build a bridge between these two groups; we aim to estimate how retirement incentives based on social security wealth affect the probability of retirement, following Oishi and Oshio’s (2000) tentative research on the option value model (see section 7.6.1). In addition, policy simulations in this paper provide useful information about how social security reform affects labor force participation for the elderly. 7.4 Data Overview Our analysis is based on the Survey on Labor Market Participation of Older Persons (MOL 1996), which was conducted in October 1996 and published in December 1997 by the MOL. The survey covers men and women of aged fifty-five to sixty-nine who were employees, company executives, self-employed, or not working. Due to data limitations, our analysis centers on those who used to be employees at age fifty-five and who had been working until 1995 (see section 7.7.1). The size of the sample we use is 4,088 out of about 22,000 in the survey. The major problem is that the data from this survey are cross-sectional and not longitudinal. What we know from the survey is an individual’s age, current working status, wage income, pension benefits, and so on at the survey date. The survey asks each individual what kind of firm (industry and size) they were working for at fifty-five, whether or not and when they would face mandatory retirement, and when they wanted to retire (if working at the time of the survey). However, any other longitudinal information, including wage profiles and the actual date of retirement, is not available: We only know from the survey simply whether or not an individual was retired or still working in the survey year of 1996. Moreover, data on an individual’s background, such as education and family situation, are limited. The most important quantitative information available from the survey

Social Security and Retirement in Japan

415

relates to an individual’s current wage earnings and their social security and other benefits, on which our incentive calculations are based. It is, however, difficult to capture the diversity of incentives in employer-based pension policies, and information about lump-sum retirement benefits is not available. Moreover, answers about the category and amount of benefits seem at times to be unreliable, probably due to inaccurate knowledge, limited knowledge, or both among respondents about social security programs. We estimate the “theoretical” value of social security benefits based on projected wage profiles, and make some adjustment if the discrepancy between theoretical and actual figures is too large to be ignored. 7.5 Earnings Histories and Projections Backward and forward projections of wage earnings are required to analyze the impact of social security incentives on retirement decisions. With little longitudinal information available and uniqueness of the wage curve in Japan, our approach differs from the norm applied to other countries. Our projections of the age-earnings profiles depend largely on the crosssectional data. Also, we use reported individual characteristics observed in the survey as well as information obtained from the Wage Census. To summarize our methodology, we use (a) current wage earnings as a benchmark, (b) average age and wage profiles obtained from the survey for the ages fifty-five to sixty-nine, and (c) cohort-specific age-earnings profiles in backward projections starting at age fifty-five and below. An additional procedure required in the case of Japan is to estimate the timing of retirement for those who have already retired at the survey year on the basis of limited longitudinal information. 7.5.1 Projections for the ages fifty-five to sixty-nine In Japan, we observe that earnings for full-time work are likely to decline with age mainly due to the transition from the primary firm with the seniority-based wages to the secondary labor market with market-based wages; thus it is not reasonable to assume zero real growth in earnings into the future. For earnings projections for the ages fifty-five to sixty-nine, we rely on average wage growth rates observed from the survey because cohort-specific information is not available. In addition, the strong sample selection bias for elderly workers in the Wage Census prevents us from applying cohort-specific age and wage profiles for the workers aged fifty-five and above. It should be noted that earnings projections of this type cannot separate age effects from cohort effects and thus are inconsistent with backward projections based on cohort-specific information. We neglect this problem for simplicity. To calculate average wage growth, we regress the logarithm of monthly earnings (separately for males and females) on an individual’s age, experi-

416

Takashi Oshio and Akiko Sato Oishi

ence of mandatory retirement, job categories, firm size at the employee’s age of fifty-five, whether or not a private or public employee at fifty-five, and residential area. All independent variables are dummies. We do not apply any linear or parabola form for a wage curve, due to its discontinuity between, before, and after the mandatory retirement age; instead we apply dummies for each age. In addition, we apply Heckman’s two-step estimation procedures to deal with the sample selection bias (the first step estimation results are not reported). Estimated parameters are summarized in table 7.5. As clearly seen from this table, wage earnings decrease sharply after age sixty, the age at which many people retire and become eligible for pension benefits. Based on this regression, we create each sample’s earnings profile for the ages fifty-five to sixty-nine using the reported current wage earnings as a benchmark. For those who have already been retired by the survey year, we predict their current wages based on their age and other characteristics and construct their earnings profile for the ages fifty-five to sixty-nine. The wage growth rate is thus set to be the same for each individual: It is calculated by taking the difference in parameters on the two subsequent age dummies. However, parameters of other dummies show individual fixed effects, which shift the earnings profile up and down for each individual. The timing of mandatory retirement, which is in most cases sixty years old, is important in projecting the earnings profile. In projecting future earnings for an individual younger than sixty, we assume that they will face mandatory retirement at sixty. 7.5.2 Backward Projections The survey shows only current wage earnings for those who are working, full time or part time, in the survey year. To construct earnings histories before age fifty-five, we rely upon cohort-specific age-earnings profiles. For this purpose, we use wage data from the Wage Census, which is conducted and published every year by the Ministry of Labor.7 The Wage Census provides average age-wage profiles by industry, firm size, and educational background. We use only information about wage profiles by firm size (categorized into three groups: manufacturing firms with more than 1,000 employees, 999–100, and 99–10) from this Wage Census. This is because (a) wage profiles are determined largely by firm size in Japan, (b) wage data in industries other than manufacturing have problems in terms of continuity and availability, and (c) no information about educational background is available for samples of the survey. We project wage earnings backwards using estimated earnings at 7. The age classes are divided into five-year increments in the Wage Census. We thus collect data from the census every five years and reconstruct them for each cohort and year with a linear interpolation.

Table 7.5

Wage Functions Men

Independent Variables (Dummies) Age (default: age 55) 56 57 58 59 60 61 62 63 64 65 66 67 68 69 Mandatory retirement (default: no experience of mandatory retirement) 55–59 60–64 65–69 Occupation at age 55 (default: clark) Specialists Managers Salespersons Service Guards Farmers Trans. & com. Blue collar Construction Firm size at age 55 (default: less than 10) 10–99 100–299 300–999 1000+ Public sector Residential areas (default: Tokyo metropolitan area) Hokkaido Tohoku Kanto2 Hokuriku Tokai Kinki 1 Kinki 2 Chugoku (continued )

Women

Coefficient

SE

Coefficient

SE

–0.020 –0.074 –0.087 –0.099 –0.160 –0.290 –0.400 –0.457 –0.470 –0.551 –0.579 –0.524 –0.597 –0.675

(0.025) (0.027) (0.028) (0.029) (0.037) (0.043) (0.045) (0.050) (0.048) (0.061) (0.061) (0.070) (0.073) (0.086)

0.056 0.150 0.126 0.016 0.073 –0.099 0.059 0.060 0.007 0.019 0.142 0.053 –0.185 –0.075

(0.050) (0.056) (0.060) (0.059) (0.075) (0.072) (0.088) (0.084) (0.106) (0.100) (0.123) (0.136) (0.149) (0.177)

–0.183 –0.214 –0.230

(0.035) (0.050) (0.066)

–0.291 0.240 0.113

(0.091) (0.119) (0.141)

0.151 0.304 –0.061 –0.163 –0.182 –0.171 –0.229 –0.194 –0.122

(0.043) (0.037) (0.044) (0.057) (0.068) (0.063) (0.041) (0.036) (0.042)

0.138 0.405 –0.297 –0.383 –0.261 –0.581 –0.304 –0.491 –0.283

(0.076) (0.091) (0.050) (0.051) (0.317) (0.099) (0.154) (0.043) (0.099)

0.069 0.082 0.054 0.181 0.018

(0.026) (0.031) (0.036) (0.028) (0.036)

0.112 0.245 0.101 0.146 0.098

(0.040) (0.043) (0.060) (0.053) (0.085)

–0.180 –0.404 –0.190 –0.268 –0.151 –0.039 –0.225 –0.300

(0.043) (0.034) (0.035) (0.045) (0.031) (0.028) (0.066) (0.043)

–0.035 –0.070 –0.060 0.017 0.029 –0.018 –0.076 –0.053

(0.081) (0.055) (0.054) (0.076) (0.046) (0.054) (0.091) (0.061)

418

Takashi Oshio and Akiko Sato Oishi

Table 7.5

(continued) Men

Independent Variables (Dummies)

Women

Coefficient

SE

Coefficient

SE

Shikoku Northern Kyushu Southern Kyushu

–0.443 –0.313 –0.473

(0.061) (0.038) (0.047)

–0.093 0.001 –0.085

(0.082) (0.059) (0.085)

Constant Inverse Mills’ ratio Log-likelihood No. of observations

3.809 –0.319 –7268.9 6,979

(0.040) (0.026)

2.765 –0.402 –4144.4 3,710

(0.056) (0.080)

Notes: SE = standard error. The dependent variable is the logarithm of monthly earnings. Wage function was estimated by means of the Heckman two-step selection correction. Additional variables that were included in the participation probit were health status, mortgage loans, public pension benefits, private pension benefits, property income, and family members’ income.

fifty-five as a benchmark and the cohort-specific wage curve. But how do we know about the earnings history for an individual who has already been retired by the survey year? First, we have to estimate at which age they retired. The survey gives some information about an individual’s working status after mandatory retirement if applicable: The survey asks, for instance, whether one has kept working at the same or another firm after mandatory retirement (but the survey does not ask for how long). Based on this information, we make a rough estimation of each sample’s retirement age (see section 7.7.1). 7.6 Incentive Variable Calculation This section describes the construction of incentive measures and provides tabulations that illustrate them by age. These incentive measures are used to capture the impact of social security programs on retirement decisions in the next section. 7.6.1 Definitions and Methodology We construct three incentive measures: benefit accrual, option value, and peak value, each of which assesses the impact of social security programs upon retirement decisions. The key concept from which these three measures are derived is social security wealth (SSW), which is the present discounted value of lifetime social security benefits. Social security wealth is gross of wage taxation, but net of income taxation. It should be noted here that the income tax system is very generous to pensioners and other elderly people (especially those aged sixty-five and above); income tax levied on them is in most cases negligible due to lower tax rates and various tax exemptions.

Social Security and Retirement in Japan

419

The three incentive measures, the latter two of which are of forwardlooking type, are defined as follows. 1. Benefit accrual is the change in SSW at each age resulting from the postponement of retirement for one additional year. If the accrual is positive, an individual may want to postpone retirement since working for an additional year will raise SSW. If it is negative, social security will provide a disincentive to work. One problem with the accrual is that it does not take into account potential large accruals in the future. 2. The option value is the (expected) gain from postponing retirement to the age when an individual’s life-cycle utility is maximized (see Stock and Wise 1990). If one retires at age r, the discount utility at the current age t is given by r1

Vt (r)  ∑ stY s  st

S

∑

st

[kBs (r)] g,

sr

where S is the maximum age, Ys is wage earnings at age s, Bs (r) is SS benefits at age s (if retired at age r), and  is the discount factor. Let r ∗( r) denote the future retirement age yielding the highest value of utility; then the option value is given by Gt (r ∗)  Vt (r ∗)  Vt (t). The individual retires if G  0; otherwise he postpones retirement. We assume that  is equal to 0.97 (a three percent discount rate),  is equal to 0.75, and k is equal to 1.5, rather than structurally estimating them. 3. The peak value is defined as the difference between SSW today and SSW at its peak; that is, the sum of all accruals from today to the year when SSW is at its maximum. This is a simpler, less structural, alternative to the option value, with utility from wage earnings neglected. After the peak point, the peak value is equal to the annual accrual. Calculations of these incentive measures have to incorporate the multiple policies reviewed in section 7.2.3. Our construction of weighted average incentive measures neglects employer-provided pension (and lumpsum retirement) benefits. The possibility that these benefits affect people’s retirement decisions cannot be ruled out, but in most cases they are paid by firms at the mandatory retirement age of sixty regardless of the employee’s working or retirement status thereafter. On the other hand, there are some cases in which employer-provided pension benefits make it profitable to retire earlier than sixty with a reduction in the discount value of benefits. This effect, however, will not be explicitly analyzed, due to limited data. Hence, weighted average incentive measures reflect the following four programs: (a) KNH benefits, (b) Zaishoku Pension benefits, (c) WS, and (d) UI benefits. A KNH participant is eligible for KNH benefits at age

420

Takashi Oshio and Akiko Sato Oishi

sixty, but they can choose to keep working with earnings-tested Zaishoku Pension benefits. The WS is paid for those who keep working between sixty and sixty-four. Beyond the age of sixty-five, only KNH benefits apply. The construction of weighted average of incentive measures is based on the actual probability of receiving each measure at each age observed from the sample. Three points should be mentioned here in estimating social security incentives. First, for those aged sixty to sixty-four, UI benefits are 50–80 percent of the wage earnings at age sixty unless current wage earnings exceed them. This means that UI benefits are usually fixed for those aged sixty to sixtyfour regardless of retirement age, because wage earnings tend to decline sharply after sixty. For those aged fifty-nine or younger, however, UI benefits usually replace 60 percent of the current-age earnings and thus postponing retirement will affect the amount of UI benefits as well as SSW and its accrual. On the other hand, one cannot apply for UI benefits at age sixty-five and after. Hence, people tend to stop working between the ages of sixty and sixty-four and receive UI benefits. Second, there appear to be many cases in which workers receive full KNH benefits, probably due to an ineffective earnings test. Also, it is unclear whether or not the samples in the survey know their own type of pension benefits; some of those who respond that they are KNH beneficiaries might actually get Zaishoku (that is, earnings-tested KNH) benefits instead. In calculating social security incentives, we assume that public pension benefits that an individual gets while employed during the ages of sixty through sixty-four are earnings-tested Zaishoku rather than full KNH. While this assumption is loyal to the law, our calculations might more or less overestimate disincentives. Third, WS is treated as a negative premium to social security, while it does not affect SSW in gross terms. 7.6.2 Summary of Incentive Measures Based on the aforementioned methodology, we obtain tables 7.6 and 7.7 which illustrate weighted average incentive measures. The results set out in table 7.6 summarize SSW, its accrual, standard deviation, and the tax or subsidy rate by age for the median, tenth, and ninetieth percentiles, compared with a previous study by Yashiro and Oshio (1999). Table 7.7 provides similar calculations for the forward-looking incentive measures: peak and option values. Tables 7.6 and 7.7 include results for men and women. Among other things, the following results are most noteworthy. First, SSW peaks at age fifty-nine. This is consistent with the fact that the eligibility age for social security benefits is sixty and that most employees exit the labor force at that age. Accrual is positive until fifty-nine and negative after that. Almost flat SSW and zero accrual beyond age sixty-five in most cases reflect the KNH formula, which allows full benefits with no

Social Security and Retirement in Japan Table 7.6

Summary of Incentive Measures (in 1998 US$) Accrual

Age Men 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 Women 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

421

Tax or Subsidy Rate

SSW Median

Median

10th Percentile

90th Percentile

SD

Median

Yashiro and Oshio (1999)a

224,314 235,188 248,940 258,654 268,532 280,562 267,989 257,437 245,849 234,591 217,702 216,390 215,686 215,273 214,753 214,739

— 7,333 10,275 6,624 8,675 11,548 –13,351 –10,839 –11,504 –11,021 –16,400 0 0 0 0 0

— 5,087 7,236 4,581 6,077 7,676 –19,605 –15,381 –16,066 –17,823 –20,250 –5,975 –6,295 –5,820 –4,948 –6

— 12,729 14,569 11,124 12,637 15,368 –7,148 –5,344 –6,164 –5,720 –10,878 709 6 6 0 0

— 2,708 3,102 2,635 2,935 3,400 4,669 3,829 3,863 4,544 4,043 3,290 2,765 2,708 1,980 1,986

— –0.235 –0.328 –0.232 –0.299 –0.424 0.602 0.552 0.670 0.668 0.921 0 0 0 0 0

— –0.195 –0.202 –0.105 –0.112 –0.138 0.338 0.340 0.342 0.340 0.204 0 0 0 0 0

176,821 184,538 195,111 201,221 208,762 214,137 206,211 197,989 189,860 184,065 182,874 182,571 180,962 180,962 180,962 180,962

— 5,331 6,583 4,847 5,402 6,375 –5,656 –4,997 –6,164 –6,379 –7,702 0 0 0 0 0

— –6,992 –7,208 –8,630 –8,981 –7,318 –19,489 –17,246 –15,476 –12,085 –14,681 –86 0 0 0 0

— 13,917 16,036 12,935 12,542 10,651 –4,948 –3,930 –5,486 –5,692 –5,576 0 0 0 0 0

— 7,303 8,037 7,979 8,351 7,362 6,920 6,182 4,287 4,624 4,319 4,544 1,827 0 0 0

— –0.326 –0.369 –0.265 –0.328 –0.313 0.501 0.523 0.550 0.405 0.655 0 0 0 0 0

— — — — — — — — — — — — — — — —

Note: SD = standard deviation. Dashes indicate that data is not relevant. a Data in this column is from Yashiro and Oshio (1999).

earnings test beyond that age. For males in the tenth percentile, small negative accruals beyond that age probably reflect negative accruals for spouses who are under sixty-five. Hence, we can conclude that social security generally works as an incentive to work until the age of fifty-nine, but turns into a disincentive at sixty, and becomes neutral beyond sixty-five. Our previous study, Yashiro

422

Takashi Oshio and Akiko Sato Oishi

Table 7.7

Summary of Forward-Looking Incentive Measures (in 1998 US$) Option Value

Age Men 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 Women 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

Peak Value

Median

10th Percentile

90th Percentile

SD

Median

10th Percentile

90th Percentile

SD

98,159 85,850 70,414 59,819 36,815 34,991 34,160 32,158 33,786 34,165 37,283 29,757 21,821 15,262 7,169

58,659 54,364 38,429 30,435 14,881 9,769 8,019 6,688 10,787 11,635 21,610 15,268 10,162 7,109 2,462

174,423 162,567 127,216 102,082 80,452 73,357 66,012 63,352 66,021 65,569 62,258 49,844 38,727 26,446 16,191

50,837 40,166 33,609 29,047 26,847 26,817 22,320 20,685 20,873 20,765 17,004 14,770 12,594 7,942 5,710

34,527 29,025 21,129 14,951 11,548 –13,351 –10,839 –11,504 –11,021 –16,400 0 0 0 0 0

23,883 21,275 14,808 11,073 7,676 –19,605 –15,381 –16,066 –17,823 –20,250 –5,975 –6,295 –5,820 –4,948 –6

51,475 48,437 33,855 23,593 15,368 –7,148 –5,344 –6,164 –5,720 –10,515 709 6 6 5 0

12,074 10,186 7,365 5,521 3,400 4,669 3,968 3,863 4,544 5,170 4,691 2,767 3,531 2,190 1,986

97,410 89,592 81,684 69,357 60,662 38,592 36,448 29,199 42,238 31,242 39,971 24,595 11,090 12,403 4,709

57,318 49,213 42,410 35,179 27,029 16,679 16,013 13,705 17,204 12,581 19,841 14,802 2,303 7,468 3,377

175,919 169,527 136,370 122,228 110,406 90,373 86,467 78,554 80,978 70,477 71,221 60,220 36,714 24,307 11,448

48,550 45,357 41,965 44,534 34,426 31,509 28,853 27,158 28,416 25,036 28,657 18,043 11,076 9,402 4,836

25,274 20,100 14,821 10,420 6,375 –5,645 –4,997 –6,164 –6,379 –7,702 0 0 0 0 0

–6,992 –5,833 –8,247 –8,981 –7,318 –19,489 –17,246 –15,476 –12,085 –14,681 –86 0 0 0 0

47,980 37,130 30,816 20,705 10,651 –4,948 –3,930 –5,486 –5,692 –5,576 0 0 0 0 0

18,850 16,100 14,377 11,770 7,389 6,923 6,218 4,327 4,624 4,319 4,948 1,899 0 0 0

Note: SD = standard deviation.

and Oshio (1999), which assumed that all (male) employees get Zaishoku benefits if they keep working, neglects UI benefits and assumed no wage growth. In the current paper, we take into account the case of receiving no Zaishoku benefits and going directly onto SS, include UI benefits, and reflect a projected reduction in wage earnings based on the cross-sectional data. As a result, the implied tax rate during the ages of sixty through sixtyfour is larger than found in our previous study. Turning to forward-looking incentive measures, the option value declines by age, suggesting that a disincentive to work tends to increase by

Social Security and Retirement in Japan

423

age. This probably can be attributed to both declining SSW and an increasing risk of lower wage earnings when postponing retirement. The option value continues to fall even beyond age sixty-five, in contrast with SS accrual which is flat beyond sixty-five; this is because SS accrual does not reflect a reduction in wage earnings beyond that age. The pattern of a change in the peak value by age is also consistent with that of benefit accrual; it is positive until age fifty-nine and then turns negative. In addition, tables 7.6 and 7.7 confirm that the peak value is simply annual SS accrual after the year when SSW is at its maximum, consistent with its definition. 7.7 Empirical Framework for Regression Analysis In this section, we describe the empirical framework for regression analysis on the impact of social security on retirement. However, we first have to estimate each sample’s previous working or retirement status, since our survey tells us only whether or not each sample is retired or working in the survey year of 1996. Hence, we first explain how to build up the quasilongitudinal data; then we address the reduced form models of retirement decisions. 7.7.1 Estimation of Retirement Age and Changes in Working Status To estimate models for incentive measures we select from the survey the individuals who are expected to have kept working until 1995, one year before the survey year, and we apply the probit model to them to explain their retirement decisions (whether or not to keep working or to retire) in 1996.8 The main problem of our analysis is that we cannot exactly identify those who were working in 1995, due to a lack of longitudinal information. Hence, we first assume that those who were working in 1996 were working in 1995 too. And for those who were already retired, we only use those whose age of retirement can be identified from their reported answers about mandatory retirement and subsequent job experience. Table 7.8 summarizes an estimated change in working or retirement status for those who are estimated to have been working in 1995. Out of the total sample, 2,629 men are estimated to have been working in 1995, and 2,296 of them kept working and 333 retired in 1996. As for women, 1,204 of 1,459 kept working and 295 retired. For men the hazard rate is very high for those aged sixty or sixty-one in 1995, roughly consistent with the actual trend of labor force participation. 8. It might be possible to take individuals who were working in 1996 out of the sample and see whether or not they will retire in 1997, since we can construct a forward-looking panel using the reported answers as to when they wish to retire. We do not do this, however, because such answers do not appear reliable enough to use for estimating retirement age.

424

Takashi Oshio and Akiko Sato Oishi

Table 7.8

Changes in Working Status of Employees (1995–1996) Males

Females

Working Status in 1996 Age in 1995 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 Total

Working Status in 1996

N

Employed

Retired (Def. I)

%

N

Employed

Retired (Def. I)

%

103 150 159 237 253 293 292 193 201 177 149 135 112 103 72 2,629

102 141 151 232 238 209 199 176 172 158 124 123 102 100 68 2,296

1 9 8 5 15 84 93 17 29 19 25 12 9 3 4 333

1.0 6.0 5.0 2.1 5.9 28.7 31.8 8.8 14.4 10.7 16.8 8.9 8.0 2.9 5.6 12.7

167 200 168 157 144 144 114 88 73 70 44 34 25 16 15 1,459

158 180 146 126 120 94 75 63 65 61 36 31 22 15 12 1,204

9 20 22 31 24 50 39 25 8 9 8 3 3 1 3 255

5.4 10.0 13.1 19.7 16.7 34.7 34.2 28.4 11.0 12.9 18.2 8.8 12.0 6.3 20.0 17.5

Source: Authors’ calculations from the SLMPOP (MOL 1996). Notes: N = number of observations. “Employed” includes executives.

7.7.2 Model Specification The dependent variable is a dummy for whether or not an individual (who is expected to have been working in 1995) retired in 1996, with retirement defined according to Definition I, which is described in section 7.2.3. We choose Definition I, the broadest definition of retirement, largely because most of the self-employed probably have been retired from the firms and their income seem to at least partly rely on public pension benefits. Then, we estimate the retirement models by probit for three incentive measures: accrual, peak value, and option value. The central issue is which controls to include in the retirement models. In particular, age itself may be very important in Japan: Most people are effectively forced to leave firms at age sixty. Meanwhile, our preliminary regressions indicate that there is relatively little value added by showing the variation in results when demographic and earnings controls are and are not incorporates in the models. Here we estimate two models: one model (M1) has all controls for earnings, demographics, sectors, and a linear age term; the other model (M2) has all these controls but replaces the linear age terms with age dummies. We estimate these two models for each incentive measure: accrual, peak

Social Security and Retirement in Japan Table 7.9

425

Summary Statistics Males

Retired (Def. I) SSW (US$ 10,000) SSA (US$ 10,000) Peak value (US$ 10,000) Option value (US$ 10,000) Property income (US$ 1,000) Health condition: not well Health condition: bad or sick PE (US$ 1,000) ALE (US$ 1,000) Square of PE Square of ALE Age in 1995 Sample N

Females

Mean

SD

Mean

SD

0.127 27.444 –0.270 0.134 5.166 0.160 0.151 0.035 2.424 2.792 8.272 11.109 60.324

0.333 9.292 1.111 1.721 4.044 1.057 0.358 0.183 1.548 1.821 11.858 17.430 3.660

0.175 22.462 –0.053 0.499 7.253 0.069 0.154 0.028 1.391 1.467 2.645 3.082 58.352

0.380 11.857 0.915 1.675 5.099 0.306 0.361 0.165 0.843 0.964 4.641 6.580 3.422

2,629

1,459

Notes: All dollar values are in 1998 US$ ($1 = ¥131.02). SD = standard deviation; PE = projected earnings; ALE = average lifetime earnings; N = number of observations.

value, and option value for men and women, separately. Each model includes SSW. Earnings controls consist of projected earnings for next year, average lifetime earnings, and the squares of each. Other controls include property income, dummies for health conditions, nine occupational dummies, dummies for four categories of firm size at age fifty-five, and eight dummies for residential area. 7.8 Estimation Results Table 7.9 shows the summary statistics for the sample, and tables 7.10 and 7.11 summarize estimation results for men and women, respectively. In tables 7.10 and 7.11, each incentive measure has two columns for M1 (with a linear age term) and M2 (with age dummies), with coefficients for controls other than earnings omitted to save space. The coefficient on each incentive measures is expected to be negative, since they should reduce the probability of retirement. The following four findings are noteworthy in assessing the impact of each incentive measure. First, coefficients on incentive measures are all negative and statistically significant for men, except for the M2 option value model. In the peak value M1 model, for example, one thousand dollar increase in the peak value would raise the hazard rate by 0.62 percent points. For women, only the M1 accrual and M1 peak value models show negative and significant coefficients on incentive measures. Men are more sensitive to incentive measures than women, probably because men’s labor participation is much

Table 7.10

Retirement Probits (male sample) Accrual Model

SSW $10,000 change Incentive measure $1,000 change Property income Health condition: not well Health condition: bad or sick PE ALE Square of PE Square of ALE Age 55 56 57 58 59 60 61 62 63 64 65

Peak Value Model

Option Value Model

Linear Age

Age Dummies

Linear Age

Age Dummies

Linear Age

Age Dummies

0.001 (0.006) 0.02 –0.399 (0.048) –0.68 0.102 (0.050) 0.265 (0.089) 0.929 (0.157) 0.054 (0.945) 0.477 (0.783) 0.059 (0.128) –0.187 (0.093) –0.032 (0.020)

–0.006 (0.006) –0.10 –0.228 (0.109) –0.37 0.103 (0.049) 0.292 (0.093) 1.011 (0.155) –1.447 (1.091) 1.817 (0.929) 0.297 (0.139) –0.368 (0.110)

–0.002 (0.006) –0.03 –0.367 (0.039) –0.62 0.101 (0.048) 0.265 (0.090) 0.958 (0.160) 0.282 (0.998) 0.343 (0.837) 0.130 (0.141) –0.246 (0.105) –0.045 (0.020)

–0.005 (0.006) –0.09 –0.204 (0.081) –0.33 0.103 (0.049) 0.290 (0.093) 1.012 (0.156) –1.479 (1.091) 1.851 (0.931) 0.318 (0.140) –0.384 (0.112)

0.008 (0.006) 0.15 –0.035 (0.015) –0.06 0.097 (0.050) 0.255 (0.088) 0.886 (0.153) –2.533 (0.854) 2.591 (0.718) 0.205 (0.118) –0.277 (0.089) –0.084 (0.020)

–0.007 (0.006) –0.11 0.019 (0.017) 0.03 0.102 (0.049) 0.283 (0.093) 1.015 (0.154) –1.365 (1.106) 1.756 (0.939) 0.252 (0.140) –0.336 (0.111)

0.912 (0.445) 0.703 (0.445) 0.383 (0.473) 1.047 (0.466) 1.421 (0.513) 1.527 (0.522) 0.604 (0.549) 0.914 (0.554) 0.539 (0.588) 1.272 (0.504) 0.864 (0.520)

0.790 (0.446) 0.483 (0.452) 0.020 (0.499) 0.541 (0.494) 0.979 (0.578) 1.086 (0.583) 0.169 (0.606) 0.479 (0.611) 0.120 (0.643) 0.816 (0.550) 0.402 (0.564)

0.858 (0.445) 0.761 (0.453) 0.399 (0.491) 1.038 (0.498) 1.957 (0.498) 2.002 (0.516) 1.099 (0.541) 1.405 (0.545) 1.143 (0.561) 1.522 (0.523) 1.147 (0.544)

Table 7.10

(continued) Accrual Model Linear Age

66

Linear Age

0.689 (0.559) 0.302 (0.605) 0.479 (0.639)

67 68 Pseudo R 2 Other controls

Age Dummies

Peak Value Model

0.172 Yes

0.225 Yes

Age Dummies

Option Value Model Linear Age

0.235 (0.604) –0.159 (0.646) 0.017 (0.678) 0.185 Yes

0.226 Yes

Age Dummies 0.969 (0.588) 0.587 (0.633) 0.772 (0.676)

0.137 Yes

0.233 Yes

Notes: Other control variables are 9 occupational dummies, dummies for 4 categories of establishment size, and 8 regional dummies. The estimated parameters on these variables are not reported. Figures in parentheses show robust standard errors. PE = projected earnings; ALE = average lifetime earnings.

Table 7.11

Retirement Probits (female sample) Accrual Model

SSW $10,000 change Incentive measure $1,000 change Property income Health condition: not well Health condition: bad or sick PE ALE Square of PE Square of ALE (continued )

Peak Value Model

Option Value Model

Linear Age

Age Dummies

Linear Age

Age Dummies

Linear Age

Age Dummies

0.007 (0.004) 0.14 –0.150 (0.058) –0.32 0.804 (0.155) 0.124 (0.115) 1.025 (0.219) –0.141 (1.797) 1.842 (1.518) –0.522 (0.524) –0.402 (0.393)

0.010 (0.004) 0.21 0.024 (0.069) 0.05 0.783 (0.149) 0.144 (0.120) 1.088 (0.227) 0.680 (2.033) 1.087 (1.754) –0.681 (0.555) –0.260 (0.424)

0.006 (0.004) 0.14 –0.101 (0.039) –0.08 0.807 (0.156) 0.127 (0.115) 1.015 (0.220) –0.310 (1.812) 1.976 (1.530) –0.468 (0.531) –0.437 (0.398)

0.011 (0.004) 0.00 0.040 (0.046) –0.14 0.781 (0.149) 0.143 (0.120) 1.090 (0.226) 0.763 (2.022) 1.024 (1.745) –0.712 (0.548) –0.241 (0.420)

0.012 (0.005) 0.26 –0.008 (0.019) –0.02 0.805 (0.155) 0.122 (0.114) 0.999 (0.220) –0.660 (1.753) 2.340 (1.490) –0.442 (0.514) –0.482 (0.387)

0.001 (0.006) 0.03 0.041 (0.019) 0.08 0.779 (0.148) 0.148 (0.121) 1.101 (0.226) 0.752 (2.039) 0.965 (1.762) –0.769 (0.555) –0.195 (0.426)

428

Takashi Oshio and Akiko Sato Oishi

Table 7.11

(continued) Accrual Model Linear Age

Age

Age Dummies

0.015 (0.016)

55

Peak Value Model Linear Age 0.007 (0.017)

0.251 (0.217) 0.511 (0.216) 0.724 (0.208) 0.524 (0.215) 1.218 (0.255) 1.092 (0.234) 0.977 (0.259) 0.359 (0.314) 0.416 (0.298) 0.768 (0.300) 0.365 (0.383) 0.305 (0.419) 0.065 (0.551) 0.761 (0.485)

56 57 58 59 60 61 62 63 64 65 66 67 68 Pseudo R 2 Other controls

0.161 Yes

Age Dummies

0.200 Yes

Option Value Model Linear Age 0.018 (0.021)

0.276 (0.219) 0.548 (0.225) 0.778 (0.222) 0.594 (0.232) 1.307 (0.252) 1.178 (0.260) 1.065 (0.283) 0.445 (0.331) 0.502 (0.319) 0.840 (0.316) 0.435 (0.395) 0.382 (0.431) 0.142 (0.559) 0.836 (0.496) 0.161 Yes

Age Dummies

0.201 Yes

0.318 (0.221) 0.641 (0.230) 0.896 (0.234) 0.744 (0.248) 1.466 (0.260) 1.336 (0.266) 0.128 (0.289) 0.627 (0.343) 0.701 (0.327) 1.089 (0.350) 0.747 (0.430) 0.686 (0.461) 0.469 (0.585) 1.160 (0.528) 0.155 Yes

0.204 Yes

Note: See table 7.10.

higher and their retirement decisions are much more linked with pension benefits. Second, compared to M1 models, coefficients on incentive measures are either smaller and less significant, have wrong signs, or both in M2 models. This result suggests that M2 specification “overfits” the data, in that age dummies absorb much of retirement incentives. Indeed, figure 7.2 illustrates, for men and women respectively, how the hazard rate at each age would rise when each age dummy is raised from zero to one, compared with the actual hazard rates. These figures show that for all cases of accrual,

Fig. 7.2

A

The retirement hazard and age dummies: A, Males; B, Females

Fig. 7.2

B

(cont.) The retirement hazard and age dummies: A, Males; B, Females

Social Security and Retirement in Japan

431

peak value, and option value models age dummies trace well the actual age pattern of hazard rates. Third, in terms of explanatory power, the peak value models look better than other models for men, while there is no big difference for women. While the fit is better in M2 models than in M1 models, coefficients on incentive measures tend to be either smaller and less significant, have wrong signs, or both in M2 as mentioned above. The explanatory power of the option value model looks relatively weak. This result seems plausible, judging by the fact that the option value monotonically declines as one gets older (as shown in table 7.7)—which is not consistent with the age pattern of hazard rates. Finally, turning to other variables, SSW itself does not seem to be important in retirement decisions; its coefficient is not significant, especially in the case of men. The incentive effect of social security benefits works largely through dynamic incentive measures rather than SSW, and the wealth effect does not seem to be large. Also, supporting intuition about income and substitution effects, average lifetime earnings tend to increase disincentive to work, while projected earnings tend to decrease it. However, the value and significance differs substantially depending on model specifications. All in all, the estimation results confirm that all dynamic incentive measures at least partially affect retirement decisions of the elderly, while the option value models show poorer performance than the others. The true impact of incentive measures is very difficult to assess, since their significance varies greatly when age dummies are in and out of the models. Thus, we should present a range of predictions based on a variety of models, instead of searching for the single best model, to predict the impact of policy changes. 7.9 Policy Simulations In this section of policy simulations, we quantitatively assess the responsiveness of retirement decisions to social security reform. We propose two simulations for reform plans, which are described in the following sections. 7.9.1 Two Reform Plans The first reform plan—referred to as the plus-three-years reform—is to raise both the early and normal eligibility age for the social security program by three years. In Japan, those ages correspond to sixty and sixty-five years, respectively. (More specifically, age sixty is the eligibility age for full benefits, but benefits are earnings-tested if one remains employed; at the age of sixty-five and over, one can get full benefits with no earnings test.) The simulation raises these threshold ages to sixty-three and sixty-eight,

432

Takashi Oshio and Akiko Sato Oishi

respectively. The eligibility ages for Zaishoku Pension benefits and WS and the age pattern of receiving UI benefits also are also raised by three years. For this reform plan, we consider three different scenarios.

• Simulation 1 (S1): increments the incentive and SSW measures and UI eligibility probabilities according to the policy changes from the model without age dummies (M1). • Simulation 2 (S2): increments the incentive and SSW measures and UI eligibility probabilities from the model with age dummies (M2), with age dummies unchanged. • Simulation 3 (S3): increments the incentive and SSW measures, UI eligibility probabilities, and age dummies from the model with age dummies (M2). It seems likely that simulations S2 and S3 will bound true responses to policy changes, with simulation S1 lying somewhere in between. The second reform plan is to implement the common reform, which has the following features: 1. An early eligibility age of sixty; 2. A normal retirement age of sixty-five; 3. A replacement rate of 60 percent (of earnings at the age of fifty-nine) at age sixty-five; 4. A six percent per-year actuarial reduction for retirement before sixtyfive and six percent actuarial increase for retirement after sixty-five; and 5. No other pathways to retirement. While this simulation allows us to compare the impact on the common reform across countries, some comments should be made regarding each component of this reform plan in Japan’s case: 1. No change is necessary because the early eligibility age is currently sixty; 2. No change is necessary because the normal eligibility age is currently sixty-five; 3. The tax rate at age sixty-five is 92 percent for men and 66 percent for women as indicated in table 7.6, suggesting that a replacement rate of 60 percent will lower the tax rate and disincentives to work at that age; 4. The net effect of this actuarial adjustment is uncertain, since existing Zaishoku benefits are to be abolished; and 5. “No other pathways to retirement” means the abolishment of Zaishoku, WS, and UI benefits. For this reform plan, we consider three different scenarios.

• S1: calculates incentive and SSW measures according to the new policy from the model without age dummies (M1).

Social Security and Retirement in Japan

433

• S2: calculates incentive and SSW measures according to the new policy from the model with age dummies (M2), with age dummies unchanged. • S3: calculates incentive and SSW measures according to the new policy from the model with age dummies (M2), and change the age dummies. The goal of this simulation would be to maintain the portion of an age dummy that reflects increasing desire for leisure as one ages, and to discard the component that reflects the effect of retirement programs (not captured by the incentive measures), with the exception of effects due to early retirement and to normal retirement eligibility. We perform these simulations by taking the estimated retirement model, plugging in new incentive measures and possibly new retirement ages in place of the existing ones and estimating for each individual a new probability of retirement. Then, we average the estimated probabilities at each age to the new age-specific retirement rates. Also, we estimate the cumulative hazard rate at each age as well as average retirement ages. 7.9.2 Simulation Results Figures 7.3 to 7.11 summarize the simulation results for men. Each figure has two graphs: The first graph compares the baseline hazard rate and hazard rate under each of the two policies for each simulation and incentive combination; the second graph shows cumulative hazard rates for the baseline and each of the two policies for the same combination. The following findings should be mentioned. First, in the case of the plus-three-years reform, S1 and S3 shift the spike of the hazard rate to age sixty-three to sixty-four from age sixty to sixty-one for the accrual and peak value models. By contrast, S2 does not show any clear shift in the spike, although it somewhat reduces the hazard rates. The policy impact in S3 is thus most probably due to a change in the age dummies, and S1 lies between the two extremes of S2 and S3. This result is also in line with the fact that the coefficient on the incentive measure is smaller if age dummies are included. Second, the common reform moderates the hazard rates across ages. This is probably because the abolishment of Zaishoku, WS, and UI benefits—together with actuarial adjustment of pension benefits—moderates the age pattern of incentive measures. At the same time, this reform fails to postpone retirement substantially; in many cases, the hazard rate becomes higher until age sixty. This is because the common reform makes social security benefits less linked to the number of contribution years and reduces both the accrual and peak value below age sixty, while the reform raises them during ages sixty and sixty-four. Third, as shown in figures 7.9 and 7.10 the option value model tends to be insensitive to policy reform (with no change in age dummies). This is

Fig. 7.3

A

Simulation S1 on males using accrual estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.3

B

(cont.)

Fig. 7.4

A

Simulation S2 on males using accrual estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.4

B

(cont.)

Fig. 7.5

A

Simulation S3 on males using accrual estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.5

B

(cont.)

Fig. 7.6

A

Simulation S1 on males using peak value estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.6

B

(cont.)

Fig. 7.7

A

Simulation S2 on males using peak value estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.7

B

(cont.)

Fig. 7.8

A

Simulation S3 on males using peak value estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.8

B

(cont.)

Fig. 7.9

A

Simulation S1 on males using option value estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.9

B

(cont.)

Fig. 7.10

A

Simulation S2 on males using option value estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.10

B

(cont.)

Fig. 7.11

A

Simulation S3 on males using option value estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.11

B

(cont.)

452

Takashi Oshio and Akiko Sato Oishi

because the coefficient on the option value is all quite small, as reported in table 7.10. The option value model, which fails to trace the age pattern of retirement, is not good at assessing the impact on retirement from policy changes. Table 7.12 shows what the model predicts will happen to average retirement ages. The current average retirement age is 60.8 for men. The plusthree-years reform increases the average retirement age to 61.8 on average. By contrast, the common reform slightly reduces the average retirement age to 60.4, largely reflecting an increase in the hazard rate before the early retirement age of sixty. Hence, the plus-three-years reform is more effective than the common reform in Japan. In particular, this kind of reform, assuming the combination of the peak value and S3, would be most efficient in postponing retirement—with the average retirement age raised by 2.4 years to 63.2. Table 7.12

Average Retirement Ages in Simulations Plus-Three-Years Reform

Actual Accrual S1 S2 S3 Average Peak value S1 S2 S3 Average Option value S1 S2 S3 Average S1 average S2 average S3 average Average Actual Peak value S1 S2 S3 Average

Common Reform

Males 60.8

60.8

61.0 61.0 63.0 61.7

60.5 60.9 60.4 60.6

62.1 61.4 62.8 62.1

59.8 60.4 60.0 60.1

61.0 60.7 63.2 61.6 61.4 61.0 63.0 61.8

60.7 60.8 60.5 60.7 60.3 60.7 60.3 60.4

Females 59.3 59.9 59.3 60.8 60.0

59.3 59.2 59.4 59.6 59.4

Note: The average retirement age is the actual retirement age plus the estimated change from the baseline in each case.

Social Security and Retirement in Japan

453

The same kind of policy simulations can be conducted for women. The bottom part of table 7.12 and figures 7.12–7.14 summarize the results of the peak value model, which seems to work best for Japanese women in terms of significance and signs of coefficients on incentive measures. The plus-three-years reform turns out to postpone retirement, but not as much as for men. This reform shifts the spike of the hazard rate by three years in the case of S3 (see figure 7.14, panel A), but it seems to be mostly due to a change in age dummies. The common reform shows no significant impact. 7.10 Concluding Remarks This paper analyzes the economic impact of social security incentives on retirement decisions, based on the micro-data from the Survey on Labor Market Participation of Older Persons (SLMPOP; MOL 1996). Our estimations confirm that the incentive measures—such as benefit accrual, the peak value, and option value—at least partially affect retirement decisions, although their impact is not easy to identify. In particular, individuals aged sixty to sixty-five face substantial disincentives to work due to public income support programs, including public pension and UI benefits. In the face of a rapidly aging population, labor force participation of elderly people is crucial for growth potential and the fiscal position of the public pension scheme. Our policy simulations quantitatively capture the potential impact of pension reforms on retirement decisions through incentive measures. For example, an increase in the early and normal eligibility ages is most likely to reduce a disincentive to work for elderly people. A three-year increment of those eligibility ages is expected to raise the average retirement age by about one year for men, while the impact varies greatly due to a choice of incentive measures and model specifications. We also find that the proposed common reform fails to postpone retirement. This is probably because the early and normal retirement ages are already sixty and sixty-five, respectively, in Japan and proposed actuarial adjustment fail to offset the impact of eliminating the existing incentives to work. The 1999 Pension Reform Act may be more aggressive than the proposed common reform, in that the act aims to completely raise the eligibility age to sixty-five and to reduce total pension benefits for employees. Our analysis centers on the supply side of the labor market for the elderly, estimating the impact on their retirement decisions of social security reform. However, the demand side matters, too. In Japan, it is not easy for employees to find a full-time job after the mandatory retirement at age sixty. If demand for elderly workers remains subdued due to institutional and other reasons, any social security reform that aims to increase incentives to work could just lower wage income for elderly workers. Future research should be directed at more comprehensively assessing impact on the

Fig. 7.12

A

Simulation S1 on females using peak value estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.12

B

(cont.)

Fig. 7.13

A

Simulation S2 on females using peak value estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.13

B

(cont.)

Fig. 7.14

A

Simulation S3 on females using peak value estimates: A, Simulated hazard; B, Cumulative probability

Fig. 7.14

B

(cont.)

460

Takashi Oshio and Akiko Sato Oishi

labor market for the elderly of policy changes, taking into account potential changes in firms’ behaviors under the effects of population aging and policy measures to stimulate demand for elderly workers.

References Abe, Y. 1998. Labor supply of the elderly male and the earnings-tested pension in the 1980s and 1990s (in Japanese). JCER Economic Journal 36:50–82. Iwamoto, Y. 2000. The social security earnings test and labor supply of the elderly (in Japanese). The Quarterly of Social Security Research 35 (4): 364–76. Management and Coordination Agency. 1999. Labor force survey 1998. Tokyo: Management and Coordination Agency. Ministry of Labor (MOL). 1996. Survey of labor market participation of older persons (SLMPOP). Tokyo: MOL. Ogawa, H. 1998. On the impact of the public pension on labor supply of the elderly (in Japanese). The Economic Review 49 (3): 245–58. Oishi, A., and T. Oshio. 2000. Social security wealth and retirement decisions (in Japanese). The Quarterly of Social Security Research 35 (4): 405–19. Oshio, T. 1997. Pension and employment insurance reforms and pension wealth (in Japanese). The Quarterly of Social Security Research 33 (3): 286–97. Seike, A. 1991. Lifetime pension wealth and retirement (in Japanese). The Economic Review 42 (1): 12–20. ———. 1993. The labor market in the aging society (in Japanese). Tokyo: ToyoKeizai Simposha. Stock, J. H., and D. A. Wise. 1990. Pensions, the option value of work, and retirement. Econometrica 58 (5): 1151–80. Takayama, N. 1998. The morning after in Japan: Its declining population, too generous pensions and a weakened economy. Tokyo: Maruzen Co., Ltd. Takayama, N., F. Funaoka, F. Ohtake, M. Sekiguchi, T. Shibuya, F. Ueno, and K. Kubo. 1990a. Estimation of human capital and income redistributional effect of the public pension (in Japanese). Economic Analysis 118. ———. 1990b. The public pension and labor supply of the elderly male (in Japanese). Economic Analysis 121. Yashiro, N., and K. Nikami. 1996. Employment insurance reform and its implications on the employment of older persons (in Japanese). JCER Economic Journal 33:177–203. Yashiro, N., and T. Oshio. 1999. Social security and retirement in Japan. In Social security and retirement around the world, ed. J. Gruber and D. A. Wise, 239–67. Chicago: University of Chicago Press. Yashiro, N., T. Oshio, N. Li, M. Matsuya, Y. Terasaki, M. Yamagishi, M. Miyamoto, and Y. Igarashi. 1997. The economics of aging (in Japanese). Economic Analysis 151.

8 Incentives and Exit Routes to Retirement in the Netherlands Klaas de Vos and Arie Kapteyn

8.1 Introduction The programs providing income to the elderly in the Netherlands may be characterized by a limited number of salient features. First, there is a distinct cutoff at age sixty-five. Broadly speaking, all individuals aged sixtyfive or over are entitled to the same basic state pension (social security; we will abbreviate social security to SS throughout). Most other benefits (e.g., disability [DI], unemployment [UI], and welfare) expire when someone turns sixty-five. Second, for people both above and below sixty-five, in addition to the public entitlement programs guaranteed by law, relatively many people who stop working are entitled to other, private benefits (e.g., occupational pensions supplementing SS for individuals over sixty-five and early retirement [ER] benefits for individuals below sixty-five). Like most other developed countries, the Netherlands is faced with an increasing share of elderly in the total population. The share of the population over sixty-five has grown from 8 percent in 1950 to 14 percent in 2000 and is expected to rise to 24 percent by the year 2035. If nothing else changes, this will cause a considerable increase in SS expenditures. Faced with this prospect, the government has recently come up with policy measures to maintain its sustainability. Still, because of the relatively large role of fully funded occupational pensions supplementing SS, the problems facing the future of the SS system in the Netherlands are less severe than in countries like Germany, France, and Italy. A greater and more immediate concern is the low labor force participation rate of individuals aged below sixty-five and the costs of the programs Klaas de Vos is a researcher at CentER Applied Research, Tilburg University. Arie Kapteyn is a senior economist at RAND.

461

462

Klaas de Vos and Arie Kapteyn

providing income to the individuals in this age group who have left the labor force. This concerns public programs, such as DI, but also private ER schemes. During the 1980s and 1990s, these programs have been used explicitly or implicitly to enable almost all employees to retire before reaching the age of sixty-five. The ER programs were originally devised to help combat high unemployment by creating job opportunities for younger individuals. Public DI and UI programs have also been used to enable employers to shed older, less-productive workers. Since the financial conditions for retirement through all of these routes were quite attractive, most of the individuals eligible chose to retire before the normal retirement age of sixty-five. All of this has led to highly increased costs, both for the public DI and UI programs as well as for the employer-financed ER schemes. In reaction, eligibility conditions for DI and UI benefits have been tightened, and, increasingly, ER schemes are replaced by flexible retirement programs offering less attractive and more actuarially fair pensions. In addition, the earlier pressure on elderly employees to vacate one’s job for a younger jobseeker has decreased dramatically with the spectacular reduction in unemployment. It remains to be seen to what extent all these phenomena will actually contribute to a reversal of the trend of decreasing labor market participation of the elderly. In this paper, using micro-data from the years 1984 to 1995, we try to assess to what extent financial incentives can be seen to determine the retirement decision. In doing so, we can also simulate the effects of possible reforms on participation rates. The remainder of the paper is structured as follows. In section 8.2, we sketch the institutional framework within which people retire in the Netherlands. Sections 8.3 and 8.4 describe the data and the way in which we use them for our analysis. Section 8.5 provides a brief summary of research on retirement in the Netherlands. Section 8.6 is devoted to the construction of incentive measures that are used in the estimation of the retirement equations. These equations are specified according to a common model that is, by and large, used for all countries represented in this study. Section 8.7 presents the estimation results for the common model, and section 8.8 gives results of some policy simulations based on the estimated common model. We find strong and statistically significant incentive effects for males. For females, the estimated effects are smaller and much less significant. Section 8.9 concludes. 8.2 Institutional Background Social security guarantees a sufficient income to virtually all individuals of sixty-five or over. Basically, SS is a flat-rate benefit equal to half the statutory minimum wage (after tax) with supplements for single individuals and for individuals with a spouse aged younger than sixty-five with a

Incentives and Exit Routes to Retirement in the Netherlands

463

low income. Social security is financed largely as a pay-as-you-go scheme administered through a payroll tax on taxable income of individuals aged below sixty-five. The 2000 associated tax rate was 17.9 percent levied on taxable income up to a maximum (of Dutch Fl 48,994 per annum). In 1999, SS benefits amounted to almost Fl 41 billion, or 5 percent of gross domestic product (GDP). Currently, about one in every five households in the Netherlands receives SS. The entitlement to SS does not require retirement from the labor force. 8.2.1 Other Public Programs A number of arrangements exist which enable individuals to stop working before turning sixty-five. The main ones are DI, UI, and various ER schemes. DI covers all employees against loss of earnings due to long-term sickness and disability. Currently, DI guarantees employees who lost more than 80 percent of their earnings capacity a benefit equal to 70 percent (80 percent before 1985) of their daily wage (up to a maximum amount). The benefit falls to a lower level after a certain period (both the length of this period and the percentage depend on age) and terminates at age sixty-five. Most employees have taken out an additional insurance to cover the risk of a DI benefit falling below 70 percent of their previous earnings.1 In the 1980s, the DI program became a very popular arrangement that employers could use to shed elderly, less-productive employees. Severe legal obstacles existed (and still exist) to lay off employees, while DI benefits were more generous than UI benefits. As a result of this, both employers and employees had a preference for the DI route to unemployment. The ensuing rise in costs of DI has induced the government to limit eligibility for DI by tightening entry conditions and reducing benefit levels. Moreover, individuals receiving DI benefits are now subject to a more rigorous screening of their loss of earnings capacity. The main reason why UI is less attractive than DI is that UI benefits are only paid for a limited period (dependent on the number of years worked before unemployment). However, most people aged sixty or above who become unemployed can expect to receive unemployment benefits equal to 70 percent of their previous earnings up to age sixty-five.2 All public benefits for individuals younger than sixty-five are only paid to the extent that an individual is not employed.3 1. It should be noted that for single earners who lost more than 80 percent of their earnings capacity, DI benefits are always at least as high as the relevant social assistance (welfare) level, which (for a couple) is approximately equal to the after-tax minimum wage. In contrast to the entitlement to social assistance, household wealth is not taken into account when determining eligibility. 2. Similar to the case for DI benefits, if necessary, the UI benefit is supplemented by welfare benefits to reach the social assistance level, without taking household wealth into account. Hence, for single earners with low wages, the replacement rate can be almost 100 percent. 3. For individuals in part-time employment, benefits may supplement their earnings.

464

Klaas de Vos and Arie Kapteyn

8.2.2 Private Transfers Next to SS, a majority of the population over sixty-five is entitled to a supplementary occupational pension. In general, if an employer offers a pension scheme, then participation in such a scheme is compulsory. Until recently, more than 99 percent of the pension schemes were of the definedbenefit type, most of them being defined on the basis of final pay. Typically, occupational pensions supplement SS to 70 percent of final pay for individuals who have worked for forty years. After tax, the replacement rate is usually substantially higher. Most large firms have their own pension fund, smaller firms usually participate in sector-wide pension funds. Usually, these private pension arrangements require that people leave the job in which they accumulate pension rights at age sixty-five at the latest. There is no earnings test, however, and people may consider looking for secondary jobs once they retire. Early retirement became increasingly common during the 1980s and was viewed as a means of reducing unemployment. Typically, the ER schemes guarantee an employee a benefit equal to 70 or 80 percent of previous earnings up to the age of sixty-five. In after-tax terms, replacement rates are even higher. Furthermore, while being in ER, one often keeps accumulating pension rights, although possibly at a lower rate than when one would be working. ER may be organized via the pension funds, which also provide the occupational pensions, or via the employer. Moreover, in contrast to pensions, ER is mostly financed as pay-as-you-go and usually requires ten years of employment with the same employer before the ER date, whereas old age pension rights remain valid if the worker changes jobs. The receipt of ER pensions usually requires a complete withdrawal from the labor market. In recent years, costs of ER have increased considerably, and many firms are currently trying to reduce these costs. In particular, as mentioned in the introduction, instead of the original arrangements that provided incentives to retire as soon as one was eligible, more and more programs are being introduced which offer flexible ER pensions of which the level depends on the retirement age. Despite these developments, the general conclusion remains that an elaborate system of income-replacing transfers exists in the Netherlands, which can be expected to act as incentives to leave the labor force on one’s sixty-fifth birthday at the latest. Moreover, it should be noted that whereas rather strict laws are in force that prevent employers from laying off younger employees, reaching the age of sixty-five is a legal reason for dismissal, and social insurances protecting against loss of earnings as a result of sickness, disability, or unemployment only cover employees younger than sixty-five.

Incentives and Exit Routes to Retirement in the Netherlands Table 8.1

465

Labor Force Participation and Benefit Receipt of Males and Females Aged 50 or Over (3,149 observations) Full-Time

Part-Time

Not Working

A. Labor Force Participation Males 50–54 55–59 60–64 65+ Females 50–54 55–59 60–64 65+

77.3 49.2 3.1 1.6

6.0 12.4 5.4 3.0

16.7 38.3 91.5 95.4

12.6 8.5 2.1 0.0

45.7 24.7 6.4 1.0

41.7 66.8 91.4 99.0

SS

PP

DI

UI/Other

B. Benefit Receipt Males 50–54 55–59 60–64 65+ Females 50–54 55–59 60–64 65+

0.0 0.0 0.0 95.7

0.6 9.0 53.6 79.6

14.4 26.3 34.4 2.1

10.9 17.3 14.3 1.1

0.0 0.0 0.0 97.3

0.6 4.1 17.5 41.6

8.6 12.2 13.9 0.5

10.6 17.0 21.1 1.3

Source: Data from the Socio-Economic Panel (SEP) for 1996. Notes: SS = social security; PP = private pension; DI = disability insurance; and UI = unemployment insurance.

Table 8.1 summarizes labor market participation and benefit program participation of individuals aged fifty and over, based on data of the SocioEconomic Panel (SEP) for 1996. The table confirms the low labor force participation rate of males aged sixty or over and, next to the high level of nonparticipation, the relatively high incidence of part-time work for females. The table also illustrates that relatively many females aged fifty or over still perform the traditional role of housewife: They do not work and do not receive an income-replacing benefit, but are likely to be dependent on their spouse. Still, as of age sixty-five, these women are entitled to SS. The role of private ER benefits is especially prominent for males aged sixty to sixty-four, while more than 80 percent of males aged over sixty-five receive a private pension supplementing SS. The incidence of DI benefits among males increases from 14 percent for fifty to fifty-four year olds to 34 percent for sixty to sixty-four year olds. Table 8.1 suggests that for males the most common route to SS is

466

Klaas de Vos and Arie Kapteyn

Table 8.2

Transition of Male and Female Income Sources Between Ages 64 and 66 Age 66 Age 64

SS only

SS+PP

Total

No work, no benefit Paid work PP (early retirement) DI UI/Other Total

Males 2.0 4.7 2.7 6.7 2.0 18.1

1.3 4.7 45.0 25.5 5.4 81.9

3.4 9.4 47.7 32.2 7.4 100.0

No work, no benefit Paid work PP (early retirement) DI UI/Other Total

Females 53.4 2.6 4.2 2.6 6.9 69.8

1.6 0.5 15.3 3.7 9.0 30.2

55.0 3.2 19.6 6.3 15.9 100.0

Source: Data from SEP for 1992 to 1996.

through ER, whereas many females have been without income or benefits of their own before turning sixty-five and become eligible for SS after sixty-five. This is confirmed by table 8.2, which divides sixty-six year olds of the last three waves of the SEP according to their income sources at age sixty-six (SS or SS and private pension) and their income situation at age sixty-four. In contrast to table 8.1, individuals receiving pensions, wages, or both, as well as other benefits, are lumped together in one group. Almost half of the males received a private (ER) pension at age sixty-four and SS plus a private pension at age sixty-six, while a quarter received a DI benefit at age sixty-four and SS plus a private pension at age sixty-six. More than half of the females received nothing at age sixty-four and SS only at age sixty-six. The route directly from work to SS is taken by just a small minority of both males and females. 8.3 Research Background Until recently, the literature on the retirement effects of SS, DI, or UI programs in the Netherlands was quite scarce, and usually descriptive and qualitative in nature. This situation has changed in the 1990s, when, due to an initiative of the Netherlands Program for Research on Aging (NESTOR),4 a substantial grant was given to a group of researchers at the University of Leiden (who subsequently called themselves CERRA, Centre for Economic Research on Retirement and Aging) to set up a panel of 4. This program is now defunct.

Incentives and Exit Routes to Retirement in the Netherlands

467

elderly households (at the time of the first wave, 1993), the head of the household had to be between forty-three and sixty-three years old). A fair amount of research of CERRA has been on retirement. One of the few examples of the earlier literature is Henkens and Siegers (1990), who provide one of the first quantitative analyses of retirement decisions of males in the Netherlands. The most prominent study is undoubtedly Aarts and de Jong (1992), who report on a project covering more than a decade of research into the determinants of disability. Next to obvious health factors, financial considerations are found to play an important role. Indeed this study was the first to document by means of quantitative analysis the fact that the DI scheme was both a financially attractive route into early retirement for the employee and a convenient way to lay off elderly employees. In view of the tightening of eligibility rules and the reduction of the benefit levels for DI, and the simultaneous introduction of various generous ER schemes, one would expect a substitution of channels into retirement. Woittiez, Lindeboom, and Theeuwes (1994) study this by modeling the probability of finding elderly individuals (defined as being between fortyeight and sixty-two years old) in one of four states: working, disabled, unemployed, or early retired. They find a significant role for financial incentives—that is, a state becomes more likely if the associated income level is higher. The authors also find evidence for stigma effects (cf. Moffit 1983), indicating that the state of unemployment is valued below the state of disability, and both are valued below early retirement. This finding is partly supported by Woittiez and Theeuwes (1997), who use self-reported measures of life satisfaction, as well as several measures of mental and somatic health, to find that, other things being equal, people who work are generally better off than nonworking people, but early retirees are a close second. The disabled are least satisfied with their life, whereas the unemployed are above the disabled and below the early retired. The key difference between the unemployed and the early retired lies in the involuntary nature of the former state, which is found to explain most of the dissatisfaction of the unemployed. In principle, also early retirement can have a nonvoluntary nature, as an employer may put pressure on an employee who is eligible for ER to leave the firm. Thio (1995) uses a competing-risks model to explain the different routes into retirement and does find some evidence for involuntary ER, although this is not significant. Nevertheless, ER remains the favorite exit route out of employment. In Kerkhofs, Theeuwes, and Woittiez (1996), transitions out of a job are analyzed by means of a duration model. They also establish a substitution pattern in the choice of exit routes. When the ER route is available, it dominates the other exit routes. As both eligibility rules and replacement rates for ER differ across firms (or sectors), one may suspect that employees and employers match to their mutual benefit.

468

Klaas de Vos and Arie Kapteyn

Workers with a preference for ER may match with firms that offer relatively low wages and the possibility to retire early. Firms (or sectors) that need healthy young workers may decide to offer generous ER schemes. Thio and Woittiez (1996) investigate this issue by estimating a hedonic price relation in which the wage offered to an individual employee is related to characteristics determining worker productivity and ER benefits. It is found that there is a trade-off between wages and ER benefits, but not one for one (i.e., the better ER benefits are not fully reflected in lower wages). This finding seems to be consistent with the behavior of employers in the Netherlands, who are increasingly anxious to change the ER rules as the current rules turn out to be much more expensive than originally anticipated. Clearly, for this type of study, the availability of data for both employees and employers is essential. Another study taking advantage of this is Theeuwes and Lindeboom (1995), who match firm and employee data to analyze the effect of exit routes on the number of elderly employees leaving the firm. They provide evidence that there is some, but not full, substitution between channels into retirement. This gives room for policy measures to reduce retirement. Based on employee data only, they find that eligibility requirements, rather than the benefit heights, determine the moment of retirement. Heyma and Thio (1994) take up the issue of explaining differences in labor force participation among elderly workers between the United States and the Netherlands, exploiting the Health and Retirement Study (HRS) and the CERRA samples. The interesting part of their analysis is where they use the U.S. estimates to predict participation in the Netherlands and vice versa. This shows that if the Dutch would have the American coefficients, labor force participation would even be higher than in the United States, whereas if the Americans would have the Dutch coefficients, labor force participation in the United States would have been even lower than in the Netherlands. This suggests that the explanation for the observed differences between the United States and the Netherlands is not a matter of different characteristics of individuals, but rather a matter of a different institutional environment—the two main features of this being financial incentives and eligibility rules. Heyma (1996) addresses both financial incentives and eligibility rules in a dynamic-programming model of retirement decisions. Having estimated the model, he simulates various policy changes, like later eligibility for ER, raising the mandatory retirement age by two years, and lower ER benefits. The effects found are substantial. For example, if the ER benefits are set equal to disability benefits, labor force participation of sixty-two-year olds easily doubles. Heyma, Lindeboom and Kerkhofs (1997) extend this model by using data on individual behavior, survival rates, private pensions and firm data. The effects are similar to the ones found in Heyma (1996). Putting emphasis on the institutional characteristics they are able to explain quite a lot of the dynamics in retirement behavior. Lindeboom (1998), estimating a competing-risk duration model that explicitly takes

Incentives and Exit Routes to Retirement in the Netherlands

469

into account eligibility rules and replacement rates, also finds that ER schemes, in particular, create strong incentives to early withdrawal from the labor market. Using both information on health and financial incentives, Kerkhofs, Lindeboom, and Theeuwes (1999) find that the effect of health on retirement depends crucially on the health measure used, but that incentive effects are relatively insensitive to alternative specifications for health. The research reviewed here provides ample evidence for the dominant role of financial incentives and eligibility rules in the explanation of the low labor force participation rate among the elderly in the Netherlands. However, no study has yet fully quantified the part of the decrease in labor force participation among the elderly that can be ascribed to the changes in incentives and eligibility rules over the last three decades. 8.4 Data Overview Most of the results presented in this paper are derived using the SEP. The SEP is a longitudinal survey administered by Statistics Netherlands (CBS) consisting of approximately 5,000 households. The survey is representative of the Dutch population, excluding those living in special institutions like nursing homes. The SEP has been launched in April 1984. The same households were interviewed in October 1984 and then twice a year (in April and October) until 1989. Since 1990, the survey has been conducted once a year in May. In order to address the problem of sample attrition, Statistics Netherlands regularly adds new households to the SEP. In the October interview, information has been collected at the respondent level on socioeconomic characteristics, income, and labor market participation. The April interviews also contain information about socioeconomic characteristics, but rather than gathering data about income, beginning in October 1987, the April questionnaire includes questions on a wide range of assets and liabilities. Since 1990, these questions are part of the annual May questionnaire. Data are available for the period 1984–1996. In the analysis, we include men and women aged fifty to sixty-four who had positive earnings in 1984. Individuals are added in later years (until 1994), as they turn fifty, subject to positive earnings at age fifty. For all individuals, we observe whether or not they retire during the next year, and if so, whether or not they receive (early) retirement pensions, DI or UI benefits, or no benefits at all. In addition, we observe education level, labor market sector, number of hours worked, and so forth. All information available for the individual is also available for their partner. In addition, we have information on assets and income from capital received by the household. However, this information is fairly unreliable and not used in our analyses. Table 8.3 presents the means of the most important variables used. As we are trying to model the individual retirement decision, we are lim-

470

Klaas de Vos and Arie Kapteyn

Table 8.3

Summary Statistics

No. of observations Retired ER/PP DI/UI No benefits Married Low education High education Agriculture Industry Noncommercial services Age 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 Earnings (dfl. before tax) Mean (standard deviation) 25th percentile Median 75th percentile

Males

Females

3,492

1,388

0.078 0.031 0.010 0.900 0.380 0.221 0.046 0.368 0.302

0.048 0.032 0.038 0.690 0.532 0.151 0.022 0.063 0.644

0.124 0.115 0.114 0.107 0.096 0.088 0.082 0.068 0.062 0.050 0.035 0.018 0.014 0.015 0.011

0.140 0.113 0.108 0.097 0.097 0.089 0.073 0.063 0.066 0.050 0.040 0.024 0.019 0.013 0.009

74,247 (52,483) 46,906 64,696 91,551

29,639 (43,290) 8,135 19,028 39,159

ited by a lack of information on the eligibility for the various exit routes out of the labor force. In particular, we do not know whether or not and at which age individuals can take ER. We also do not know whether or not they might be entitled to DI benefits—and we have insufficient health information to be used as a proxy (although health would not present the full picture anyway). For the eligibility for ER we have used information based on one wave of an alternative panel, the CentER panel,5 in which employed individuals were asked whether or not they were participating in an occupational pension plan and at which age that would allow them to retire. The probability of being eligible for ER at a certain age is approximated by multiplying the estimated probability of participation in an occupational pen5. The CentER panel comprises about 2,000 households and is run by CentER data, a survey research institute affiliated with Tilburg University.

Incentives and Exit Routes to Retirement in the Netherlands Table 8.4

471

Logit Equation for Eligibility for Private Pension

Variable Sector Industry Noncommercial services Age Age squared Full time (
32 hrs/week) Female Constant

Estimate

SE

Significance Levela

0.933 1.595 0.230 –0.003 0.734 –1.039 –3.417

0.361 0.270 0.104 0.001 0.315 0.324 2.277

0.010 0.000 0.026 0.037 0.020 0.001 0.133

Note: SE = standard error. a Test of null hypothesis that the parameter in question is zero. Table 8.5

Sector Age 55 56 57 58 59 60 61 62 63 64 65

Distribution of Age of Eligibility for Early Retirement Given Eligibility for Private Pension

Industry

Agriculture, Commercial Services

Noncommercial Services

0.043 0.012 0.067 0.037 0.018 0.244 0.220 0.244 0.012 0.000 0.104

0.044 0.016 0.027 0.022 0.055 0.224 0.126 0.137 0.055 0.000 0.295

0.095 0.011 0.046 0.052 0.032 0.204 0.187 0.236 0.043 0.006 0.089

sion plan by the fraction of individuals in an occupational pension plan who say they can retire at that age. The estimated probability to be in an occupational pension plan is based on a logit specification with sector of employment, age, sex, and whether or not they are working full-time as explanatory variables. The logit estimates are presented in table 8.4. The eligibility age distribution for ER is differentiated by sector of employment (table 8.5). 8.5 Earnings Histories and Projections As SS is a flat-rate benefit, whereas ER, DI, and UI benefits and occupational pensions are (usually) based on final pay, information on earnings histories is not needed to determine Social Security Wealth (SSW), the actuarially discounted sum of future benefits. The benefits include not only SS benefits, but also private pension (PP), DI, and UI benefits wherever appropriate. Only information on the number of years in pensionable employment together with information on the final earnings would be suffi-

472

Klaas de Vos and Arie Kapteyn

cient to determine the benefit level to which the individual is entitled. The number of years in pensionable employment is generally unknown, but in the Dutch system this number, although clearly affecting SSW, generally has only a marginal effect on most of the incentive variables to be included in the retirement decision (accrual, peak value, and option value; to be defined later): The effect of working an additional year is, by and large, constant over a rather wide interval of years. As the jumping-off point for the forward projections we use actual earnings, assuming constant real earnings. We do not calculate three-year averages because this would limit the number of observations that could be included in the analysis.

8.6 Construction of Incentive Measures In our calculations we do not distinguish UI benefits from DI benefits (both are received until the age of sixty-five), and we assume that, like SS, which is received after age sixty-five, the benefits do not depend on the age of retirement. After becoming unemployed or disabled, the older worker can expect to keep the same level of benefits up to age sixty-five. After age sixty-five, SS is independent of work history. Hence, if we would limit ourselves to these three benefit types, the implicit tax or subsidy rate on retiring, which is the change in the worker’s future benefits, relative to what he would earn in the coming year, would be equal to the replacement rate (the level of benefits in the coming year relative to their earnings in the coming year). The only way in which an employee’s future income (after the coming year) may be affected by retiring one year earlier is via their private pension. Retiring before the age of sixty-five may affect the level of PP to be received after age sixty-five by reducing the number of years counting towards pension benefits. Moreover, if an employee were to retire before his ER age, he would no longer be eligible for ER benefits. In this section we describe how SSW, accrual rates, implicit tax or subsidy rates, option values, and peak values are calculated. As in Gruber and Wise (1999), accrual rates are defined as the change in the worker’s SSW relative to the SSW if they would retire one year earlier, and tax or subsidy rates are defined as the change in the SSW relative to what they would earn over the coming year. As mentioned above, SSW is calculated as the actuarially discounted sum of future benefits. In contrast to the earlier volume (Kapteyn and de Vos 1999), contributions paid toward the various benefit and pension programs during the remaining working life are no longer deducted from SSW. In our incentive calculations we distinguish the following cases. 1. Eligibility for early retirement at a certain age between fifty-five and sixty-five (eligible individuals will receive a PP in addition to SS once they turn sixty-five)

Incentives and Exit Routes to Retirement in the Netherlands

473

2. Eligibility for disability or unemployment benefit upon retirement before age sixty-five and receipt of a PP in addition to SS as of age sixty-five 3. Eligibility for SS only (as of age sixty-five). For all entitlements we assume zero growth in real terms after 1995.6 For survival probabilities, we use sex- and age-specific survival tables of Statistics Netherlands (1992). We assume independence between the mortality rates of the worker and their spouse. We use a real discount rate of 3 percent. To compute net-benefit and pension levels, we subtract payroll and income taxes. For the years after 1995, we use the tax schedule for 1995, keeping tax rates and brackets fixed in real terms. For individuals with working spouses, we assume that the spouse will stop working at age sixtyfive. In our calculations, we take into account that if an individual retires before age sixty-five and is not entitled to any benefit or pension, the spouse (if older than sixty-five) may be entitled to a supplement to his SS benefit. The option value of postponing retirement is approximated as: Gt (r ∗)  Vt (r ∗)  Vt (t),

(1)

where Vt (t) represents the utility of retiring now, and Vt (r ∗) represents the highest feasible utility (obtained by retiring at age r ∗). Vt (r) is calculated as: r1

(2)

Vt (r) 

∑ st

S

stY s 

∑

st [kBs (r)] ,

sr

where Ys represents earnings in the years before retirement, and Bs (r) represents benefits received in the years after retirement at age r. We use k equal to 1.5,  equal to 0.97, and  equal to 0.75. The incentive measures used in the estimations in the next section are weighted averages of the incentives for the various exit routes (DI, ER, and SS only), where the weights are determined by the empirical take-up rates differentiated by age. Notably, in these calculations, the fact that persons may be eligible for ER at a later age is taken into account by using the probabilities for eligibility by age based on results of the CentER panel, as described earlier, multiplied by the empirical take-up rate of continued work. In this way, the weighted SSW measure of individuals aged sixty, for example, is calculated as (3) SSW60  pER,60  SSWER,60  pDI,60  SSWDI,60  pexit,60  SSWSSonly,60  (1  pER,60  pDI,60  pexit,60 )  ( pp  ( pER,61  SSWER,61  pER,62  SSWER,62  pER,63  SSWER,63  pER,64  SSWER,64  pER,65  SSWER,65 )  (1  pp)  SSWSSonly,60 ), 6. For DI, SS, and UI benefits, this is more or less in line with current government policy.

474

Klaas de Vos and Arie Kapteyn

Table 8.6

Age Males 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 Females 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

The Distribution of One-Year Accrual: Weighted Accrual as Used in Estimations

N

Median SSW

Accrual Median

10th Percentile

90th Percentile

SD

433 402 399 375 334 306 288 238 216 174 123 62 50 53 39

317,375 335,555 344,972 358,162 387,512 389,206 394,186 385,403 434,945 424,633 447,151 342,854 365,585 421,203 371,499

2,835 3,592 3,390 3,670 16,405 5,946 9,857 6,231 1,597 17,338 12,466 12,680 4,421 –34 –8,659

–8 –4 0 –9 6,943 390 218 –13 –1,665 2,441 –26 –18 –30 –1,708 –12,668

7,458 7,910 7,696 7,683 40,604 10,597 19,843 16,919 7,710 34,294 32,026 65,061 22,173 3,576 –14

3,588 3,879 4,001 4,481 15,031 4,237 7,346 7,934 4,790 12,047 12,913 24,192 8,073 5,763 5,532

195 157 150 134 134 123 101 88 92 70 55 33 26 18 12

300,180 297,623 288,047 293,246 275,090 294,974 313,990 285,680 291,082 290,916 285,620 295,074 293,533 305,938 305,730

–123 –74 –116 –133 3,967 139 1,383 657 –537 2,785 1,040 5,503 844 –781 –3,522

–362 –269 –373 –501 0 –91 –206 –555 –2,512 –564 –1,598 –1,433 –7,072 –2,207 –15,404

2,980 3,185 3,918 3,520 18,596 6,373 11,201 9,622 4,061 22,175 21,197 48,767 10,171 616 –47

1,914 2,092 2,943 3,922 8,865 2,805 11,561 4,730 2,545 9,026 7,883 16,851 5,696 1,320 5,297

Note: N = number of observations; SD = standard deviation.

where pER,60, pDI,60 and pexit,60 are empirical take-up rates, and pp and pER,61, . . . , pER,65 represent sex- and sector-specific eligibility probabilities (for PP and ER at age sixty-one through sixty-five, respectively, given eligibility for PP) based on the CentER panel data. The same weighting scheme is used for the calculation of accrual, option, and peak values. Tables 8.6 and 8.7 summarize the (weighted) incentive measures differentiated by age and sex. It should be noted that these weighted figures do not represent the incentives as faced by individuals, since one usually knows whether or not and at which age one is eligible for ER. In tables 8.8 and 8.9, the incentives for males are shown separately for the cases in which individuals are eligible for ER at age sixty and SS supplemented by PP as

Incentives and Exit Routes to Retirement in the Netherlands Table 8.7

475

The Distribution of Peak and Option Value: Weighted Incentives as Used in Estimations Peak Value

Option Value

Age

Median

10th Percentile

90th Percentile

SD

Median

10th Percentile

90th Percentile

SD

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 Females 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

122,504 131,172 126,521 125,900 122,944 111,062 94,440 79,704 66,113 50,606 35,431 20,383 7,518 9 –8,659

46,158 45,290 51,930 49,236 58,930 31,514 12,186 18,756 14,704 6,532 –15 –18 –29 –1,203 –12,668

220,293 221,813 213,795 221,577 203,457 187,141 166,165 149,364 127,589 100,447 78,583 76,114 32,652 11,983 –14

72,518 84,788 80,833 96,542 75,628 62,055 56,703 63,207 49,935 33,820 29,855 28,758 12,461 14,382 5,532

40,195 39,556 37,139 35,387 31,439 29,343 23,077 19,174 15,609 11,640 9,524 8,343 6,977 5,136 1,969

27,513 26,219 24,858 23,100 21,379 18,413 14,846 11,594 8,989 5,971 3,017 1,211 1,562 1,467 619

59,058 55,427 51,319 48,804 44,119 41,019 35,094 31,703 25,135 18,385 16,142 14,942 12,318 7,917 3,285

14,444 15,933 14,263 14,412 12,494 10,449 8,753 8,963 7,795 5,002 5,225 5,150 3,630 4,311 903

23,015 25,900 24,409 24,134 20,690 21,489 26,730 16,955 12,131 9,823 5,140 6,473 1,053 –482 –3,522

1,140 161 403 0 0 64 598 –33 –294 –87 –600 –1,114 –7,072 –2,207 –15,404

132,476 130,944 153,484 135,627 91,100 110,312 108,188 88,796 70,873 66,793 54,316 59,557 13,990 3,426 –47

53,649 54,716 71,471 92,281 43,028 46,371 97,856 39,060 32,284 25,473 18,696 20,392 7,381 2,702 5,297

15,427 15,068 14,718 12,676 10,377 9,515 9,750 7,189 5,563 4,270 2,838 2,491 2,104 1,220 101

6,055 4,809 5,367 4,302 3,798 3,004 2,253 1,669 1,402 804 668 557 0 0 0

39,144 37,080 38,902 32,823 23,909 28,698 25,230 23,785 19,118 15,155 11,513 11,666 7,693 5,666 1,433

13,129 12,969 16,998 15,578 9,176 9,603 14,571 8,053 6,646 5,509 4,497 3,854 2,711 2,031 512

Note: SD = standard deviation.

of age sixty-five; for DI now and SS supplemented by PP as of age sixtyfive; and for SS only (as of age sixty-five). Table 8.8 also compares the median implicit tax rates (“tax rate 1”) with the implicit tax rate for the median worker as presented in Kapteyn and De Vos (1999; “tax rate 2”). 8.7 Estimation Results for the Common Model Tables 8.10 and 8.11 present the estimation results for males and females, respectively. The results for males are (statistically) much more significant than for females. To a considerable extent, this may be due to the

Table 8.8

The Distribution of One-Year Accrual for Males (assuming eligibility)

N

Median SSW

Accrual Median

10th Percentile

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

433 402 399 375 334 306 288 238 216 174 123 62 50 53 39

317,103 339,826 347,457 365,761 379,949 386,860 388,246 379,782 409,105 405,987 587,764 480,417 494,865 542,783 537,045

4,008 4,640 4,629 5,541 5,606 5,386 4,562 4,137 4,148 138,109 –28,384 –25,061 –27,397 –27,913 –28,897

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

433 402 399 375 334 306 288 238 216 174 123 62 50 53 39

713,792 719,464 701,801 688,347 677,054 659,681 623,541 577,718 588,773 544,953 567,890 466,454 483,669 534,627 532,769

–27,687 –28,041 –27,457 –27,331 –26,927 –26,863 –26,135 –24,600 –24,517 –23,418 –23,926 –21,056 –23,537 –23,294 –23,921

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

433 402 399 375 334 306 288 238 216 174 123 62 50 53 39

200,054 205,284 210,964 217,441 224,578 228,552 234,374 245,483 252,475 259,868 267,729 276,134 285,173 294,950 311,367

–12 –13 –14 –15 –16 –17 –19 –19 –22 –23 –26 –22 –24 –31 –31

Age

90th Percentile

SD

Median Tax Rate 1

Tax Rate 2

A. ER(60) + SS + PP 671 8,987 802 9,041 1,127 9,104 1,273 9,374 1,368 9,032 789 9,367 619 9,354 94 9,383 –1 9,718 58,093 246,498 –45,276 –8,040 –42,949 –5,082 –43,750 –5,282 –44,591 –5,707 –45,350 –6,471

3,902 4,201 4,361 4,803 4,196 3,959 3,935 4,922 5,284 72,501 14,211 15,721 13,703 22,635 12,339

0.458 0.468 0.465 0.453 0.449 0.431 0.430 0.415 0.417 –3.603 1.353 1.264 1.320 1.310 1.295

0.687 0.650 0.612 0.578 –3.777 1.410 1.384 1.339 1.282 1.222

B. DI + SS + PP –33,110 –19,653 –32,973 –19,255 –32,850 –19,514 –32,688 –19,435 –32,827 –18,877 –32,651 –18,475 –32,514 –18,019 –32,081 –16,010 –32,815 –14,682 –31,646 –12,076 –31,623 –7,601 –31,355 –4,438 –31,634 –4,565 –31,208 –3,584 –31,445 –5,604

6,758 6,766 6,272 6,330 6,381 6,839 6,589 7,020 7,841 7,663 8,317 10,225 9,599 12,874 7,842

1.263 1.260 1.263 1.240 1.242 1.226 1.223 1.191 1.203 1.194 1.202 1.096 1.175 1.164 1.150

1.478 1.428 1.379 1.338 1.269 1.184 1.160 1.121 1.099 1.009

550 551 755 667 10 10 11 12 13 14 1,432 251 16 1,437 193

0.369 0.374 0.372 0.371 0.365 0.354 0.361 0.344 0.349 0.329 0.332 0.283 0.337 0.339 0.347

0.475 0.464 0.447 0.436 0.407 0.421 0.415 0.431 0.410 0.380

C. SS Only –19 –21 –23 –26 –29 –31 –34 –37 –38 –40 –47 –49 –45 –53 –47

–6 –7 –7 –6 –4 0 0 –6 0 0 –8 0 0 0 0

Note: N = number of observations; SD = standard deviation.

Table 8.9

The Distribution of Peak and Option Value for Males (assuming eligibility) Peak Value

Option Value

Age

Median

10th Percentile

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

150,693 163,625 166,593 166,799 169,749 167,100 157,076 138,857 144,377 138,109 –28,384 –25,061 –27,397 –27,913 –28,897

80,603 80,051 86,275 91,733 92,502 86,411 84,069 68,104 69,874 58,093 –45,276 –42,949 –43,750 –44,592 –45,350

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

–27,687 –28,042 –27,457 –27,331 –26,927 –26,863 –26,135 –24,600 –24,517 –23,418 –23,926 –21,056 –23,537 –23,294 –23,921

–33,110 –32,973 –32,850 –32,688 –32,827 –32,651 –32,514 –32,081 –32,815 –31,646 –31,612 –31,355 –31,634 –31,208 –31,445

–19,653 –19,225 –19,514 –19,435 –18,877 –18,475 –18,029 –16,010 –14,682 –12,076 –7,601 –4,438 –4,565 –3,584 –5,604

50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

–12 –13 –14 –15 –16 –17 –19 –19 –22 –23 –26 –22 –24 –31 –31

–19 –21 –23 –26 –19 –31 –34 –37 –38 –40 –47 –49 –45 –53 –47

–6 –7 –7 –6 –4 0 0 –6 0 0 –8 0 0 0 0

Note: SD = standard deviation.

90th Percentile

10th Percentile

90th Percentile

SD

A. ER(60) + SS + PP 258,249 79,438 41,873 253,427 96,412 41,725 249,882 92,394 39,752 255,650 103,591 36,915 251,471 92,974 34,383 257,709 82,074 31,751 261,287 77,581 27,882 257,614 88,716 22,632 257,907 94,570 20,584 246,498 72,501 17,465 –8,040 14,211 0 –5,082 17,087 0 –5,282 13,703 0 –5,707 22,635 0 –6,471 12,339 0

28,894 27,051 26,009 24,978 23,380 20,319 18,033 13,739 12,243 9,070 0 0 0 0 0

61,469 56,668 52,663 50,721 46,917 43,988 40,014 36,216 31,974 26,812 221 221 261 114 111

14,903 16,289 14,575 14,701 13,017 11,365 9,944 9,915 9,725 6,898 379 165 127 841 50

B. DI + SS + PP 6,758 799 11,882 1,017 10,772 1,132 21,742 1,127 6,501 1,063 6,841 906 6,591 699 11,766 259 8,933 386 7,663 339 8,317 284 10,232 0 9,599 71 17,377 306 7,842 129

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

14,430 12,652 10,878 11,703 11,121 10,132 9,354 8,128 7,163 5,454 6,329 4,027 3,216 2,235 1,249

8,420 10,855 10,014 11,065 9,279 6,813 5,651 6,120 6,381 2,979 3,974 2,366 1,507 3,971 558

24,681 23,640 22,483 21,506 19,773 17,066 14,913 11,530 8,140 7,162 4,057 1,376 1,569 1,285 858

49,388 45,871 42,589 40,658 37,626 34,764 31,873 28,695 25,156 20,979 18,603 14,536 11,221 7,682 4,031

11,327 12,271 10,914 10,833 9,582 8,579 7,546 7,346 7,271 5,264 5,722 4,953 3,306 3,554 1,033

SD

C. SS Only 550 551 755 667 10 14 15 12 17 15 1,435 254 17 1,437 193

Median

34,469 32,854 30,916 28,881 26,519 24,168 21,618 18,459 16,421 13,813 12,092 8,593 7,031 5,110 2,664

478

Klaas de Vos and Arie Kapteyn

Table 8.10

Retirement Probits for Males (marginal effects)

Variable SSW Accrual

Accrual 1.40e-07 (2.51) 1.59e-06 (3.46)

9.48e-08 (1.68) 1.87e-06 (2.84)

Peak value

Peak Value 2.50e-07 (4.65)

2.03e-07 (3.69)

–4.41e-07 (3.34)

–3.37e-07 (2.46)

Option value Earnings Earnings2 Earnings, 2(1) Earnings partner Earnings partner 2 Earnings partner, 2(1) Age Age2 Age 2(1) Two earners (yes = 1) Partner? (yes = 1)

–1.82e-06 (3.10) –1.07e-12 (.32) 9.60** 1.46e-06 (.97) –1.78e-11 (.70) .94 .062 (1.68) –3.77e-04 (1.14) 160.7** –.022 (1.20) –.0093 (.46)

Age dummies, 2(14) Log-likelihood Pseudo R 2

–1,050.8 .174

–1.77e-06 (2.96) 1.17e-12 (.37) 8.77** 1.26e-06 (.80) –1.45e-11 (.54) .64

–.019 (.98) –.0029 (.14) 221.1** –1,023.0 .197

–6.07e-07 (.98) –7.13e-12 (1.89) .97 1.53e-06 (.86) –2.80e-11 (.82) .74 .11 (3.12) –8.66e-04 (2.70) 58.1** –.018 (.92) –.031 (1.50) –1,050.9 .175

–7.19e-07 (1.15) –4.17e-12 (1.10) 1.32 1.38e-06 (.73) –2.49e-11 (.67) .54

–.016 (.78) –.022 (1.10) 112.5** –1,023.8 .196

Option Value 1.54e-07 (3.22)

1.32e-07 (2.7)

–5.28e-06 (5.80) 1.78e-06 (2.31) –9.87e-12 (2.69) 5.35** 1.67e-06 (.98) –2.45e-11 (.74) .95 .07 (2.06) –5.75e-04 (1.92) 6.95** –.016 (.83) –.0099 (.54)

–4.27e-06 (4.40) 1.25e-06 (1.55) –6.75e-12 (1.80) 2.41 1.53e-06 (.85) –2.24e-11 (.64) .72

–1,039 .184

–1,016.9 .201

–.015 (.75) –.0072 (.39) 49.9**

Notes: Numbers in parentheses are t-values. 3,492 observations. **Significant at the 5 percent level.

larger number of observations for males than for females (the number of observations for males is about three times larger than for females), but the fit is also better for males than for females. We will discuss the results for males and females consecutively. Table 8.10 shows that the inclusion of option value as an incentive variable gives a substantially better fit than the inclusion of accrual or peak value. Furthermore, the results with accrual included are economically implausible. The coefficient of the accrual variable has the wrong sign and is significantly different from zero. In the specification without age dummies, the incentive variables are more significant and bigger in absolute value than in the specifications with the age dummies included. The age functions for the specification without age dummies have been specified as a

Incentives and Exit Routes to Retirement in the Netherlands Table 8.11

Retirement Probits for Females (marginal effects)

Variable SSW Accrual

Accrual 2.74e-08 (.22) –6.52e-07 (.44)

–3.70e-09 (.03) 5.48e-07 (.31)

Peak value

Peak Value 3.01e-08 (.25)

1.66e-08 (.14)

–2.27e-07 (.64)

–1.64e-07 (.46)

Option value Earnings Earnings2 Earnings, 2(1) Earnings partner Earnings partner 2 Earnings partner, 2(1) Age Age2 Age 2(1) Two earners (yes = 1) Partner? (yes = 1)

1.45e-06 (.94) –4.51e-11 (.88) 1.47 –3.67e-07 (.29) –2.04e-12 (.22) .08 –.022 (.37) 3.67e-04 (.70) 64.6** .0070 (.19) .0165 (.67)

1.53e-06 (.99) –4.80e-11 (.90) 1.56 –1.26e-07 (.10) –1.44e-12 (.26) .01

–448.9 .110

–442.5 .122

Age dummies, 2(14) Log-likelihood Pseudo R 2

479

–1.84e-04 (.01) .0249 (1.03) 75.2**

1.64e-06 (1.05) –4.34e-11 (1.09) 1.43 –3.84e-07 (.31) –1.77e-12 (.25) .10 –.023 (.39) 3.67e.04 (.71) 49.9** .0077 (.22) .0152 (.61)

1.68e-06 (1.07) –4.44e-11 (1.16) 1.49 –2.30e-07 (.19) 1.62e-12 (.28) .04

–448.7 .110

–442.5 .122

.0041 (.12) .0197 (.80) 60.4**

Option Value –2.56e-08 (.23)

–2.78e-08 (.25)

–2.29e-06 (1.24) 2.44e-06 (1.42) –4.13e-11 (2.01) 1.40 4.79e-08 (.04) –1.53e-12 (.29) .00 –.033 (.56) 4.45e-04 (.86) 34.6** 7.22e-04 (.02) .0185 (.80)

–2.00e-06 (1.08) 2.42e-06 (1.40) –4.23e-11 (1.96) 1.46 1.35e-07 (.11) –1.53e-13 (.32) .01

–448.2 .111

–442.0 .123

–.0015 (.04) .0216 (.94) 45.3**

Notes: Numbers in parentheses are t-values. 1,388 observations. **Significant at the 5 percent level.

quadratic. The estimated age functions are shown in figure 8.1 for all specifications given in tables 8.10 and 8.11. With one exception (the specification with option value for males), the age functions are all monotonically increasing. Of course, a quadratic specification may be too restrictive to capture incentive effects at particular ages. Figure 8.2 therefore presents the estimated age dummies for the specifications with the three different incentive variables. Furthermore, figure 8.2 also shows the empirical retirement hazards. In all specifications (and in the empirical hazard), we observe a peak at age sixty. Up until sixty, the probability of retirement increases with age. After age sixty, it falls and then gradually goes up again until the age of sixty-five.

Fig. 8.1

Age effects on retirement

Fig. 8.2

Age dummies and hazard rates, males

Incentives and Exit Routes to Retirement in the Netherlands

Fig. 8.3

481

Earnings effects on retirement, no age dummies, males

Figures 8.3 and 8.4 present the influence of own and partner’s earnings (if present) on the retirement index. Although the pictures suggest a negative effect of both own and partner’s earnings on the probability of retirement in any given year, we should be aware of the fact that the earnings terms are rarely significant. Partner’s earnings are never significant at the 5 percent level. Own earnings are significant for the accrual and option value specifications. In table 8.11 (females) very few coefficients are significant. The t-values corresponding to the incentive variables are quite low. The age dummies are jointly highly significant and exhibit a pattern that is similar to that for males (cf. figure 8.2). The same is true for the quadratic age functions. Neither own earnings nor partner’s earnings appears to exhibit a significant effect on the probability of retirement. We abstain therefore from presenting these earnings functions. It appears that the incentive variables used

482

Fig. 8.4

Klaas de Vos and Arie Kapteyn

Earnings effects on retirement, age dummies, males

are not very successful in capturing the motivating forces behind the retirement decisions of females. 8.8 Simulations We consider a number of incentive changes and their effects on retirement probabilities. The reforms and their implications for retirement are discussed consecutively. In view of the difference in statistical significance between the results for males and females, we mainly concentrate on the results for males. 8.8.1 A Three-Year Increment in Eligibility Ages Since eligibility is not directly observed and the computation of the incentives is based on actual retirement behavior, there are different ways in

Incentives and Exit Routes to Retirement in the Netherlands

483

which one can implement such a reform in the context of the model presented so far. We choose a particularly straightforward approach by calculating for every individual in the sample (and their partner, if any) the incentive variables as if this individual were three years younger. That is, we assign eligibility probabilities (but not mortality rates) as if the individual is three years younger and then recalculate the incentive variables. We present three different types of simulations. The first and second simulations simply use the models with and without dummies and replace the incentive variables by the new ones based on delayed eligibility. Regarding the model with age dummies, this may be less than appropriate, as the age dummies probably partly reflect variations in eligibility across age that are not captured by the incentive variables. Thus a third simulation shifts the age dummies backward by three years. That is, for ages fifty-one, fifty-two, and fifty-three, the age dummies are set equal to zero. The new age dummy for age fifty-four is set equal to the estimated age dummy for age fifty-one, the new age dummy for age fifty-five is set equal to the estimated age dummy for age fifty-two, and so forth. Notice that our data set only comprises individuals between fifty and sixty-four, so we cannot simulate the effect of a change in policy beyond the age of sixty-four. 8.8.2 The Common Reform The common reform involves early retirement at age sixty and normal retirement at age sixty-five. The replacement rates depend on age. The replacement rate at age sixty-five is equal to 60 percent of wages that one would have earned at age sixty. At other ages an actuarial adjustment of 6 percent is applied. That is, when retiring at age sixty, an individual receives 70 percent of 60 percent of wages; when retiring at age sixty-one, the individual receives 76 percent of 60 percent of age sixty wages; and so on. This also applies to retirement ages higher than sixty-five. For instance, when retiring at age seventy, an individual receives 130 percent of 60 percent of age sixty wages. However, as mentioned previously, since we have no individuals over sixty-four in the sample, the latter part of the reform cannot be simulated. Again, we consider three different types of simulations. The first two take the model without and with age dummies and change the incentive variables. The third simulation also adapts the age dummies. Between ages fifty and sixty, the age dummies are linearly interpolated between zero and the estimated age dummy for age sixty. Between ages sixty-one and sixtyfour, the age dummies are linearly interpolated between the estimated age dummies for ages sixty and sixty-four. 8.8.3 Simulation Results Figures 8.5 through 8.13 provide a graphical representation of the simulation outcomes for males. In each graph, the top panel shows hazards

Fig. 8.5

Accrual, males, no dummies

Fig. 8.6

Accrual, males, with dummies

Fig. 8.7

Accrual, males, with dummies

Fig. 8.8

Peak value, males, no dummies

Fig. 8.9

Peak value, males, with dummies

Fig. 8.10

Peak value, males, with dummies

Fig. 8.11

Option value, males, no dummies

Fig. 8.12

Option value, males, with dummies

Fig. 8.13

Option value, males, with dummies

Incentives and Exit Routes to Retirement in the Netherlands Table 8.12

493

Average Retirement Ages under Different Policies Delayed Eligibility

Common Reform

Males Baseline Accrual No dummies With dummies Adjusted dummies Peak value No dummies With dummies Adjusted dummies Option value No dummies With dummies Adjusted dummies

57.8

58.2

58.2 58.1 59.9

58.0 58.0 61.2

58.6 58.4 60.0

60.7 61.6 62.5

59.8 59.4 60.1

62.2 62.4 57.3

Females Option value No dummies With dummies Adjusted dummies

57.6 57.6 59.0

58.7 58.6 58.8

and the bottom panel shows the corresponding cumulative distribution functions (CDFs).7 Within each top panel we show the empirical hazards, the hazard as predicted by the fitted model, and then the hazards according to the two policies considered (delayed eligibility and common reform). The bottom graphs show the corresponding CDFs. Notice that we have nine graphs in total: three incentive measures and three simulations per policy measure. In addition table 8.12 presents the average retirement ages under the different policies and for the different model specifications. The average retirement ages have been calculated under the assumption that everyone retires at age sixty-five at the latest.8 This involves an underestimation of the true average; probably not by much for actual retirement (labor force participation after sixty-five is less than 5 percent in the Netherlands), but for simulations where participation is still substantial at age sixty-four, our calculation of mean retirement may be off by a nonnegligible amount. 7. The CDFs are related to the hazards as follows. Let the hazards be 50, 51, 52, and so forth. Then the corresponding values of the CDF are c50  50 ; c51  c50  51(1 – c50 ); c52  c51  52(1 – c51 ); and so forth. 8. The calculation of the CDF (see note 7) yields values of the retirement density equal to d50  50 ; d 51  51(1 – c50); d 52  52(1 – c51); and so forth. The mean retirement age is computed as Σ(t  50   64) d

 (1 – c64)65. The last term is based on the assumption that at age sixty-five everyone retires.

494

Klaas de Vos and Arie Kapteyn

As noted before, the estimation results for accrual are economically implausible and inferior to the results for option values on statistical grounds. Although the results for accrual have been included for completeness, we will not discuss them any further. First, then, consider the outcomes for peak value. Not surprisingly, the specifications including age dummies do a better job in reproducing the observed hazard rates than the specifications without the age dummies: For the specifications with age dummies, observed and fitted hazard rates are essentially equal. Clearly, the common reform has a much more dramatic effect on the average retirement age than the delayed-eligibility policy. The graphs show that under the common reform, the retirement hazard jumps at age sixty, except for one case (figure 8.10): The CDF for the common reform remains uniformly below the CDF for the delayed-eligibility policy. Although the specifications with age dummies show smaller policy effects than the specifications without dummies, the differences are not large (at least not for our preferred incentive measure, the option value). If, in the specification with age dummies, we adjust the dummies, the size of the policy effects increases somewhat. Since the estimation results for women were generally not significant, we only pay limited attention to the simulation results for females. The bottom part of table 8.12 and figures 8.14–8.16 present the results for option value as an incentive variable. By and large, the policy effects are minor, in line with the small and statistically insignificant effects found in estimation. 8.9 Conclusion In this paper, we have tried to estimate the incentive effects of the various programs providing earnings-replacing benefits upon the decision to retire in the Netherlands. Despite the fact that we are hampered by a lack of information with respect to the actual eligibility for these programs, we find significant effects for the peak and option values for males, in particular. Moreover, we think these results can be improved upon by a more sophisticated method that takes into account the (unknown) eligibility for the various programs. Preliminary simulations show that changing the incentives can have considerable effects on the average retirement age, for example. This is important for a period in which increasing the labor force participation of the elderly is a major policy objective, both for decreasing the mounting shortages on the labor market and for guaranteeing the sustainability of the payas-you-go SS system.

Fig. 8.14

Option value, females, no dummies

Fig. 8.15

Option value, females, with dummies

Fig. 8.16

Option value, females, with dummies

498

Klaas de Vos and Arie Kapteyn

References Aarts, L., and P. R. de Jong. 1992. Economic aspects of disability behaviour. Amsterdam: North-Holland. Gruber, J., and D. Wise, eds. 1999. Social security and retirement around the world. Chicago: University of Chicago Press. Henkens, K., and J. J. Siegers. 1990. The decision to retire: The case of Dutch men aged 50–64. Working Paper. The Hague: Netherlands Interdisciplinary Demographic Institute (NIDI). Heyma, A. 1996. Retirement and choice constraints: A dynamic programming approach. Research Memorandum no. 96.03. Leiden: University of Leiden, Centre for Economic Research on Retirement and Aging (CERRA). Heyma, A., M. Lindeboom, and M. Kerkhofs. 1997. Retirement from the labour force: On the relative importance of supply and demand. University of Leiden, Centre for Economic Research on Retirement and Aging (CERRA). Working Paper. Heyma, A., and V. Thio. 1994. Labour force participation of elderly men. University of Leiden, Centre for Economic Research on Retirement and Aging (CERRA). Working Paper. Kapteyn, A., and K. de Vos. 1999. Social security and retirement in the Netherlands. In Social security and retirement around the world, ed. Jonathan Gruber and David A. Wise, 269–304. Chicago: University of Chicago Press. Kerkhofs, M., M. Lindeboom, and J. J. M. Theeuwes. 1999. Retirement, financial incentives and health. Labour Economics 6 (2): 203–227. Kerkhofs, M., J. J. M. Theeuwes, and I. Woittiez. 1996. Mobility and the older worker: A substitution of routes. University of Leiden, Centre for Economic Research on Retirement and Aging (CERRA). Working Paper. Lindeboom, M. 1998. Microeconomic analysis of the retirement decision: The Netherlands. Organization for Economic Cooperation and Development (OECD) Working Paper no. 206. Paris: OECD. Moffit, R. 1983. An econometric model of welfare stigma. American Economic Review 73 (5): 1023–35. Statistics Netherlands. 1992. Life tables for the Netherlands, 1986–1990. Voorburg: Statistics Netherlands. Theeuwes, J. J. M., and M. Lindeboom. 1995. Arbeidsmarkt en Uittreding (Labor market and retirement). In Ouderen, wetenschap en beleid II (The elderly, science and policy), ed. B. C. M. Nitsche, 115–34. Utrecht: Netherlands Institute of Gerontology (NIG). Thio, V. 1995. Retirement; the distinction between quits and layoffs. University of Leiden, Centre for Economic Research on Retirement and Aging (CERRA). Working Paper. Thio, V., and I. Woittiez. 1996. Wages and early retirement; a hedonic price approach. University of Leiden, Centre for Economic Research on Retirement and Aging (CERRA). Working Paper. Woittiez, I., M. Lindeboom, and J. J. M. Theeuwes. 1994. Labour force exit routes of the Dutch elderly; a discrete choice model. In The economics of pensions: The case of the Netherlands, ed. L. Bovenberg, 1–23. Rotterdam, the Netherlands: Erasmus University, Research Centre for Economic Policy (OCFEB). Woittiez, I., and J. J. M. Theeuwes. 1997. Well-being and labour market status. In The distribution of welfare and household production: International perspectives (in honor of Aldi Hagenaars), ed. S. Jenkins, A. Kapteyn, and B. M. S. van Praag, 211–30. Cambridge: Cambridge University Press.

9 Micro-Modeling of Retirement Behavior in Spain Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

9.1 Introduction For the average Spaniard receiving a pension still means receiving a public pension. Among retired individuals, those drawing more than 10 percent of their annual income from a private pension plan are a negligible fraction (less than one percent). The situation, while slowly evolving, will not change substantially for another two decades or more. In 1990, the total number of participants in all kinds of private pension plans was of 600,000—less than 5 percent of total employment at the time. Since then, participation in pension funds has increased rapidly but not exceptionally, reaching a total of 4 million at the end of 1999. This is slightly less than 30 percent of current total employment and is mostly composed of individuals that are at relatively early stages of their working life. It is therefore reasonable to expect that, at least for the next two to three decades, the public pension system will remain the fundamental provider of old age income for Spanish citizens. Over the course of year 2000, the number of Spanish workers covered by the Social Security Administration (the general regime plus the special regimes) is estimated at around 14.9 million (it was 14.6 million at the end of 1999). Of these, more than 70 percent were covered by the Regimén GenMichele Boldrin is professor of economics at the University of Minnesota. Sergi JiménezMartín is associate professor in the department of economics at the Universitat Pompeu Fabra and Universitat Carlos III de Madrid. Franco Peracchi is professor of economics at the University of Rome, Tor Vergata. Financial support from the NSF, the Ministerio de Ciencia y Tecnologia (MCYT; Ministry of Science and Technology) project number BEC 2002-04294-C02-01, and the Ministerio de Educación (MEC; Ministry of Education) project number HPSE-CT-1999-00037 is gratefully acknowledged.

499

500

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

eral de la Seguridad Social (RGSS) and the rest by the special regimes. This corresponds to the practical totality of Spanish private-sector workforce (see Boldrin, Jiménez-Martín, and Peracchi [2001] for further details). Enrollment in the Spanish social security (SS) programs grew remarkably fast during the last three years, in fact, much faster than actual employment as estimated by the Spanish labor force Survey, Encuesta de Población Activa (EPA; Instituto Nacional de Estadística [INE] 2002). More precisely, the number of individuals enrolled in the SS administration grew at annual rates of 3.5, 5.7, and 6.0 percent in 1997, 1998, and 1999, respectively. The growth rate for 2000 is estimated to be around 3.7 percent. Such exceptional growth in the ranks of the SS administration is only partly explained by the favorable cyclical conditions during the period 1996–2000. A substantial portion of the increase in the number of contributors is likely to reflect a one-time emersion from the underground economy. This should be attributed partly to structural changes in the Spanish labor market and partly to an effort of the current administration to reduce contributive fraud. Hence, we should not extrapolate these growth rates in the future, even assuming that the current growth rate of the economy could be maintained, which, as 2001 and 2002 have shown, it cannot. The total number of contributive pensions, as of the end of 1999, was 7.6 million. The growth rate in the number of contributive pensions was 2.6 percent in 1996, 1.96 percent in 1997, 1.52 percent in 1998, 1.26 percent in 1999, and 1.24 percent (estimated) in 2000. The functional composition of the stock of contributive pension payments is the following: Retirement pensions are 59 percent of the total, survival pensions 26 percent, disability 11 percent, orphans 3 percent, and those to other family members 1 percent. The division by SS regimes gives 55 percent to the RGSS, followed by the farmers special regime (RETA) with 21 percent, and the selfemployed regime (REA) with 11 percent. The other, much smaller, special regimes share the residual 10 percent. For the year 2000, expenditure on public pensions in Spain is expected to equal PTA9,229.7 billion, approximately 9.5 percent of gross domestic product (GDP). By adding noncontributive and welfare pensions, total pension payments reach about 10.2 percent of GDP. This number is lower than previously forecasted and, in fact, lower than it was in 1996 and 1997. This is due to the strong growth performance of the Spanish economy over the four years from 1997 to 2000, which was continuing on into 2001. SS contributions in 1999 amounted to 9.4 percent of GDP and were expected to remain at about the same percentage level in 2000. The functional distribution of payments reflects the difference in average pension between old age and other functions. So, expenditure for old age is 68 percent of the total, while survival benefits are 19 percent, and disability is 12 percent of the total. During the year 2000, the total expenditure for contributive pensions amounted to PTA8,300 billion (about US$

Micro-Modeling of Retirement Behavior in Spain

Fig. 9.1

501

Social security and pension expenditures and GDP, 1980–1998

45 billion at the exchange rate of late November 2000), which corresponds to slightly less than 10 percent of Spanish GDP. Figure 9.1 reports the evolution of the ratio of pension expenditures P to GDP, during the last two decades. We also report, on the right vertical axis, the ratio between total expenditure of the Spanish SS system and GDP. Beside pensions, total expenditure includes expenditure for temporary illness, maternity or paternity, family support, public health services, and social services. Data show that the tendency to grow is not limited to pension expenditures: While the ratio between pension expenditure and GDP grew by 52 percent between 1980 and 1998, the one between total expenditure and GDP grew by 53 percent over the same period. The rule, at least in Spain, seems to be that when pension expenditure accelerates, total expenditure of the SS system accelerates as well. It is hard to judge whether the recent flattening of the trend line of the P-GDP ratio, should be considered as the beginning of a new phase or instead just a cyclical event without substantial long-run implication. The growth rate of the number of pensions does not appear to slow down. The same goes for the size of the average pension. Our analysis leads to the conclusion that the second interpretation is more likely, even if some small structural changes have taken place. Before going on to illustrate the longrun tendencies and their effects, we should briefly mention the steps that were taken in a new direction. The Spanish labor market is more flexible after the 1997 agreement be-

502

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

tween the government and the trade unions. Firing costs are slightly lower, and contracts that allow for dismissal can be signed. This shows positive effects on registered employment, which has been growing at near historical record rates since 1997. Labor force participation rates among women keep increasing. This is strictly linked to the higher school-attendance rates women have experienced since the middle 1980s, which appear to continue. This structural change in educational attainments has become an important factor, which may have a dramatic impact on the Spanish labor market during the next two or three decades. It is likely to increase labor force participation rates at all ages and for women, in particular. Meanwhile, if legislation is not modified, (which we find unlikely in light of the political-economy mechanism driving public pension systems) by 2005–2010, the number of workers having the right to retire before the age of sixty-five will dwindle to almost zero. Current legislation allows this privilege only to those who started contributing before 1967.1 Currently this constitutes the bulk of the retirees or near-retirees. Within a decade, this may provide a one-time boost to labor force participation rates for the age group sixty to sixty-five. For a long time, an alarming feature of the retirement behavior of Spanish workers has been the large fraction retiring and drawing pensions before the normal retirement age of sixty-five (see Boldrin, Jiménez-Martín, and Peracchi 2001 for detailed time series). The percentage of new pensioners strictly younger than sixty-five seems to have peaked at 71.6 percent in 1995, decreasing slowly in subsequent years. It was 68.2 percent in 1998. The percentage of those retiring at sixty may have peaked at 46.5 percent in 1997 and decreased marginally in 1998, the last year for which such data are available. This pattern is linked to the phasing out of the special legislation mentioned in the previous paragraph. On the other side, there are a number of facts that should make us pause. The large and persistent discrepancies reported between enrollment rates in the SS programs and the employment levels estimated by the EPA suggest that the record increase in SS enrollment is partly due to the emergence of the black labor market. This has two implications. First, it cannot be excluded that in a recession period the same workers would go “underground” again and stop contributing. Secondly, it implies that most of the 1997–2000 growth of employment is actually a statistical artifact. This fits well with the very low growth rate of labor productivity (0.2–0.5 percent) 1. In fact, following a new pact between the government and trade unions, legislation is being modified at the time of last writing (Spring 2002) to extend the right of early retirement at sixty to workers who enrolled in the SS system after 1967. According to the pact, early retirement for workers between the age of sixty and sixty-five will be possible under conditions essentially identical to those reported in the sequel of this paper. This fact vindicates our earlier pessimistic views about early retirement (see, e.g., Boldrin, Jiménez-Martín, and Peracchi 2001).

Micro-Modeling of Retirement Behavior in Spain

503

measured over the last four years. Should this pessimistic interpretation turn out to be correct, we would just be facing a (once-and-for-all) jump in SS enrollment levels. Furthermore, recent legislation (fall 1999) has once again increased the real value of minimum pensions by about 3 to 5 percentage points. Besides raising the overall pension burden, this policy has negative effects on participation rates. As clearly shown in Boldrin, Jiménez-Martín, and Peracchi (1999), minimum pensions are one of the major determinants of early retirement, especially among low earners. This suggests that the political willingness to increase pension expenditure to please special interest groups and to maintain the distortionary effects of current legislation has not been reduced by the arrival of a new government. A few other basic structural factors are illustrated next; for detailed statistics about Spanish demographic evolution and labor market, we refer to Boldrin, Jiménez-Martín, and Peracchi (2001). Table 9.1 presents the evolution of the structure of the Spanish population in the 1950–2050 period. We break down the total population in three representative age groups: zero to fourteen, fifteen to sixty-four, and sixtyfive and older. There are two salient features. On the one hand, the young population (zero to fourteen), after peaking in 1970, has steadily decreased. Currently the size of each generation of newborns is half the size of the same generation twenty years ago. On the other hand, the portion of the population sixty-five and older has continuously increased and will Table 9.1

Structure of the Spanish Population and Life Expectancy Life Expectancy Structure of the Population

At 0

At 65

Year

0–14

15–64

65+

Male

Female

Male

Female

1950 1970 1981 1986 1991 1996 2000

26.23 27.79 25.70 22.12 18.63 15.81 14.58

66.54 62.54 63.06 65.60 67.43 68.54 68.54

7.23 9.67 11.24 13.94 13.94 15.65 16.88

59.81 69.17 70.40 73.27 73.40 74.30 75.30

64.32 74.69 16.19 79.67 80.49 81.6 82.4

11.83 13.25 13.58 15.10 15.53 16.00 n.a.

13.48 15.89 16.44 18.43 19.17 20.10 n.a.

2010 2020

13.13 11.16

68.73 68.23

18.14 20.61

76.40 77.20

83.4 84.0

n.a. n.a.

n.a. n.a.

2030 2040 2050

12.5 12.6 12.8

62.70 57.10 53.90

24.80 30.30 33.30

77.80 78.20 78.50

84.5 84.8 84.9

n.a. n.a. n.a.

n.a. n.a. n.a.

Source: INE (1995) and Cordón (1999). Note: n.a.
not available.

504

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

reach 20 percent of the population before 2020. Furthermore, life expectancy at zero and at sixty-five has been growing steadily since 1950 and is expected to grow considerably also in the forthcoming years. In 2020, life expectancy at birth is expected to reach 76.0 and 83.7 for men and women, respectively. Historically, Spain has an extremely low ratio of contributors to pensioners (oscillating around 2 and equal to 2.02 and 2.06 in 1998 and 1999, respectively). Its long-run value could be somewhat higher for purely compositional reasons, as the ratio for RGSS and RETA was 2.52 and 2.84, respectively, in 1998, while it was well below unity for all other downsizing regimes (from 0.29 for miners to 0.77 for fishermen). The pattern of labor force participation in Spain is similar, from a qualitative viewpoint, to the rest of continental and non-Scandinavian Europe (France, Italy, Germany, and Belgium). Although Spain has higher unemployment and a lower labor force participation rate (LFPR), especially among females, the dynamics have been very similar since Spain joined the European Union (EU) in 1986. As a matter of fact, recent trends (post1996) are better than those in the other large European countries; most likely some form of convergence in labor market variables is taking place alongside convergence in per capita GDP and labor productivity. Immigration policies and outcomes are an exception to the former statement: Immigrants flow into Spain at a much lower rate than in the rest of continental Europe, and Spain does not seem to be prone to policies to encourage immigration (recent political debate confirms this posture). This is confirmed by the extremely low number of working permits issued to foreigners over the last five years: 88.620 (1994), 100.290 (1995), 126.407 (1996), 86.841 (1997), and 40.440 (1998). On average this represents about 0.5 percent of the workforce. The remarkable decline in number of working permits since the Popular Party (PP) came to power in 1996, in spite of the increased pressure from illegal immigrants landing on the Spanish coast, is particularly noticeable. The rest of this paper is organized in seven sections. The next one, section 9.2, provides an overview of the technical rules underlying the Spanish public pension system. Section 9.3 describes our micro-economic data set, illustrates its main limitations, and outlines the steps taken to overcome such limitations. (Additional material complementing sections 9.2 and 9.3 can be found in the appendix.) In section 9.4, we proceed to characterize the sample distribution of earnings and our estimated projections. Section 9.5 uses such projections to compute the various measures of retirement incentives used in this study. Such retirement incentive measures are the inputs for the estimation of the retirement models, which is undertaken in section 9.6. Section 9.7 studies three simple policy reforms and evaluates their differential impact upon the retirement incentives; section 9.8 briefly concludes.

Micro-Modeling of Retirement Behavior in Spain Table 9.2

505

Public Programs at Older Ages

Age

Unemployment Insurance

Disability Insurance

Private Pension Plan

50 52 55 60 65

Cont. from 45+ Cont. from 52+ Cont. from 52% Cont. —

Cont./Noncont. Cont./Noncont. Cont./Noncont. Cont./Noncont. —

Yes Yes Yes Yes Yes

Social Security Benefits a a a

ER: Cont. NR: Cont./Noncont.

Notes: Cont.
contributory; Noncont.
noncontributory; ER
early retirement; NR
normal retirement; 45+ and 52+ indicate a special UI program for 45+ and 52+ workers enrolled in the RGSS. All public programs provide benefits for dependant. Dashes indicate that data are not relevant. a There are age bonuses for certain professions, allowing for retirement before 60.

9.2 Institutional Features 9.2.1 Public Programs for Old-Age Workers Table 9.2 summarizes the programs available after age fifty. Besides private pensions, there are three other public programs that affect the behavior of old age workers: unemployment benefits, disability benefits, and retirement pensions. Unemployment benefits (UB) are generally conditional on a previous spell of contributions (see appendix) and are available only for workers in the RGSS.2 There are two continuation programs for those who have exhausted their entitlement to contributory UB: one for those aged forty-five and older (UB45 program) and another for those aged fifty-two and older (UB52 program). The latter program is a special subsidy for unemployed people that are older than fifty-two, lack income sources, have contributed to unemployment insurance for at least six years and, except for age, satisfy all requirements for an old age pension. To avoid cluttering the main text, we describe the various provisions of both programs in the appendix. The SS system provides insurance against both temporary and permanent illness as well as disability. Contributory disability (DI) benefits are far more generous than any other old age program since they are not subject to penalties for young age or insufficient years of contribution.3 The DI benefits are subject to approval by a medical examiner (the tightness of the 2. People enrolled in RESS have either no access to UB (self-employed and household employees) or have special unemployment (farmers and fishermen). 3. For a discussion of noncontributory disability pensions and other marginal insurance schemes (which are not relevant to the following analysis and have little or no impact on the retirement decisions of the workers we are considering) see Boldrin, Jiménez-Martín, and Peracchi (1999).

506

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

admissibility criteria used by examiners varies notoriously both over time and across regions) and, since the early 1990s, they have become harder to obtain at older ages. In fact, and contrary to the practice prevailing during the 1980s, it is now uncommon to access permanent DI benefits after age fifty-five. This has mainly been achieved by extending the disability evaluation process for the temporary illness program, Incapacidad Laboral Transitoria (see the appendix for a description), which, in the past, was most often used as a bridge to retirement. Both the unemployment and the disability plans offer, as we will argue momentarily, a pathway to early retirement alternate to the normal one (with early retirement at sixty and normal retirement at sixty-five). Such alternative pathways are taken in due account in our estimation and simulation procedures. The retirement program has two options: early retirement and normal retirement. Early retirement is possible from age sixty, but it only applies to workers who started their contributive career before 1967 (see note 1 and recall that this privilege is soon to be extended also to workers who enrolled after 1967). The normal retirement age is sixty-five, although some special professions have lower normal retirement ages (miners, military personnel, policemen, and fishermen are the main ones). Collective wage settlements often impose mandatory retirement at age sixty-five, facilitate retirement at sixty-four with full benefits, or encourage retirement between sixty and sixty-three through lump-sum amounts. 9.2.2 Social Security Regimes and Their Rules Under current legislation, public contributory pensions are provided by the following programs.

• The general SS scheme, Régimen General de la Seguridad Social (RGSS), and special SS schemes, Regímenes Especiales de la Seguridad Social (RESS), cover all private-sector employees, self-employed workers and professionals, members of cooperative firms, employees of most public administrations, other than the central government, as well as unemployed individuals who comply with the minimum number of contributory years when reaching sixty-five. The RESS include five special schemes: 1. Régimen Especial de Trabajadores Autónomos (RETA) for the self-employed. 2. Régimen Especial Agrario (REA) for agricultural workers and small farmers. 3. Régimen Especial de Empleados de Hogar (REEH) for domestic workers. 4. Régimen Especial de Trabajadores del Mar (RETM) for sailors. 5. Régimen Especial de la Minería del Carbón (REMC) for coal miners.

Micro-Modeling of Retirement Behavior in Spain

507

• Government employees scheme, Régimen de Clases Pasivas (RCP) includes public servants employed by the central government and its local branches. In this study, we do not consider this regime. Summary information about its structure and rules are reported later in this section. Legislation approved by Parliament in 1997 establishes the progressive elimination of all the special regimes by the end of year 2001. Aside from the pension scheme for public employees (RCP), the Spanish SS system will then be structured around only two schemes for the private sector: one for employees and one for the self-employed. It is not clear, at this point, if such reform will be completed on time.4 9.2.3 The General Regime This section describes the rules governing old age and survivors’ pensions under the general scheme (RGSS) until 1997. The main changes introduced by the 1997 reform will be illustrated as we go along. A summary of the basic technical aspects of the pre- and post-1997 systems can be found in table 9.3. Our focus on the RGSS is justified by the fact that this is the main SS program in Spain and the benchmark for our simulations. Financing and Eligibility The RGSS is a pure pay-as-you-go scheme. Contributions are a fixed proportion of covered earnings (defined as total earnings and excluding payments for overtime work) between a floor and a ceiling that vary for each of the broadly defined professional categories. Currently, eleven categories are distinguished, each one with its covered-earnings ceiling and floor. The current RGSS contribution rate is 28.3 percent, of which 23.6 percent is attributed to the employer and the remaining 4.7 percent to the employee. A tax rate of 14 percent is levied on earnings from overtime work. Entitlement to an old age pension requires at least fifteen years of contributions. As a general rule, recipiency is conditional on having reached age sixty-five and is incompatible with income from any kind of employment requiring affiliation to the SS system. Benefit Computation If eligibility conditions are met, a worker retiring at age sixty-five receives an initial monthly pension Pt equal to Pt
nBRt , where the benefit base (base reguladora) BRt is a weighted average of monthly covered earnings over a reference period that consists of the last eight years before retirement: 4. They have not, at least as of Spring 2002.

j
1

24

96 It–25  ∑ BCt–j  It–j j
25



Early retirement ()

j
1

24



Changes if n  40: 0 if a  60 .65 + .07(a – 60) if 60  a  65 1 if 65  a





if n  15 if 15  n  25 if 25  n  35 if 35  n

180 It–25  ∑ BCt–j  It–j j
25

0 .5 + .03(n – 15) .8 + .02(n – 25) 1

t–j


∑ BC

15 years 13 years

1  180

RGSS System after 1997

Provisions Affecting Particular Individuals C. Income tax exemptions Maximum pension exempted kt minimum wage kt minimum wage Maximum income exempted kt minimum wage kt minimum wage D. Contributions Minimum level of contribution (specific for 12 group) Maximum level of contribution (specific for 12 group) E. Pensions Minimum pension linked to minimum wage linked to minimum wage Maximum pension 4.3 minimum wage (in 1995) 4.3 times minimum wage F. Age bonuses Yes (occupation specific) Yes (occupation specific) G. Survivor benefits 0.45Pd 0.45Pd bt
max{min{b˜t [n, e, BR(BC, I ), bt}, b t}, where b˜t is the pension in A + B, and bt and b t are respectively the maximum and minimum pension

0 if a  60 .6 + .08(a – 60) if 60  a  65 1 if 65  a





Inadequate contribution ()

[progressive] linear (regime specific)

t–j


∑ BC

8 years 6 years

1  96

0 if n  15 .6 + .02(n – 15) if 15  n  35 1 if 35  n

Contribution period Fraction actualized A2. Fiscal system Income tax Labor tax B. Penalties

A1. Benefit base formula

Provisions Affecting All Individuals

RGSS System after 1985

Pension Provisions, Institutions, and Systems

A. Basic ingredients

Institutions

Table 9.3

Micro-Modeling of Retirement Behavior in Spain

1 BRt
 112




24

∑W

t j

j
1

509



96 It 25  ∑ Wt j  , It j j
25

where Wt–j and It–j indicate, respectively, earnings and the consumer price index in the j-th month before retirement. The replacement rate n depends on the number of contributive years and is equal to



0 n
.6  .02(n 15) 1

if n  15 if 15  n  35 if 35  n.

It may be further adjusted in the case of early retirement as described later. Starting from 1997, the number of reference years has been increased by one every year until 2001 and should soon be increased further to fifteen years. Moreover, the formula for computing n has been changed to the following. n




0 .5  .03(n 15) .8  .02(n 25) 1

if n  15 if 15  n  25 if 25  n  35 if 35  n.

In all of our simulations we use the pre-1997 formula, which was in place over the relevant sample period. We consider the impact of the 1997 reform (R97) when examining alternative policies (see discussion of R97 in section 9.7). Outstanding pensions are fully indexed to price inflation, as measured by the consumer price index. Until 1986, pensions were also indexed to real wage growth. Early Retirement The normal retirement age is sixty-five, but early retirement at age sixty is permitted for those who became affiliated to the SS before 1967. The current legislation distinguishes between two cases. The first one, representing the vast majority of those currently retiring between age sixty and sixtyfive, is the case of workers who started contributing as dependent employees to some Mutualidad Laboral (Workers’ Mutual) before 1967. In this case, the replacement rate is reduced by 8 percentage points for each year under age sixty-five. Starting from 1997, workers who retire after the age of sixty with forty or more contributive years are charged a penalty of only 7 percent for each year under age sixty-five. The second case, representing about 10 percent of the early retirees, is the case of workers with dangerous or unhealthy jobs (e.g., bullfighters, employees of railroads, airlines and public transportation companies, and so forth), or workers who were laid off for industrial restructuring regu-

510

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

lated by special legislation. In this case, no reduction applies. Notice that these exemption rights are “portable” in the following sense: They apply to individuals who were previously employed in one of the sectors deemed dangerous or unhealthy, but the minimum retirement age without penalty is proportional to the number of years of work spent in such sectors. Unless there are collective agreements that prescribe mandatory retirement, individuals may continue working after age sixty-five. In our empirical work, we try to estimate the impact of such special arrangements. Maximum and Minimum Pension Pensions are subject to an annually legislated ceiling, roughly equal to the ceiling on covered earnings. The 2000 ceiling corresponds to about 4.3 times the minimum wage (salario mínimo interprofesional; SMI) and about 1.6 times the average monthly earnings in the manufacturing and service sectors. If the computed old age pension is below a minimum, then a person is paid a annually legislated minimum pension. Other things being equal, minimum pensions are higher for those who are older than sixty-five or have a dependent spouse. In the last decade, minimum pensions grew at about the same rate as nominal wages, whereas maximum pensions grew at a lower rate that was about equal to the inflation rate. The ratio between the minimum old age pension and the minimum wage has been increasing steadily from the late 1970s (it was 75 percent in 1975) until reaching almost 100 percent in the early 1990s. On the other hand, the percentage of pensioners of the RGSS receiving a minimum pension has been declining steadily, from over 75 percent in the late 1970s to 27 percent in 1995. Family Considerations A pensioner receives a fixed annual allowance for each dependent child that is younger than eighteen or disabled. In 2000, this allowance is equal to PTA 48,420 for each child under eighteen and to PTA 468,720 (45 percent of the annualized minimum wage) for each disabled child. In addition, the minimum pension is increased by a fixed amount if a pensioner has a dependent spouse. Survivors (spouses, children, or other relatives) may receive a fraction of the benefit base of the deceased if the latter was a pensioner or died before retirement after contributing for at least 500 days in the last five years. The benefit base is computed differently in the two cases. If the deceased was a pensioner, the benefit base coincides with the pension. If the deceased was working, it is computed as an average of covered earnings over an uninterrupted period of two years chosen by the beneficiary from among the last seven years immediately before death. If death occurred because of a work accident or a professional illness, then the benefit base coincides with the last earnings.

Micro-Modeling of Retirement Behavior in Spain

511

The surviving spouse gets 45 percent of the benefit base of the deceased. In case of divorce, the pension is divided between the various spouses according to the length of their marriage with the deceased. Such a pension is compatible with labor income and any other old age or DI pension, but is lost if the spouse remarries. Surviving children get 20 percent each of the benefit base of the principal as long as they are less than eighteen or unable to work and stay unmarried. An orphan who is a sole beneficiary may receive up to 65 percent of the benefit base. If there are several surviving children, the sum of the pensions to the surviving spouse (if any) and children cannot exceed 100 percent of the benefit base. A Spanish peculiarity is the “pension in favor of family members.” This pension entitles other surviving relatives (e.g., parents, grandparents, siblings, nephews, and so forth) to 20 percent of the benefit base of the principal if they satisfy certain eligibility conditions (older than forty-five, do not have a spouse, do not have other means of subsistence, and have been living with and depending economically upon the deceased for the last two years). To this pension, one may add the 45 percent survivors’ pension if there is no surviving spouse or eligible surviving children. 9.2.4 Special Schemes In this section, we sketch the main differences between the general and the special schemes. Whereas rules and regulations for sailors and coal miners are very similar to the ones for the general scheme, special rules apply to self-employed, farmers, agricultural workers, domestic helpers, and a few other categories not discussed here, such as part-time workers, artists, traveling salespeople, and bullfighters. Beside differences in the SS tax rate and the definition of covered earnings, an important difference is the fact that those affiliated with the special schemes have no early retirement option (although exception is made for miners and sailors). The rest of this section focuses on the special schemes for self-employed workers (RETA) and farmers (REA), which together represent 93 percent of those affiliated with the special schemes and 86 percent of the pensions that are paid out. Self-Employed While the SS tax rate is the same for the RETA and the general scheme (28.3 percent in 2000), covered earnings are computed differently, as the self-employed are essentially free to choose their covered earnings between a floor and a annually legislated ceiling. Not surprisingly in the light of the strong progressivity of Spanish personal income taxes, a suspiciously large proportion of self-employed workers report earnings equal to the legislated floor. In 2000, the floor and the ceiling were equal to PTA 116,160 and PTA

512

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

407,790 per month, respectively, corresponding in annualized terms to 1.4 and 5 times the minimum wage, and 0.5 and 1.9 times the average earnings in manufacturing and services. For a self-employed aged fifty and older, the ceiling was only about half—PTA 219,000 per month—which was about equal to the average monthly earnings. A crucial difference with respect to the general scheme is that, under the RETA, recipiency of an old age pension is compatible with maintaining the self-employed status. The implications of this provision for the retirement behavior of self-employed workers are discussed later. Some other important provisions are the following. In order to qualify for survivors’ pensions, RETA only requires at least five years of contribution in the ten years immediately before the death of the principal. Under RETA, the latter is 50 percent of the benefit base. If the principal was not a pensioner at time of death, the benefit base is computed as the average of covered earnings over an uninterrupted period of five years chosen by the beneficiary among the last ten years before the death of the principal. Farmers In this case, both the SS tax rate and the covered earnings differ with respect to the general scheme. Self-employed farmers pay 19.75 percent of a tax base that is legislated annually and is unrelated to actual earnings. In 2000, this is equal to PTA 91,740 per month, corresponding to 1.24 times the minimum wage and about 40 percent of the average monthly earnings in the manufacturing and service sectors. Farm employees, instead, pay 11.5 percent of a monthly base that depends on their professional category and is legislated yearly. In addition, for each day of work, their employer must pay 15.5 percent of a daily base that also varies by professional category and is legislated annually. 9.2.5 Government Employees We now describe briefly the main differences between the general scheme and the RCP, the pension fund for the employees of the central government. Public servants are divided into five categories, labeled from (a) to (e), corresponding loosely to decreasing schooling levels: (a) for college graduates (doctor, licenciado, arquitecto o equivalente), (b) for people holding certain kinds of college diplomas (ingeniero técnico, diplomado, and so forth), (c) for high school graduates (bachiller o equivalente), (d) for junior high school diplomas (graduado escolar o equivalente), and (e) for individuals with lower education levels (certificad o de escolaridad ). For each of these categories, the budget law defines every year a theoretical SS wage (haber regulador), which is used to compute SS contributions and pensions. The implied wage scale has remained relatively constant since 1985. The top to bottom ratio never exceeded 2.5.

Micro-Modeling of Retirement Behavior in Spain

513

Social security contributions are the sum of three parts, each proportional to the legislated covered wage, according to proportionality factors legislated annually: (a) derechos pasivos (currently 3.86 percent), (b) cuota mensual de mutualidades (monthly mutual premium; 1.89 percent), and (c) aportacíon del estado (paid by the government, it varies between 6 and 10 percent depending on the sector of the administration). To parallel this three-part contribution structure, actual pensions are computed by adding up three sources of benefits: (a) the basic pension (derechos pasivos), (b) a portion directed to the pensioner’s family (ayuda familiar), and (c) a complementary portion coming from the various mutualidades—Instituto Sociale de las Fuerzas Armadas (ISFAS), Mutualidad Funcianarios de la Administración Central del Estado (MUFACE), and Mutualidad general de Empleados de la administración Judicial (MUGEJU). The basic monthly pension of a public servant who retires in month t after contributing for n years to RCP is computed as Pt
n BRt , where the dependence of n upon the numbers of years worked has changed frequently over time. For n  15, the last table of proportionality factors, legislated in 1990, can be reasonably (but not exactly) approximated by n
min[1, 1 .0366(35 n)]. The differences with respect to the general scheme are various. First, while the entitlement to a pension still requires at least fifteen years of contributions, the replacement rate (the ratio of the pension to the benefit base) increases somewhat irregularly with seniority, up to 100 percent after thirty-five years. So, for example, fifteen years of service gives pension rights equal to only 26.92 percent of the benefit base, against 60 percent of the general scheme. After thirty years the same ratio has increased to 81.73 percent, against 90 percent for the general scheme. Second, the benefit base is computed as a weighted average of covered earnings (to which the worker paid contributions) with weights equal to the percentage of the career spent at each level; that is, BRt
∑ pi Hit , t

where pi is the fraction of the career spent on level i and Hit are the covered earnings corresponding to level i, as determined by the current law at time t. Third, unlike the general scheme, the RCP imposes mandatory retirement at age sixty-five. Exceptions are made for a few special categories, such as university professors and judges. On the other hand, the RCP allows for early retirement at the age of sixty, without any penalty for public servants with at least thirty years of service (twenty years for military personnel). A fourth important difference with respect to the general scheme is compatibility between RCP pensions receipt and income from continued work.

514

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

Table 9.4

Pathway

Pathways to Retirement RGSS and Assimilates

1. NR at 65 2. ER at 60 3. UB then ER 4. ER through DI

✓ ✓ ✓ ✓

RETA ✓ ✓ ✓

RCP (Public Employees) ✓ ✓ (Without penalty with 35 years service) Not relevant ✓

Note: See text for explanation of abbreviations.

In a number of special cases, RCP pensioners are allowed to keep a publicsector occupation, as long as this does not provide them with a regular flow of income (for example, the case of members of legislative bodies). More importantly, the legislation allows RCP pensions to be cumulated with earnings from employment in the private sector. When a public servant is dismissed because of disability (and therefore starts drawing a DI pension) or dies (and the survivors are therefore entitled to a pension), the missing years between the person’s age at death and sixty-five are counted as actual years of service in the computation of either the DI or the survivors’ pension. Should the disability be caused by an accident while on duty, the DI pension is doubled. 9.2.6 Pathways to Retirement This brief illustration of the public pension system clarifies that more than one pathway to retirement is available to Spanish workers. We have identified four of them in more detail: early retirement, normal retirement, temporary illness or disability and the UB program (specifically, UB52). In table 9.4 we provide a brief listing of which programs are available according to the SS regime in which one is enrolled. 9.3 Data Description 9.3.1 The Main Data Set Our main micro-economic data set is based on administrative records from the Spanish SS administration (Historiales Laborales de la Seguridad Social; HLSS). The sample consists of 250,000 individual work histories randomly drawn from the historical files of SS affiliates (Fichero Histórico de Afiliados; FHA). The sample includes only individuals aged forty and older on 31 July 1998, the date at which the files were prepared. The sample contains individuals from the RGSS and the five special regimes—RETA, REA, REEH, RTMC and RTMAR. As we mentioned above, civil servants and other central government employees are not covered by the SS administration and are not considered in this study. The data set consists of three files. The first file, the “history file” (H file),

Micro-Modeling of Retirement Behavior in Spain

515

contains the work history of the individuals in the sample. Each record in this file describes a single employment spell of the individual. As we argue later, the work histories are very accurate for employment spells that began after the mid-1960s. The second file, the “covered earnings file” (CE file), contains annual averages of covered earnings (bases de cotización) from 1986 to 1995. The third file, the “benefits file” (B file), contains information on the lifetime SS benefits received by the individuals in the sample. Benefits are classified by function (retirement, disability, survival, and so forth) and initial amount received. To be more precise, the B file contains the initial benefit amount and the length of the period during which the benefit was received. A fourth file, the “relatives file” (R file), is also available; it reports some benefits paid to relatives who were members of the individual’s household. For each individual in the sample who contributed to SS during the 1986–1995 period, the CE file reports the annual average of covered earnings together with the contributions paid. For individuals enrolled in either the RGSS or the RTMC, covered earnings are a version of earnings that are doubly censored (from above and below). What this means is that covered earnings have both ceilings and floors: Contributions must be paid over some legislated minimum wage, no matter what actual earnings are. Furthermore, earnings above certain legislated ceilings are not covered; that is, they do not generate any future right and, as such, are not reported in the SS administration files. Notice, although, that they are taxed for contributions, which matter for retirement incentives. For people enrolled in SS regimes other than RGSS and RTMC, covered earnings are chosen by the individual within given ceilings and floors (see section 9.2 for details) and, consequently, there is no clear link between covered and actual earnings in this case. For each employment spell in the H file, we know age, sex, and marital status of the person, the duration of the spell (in days), the type of contract (in particular, we can distinguish between part-time and full-time contracts), the social security regime, the contributive group, the cause for the termination of the spell, the sector of employment (four-digit standard industrial classification [SIC]), and the region of residence (fifty-two Spanish provinces). For each individual in the H file who has received some benefits at any point in time, we know most of the information that the SS administration uses to compute the monthly benefits to be paid. In particular, we know the initial and current pension, the benefit base (base reguladora), the number of contributive years, the current integration toward the minimum pension (complementos por el mínimo), the date pension was claimed, the date it was awarded, the type of benefits, and so on. We refer to Martínez (1999) for a detailed description of the variables and for summary statistics of the history, covered earnings, and benefits files. The distribution of the HLSS sample, by activity, regime and status of the individuals therein recorded is summarized in table 9.5.

Table 9.5

Distribution of Sample, by Activities in 1997 Activity Status

Age and Gender

Working Full-Time

Working Part-Time

Not Working

Temporary Illness

Died While Active

Out of LF

Total

All the Regimes Male 50–54 55–59 60–65 65–69 Female 50–54 55–59 60–65 65–69

72.74 69.76 39.72 5.22

0.99 0.52 0.31 0.07

4.16 2.89 1.23 0.21

1.79 1.49 1.51 1.80

1.82 2.49 3.54 4.12

18.50 22.85 53.68 88.58

20794 20878 22813 19304

57.56 56.32 45.72 12.97

5.54 4.21 2.45 0.48

3.51 2.40 1.22 0.28

3.85 4.01 3.12 2.69

0.58 0.85 1.44 1.53

28.95 32.21 46.05 82.06

12409 9385 9237 7142

RGSS Male 50–54 55–59 60–65 65–69 Female 50–54 55–59 60–65 65–69

71.47 68.73 35.78 3.10

1.44 0.78 0.49 0.13

4.80 3.49 1.61 0.27

2.50 2.02 2.08 2.60

1.67 2.32 3.16 3.90

18.13 22.66 56.88 90.00

14222 13887 14327 11127

46.41 46.07 34.88 6.54

10.47 8.94 5.85 1.30

4.96 3.35 2.02 0.27

7.00 7.63 6.24 6.27

0.66 0.95 1.16 1.26

30.51 33.05 49.86 84.37

6555 4417 3865 2616

RETA Male 50–54 55–59 60–65 65–69 Female 50–54 55–59 60–65 65–69

77.21 72.72 52.23 8.03

— — — —

3.05 2.21 1.17 0.24

0.11 0.09 0.22 0.00

1.38 2.64 4.68 5.27

18.25 22.34 41.70 86.46

3770 3442 3163 2466

60.08 53.99 45.39 15.24

— — — —

3.00 2.47 0.91 0.53

0.11 0.43 0.40 0.07

0.37 0.66 1.82 1.72

36.43 42.45 51.47 82.44

2698 2106 1976 1509

RETA Male 50–54 55–59 60–65 65–69 Female 50–54 55–59 60–65 65–69

58.80 40.02 15.32 10.31

0.62 0.25 0.15 0.06

34.13 49.43 22.59 0.09

0.00 0.00 8.27 20.15

0.34 2.43 46.71 62.11

6.11 7.86 6.97 7.29

3847 4770 12246 17100

80.68 66.66 41.02 21.28

5.12 8.34 10.04 6.98

11.72 20.64 21.77 0.14

0.00 0.00 7.45 22.97

0.03 0.36 14.81 44.36

2.45 4.00 4.91 4.28

3593 3023 4254 5861

Micro-Modeling of Retirement Behavior in Spain

517

(continued)

Table 9.5

Benefits (as a fraction of people out of the LF) Without SS Benefits

Survival

DI

Retirement After DI

Retirement

Died While Receiving

Total

All the Regimes Male 50–54 55–59 60–65 65–69 Female 50–54 55–59 60–65 65–69

58.80 40.02 15.32 10.31

0.62 0.25 0.15 0.06

34.13 49.43 22.59 0.09

0.00 0.00 8.27 20.15

0.34 2.43 46.71 62.11

6.11 7.86 6.97 7.29

3847 4770 12246 17100

80.68 66.66 41.02 21.28

5.12 8.34 10.04 6.98

11.72 20.64 21.77 0.14

0.00 0.00 7.45 22.97

0.03 0.36 14.81 44.36

2.45 4.00 4.91 4.28

3593 3023 4254 5861

RGSS Male 50–54 55–59 60–65 65–69 Female 50–54 55–59 60–65 65–69

56.25 40.48 14.48 10.66

0.74 0.29 0.11 0.06

35.61 47.63 19.45 0.12

0.00 0.00 6.87 17.39

0.19 3.18 51.74 64.20

7.21 8.42 7.35 7.58

2578 3147 8149 10059

81.70 67.81 39.91 23.02

5.30 9.38 10.02 7.16

10.10 17.19 16.97 0.09

0.00 0.00 5.35 17.72

0.05 0.68 21.90 47.30

2.85 4.93 5.86 4.71

2000 1460 1927 2207

RETA Male 50–54 55–59 60–65 65–69 Female 50–54 55–59 60–65 65–69

72.82 52.28 26.91 12.90

0.73 0.26 0.45 0.14

23.29 42.13 26.91 0.09

0.00 0.00 9.40 18.15

0.15 0.91 31.24 62.24

3.05 4.42 5.08 6.47

688 769 1319 2132

86.78 78.64 57.23 31.67

4.48 7.72 12.29 10.29

7.53 12.75 14.85 0.08

0.00 0.00 4.42 14.95

0.00 0.00 8.85 39.07

1.22 0.89 2.36 3.94

983 894 1017 1244

Note: Dashes indicate that data is not relevant.

9.3.2 Data Problems in Historiales Laborales de la Seguridad Social (HLSS) We face several problems with the HLSS files, most of which are inherent to the structure of the Spanish SS record-keeping procedures. We shall illustrate some of the problems we encountered by comparing sample statistics from the HLSS with those obtained from other data sources that are,

518

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

presumably, more representative of the working population under study, at least along the dimensions considered here.

• Overrepresentation of some regions or sectors: The proportion of individuals from some geographical regions or industrial sectors is much larger in our file than in either the census or the EPA labor force survey. Carrying out inference conditional on region and industrial sector is therefore essential. • Mortality data: We have limited information on mortality. In principle, information for those who are active is good enough in the sense that the data report whether or not a person is alive or dead and, if dead, when the event occurred. However, information for the retirees is incomplete, since we only know whether a person is still alive at the reference date (31 July 1998). In other words, for those in the B file, we know if they stopped receiving benefits because of death before 31 July 1998, but not the exact date when this happened. • Left-censored histories: Early spells (those which started before the mid-1960s) are very poorly recorded because the current structure of the Spanish SS administration was set up only in the second half of the 1960s. Hence, we have incomplete records of the work histories of individuals older than fifty-five or sixty. This is a major problem for our computation of expected pensions, since the current system establishes a clear formula to impute years of contribution before 1967, specifically, IDC
[1967 ( yb  210)] 250, where IDC stands for imputed days of contribution, and yb stands for the year of birth. For each individual, the IDC is then compared with the actual days of contribution before 1967 as reported in the H file. The largest of these quantities is chosen as the individual’s contributive history before 1967. • Marital status: This variable is very poorly recorded, especially for individuals who are still active. The reason is that marital status does not affect contribution rates but may affect SS benefits. Hence, individuals have no incentive to adjust their records while active. They do so only when pressured, which is not frequent, or when they change jobs. Most often, people adjust their marital status in the SS records at the time of retirement in order to draw benefits for the spouse. Information on marital status is therefore incorrect for many individuals (in 1997, only 27.50 and 10.31 percent of men and women, respectively, report to be married). To fill in the gap we use the EPA. From this survey, we can extract the following information: 1. Working men are married with women three years younger, whereas working women are married with men four years older.

Micro-Modeling of Retirement Behavior in Spain

519

2. Male workers are very likely to be married. The fraction of working males who are married, from EPA, is 95.10, 94.6, and 92.7 percent for males aged fifty to fifty-four, fifty-five to fifty-nine, and sixty to sixty-four, respectively. Female workers are less likely to be married (between 60 and 70 percent) and more likely to be widowed (between 10 and 20 percent), depending on the level of education. 3. For low-educated male workers fifty-five and older, the fraction of working spouses is very low (less than 15 percent) and decreases with age; whereas for highly educated male workers, the fraction of working spouses is much higher (35 percent). 4. For low-educated female workers fifty-five and older, the fraction of working spouses is low (less than 35 percent for age group fiftyfive to fifty-nine and 25 percent for the age group sixty to sixty-four. For highly educated female workers, the fraction of working spouses is also much higher (45 percent and 30 percent, respectively, for the age groups fifty-five to fifty-nine and sixty to sixty-four). • Family data: There is no information on family size or its structure. The Spanish SS simply does not keep this kind of records because family size does not affect either contribution rates (like marital status) or pensions (as marital status does). 9.3.3 Sample Selection Rules We distinguish between a “wage sample,” used to study earnings dynamics, and a “participation sample,” used to study labor force participation and exit into retirement. In either case, the analysis is carried out separately for men and women born between 1916 and 1958 (about 160,000 men and 84,000 women). In figure 9.2, we show the distribution of the sample by sex and year of birth. Note the upward jump between 1918 and 1919 and the reduction during 1938 and 1939 followed by a spike in 1940, which was a direct consequence of the Spanish Civil War. While we place practically no restrictions on the H-file derived participation sample, which covers all the Spanish SS regimes (see table 9.6 for a

Fig. 9.2

Distribution of the participation sample by birthdate and sex

520

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

Table 9.6

Distribution of the Participation Sample, by Regime

Year

RGSS

RETA

REA

RTMAR

RTMC

REEH

N

1985 1989 1993 1997

59.66 61.36 63.43 62.00

15.98 16.97 17.01 17.58

18.48 16.24 14.32 14.93

2.97 3.08 3.40 3.90

0.07 0.07 0.07 0.08

2.87 2.28 1.77 1.51

213,238 196,254 168,560 132,587

Notes: N = number of observations. See text for explanation of abbreviations.

description of the sample distribution by regime), the CE-file wage sample is restricted to individuals in the RGSS, the RTMC, or RETA. We have excluded individuals enrolled in the REA, RTMAR, or REEH because of their discontinuous careers and very low reliability of the earnings reported. 9.3.4 Key Elements in Data Handling Definition of Retirement In any given year t, being retired can be characterized by a number of different events. Correspondingly, we have the following four different definitions of retirement, of which the first is the broadest, while the others cannot be directly compared. 1. Not having an open spell after year t 2. In addition to 1, not contributing in years t  k, k  0 3. In addition to 1, a retirement recorded in the entry “cause of termination of the spell” 4. In addition to 1, a retirement or termination by temporary illness recorded in the entry “cause of termination of the spell” In our sample, we obtain practically the same fraction of retired people under either definition 1 or 2. More precisely, 99.15 percent of males retired under definition 1 are also retired under definition 2. Likewise, 96.01 percent of females retired under definition 1 are retired under definition 2. The definitions using the cause for the termination of the last employment spell (either definitions 3 or 4) are much stricter and probably too much so. Only a fraction of those considered retired under definition 1 are as such under either definitions 3 or 4 (48.82 and 30.82 of the male and female sample, respectively). This lack of coincidence is sharply reduced when considering individuals aged sixty to sixty-four, but is still very important (well above 40 percent in both cases). Most likely, this large discrepancy is due to omission of the cause for termination of spell and does not reflect a different status. A number of other technical details should be kept in mind. 1. Eligibility: The early retirement age in Spain is sixty (see section 9.2 for a description of restrictions and related penalties). Some exceptions are

Micro-Modeling of Retirement Behavior in Spain

521

still possible at age fifty-eight for workers of distressed firms. We can follow individuals from either age or even earlier when appropriate. 2. Transition from unemployment to retirement: This is an option open to workers older than fifty-two who are eligible for a special kind of UB. This special kind of UB is not reported in the current version of our data set. Still, we need to consider that, for individuals older than fifty-two, this is a possible path to retirement. In particular, we must decide whether or not unemployed people older than fifty-two should be considered in or out of the labor force. In this work, we have decided that a worker is in the labor force as long as they are contributing to the SS administration. 3. Transition to and from disability: These transitions are hard to capture using the available data. Considerations similar to those developed for the case of transition from unemployment apply to working histories involving disability or long-term illness. In order to classify people, we follow the criteria just outlined for those unemployed: A worker in the temporary illness program is in the labor force as long as they keep contributing to the SS. Demographic Characteristics of the Sample Our sample reports information on age, sex, contributive group (from which we can extract a proxy of the level of education), marital status (very imprecise, as mentioned earlier), sector of employment (four-digit SIC classification), and province of residence. Other available information is part-time work and length of tenure in the current job and in the labor market. Education and Participation (RGSS) No direct measure of educational attainment is available. For individuals enrolled in the RGSS, a good proxy for the level of education may be constructed using the information on the contributive group of the subject since, for these workers, a variety of labor market regulations force a close relationship between educational level and contributive group. The criteria we adopted are the following. All individuals in contributive group 1 were assigned to the college level of the educational variable. Individuals in contributive groups 2, 3, and 4 were assigned to the high school (diploma) level. People in all other contributive groups were assigned to a generic class labeled as “less than high school.” Table 9.7 compares the resulting distribution of educational levels in the HLSS sample in 1997 with the corresponding distribution obtained from the EPA, which reports educational levels directly. Results are mixed. If we take the EPA as a correct estimate of the population distribution, then our sample overestimates the number of educated men (summing together those at the college and diploma levels) and underestimate the proportion of educated women. This is not altogether surprising. We are inferring the

522

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

Table 9.7

Distribution by Education in HLSS and EPA in 1997 (employed individuals, born 1916–58, enrolled in the RGSS) HLSS Sample

EPA IV

Education

Male

Female

Total

Male

Female

Total

College Diploma Other Total

9.55 10.18 80.27 100.00

6.38 10.51 83.11 100.00

8.58 10.28 81.13 100.00

7.99 6.81 85.20 100.00

7.48 14.27 78.25 100.00

7.82 9.22 82.96 100.00

educational level from the professional rank in which individuals are classified for contributive purposes. The Spanish labor market is notoriously characterized by a substantial amount of discrimination by sex. This forces a large fraction of working women into occupational profiles lower than those of men with similar educational attainment and qualification. Various empirical studies have documented this fact, which is clearly reflected also in the SS administrative records that we use. The bottom panels of figure 9.3 show the pattern of participation by educational level and age for individuals enrolled in the RGSS. The large variations in the participation rate of women with either a college degree or a diploma at ages higher than fifty should be interpreted as noise generated by the very limited number of observations available for female in those age groups. The table confirms the well-documented finding that, for both men and women, higher education is associated with higher participation rates in general and at older ages, in particular. Notice also how retirement patterns for individuals with high education are much more sharply defined: Hazard rates before the normal retirement age of sixty-five are lower than for the rest of the labor force, while they become much higher at the normal retirement age. Economic Characteristics of the Sample In table 9.8 we show, for selected years, the sample distribution by contributive groups of workers enrolled in the RGSS and the RTMC. As mentioned earlier, the contributive group may be regarded as a combination of education, skills, and type of contract. The distribution by contributive groups in our sample seems quite stable over the whole period, except for blue-collars. The fraction of skilled blue-collars increases (from 21.6 to 25.4 percent), whereas the fraction of semiskilled and unskilled decreases. These findings reflect accepted modifications in the skill distribution of the Spanish labor force over the sample period. Table 9.9 shows the sample distribution of workers in the RGSS by broad industry categorization (1-digit SIC classification). For both men and

Fig. 9.3

A

Participation, employment, and education level, all regimes

Fig. 9.3

B

(cont.) Participation, employment, and education level, all regimes

Table 9.8

Distribution of the RGSS Plus the RTMC Sample, by Group of Contribution Year

Group of Contribution White-collar 1. Engineer and College 2. Technical engineer 3. Supervisor and foreman 4. Assistant without grade 5. Clerk 6. Janitor 7. Clerk assistant Blue-collar 8. Skilled (1st and 2nd class) 9. Semiskilled (3rd and specialized) 10. Unskilled 11. Worker (17 yrs old) 12. Worker (16 yrs old) 13. Other Total

Table 9.9

1985

1989

1993

1997

6.08 4.34 3.97 3.39 9.35 5.73 7.04

7.10 4.68 4.53 3.70 9.68 5.87 6.58

7.75 5.11 5.13 3.91 10.13 6.21 6.67

7.75 5.87 5.36 4.09 10.70 6.70 7.08

21.57 14.43 20.83 0.29 0.16 2.81

24.05 12.97 20.13 0.16 0.07 0.38

24.89 11.43 18.50 0.09 0.01 0.17

25.42 10.26 16.58 0.04 0.00 0.16

127,356

120,568

107,032

82,313

Sample Distribution in 1985 and 1997, by Sector (RGSS and RTMC) 1985

Sector

Male

Female

Agriculture, fishing Energy Minerals, chemical Mechanical, engineering Other manufacturing Industry Construction Retail Transportation Communication and financial Other services Administration

0.32 2.29 7.12 10.52 11.19

0.23 0.50 1.75 2.29 8.58

11.35 14.65 7.15 8.16 6.08 8.92

Code 9130 Temporary illness Other (Codes 0000 & 9990) Total

1997 Total

Male

Female

Total

0.30 1.79 5.61 8.21 10.46

0.26 1.99 4.78 6.94 7.67

0.16 0.34 1.25 1.54 5.32

0.23 1.51 3.75 5.36 6.98

1.93 19.51 2.81 6.72 16.37 23.74

8.71 16.01 5.93 7.75 8.97 13.08

10.30 13.50 7.15 8.26 9.19 10.23

1.18 15.99 2.71 7.14 24.85 20.47

7.63 14.23 5.85 7.94 13.74 13.23

2.76 8.66 0.79

5.17 8.97 1.42

3.44 8.75 0.97

9.89 8.10 1.75

10.03 8.01 1.00

9.93 8.08 1.54

69,682

27,219

96,901

53,134

22,006

75,140

526

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

Table 9.10

Length of Firm and Market Tenure, by Years All SS Regimes Spell

RGSS and RTMC

Experience

Spell

Experience

Year

Male

Female

Male

Female

Male

Female

Male

Female

1985 1988 1991 1994 1997

8.3 7.7 7.4 7.2 6.8

5.0 4.8 4.6 4.5 4.4

16.3 16.0 15.9 15.8 15.3

8.1 8.1 7.9 7.9 7.9

6.5 5.9 5.7 5.9 5.8

4.7 4.4 4.2 4.4 4.5

16.7 16.2 16.2 16.2 15.9

10.1 9.9 9.7 9.9 10.1

Note: History spell is uncorrected for left-censored histories.

women, the most important sector of employment is administration and other services, followed by retail services. Note also the important fraction of men in a condition of temporary illness. We do not report the distribution for individuals enrolled in the RESS (RETA, REA, or REEH) because the data set does not indicate the industry in which they should be classified. In table 9.10, we present the average length of the last spell of work (a good proxy for the length of tenure with the current firm) and the average work history (a good proxy for labor market tenure). Notice that both measures remain fairly stable during the sample period and that, for obvious reasons, work spells with a firm are much longer for self-employed than for employees. Both firm and market tenures are much longer for men than for women. 9.4 Earnings Distribution, Earnings Histories, and Projections As commented in section 9.3.1, we do not observe earnings directly, but instead only use covered earnings. Covered earnings are a doubly censored version of earnings for workers in the RGSS or RTMC, and they are very weakly related to true earnings for workers in the RESS because of the presence of both legislated tariffs and widespread tax fraud. 9.4.1 RGSS and RTMC In figure 9.4 we present the distribution of the log of real covered earnings for workers enrolled in the RGSS or RTMC for the years 1986 (top panels) and 1995 (bottom panels), respectively. We distinguish by sex and report only two contributive groups (1 and 8). Two patterns arise. First, the increase over time in the fraction of top-censored observations for both the first (workers with college degrees) and the eighth (skilled workers) contributive group. For both men and women, the increase is quite pronounced for the first group, which corresponds to the highest wages. In the

Fig. 9.4

A

Distribution of covered earnings, RGSS by sex and group of contribution: A, 1986; B, 1995

Fig. 9.4

B

(cont.) Distribution of covered earnings, RGSS by sex and group of contribution: A, 1986; B, 1995

Micro-Modeling of Retirement Behavior in Spain

529

other group, which corresponds to median and below-median wages, it is relevant only for men and not for women. This asymmetry suggests that the gender bias characterizing the Spanish labor market is actually weaker or weakening in the top segments of the wage distribution but is still quite strong in the lower ones. The presence of increasing top censoring is also evidence of the inability of legislated ceilings to keep up with real wage dynamics: Ceilings on covered earnings are adjusted only to price inflation and do not track growth in real wages. The second important observation is that bottom censoring also increases in the eighth contributive group for both men and women, which is quite surprising. This suggests that the wage distribution has become more spread out over time. Notice that the Spanish government has also followed the policy of progressively reducing the relative size of the floor-to-ceiling bands by increasing the floor faster than the ceiling, which helps explaining the increasing number of bottomcensored individuals. Overall the evidence reported suggests that we should invest a considerable effort in recovering true earnings from covered earnings for people enrolled in the RGSS or RTMC. Also, given the purpose for which we need to uncover true earnings, eliminating the effect of top censoring is the important goal. In our analysis, true earnings are used to project or forecast future wages for workers (sixty-five years or older) that are making the choice between retiring and continuing to work. In any given contributive group, it is most unlikely that such workers would be at the bottom of the wage distribution and would look forward to an ever-decreasing salary if they kept working. Both, the skill-acquisition process and the existence of seniority pay (still important in Spain) suggest that old workers should be found in the upper tail of the distribution of salaries. This intuition is confirmed, but only partially, by the data. In table 9.11, we report the percentage of workers, men or women, that are forty or older and are either at the bottom or at the top of the distribution of covered earnings for each one of the ten contributive groups. As expected, the frequency of observations in the bottom-censored groups is substantially lower than at the top-censored groups. Still, it is higher than one would expect, especially for people in the contributive groups with lower salaries (7 and higher-index numbers), and it does not seem to decrease with age. These anomalies in the data notwithstanding, we find it rather unlikely for the near retirees to be found in any significant proportion at the bottom of the distribution of wages and looking forward to further decreases in the wage itself. Hence, we have elected not to bother getting rid of the bottom censoring and to concentrate on the top-censoring problem. To deal with the top-censoring problem, we proceed as follows. First we estimate a tobit model for covered earnings. Then we use the estimated parameters to impute the earnings of the censored observations and estimate an earning function using imputed earnings for those affected by the ceil-

530

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

Table 9.11

Relevance of Floors and Ceilings in Covered Earnings (RGSS, RTMC, and RETA) 1986 Male

Group

1 2 3 4 5 6 7 8 9 10

60–64

Female

Censored

40–54

55–59

65+

Total

B A B A B A B A B A B A B A B A B A B A

5.5 37.9 1.4 37.8 1.2 34.8 0.6 34.1 1.2 37.0 2.8 17.6 8.6 25.2 3.2 13.2 2.3 14.8 12.7 9.6

5.0 33.2 4.1 22.1 1.6 32.3 0.9 20.3 2.2 36.1 2.8 12.8 10.5 18.8 3.3 15.1 2.8 11.2 16.5 4.9

RGSS and RTMC 5.4 19.4 5.6 30.4 25.8 36.6 3.3 9.5 1.9 21.3 9.5 34.8 1.9 4.8 1.3 25.0 14.3 33.5 0.9 0.0 0.7 19.7 13.3 31.3 0.4 0.0 1.3 30.0 22.9 36.3 3.1 10.2 3.0 8.2 6.8 14.8 10.3 5.6 8.9 10.3 5.6 22.6 2.2 2.9 3.1 10.8 0.0 13.3 3.1 17.1 2.5 8.9 0.0 13.8 10.1 21.7 13.3 6.0 8.3 8.2

B A

93.2 0.0

90.8 0.0

83.8 0.1

RETA 83.6 91.6 0.0 0.0

40–54

55–59

60–64

65+

Total

9.0 14.6 3.1 4.8 5.1 18.2 1.8 14.5 3.2 22.2 6.9 3.1 9.1 12.1 5.1 4.4 15.8 3.8 24.3 3.1

11.1 5.6 6.5 4.8 5.3 10.5 7.1 0.0 2.7 20.5 9.8 2.5 12.3 7.0 9.3 2.8 13.2 3.6 19.9 2.2

0.0 13.3 5.3 0.0 0.0 10.0 0.0 12.5 4.0 12.0 11.6 5.8 3.1 3.1 3.8 1.9 13.5 1.9 22.9 0.9

0.0 0.0 0.0 11.1 0.0 25.0 — — 0.0 33.3 0.0 11.1 100. 0.0 0.0 0.0 11.1 0.0 39.3 0.0

8.5 13.8 3.6 4.6 4.7 17.1 2.3 12.9 3.2 21.7 7.7 3.3 9.2 10.9 5.7 3.9 15.3 3.6 23.5 2.7

97.6 0.0

96.3 0.0

93.3 0.0

93.6 0.0

96.5 0.0

ings. Finally, we generate “true earnings” for all the individuals in the topcensored groups by using the estimated regression function and adding an individual random noise component. The first two steps of the above procedure are detailed in the appendix, the latter is described in section 9.4.1. From the individual profile of covered earnings ct between year T – k and year T, we impute the individual profile of real true earnings (wt , t
T – k, . . . , T ). Given this information, we project earnings forward and backward in the following way:

• Forward: (zero real growth) wˆTm
wˆt for m
1, . . . , M; • Backward: wˆT–k–

wT–k  g (aT–k) for

1, . . . , L. The function g( ) corrects for the growth of earnings imputable to age a and is defined as g(aT k
)
1 aT k
 2 a2T k
1 aT k 2 a2T k . The s are the estimated coefficients from a fixed-effects earnings equation, the details of which are available upon request. The correction is specific for each combination of sex and contributive group.

Micro-Modeling of Retirement Behavior in Spain

531

(continued)

Table 9.11

1995 Male Group

1 2 3 4 5 6 7 8 9 10

60–64

Female

Censored

40–54

55–59

65+

Total

B A B A B A B A B A B A B A B A B A B A

3.8 54.6 2.8 36.9 3.0 33.6 2.6 20.5 4.4 35.1 5.0 18.6 10.8 19.1 9.8 9.4 10.1 8.3 25.8 2.9

3.7 54.2 4.9 39.1 6.0 29.1 8.4 15.4 9.6 29.9 9.0 11.7 12.0 19.5 18.5 9.1 16.3 8.1 41.5 3.2

RGSS and RTMC 4.0 7.8 3.9 48.1 39.2 53.5 4.1 3.0 3.2 25.1 18.2 36.0 5.5 6.8 3.9 20.9 25.0 31.4 7.7 15.2 4.4 11.6 12.1 18.5 6.2 8.7 5.5 23.5 11.6 32.7 6.2 5.8 6.1 4.8 4.3 14.4 16.1 34.6 12.3 13.5 1.9 18.0 14.3 18.3 12.1 5.5 2.4 8.9 18.4 23.6 12.7 3.5 0.0 7.6 39.5 58.2 32.3 2.1 0.7 2.8

B A

91.9 0.1

85.6 0.3

80.8 0.4

RETA 83.9 89.1 0.7 0.2

40–54

55–59

60–64

65+

Total

9.6 31.7 3.8 5.2 4.3 14.4 16.1 5.9 12.5 19.8 11.4 5.3 21.8 5.0 27.2 2.0 36.9 1.9 49.3 0.8

7.2 39.1 5.5 5.5 8.9 8.9 9.1 5.5 15.0 13.1 14.6 5.1 30.1 2.8 27.2 2.5 39.6 1.3 45.2 0.5

8.0 44.0 5.3 7.4 7.7 0.0 18.2 9.1 16.3 14.1 11.8 0.8 20.2 3.2 23.4 0.0 38.0 0.5 41.0 1.0

0.0 15.4 17.6 0.0 12.5 12.5 16.7 0.0 16.0 4.0 13.5 0.0 28.6 0.0 53.3 0.0 29.6 0.0 42.5 2.1

9.2 32.6 4.2 5.3 5.2 12.8 15.4 6.0 13.0 18.5 11.9 4.7 22.6 4.7 27.3 1.8 37.2 1.7 47.6 0.8

96.5 0.0

95.6 0.1

94.9 0.1

93.9 0.0

96.0 0.0

Note: A
above; B
below.

9.4.2 RESS As already pointed out, for individuals enrolled in the RESS, covered earnings are very weakly related to true earnings. In particular, the selfemployed are free to choose their benefit base between an annual floor and a ceiling. Practically all of them choose the floor, as confirmed by table 9.11, which displays the fraction of self-employed contributing the minimum (censored from below) or the maximum (censored from above) for the years 1986 and 1995, respectively. This implies that there is no way in which true earnings for the self-employed can be recovered from the HLSS data set. We have therefore assumed that the earnings and the contributive profile coincide.5 Thus, we project (real) earnings given the observed profile of (real) contributions. 5. An alternative solution to this problem is to impute to self-employed an average earnings profile obtained from alternative sources (the recent European Community Household Panel [ECHP] constitutes an excellent example; see Peracchi 2002 for a description).

532

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

• Backward: wt–k–

ct–k , for

1, . . . , L • Forward: wtm
ct (1  g)m, for m
1, . . . , M with g
0.005 In other words, we assume that contributions were constant up to the first time they are observed, while they grow at a constant annual rate of 0.5 percent thereafter. It is important to recall, from section 9.2, that current Spanish legislation allows the self-employed to begin drawing retirement pensions without retiring, at least as long as they keep managing their own business. Hence, in the dynamic choice of the self-employed, the opportunity cost of retiring is not measured by the loss of future earnings but, instead, by the fact that contributions cannot longer be accumulated to increase future pensions and that marginal income taxes must be paid on pensions. This implies that, for the self-employed, maximization of the (net of taxes) SS payoff is a very reasonable objective function. 9.5 Evaluation of Social Security Incentives 9.5.1 Assumptions Made in the Computations For every male worker in the wage sample enrolled in either the RGSS or the RETA we assume that (a) he is married to a nonworking spouse, (b) his wife is three years younger, and (c) his mortality corresponds to the baseline male mortality from the most recent available life tables (INE 1995). For every female in the wage sample, we assume that (a) she is married to either a retiree or a worker entitled to retirement benefits, (b) her husband is four years older, and (c) her mortality is the baseline female mortality from the most recent available life tables (INE 1995). For both men and women we further assume that, starting at age fiftyfive and until a person reaches age sixty-five, there are three pathways to retirement: the UB52 program, DI benefits, and early retirement. At each age, an individual has an age-specific probability of entering retirement using any of these three programs. However, we must take into account the following restrictions. 1. No person has access to early retirement before age sixty 2. After age sixty, a person cannot claim UB52 and can only claim early retirement or DI benefits 3. A self-employed person enrolled in RETA can never claim UB52 benefits 9.5.2 Calculating Social Security (SS) Incentives For a worker of age a, we define social security wealth (SSW) in case of retirement at age h  a as the expected present value of future pension benefits S

SSWh


∑

s
h1

s Bs (h).

Micro-Modeling of Retirement Behavior in Spain

533

Here, S is the age of certain death, s
s–a s , with  denoting the pure time-discount factor, s is the conditional survival probability at age s for an individual alive at age a, and Bs (h) the pension expected at age s  h  1 in case of retirement at age h. Given SSW, we define three incentive variables for a worker of age a. 1. Social security accrual (SSA) is the difference in SSW from postponing retirement from age a to age a  1: S

∑

SSAa
SSWa1 SSWa


s [Bs (a1) Bs (a)] a1Ba1(a)

s
a2

The SSA is positive if the expected present value Σ Ss
a2 s [Bs (a1) – Bs (a)] of the increment in the flow of pension benefits is greater than the expected present value a1Ba1(a) of the pension benefit foregone by postponing retirement. If the increments Bs (a  1) – Bs (a) are small, as it is usually the case, then the SSA is negative. The rescaled negative accrual a
–SSA /Wa1, where Wa1 equals expected net earnings at age a  1 based on the information available up to age a, is called the implicit tax or subsidy on postponing retirement from age a to age a  1. 2. Peak value PVa
maxh(SSWh – SSWa ), h
a  1, . . . , R, where R is a mandatory retirement age (which, strictly speaking, does not exist in Spain; given the retirement evidence we find it reasonable to assume R equals 70). Thus, the peak value is the maximum difference in SSW between retiring at any future age and retiring at age a. 3. Option value OVa
maxh(Vh – Va ), h
a  1, . . . , R, where S

Va


∑

s [kBs (h)] 

s
a1

is the total expected utility of retiring at age a, and h

Vh


∑

S

s W s 

s
a1

∑

s [kBs (h)] 

s
h1

is the total expected utility of retiring at age h  a. Thus, the option value is the maximum utility difference between retiring at any future age and retiring at age a. We parameterize the model by assuming  equals 0.97,  equals 1, and k equals 1.25. Under our assumptions, Va equals 1.25 SSWa and h

Vh


∑

sWs  1.25SSWh .

s
a1

If expected earnings are constant at Wa (as assumed by our earnings model), then h

Vh Va
Wa

∑

s
a1

s  1.25(SSWh SSWa ),

534

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

Table 9.12

Unconditional Disability Take-Ups, by Regime, Sex, and Age (1985–94) RGSS

RETA

Age

Male

Female

Male

Female

55 56 57 58 59

1.26 1.40 1.41 1.60 1.61

0.87 1.17 1.34 1.40 1.15

0.83 1.21 1.31 1.56 1.64

0.75 0.92 1.08 1.29 1.43

60 61 62 63 64

1.92 2.00 1.96 1.99 1.61

1.74 2.21 1.89 2.18 2.81

2.22 2.08 2.32 2.27 2.75

1.65 1.29 2.05 2.10 2.01

65

1.10

1.87

1.15

1.88

66 67 68 69 70

1.61 1.34 1.65 2.21 2.25

1.08 3.10 2.14 3.61 0.00

1.65 2.66 1.26 1.69 0.81

2.39 2.20 1.37 2.26 1.65

that is, the peak value and the option value are proportional to each other except for the effect due to the term Σhs
a1 s . The restrictions embodied in the fourth assumption require us to combine the incentive measures Ij from the various programs ( j
UB, DI, R, where UB denotes unemployment benefits; DI, disability benefits; and R, the retirement programs) as follows:



DI IDI p DI if 55  a  60 a  IUB(1 p a ) DI DI I
IDI p a  IR(1 pa ) if 60  a  65 I if 65  a,

where p DI a denotes the probability of observing a transition from employment into disability at age a. Since the self-employed have no access to UB52 benefits, the combined incentives from age fifty-five to age fiftynine for members of this group change to DI I
IDI p DI a  IR(1 p a ),

55  a  59.

We follow a regression-based approach to compute the unconditional probability of qualifying for a disability pension (see table 9.12 for summary statistics by regime, sex, and age).6 The model is estimated, separately 6. We decided to model the unconditional probability of qualifying for a disability pension because it is the option that best captures the tightness of the SS system in the concession of disability pensions.

Micro-Modeling of Retirement Behavior in Spain Table 9.13

SSW, Accrual, and Tax Incentive Measures (1985 system, 1995 sample; in 1995 US$) SSA N

Median SSW

P10

2,609 1,593 1,772 1,981 1,975 1,734 1,166 1,063 969 717 512 12

95,311 95,005 95,980 100,033 104,421 112,619 126,567 130,285 134,383 134,735 131,576 117,295

–3,538 –3,553 –3,525 –3,526 –3,513 –4,527 –4,449 –4,462 –4,464 –5,023 –14,917 –10,582

55 563 56 414 57 416 58 430 59 467 60 422 61 374 62 346 63 299 64 283 65 219 69 15 (continued )

77,772 80,379 82,930 85,611 88,307 91,132 89,514 87,907 86,471 84,836 87,132 76,862

–24 –29 –40 –36 –43 –4,388 –4,381 –4,360 –4,371 –4,907 –4,967 –4,976

Age

55 56 57 58 59 60 61 62 63 64 65 69

535

Median

Median of Tax P90

SD

Sample

Simulated

RGSS Male Sample –3,265 15,466 –2,120 15,674 1,420 13,418 2,363 13,282 3,507 13,552 5,910 13,619 6,559 12,995 5,289 11,806 3,876 10,090 2,797 9,321 –5,437 –879 –5,174 –1,721

7,231 7,626 7,294 7,334 7,369 9,102 8,385 8,098 7,111 7,046 7,622 4,312

26.8 19.3 –10.9 –23.6 –36.0 –47.4 –50.4 –42.2 –33.9 –25.8 61.9 60.4

21.6 10.8 15.3 36.2 28.6 –14.9 –12.0 –11.0 4.6 16.0 77.5 70.0

RETA Male Sample –14 –3 –22 –13 –26 –12 –33 4 –38 67 –4,345 –4,178 –4,308 7,726 –3,925 8,339 –1,279 3,529 2,610 5,006 –2,277 –971 –2,791 –2,791

1,670 1,320 2,446 2,269 2,165 4,598 6,859 5,616 5,769 3,995 2,229 1,053

0.3 0.4 0.5 0.6 0.7 83.9 83.1 75.7 24.7 –50.4 43.9 53.9

41.6 40.1 39.0 37.7 35.3 106.5 94.5 48.7 15.3 26.8 100.7 95.2

by sex and regime, using the data from the HLSS for the period 1985–1994. The set of regressors include age and region dummies and a cubic time trend for all the regimes. For people in RGSS we also consider industry and group of contribution dummies. 9.5.3 Results under the 1985 System Tables 9.13 and 9.14 present the estimates of SS incentives by age (omitting ages sixty-six to sixty-eight) for the combined set of options (UB, DI, or R) described earlier. Incentives are computed separately by sex, and earnings projections are based on the methodology and the assumptions described in section 9.4.1. Table 9.13 presents median values for SSW, SSA, and the implicit tax or subsidy to work, as well as the first and ninth decile and the standard deviation of the accrual. For comparison purposes, the column labeled “simulated” reports the age profiles of the implicit tax constructed for synthetic

536

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi (continued)

Table 9.13

SSA Age

N

Median SSW

55 56 57 58 59 60 61 62 63 64 65 69

569 346 375 445 409 381 311 294 276 194 167 14

75,376 74,441 73,605 79,449 82,798 83,095 98,829 96,258 96,240 104,197 95,158 65,969

–3,469 –3,469 –3,479 –3,466 –3,457 –18,703 –18,666 –3,803 –3,797 –4,300 –9,839 –9,099

55 56 57 58 59 60 61 62 63 64 65 69

240 158 168 182 209 207 165 137 177 122 110 28

59,685 61,714 63,756 65,883 68,081 70,532 68,662 67,369 65,788 64,202 66,026 58,198

–32 –25 –46 –45 –53 –3,782 –18,358 –3,745 –3,745 –4,193 –4,214 –2,255

P10

Median

Median of Tax P90

SD

Sample

Simulated

RGSS Female Sample –3,440 8,564 –3,430 8,285 –3,404 8,268 –3,262 9,245 –2,739 9,692 –1,305 13,096 5,388 13,557 5,501 12,800 5,226 10,295 3,858 9,749 –3,340 1,075 –4,003 –2,358

4,953 5,083 5,159 5,661 5,615 12,674 14,004 10,339 9,034 12,840 6,041 2,212

37.9 37.2 32.1 31.5 23.6 14.7 –52.8 –54.6 –54.1 –41.5 48.5 65.7

43.2 41.9 40.9 39.5 29.4 –4.3 –55.9 –27.2 –8.9 –0.3 78.7 69.9

RETA Female Sample –5 –2 –8 –3 –12 –5 –17 –7 –23 –10 –3,698 –3,107 –3,722 –858 –3,425 3,918 –2,289 5,866 1,932 5,466 –56 164 –742 –552

1,318 1,494 1,039 885 1,158 7,198 11,749 10,440 7,640 9,036 4,571 2,261

0.1 0.2 0.2 0.3 0.4 71.4 71.8 66.1 44.2 –37.3 1.1 14.3

41.8 40.5 39.7 38.5 36.4 97.1 92.8 81.4 –23.5 –14.6 68.3 76.4

Note: N = number of observations; P10
tenth percentile; P90
ninetieth percentile; SD = standard deviation.

individuals using the criteria described in Boldrin, Jiménez-Martín, and Peracchi (1999). More precisely, we consider the following cases. 1. Male in RGGS: base case as in Boldrin, Jiménez-Martín, and Peracchi (1999, 338) 2. Male in RETA: same as above, but with thirty-two years of contributions at age sixty and contributing to the minimum 3. Female in RGSS: twenty years of contributions at age sixty, without dependent spouse and receiving 60 percent of the sample average wage 4. Female in RETA: same as above, but with twenty-two years of contribution at age sixty, without a dependent spouse, and always having contributed to the minimum For men in the RGSS, the SSW starts off at $95,311 (1995 exchange rate) and peaks between sixty-three and sixty-four years of age at $134,735. The

Micro-Modeling of Retirement Behavior in Spain Table 9.14

Peak and Option Value Incentive Measures (1985 System, in 1995 US$) Peak Value

Age

P90

Option Value

N

P10

P50

2,609 1,563 1,772 1,981 1,975 1,734 1,166 1,063 969 717 512 12

–3,529 –3,546 –3,518 –3,525 –3,513 –4,508 –4,439 –4,453 –4,458 –5,023 –14,710 –10,582

18,384 18,228 19,871 15,379 13,827 13,384 12,949 9,800 6,511 2,806 –5,388 –5,174

RGSS Male Sample 86,697 35,220 75,456 31,228 60,505 25,333 50,096 21,085 41,647 18,759 35,447 16,834 28,876 14,286 23,764 12,001 17,902 9,459 9,798 7,815 –687 7,745 –1,721 4,312

55 563 56 414 57 416 58 430 59 467 60 422 61 374 62 346 63 299 64 283 65 219 69 15 (continued )

–4 –20 –29 –35 –42 –4,388 –4,378 –4,360 –4,365 –4,907 –4,967 –4,976

1 –14 –25 –33 –38 –4,345 –4,307 –1,112 1,312 2,610 –2,277 –2,791

55 56 57 58 59 60 61 62 63 64 65 69

537

SD

P10

P50

P90

SD

3,319 2,690 2,049 1,382 702 0 4,023 3,337 593 0 0 0

124,264 111,068 103,772 85,871 73,011 65,032 58,829 46,348 33,269 19,916 3,567 2,432

311,883 284,558 245,406 215,663 187,562 162,590 143,823 116,935 82,473 67,644 39,548 12,755

119,742 109,823 91,044 80,676 70,260 59,415 49,550 40,827 31,124 29,071 21,755 7,527

RETA Male Sample 48 12,234 31,203 6 9,960 23,017 217 12,207 21,707 964 10,878 14,391 8,023 10,760 10,433 2,920 10,031 5,272 19,124 13,007 5,460 18,030 9,728 4,308 6,235 7,506 2,058 5,006 4,371 2,693 –971 2,365 90 –2,791 1,053 34

38,383 33,906 29,294 24,528 19,715 14,216 13,947 13,413 11,824 10,449 4,184 1,693

40,927 34,129 35,741 34,185 45,954 31,141 52,692 49,320 23,436 20,114 8,863 1,693

28,280 23,042 27,015 23,836 23,273 18,699 22,595 18,960 14,432 10,981 6,352 889

tenth percentile of the accrual is negative at all ages. The median accrual is negative until age fifty-six, becomes positive between fifty-seven and sixtyfour, and then negative at older ages. Notice that, except after age sixtyfive, there is little agreement between our median or average tax rate and the simulated base case in Boldrin, Jiménez-Martín, and Peracchi (1999). Part of this discrepancy is due to a technical correction in the set of assumptions made in the computation of incentives before normal retirement age. In Boldrin, Jiménez-Martín, and Peracchi (1999), we assumed that when a person stops working between age fifty-five and fifty-nine, their pension is computed considering earnings until that age, even if they start receiving the pension only at age sixty. In this present matter, for an individual aged between fifty-five and fifty-nine, we assume instead that

• They receive unemployment benefits until age sixty and retirement benefits thereafter; and

538

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi (continued)

Table 9.14

Peak Value Age

N

P10

P50

55 56 57 58 59 60 61 62 63 64 65 69

569 346 375 445 409 381 311 294 276 194 167 14

–3,460 –3,465 –3,464 –3,456 –3,451 –18,654 –18,585 –3,794 –3,790 –4,300 –9,135 –9,099

4,363 –745 12,030 11,121 16,531 13,871 13,840 11,547 9,757 3,858 –3,294 –4,003

55 56 57 58 59 60 61 62 63 64 65 69

240 158 168 182 209 207 165 137 177 122 110 28

–28 –25 –46 –45 –45 –3,769 –18,358 –3,745 –3,742 –4,193 –4,214 –2,255

180 –8 –12 –9 –23 –2,564 –2,423 2,411 2,805 2,011 –56 –742

P90

Option Value SD

P10

P50

P90

SD

RGSS Female Sample 64,805 28,777 1,283 59,539 26,634 4,007 49,515 22,512 0 49,837 21,830 0 41,702 19,911 781 39,334 21,953 0 36,439 22,330 1 28,191 16,147 0 19,173 12,009 294 10,509 12,910 0 1,586 7,266 0 –2,358 2,212 0

90,648 80,670 93,303 74,859 78,794 62,281 56,892 46,127 34,382 18,913 6,217 1,372

233,563 207,748 171,444 177,974 142,584 128,809 118,919 93,063 65,189 49,401 37,962 6,180

91,939 83,268 71,615 74,338 60,987 55,387 51,049 39,942 28,377 28,260 18,361 2,845

RETA Female Sample 8,934 10,398 36,029 8,756 11,312 31,578 8,431 7,212 26,115 8,117 7,047 10,167 7,810 6,295 5,273 7,104 12,086 12,670 9,862 14,296 113 11,686 13,087 10,576 10,293 9,122 8,647 5,466 9,106 10,341 164 4,646 6,422 –552 2,261 2,364

44,058 39,105 34,216 32,113 26,156 22,642 20,000 22,058 20,497 17,729 13,573 4,255

55,356 51,242 46,872 42,401 37,817 32,810 31,882 29,615 26,610 21,394 15,823 4,493

22,158 23,707 15,544 16,533 14,472 17,505 10,470 12,153 7,544 6,350 5,399 1,197

Note: See table 9.13.

• The pension is computed at age sixty. From age a to age sixty, the individual contributes the mandatory minimum level of contributions to their pension. For the median worker, this modification introduces incentives to keep working until the early retirement age. However, for the seventy-fifth percentile or higher, we still find strong incentives to stop working. For men in RETA, the SSW reaches a peak ($91,132) at sixty but is flat between fifty-nine and sixty-five. The median accrual is negative at all ages except age sixty-four, whereas the opposite occurs with the median implicit tax. The age-incentive profiles for women in the RGSS or the RETA are similar to those for men, although the median values of our incentive measure are higher than those for men in the age range sixty to sixty-four. This is because women have shorter careers and, on average (or at the median), do not qualify for a full pension in that age range. Table 9.14 presents the age profile of the median, tenth percentile, and

Micro-Modeling of Retirement Behavior in Spain

539

ninetieth percentile and the standard deviation of the peak and option values for men and women in the RGSS and the RETA. In all the cases considered, the peak value and the accrual (presented in the previous table) show very similar profiles. However, from age fifty-five to age sixty-four, the median peak value is much higher, thus reinforcing retention incentives in that age range. From age sixty-five, they are identical in practically all the cases. The option value of retiring starts at a very high level for individuals enrolled in the RGSS and decreases continuously with age. Note that the tenth percentile is close to zero at practically all ages, revealing strong retirement incentives for those people. For the individuals in the RETA, the fact that we have used the contributive profile to approximate earnings explains why the option value of retirement is very low at all ages compared to that of people from RGSS. Furthermore, the fact that most of the people enrolled in the RETA contribute the minimum amount explains why the tenth and ninetieth percentiles of the option value are very similar.7 9.6 Retirement Models for the Year 1995 This section investigates the explanatory power of our incentive measures (accrual, peak value, and option value) for retirement behavior. Before presenting the estimates of our model for the probability of retirement, we review the available sample evidence for the RGSS (including RMTC) and the RETA. 9.6.1 Sample Evidence Figure 9.5 shows the patterns of retirement in 1995. The top panels show the age profile of the exit rate from the labor force by sex and SS regime (RGSS and RETA on the top-left and top-right panel, respectively). For men in the RGSS, the age profile of exit rates shows two peaks, at age sixty and sixty-five (respectively, the early and normal retirement ages), whereas for women in the RGSS and both men and women in the RETA, only the peak at age sixty-five is evident. The bottom-left panel plots, for those enrolled in the RGSS, the exit rate from the labor force at the early retirement age of sixty against the quantiles of expected earnings at the same age (in 1995 pesetas). Who is leaving the labor force at age sixty? The answer, especially for men, is clear from the figure: those with relatively low wages, in particular, those with wages below the twenty-fifth percentile. As shown in Jiménez-Martín and Sánchez (2000), the main cause of retirement for this group is the interaction between age, the penalties for insufficient contributions, and the minimum pension provision. In addition, it can be shown that exit rates from the labor force for women with relatively low earnings are 7. We must also take into account our lack of information about true earnings for people enrolle d in RETA. To compute the option value, we instead used information on contributions.

Fig. 9.5

A

Retirement patterns by sex, age, and expected wage, 1995

Fig. 9.5

B

(cont.)

542

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

already nonnegligible at age fifty-five. Finally, the right-side panel plots the exit rate at the normal retirement age of sixty-five. It is evident that exit rates at this age are largely independent of expected wages. 9.6.2 Retirement Models We follow a regression-based approach to model the effect of SSW, incentive measure (either accrual, peak value, or option value), and individual demographic characteristics on the decision to retire in year 1995 conditional on being active at the end of 1994. Retirement probabilities are assumed to have the probit form Pr(R i
1)
(1SSWi  2 Ii  3 X i ), where R is a binary indicator of retirement,  is the distribution function of a standard normal, I denotes the incentive measure, and X is a vector of predictors that include individual earnings and sociodemographic characteristics.8 For each incentive measure we present, separately by sex and SS regime, the results obtained for the following specifications.

• M1 is basic specification and includes three sets of predictors. M1A includes the incentive measure (accrual, peak value, or option value), an eligibility dummy for attainment of a minimum of fifteen years of contributions, and three industry-specific variables—the fraction of collective wage settlements having a clause favoring early retirement, the presence of rules permitting retirement at age sixty-four without any age penalty, and the existence of mandatory retirement at age sixty-five (see the appendix for a brief description of the data source). M1B includes a linear age trend, the length of the current employment spell and its square, the number of years of contribution and its square, the number of years of potential experience, dummies for schooling level and the contributive group (only for people in the RGSS), and dummies for part-time work and the sector of occupation (only for people in the RGSS). M1C includes controls for earnings (expected wage, pension, and average lifetime income and their squares) and the net present value of expected wages until the year in which either the peak value or the option value reach their maximum. • M2 is the same as M1 but age dummies replace the linear age trend. In table 9.15, we present the main results obtained by fitting our two models to the observed transitions between 1994 and 1995. We show, for 8. The socioeconomic and earnings information is richer for the RGSS. Results for the RETA should be taken with caution.

.00344 .00128 .00033 .00012 –.00906 .00430 –.00088 .00042 –1.642 .50046 .336 –3,791

.00970 .00325 .00090 .00030 –.0092 .00710 –.00086 .00066 –.4766 .64579 .327 –897.8

SSW Marginal effect Incentive Marginal effect Constant R2 Log-likelihood (continued )

Coefficient

SE

Coefficient

M2

Male RGSS (16,191 observations) .00152 .00871 .00149 .01387 .00012 .00087 .00015 .00136 .00489 .00147 .00245 .00448 .00046 .00015 .00024 .00044 .53053 –1.495 .49230 –1.273 .341 .380 –3,766 –3,544

SE

M1

Peak Value

.00170 .00017 .00254 .00025 .52863

SE

Female RGSS (3,852 observations) .01812 .00419 .01138 .00345 .02022 .00438 .00162 .00038 .00107 .00033 .00185,00040 –.00580 .00755 .00135 .00490 .00393 .00527 –.00053 .00068 .00013 .00046 .00036 .00048 –.2204 .74217 –.3112 .64244 –.2072 .74880 .355 .327 .356 –.860.1 –897.7 –858.5

.00749 .00071 –.00130 –.00012 –1.197 .373 –3,579

Coefficient

Coefficient

SE

M2

M1

Accrual

Probit Models of the 1995 Retirement Rates

SSW Marginal effect Incentive Marginal effect Constant R2 Log-likelihood

Table 9.15

.00165 .00017 .00111 .00011 .49665

SE

.01176 .00381 .00111 .00036 .00247 .00202 .00023 .00019 –.3301 .64892 .326 –897.9

.01080 .00109 .00884 .00089 –1.360 .342 –3,758

Coefficient

M1

M2

.02175 .00199 .00361 .00033 –.3375 .356 –858.5

.01627 .00161 .01032 .00102 –1.262 .381 –3,534

Coefficient

Option Value

.00477 .00044 .00210 .00019 .75922

.00186 .00018 .00115 .00011 .53657

SE

.00316 .00643 .00047 .00095 .01813 .01096 .00268 .00162 –3.358 3.5687 .142 –638.5

SSW Marginal effect Incentive Marginal effect Constant R2 Log-likelihood

Note: SE
standard error.

.00870 .00496 .00117 .00067 –.04703 .01212 –.00630 .00162 –2.079 .68022 .168 –1201.

SE

SE

Coefficient

M2 SE

Male RETA (4,355 observations) –.00068 .00695 .00992 .01238 –.00009 .00092 .00131 .00163 –.02915 .00900 .01432 .01056 –.00385 .00119 .00188 .00139 –1.848 .72708 –1.6444 1.2819 .166 .253 –1203. –1078.

Coefficient

M1

Peak Value

Female RETA (2,051 observations) –.00176 .01113 .00188 .00732 –.00248 .01119 –.00025 .00156 .00018 .00108 –.00035 .00157 .02538 .01207 .00849 .00979 .01824 .01039 .00355 .00169 .00126 .00145 .00256 .00146 –3.678 3.7786 –3.175 3.5836 –2.457 3.8070 .197 .141 .196 –597.9 –639.4 –598.5

.00726 .01174 .00096 .00155 .01050 .01440 .00138 .00190 –1.542 1.2772 .252 –1079.

Coefficient

Coefficient

SE

M2

Accrual

M1

(continued)

SSW Marginal effect Incentive Marginal effect Constant R2 Log-likelihood

Table 9.15

.00938 .00124 .00729 .00097 .70436

SE

.00342 .01334 .00051 .00199 .00241 .01448 .00036 .00215 –3.259 3.6326 .140 –639.8

.00757 .00100 –.00920 –.00122 –2.107 .167 –1202.

Coefficient

M1

M2 SE

–.01475 .01781 –.00207 .00250 .00739 .01736 .00104 .00244 –1.876 3.9571 .195 –598.9

.00501 .01451 .00066 .00191 .00187 .00758 .00025 .00100 –1.324 1.283 .253 –1079.

Coefficient

Option Value

Micro-Modeling of Retirement Behavior in Spain

545

each combination of sex and regime, the estimates of the probit coefficients, their estimated standard errors, and the implied probability effect. Complete definitions, data sources, as well as summary statistics for all variables employed are presented in the appendix. Since we report the results from a large number of models, we concentrate on the variables of interest. The complete set of results is available from the authors upon request. On the one hand, we find that the basic specification with only demographic and earnings controls (M1) explains, in the case of the RGSS, an important fraction of the retirement peaks at the early and normal retirement ages. In contrast, this specification seems to be unable to capture the retirement peak at age sixty-five for workers in the RETA. This is partly due to the fact that the socioeconomic information for individuals enrolled in the RETA is poorer than for people enrolled in the RGSS. The SSW term is positive and significant in all cases. Contradictory results are obtained instead for the incentive variable. In fact, while the accrual usually shows the expected (negative) sign, both the peak and the option value show the wrong (positive) sign. Alternatively, neither SSW nor the incentive variables are significant for people enrolled in RETA, indicating that the SSW and the financial variables do not capture retirement incentives for individual enrolled in RETA. On the other hand, the introduction of age dummies (specification M2) always increases the coefficient of both the SSW and the incentive variables and substantially improves the fit of the model. For men in the RGSS, for example, the pseudo
R 2 for the model with the accrual as the incentive variable goes from 33.6 percent (model M1) to 37.3 percent (model M2). For men in RETA, the pseudo-R 2 goes from 16.8 to 37.3 percent. The pattern for the other incentive variables (accrual, peak value, or option value) is very similar. Figure 9.6 compares the age profile of the empirical hazard rate with those of the coefficients on the age dummies for the three versions of model M2 (that is, with the accrual, the peak value, and the option value as the incentive variable), estimated separately by sex and SS regime. In all cases, the age dummies have been rescaled to the empirical hazard scale. For males enrolled in the RGSS (left panels) and all three models (either accrual, peak value, or option value), the profile of the age dummies resembles the hazard profile, although there are some discrepancies between the two profiles at ages sixty to sixty-four and from age sixty-seven onward. For females enrolled in the RGSS, the discrepancies are more evident at all ages from fifty-five to sixty-four, with age fifty-nine as an exception. For individuals enrolled in RETA, the profiles are quite different from those estimated for the RGSS. In particular, important discrepancies between ages fifty-five and fifty-nine and between models are detected.

Fig. 9.6

Evaluation of the explanatory power of incentive measures in M2: Empirical hazard versus age effects (rescaled)

Micro-Modeling of Retirement Behavior in Spain

547

9.7 Policy Simulations In this section we use our estimates to simulate retirement behavior under alternative institutional settings. 9.7.1 Description of the Simulations In the simulations we consider three policies, of which the third is specific to the Spanish case. R1: A reform of the existing system consisting of a three-year increase in both the early and the normal retirement age, while keeping all other aspects of the Spanish SS system unchanged. R2: A reform common to all countries considered in this volume including (a) early entitlement age at sixty; (b) normal retirement age at sixty-five; (c) a replacement rate at age sixty-five equal to sixty percent of the earnings at age fifty-nine; and (d) an actuarial adjustment of 3.4 percent per year from age sixty to age seventy (this implies a replacement rate of 42 percent at age sixty and 78 percent at age seventy). Notice that (a) and (b) correspond to the current Spanish system, whereas the actuarial adjustment for retirement before age sixty-five is less favorable than the one currently used in Spain. The current Spanish system is more generous for retirement at age sixty-five and has no actuarial adjustment for postponing retirement after that age. R97: The regime created by the 1997 Spanish reform and currently in place. We recall that the 1997 reform, described in section 9.2, implies the following changes in the basic benefit formula and in the penalties related to age and contributive history: (a) the number of years of contribution used to construct the benefit base is increased from eight, as prescribed by the 1985 legislation, to fifteen; (b) workers retiring after the age of sixty with forty or more contributive years are charged an actuarial adjustment of only 7 percent (instead of 8 percent) for each year under age sixty-five; and (c) the penalty for insufficient contributions, expressed as a fraction of BR to be received (see section 9.2), is changed to n




0 .5  .03(n 15) .8  .02(n 25) 1

if n  15 if 15  n  25 if 25  n  35 if 35  n.

For each of the three policies we carry out three simulation exercises. S1: Starts from the model without age dummies (M1), we modify the SSW and incentive measures in accordance with the new policy. Specifically, in the calculation of SSW, we increase by three years the early and the

548

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

normal retirement ages and shift by three years the age-specific probability of receiving DI or UI benefits. S2: Starts from the model with age dummies (M2), we modify the SSW and incentive measures according to the assumed policy changes. We also change the probabilities of receiving DI benefits, as in S1, but not the coefficients on the age dummies. S3: Applies only for reform R1 and starts from the model with age dummies (M2). In addition to the changes described in S2, we also shift the coefficients on the age dummies by three years, so that the entire ageprofile of the retirement hazard shifts forward by three years. 9.7.2 Results for Male Workers in the RGSS Figures 9.7 to 9.15 show the simulated retirement probabilities by age for male workers in the RGSS. The results are presented separately for each combination of simulation (S1 to S3) and incentive variable (accrual, peak value, and option value). Each graph presents both the age profile of estimated conditional retirement probabilities or the “hazards,” and the cumulative distribution function (CDF) of retirement age.9 In general, because all incentive measures explain little of the variation in retirement ages across individuals, it is hard to detect the impact of changes in the incentive measures on individual retirement behavior. When the coefficients estimated under the specification M1 are employed in the simulations, all the reforms reduce retirement hazard at ages sixty and age sixty-five. Although in some cases an increase in the hazard at ages sixty-three or sixty-four is observed, in general, the CDF of retirement age is shifted to the right. The reduction of the hazard at sixty is more important for the Spanish reform (R97), while the reduction at age sixtyfive is more important for the common reform (R2). When, instead, the coefficients estimated under the specification M2 are used in the simulations but the age dummies coefficients are not shifted (S2), only the R2 reform seems to sensibly reduce the hazard at the key age range and thereby reduce the CDF at, for example, age sixty-five. In fact, both the R1 and the R97 reform mildly increase the average retirement hazard at age sixty. This appears to be largely a consequence of the minimum pension rules in effect in Spain. Apart from this, the reduction is more important when retirement incentives are measured by the option value. As expected, when the age dummies are shifted by three years (S3; figures 9.13 to 9.15), the whole hazard for R1 shifts to the right by three years, and consequently, the CDF is reduced substantially both at age sixty (by 50, 39, and 37 percent for the accrual, peak value, and option value specifications, respectively) and at age sixty-five (36, 30, and 28 percent for the accrual, peak value, and option value specifications, respectively). 9. The CDF F(a) at age a is obtained from the conditional retirement probabilities h(a) through the recursion F(a)
F(a – 1)  [1 – F(a – 1) h(a)], starting from F(54)
0.

Fig. 9.7

Male workers in the RGSS: S1, accrual

Fig. 9.8

Male workers in the RGSS: S1, peak value

Fig. 9.9

Male workers in the RGSS: S1, option value

Fig. 9.10

Male workers in the RGSS: S2, accrual

Fig. 9.11

Male workers in the RGSS: S2, peak value

Fig. 9.12

Male workers in the RGSS: S2, option value

Fig. 9.13

Male workers in the RGSS: S3, accrual

Fig. 9.14

Male workers in the RGSS: S3, peak value

Fig. 9.15

Male workers in the RGSS: S3, option value

558

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

9.7.3 Results for Female Workers in the RGSS Figures 9.16 to 9.18 present the simulated retirement probabilities by age for female workers in the RGSS. Results are presented separately for each simulation exercise S1 to S3. When the coefficients estimated under the specification M1 are employed, only the R2 reform reduces the hazard of retirement at both ages sixty and sixty-five (figure 9.16). The reduction is similar for all incentive measures (about 35 and 37 percent at ages sixty and sixty-five, respectively). As a consequence, the CDF of retirement age is substantially lower at age sixty-five than in the sample. The reduction is slightly bigger under the peak or the option value (17 percent) than under the accrual (16 percent). On the other hand, both the R1 and R97 reforms are ineffective in reducing the retirement hazard in the relevant age range. Very similar results are obtained under S2. In S3, on the other hand, the results for R1 change substantially (figure 9.18). The CDF at age sixty reduces by 22.5 percent in all cases and by 33.6, 35.1, and 35.6 percent at age sixty-five for accrual, peak value, and option value, respectively. 9.7.4 Results for Individuals in the RETA In figures 9.19 to 9.22, we present the simulated retirement probabilities by age for the self-employed or individuals enrolled in the RETA. However, in this case, we do not report any result based on model M1, since this model is unable to capture the retirement peaks at sixty and sixty-five that are evident in the data. Thus, in each figure, we present results for combinations of sex and simulation exercise (S2 and S3) jointly for all the incentive variables (accrual, peak value, and option value). When the coefficients estimated under the specification M2 are used in the simulations but the age-dummies coefficients are not shifted (S2), the R2 reform seems to considerably reduce the hazard over the key age range and thereby reduce the CDF at, for example, age sixty-five. In fact, the Spanish reform (R97) substantially increases the retirement hazard at all ages below the normal retirement age (see figures 9.19 and 9.21 for men and women, respectively). When the age dummies are shifted by three years (S3), the whole hazard for R1 shifts three years toward the right, and consequently, the CDF of retirement age is reduced substantially in all cases (see figures 9.20 and 9.22 for men and women, respectively). The reduction of the CDF at ages sixty and sixty-five for men is much more important when the peak value is employed than it is for either the accrual or the option value. For example, the reduction for the accrual specification is 19.5 and 34.9 percent for men at ages sixty and sixty-five, while for the peak value, it reaches 28.1 and 38.7 percent, and under the option value, it lowers to 10.7 and 31.4 percent. For women, instead, the reduction is more

Fig. 9.16

Female workers in the RGSS: S1

Fig. 9.17

Female workers in the RGSS: S2

Fig. 9.18

Female workers in the RGSS: S3

Fig. 9.19

Male workers in the RETA: S2

Fig. 9.20

Male workers in the RETA: S3

Fig. 9.21

Female workers in the RETA: S2

Fig. 9.22

Female workers in the RETA: S3

566

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

Table 9.16

Average Retirement Ages from Simulated Policies Male (observed 62.00) R1

R2

Female (observed 62.31)

R97

R1

R2

R97

RGSS S1 Accrual Peak value Option value S2 Accrual Peak value Option value S3 Accrual Peak value Option value

62.19 61.58 61.41

61.87 62.74 62.74

61.98 61.99 62.10

62.90 62.53 62.51

63.00 63.34 63.27

62.26 62.31 62.30

61.98 61.39 61.15

61.81 62.61 62.54

61.92 61.90 61.98

62.70 62.24 62.12

63.04 63.39 63.33

62.18 62.22 62.20

63.58 62.96 62.69

— — —

— — —

64.00 63.54 63.49

— — —

— — —

Male (observed 62.88) R1

R2

Female (observed 62.35)

R97

R1

R2

R97

RETA S2 Accrual Peak value Option value S3 Accrual Peak value Option value

63.40 63.51 63.69

63.12 62.99 62.81

63.03 63.03 63.06

62.27 62.56 61.96

62.58 63.00 62.79

62.47 62.63 62.63

65.38 65.59 65.32

— — —

— — —

64.75 64.38 63.53

— — —

— — —

Note: Dashes indicate that data are not available.

important for the accrual specification (55.4 and 53.7 percent at ages sixty and sixty-five, respectively) than for either the peak or the option value specifications. 9.7.5 Effects of Reforms on Average Retirement Ages In table 9.16, we summarize the impact of the proposed reforms on the average retirement age.10 In almost all cases, the impact of the reform varies considerably across both simulation exercises and choice of the incentive measures (accrual, peak value, and option value). An exception is the 1997 Spanish reform, and its effect on the average retirement age is very small in 10. The average retirement age is obtained as a
Σ70 a
55 a f (a), where f (a)
F(a – 1) is the unconditional probability of retiring at age a.

Micro-Modeling of Retirement Behavior in Spain

567

general and if anything negative, thus confirming previous evaluations in Jiménez-Martín and Sánchez (2000) or Jiménez-Martín (1999). For male employees in the RGSS (top-left panel of table 9.16) the impact of R1 and R2 on the average retirement age is unclear. However, we detect the following regularities. First, R1 has little impact in S1 or S2. Second, R1 in S3 implies an increase of the average retirement age between 0.69 years (option value specification) and 1.58 years (accrual specification). Finally, R2 has a similar impact in either simulation, the impact for the accrual specification being much smaller than for either the peak or the option value. For female employees in the RGSS (right top panel of table 9.16), the effect is much more consistent across specifications of the incentive variable. Again, for S1 and S2, we are able to show a significant increase of the average retirement age (between 0.70 and 1.0 years) only for R2, depending on the specification of the incentive variable. The R1 reform visibly increases the average retirement age in S3 (between 1.2 and 1.7 years) in all cases. For a self-employed male in the RETA (bottom-left panel of table 9.16), the results are very different than those obtained for males in the RGSS, since only R1 is able to significantly increase the average retirement age in either S2 or S3 (between 0.46 and 1.37).11 The R2 reform, instead, reduces the average retirement age in all cases. Finally, for females in the RETA (bottom-right panel of table 9.16), the results vary sharply across incentive specifications. In S2, only R2 slightly increases retirement age. The effect of R1 is substantial only in S3, but, in the case of female self-employed, its effect varies considerably across specifications of the incentive variable. 9.8 Final Remarks Summarizing the large number of findings reported in this study is not an easy task. We will therefore limit ourselves to the most important results, paying special attention to those that appear to have at least potential implications for actual policy. The first important result is that, while economic and financial measures of retirement incentives can go a long way to explain and quantify retirement behavior, a substantial portion of the latter still remains unexplained. Various specifications of the basic model do decently well for workers enrolled in the general regime (RGSS) but rather poorly for the self-employed regime (RETA). This may be attributable, on the one hand, to the poor quality of the socioeconomic information available and, on the other hand, to the amply discussed unreliability of the earnings re11. Again, we do not present results of S1 for the self-employed.

568

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

ported by self-employed Spanish workers. This makes the evaluation of true expected earnings and opportunity costs of retiring most difficult, if not impossible. Of the three quantitative indicators tested, SSW seems to perform uniformly better. This is somewhat comforting, as it is the simplest measure of forward-looking behavior, the easiest to compute, and, quite likely, also the most reliable, given the available data. The great relevance of age dummies suggests that institutional factors and coordination incentives play a major role in determining workers retirement decisions. This lends credibility to the view that a proper mix of economic incentives and institutionally mandated constraints may provide the most effective way to modify and, in the light of the increasing life expectancy, push forward in time retirement decisions. Nevertheless, further and more detailed analysis appears to be necessary in order to properly and safely design much-needed reforms. Because the financial incentives explain only a small portion of the variation in retirement ages across individuals, it is hard to detect the impact of changes in the incentive measures on individual retirement behavior. This view, which calls for additional investigation, is confirmed by the results of our policy simulations. None of the three reforms considered predicts a major shift of the distribution of retirement ages in the desired forward direction. In particular and, in our view, most importantly (given that this reform has recently been implemented in Spain), we confirm that R97 has very little impact on retirement incentives and, consequently, on the average retirement age. In fact, as predicted by earlier and much simpler studies, it may even shift the distribution of retirement ages in the wrong direction, especially for low earners. It seems hard to rank the other two stylized reforms, at least in the light of our findings. Both of them seem to move the average retirement age in the right direction, but only by fairly small amounts (less than two full years). The R1 reform, definitely the simplest, tends to perform better than the R2 reform, but this result is not uniform across regimes and sexes. On the other hand, R2 (which is designed to be common to all countries considered in this volume) modifies current Spanish legislation only slightly. As pointed out above, the early and normal retirement ages selected by R2 coincide with those already in place in Spain. The only difference is the actuarial adjustment for retirement before age seventy, which is considerably less favorable than the one currently used in Spain. Our finding that even a relatively sizeable reduction in SSW would increase the average retirement age by only about one-and-a-half year is a direct consequence of the fact that SSW explains much less than half of the total variability in Spanish retirement age. Uncovering the socioeconomic factors explaining the residual half of such variability is therefore crucial for designing an effective reform.

Micro-Modeling of Retirement Behavior in Spain

569

Appendix From Covered Earnings to Earnings; Workers in the RGSS The relationship between earnings wt, covered earnings (base de cotización) et, and benefit base (base reguladora) Bt follows. The value of et is a doubly censored version of wt , specifically, et
max[
t , min(wt , u t )], where
t and u t are lower and upper ceilings mandated annually, while Bt is a weighted average of current and past covered earnings Bt
∑ bj et j j0

for a suitable set of weights bj . For each person i in our sample, we observe covered earnings eit at each year during the period 1986–1995 in which the person works. From these data, we have to compute Bit using formula (1) and impute wit , which is not observed whenever it exceeds ut or falls below
t . Clearly, imputation is only needed for those observations such that eit

t or eit
u t , and not for the others that are fully observed. To impute wit , we proceed as follows. We assume that the marginal distribution of the logarithm of earnings, ln wit , is normal with mean it
t  t⊥Xit and variance 2t , where Xit is a vector of observable individual characteristics. The model parameters t , , and 2t are then estimated using maximum likelihood. To perform these calculations, we neglect left censoring. When eit
ut (the observation is right censored), a naive imputation of wit is ˆ ⊥t Xit , wˆit
 ˆ it
 ˆ with the estimated mean of wit under the tobit model. Since we know that wit is at least equal to ut , a better approach is to use instead the estimated conditional mean of wit , given that wit  ut , specifically, wˆit
 ˆ it   ˆ (ct )  εˆit , where ct
(ut –  ˆ it )/ ˆ , (ct )
(ct )/ [1 – (ct )], and ( ) and ( ) denote, respectively, the density and the distribution function of a standard normal. Replacing in the original data set the censored values eit with the imputations wˆit gives a set of “completed data” that may be treated (to a first approximation) as the true earnings. With the complete data we may estimate a fixed-effects model for the level of earnings, use the estimates from this model to project earnings for-

570

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

ward and backward; and use projected earnings to compute projected benefit base (necessary for computing projected payroll taxes and projected pension amounts) and the implicit tax for working one additional year. Unemployment Benefits The Spanish Social Protection system provides contributory and noncontributory coverage against unemployment spells through the Instituto Nacional Empleo (INEM). Contributory Unemployment Benefits A program exists protecting employees against a nonvoluntary unemployment spell. Duration of benefits ranges from 120 to 720 days, increasing at the rate of 120 days per year of contribution within the previous six years. The amount of benefits is a function of the benefit base, which is the average of the contributive bases during the 180 days preceding the unemployment spell. The minimum benefit amount in 1999 was PTA 69,611, or $405 (PTA 150) equals $1). The maximum benefit amount is a function of the number of dependent children. Without children, it equals PTA 137,395 or $916. With two or more children, it equals PTA 177,793, or $1,185. Unemployment benefits are subject to both SS contributions and income taxes. Subsidy for Workers Aged Fifty-Two and Older A special unemployment scheme exists for those workers fifty-two and older who (a) are otherwise eligible for a retirement pension, except for their age and (b) have an income below 75 percent of the monthly minimum wage, which is PTA 51,952. The benefit amounts to 75 percent of the monthly minimum wage. It can be collected until the person reaches a retirement age, either early or normal. Disability Pensions The SS system provides insurance against both temporary and permanent illness or disability. Temporary Illness or Disability The subsidy for temporary illness or disability (incapacidad laboral transitoria) was not regulated by the 1985 reform, and its terms of provision have undergone frequent changes. Eligibility requires affiliation to the SS system for a minimum period that depends upon the nature of the covered risk. Common illness requires only 180 days of contributions during the last five years, and paid maternity or paternity leave requires at least nine months before the date of delivery and 180 days during the last twelve months, whereas no minimum eligibility criterion is imposed for work-related accidents or illnesses.

Micro-Modeling of Retirement Behavior in Spain

571

The benefit base depends on actual earnings during the last twelve months. In case of a common illness or a non-work-related accident, the subsidy is equal to 6 percent of the benefit base for each day of absence between the fourth and the twentieth and to 75 percent of the benefit base afterwards until the maximum period is reached. It is always equal to 75 percent in case of work-related accident or illness and in case of maternity or paternity leave (only one of the parents being allowed to use the subsidy for each child). The maximum period for which the subsidy can be received is eighteen months, after which the worker must either return to work or be classified as “permanently disabled.” Contributive Disability Pensions Permanent disability pensions have played an important role in allowing Spanish workers to retire at ages earlier than sixty. In particular, they have been used extensively during the late 1970s and early 1980s as an early retirement mechanism for workers in restructuring industries (shipbuilding, steel, mining, and so forth) or as substitutes for long-term unemployment subsidies in depressed regions. The total disability rate (as a percentage of the workforce) doubled in less than ten years, from about 0.7 percent in 1975 to 1.5 percent in 1983. The 1985 reform, by tightening the requirements, managed to bring the phenomenon under partial control. Disability rates have since decreased, stabilizing around 0.6 percent (see table 9A.1 for an illustration). Table 9A.1

Percentage Ratio Between the Number of Disability Pensions Paid and the Number of Workers Covered by the Various SS Programs, 1981–1994

Year

RGSS

RETA

REAa

REAb

REMC

RETM

REEH

Total

1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994

0.79 1.15 1.31 1.17 0.72 0.62 0.55 0.52 0.43 0.44 0.41 0.47 0.47 0.44

1.06 1.06 1.03 0.83 0.58 0.57 0.51 0.51 0.43 0.51 0.57 0.64 0.68 0.77

2.29 3.17 3.02 2.41 1.61 1.67 1.34 1.21 1.13 1.21 1.30 1.37 1.25 1.35

2.14 2.34 2.33 2.14 1.80 1.97 1.84 2.06 1.95 2.38 2.58 2.53 2.15 1.91

2.33 3.61 3.21 2.91 1.52 1.80 1.42 1.69 1.64 2.36 2.18 2.37 2.29 2.03

— — — — — 1.58 1.34 1.45 1.12 1.22 1.18 1.26 1.25 1.24

2.32 2.79 2.88 2.57 2.48 1.93 2.00 2.21 2.25 2.90 3.30 3.12 2.85 2.75

1.10 1.45 1.54 1.33 0.90 0.83 0.72 0.70 0.60 0.62 0.62 0.67 0.64 0.61

Notes: General Fund (RGSS); self-employed (RETA); agricultural employees (REAa); farmers (REAb); coal miners (REMC); sailors (RETM); domestic workers (REEH). Dashes indicate that data are not available.

572

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

Disability pensions are distinguished into contributory and noncontributory. We limit ourselves to the contributory pensions. Eligibility and pension amounts depend on the level of disability. The 1985 reform distinguished four levels of permanent disability characterized by increasing severity. Since then, the legislation has formally reduced them to three, but has also created a special subcase of the first level with the explicit purpose of using the disability funds to subsidize the dismissal of old workers from certain sectors or geographic areas. The first level (incapacidad permanente total para la profesión habitual; IPT) corresponds to inability to do the usual job. A special subcase (incapacidad permanente total cualificada para la profesión habitual; IPTC) applies only to employees older than fifty-five that are in particular socioeconomic situations. The second level (incapacidad permanente absoluta; IPA) corresponds to inability to do any kind of job. The third level (gran invalidez; GI) requires, in addition to the inability to perform work, continued attendance by other persons in order to carry out the basic vital functions. When disability is caused by an ordinary illness, eligibility to a pension requires from five to fifteen years of contributions, depending on the age when the person fell ill and the seriousness of the disability. There is no contributive requirement when the disability is caused by an accident (whether work related or not) or a professional illness. Table 9A.2 shows the fraction of new disability pensions awarded to individuals aged fifty-six and older by program and regime. Eligibility requirements are fairly complicated. We try here to streamline their presentation. In the cases of IPA or GI, fifteen years of contributions are required, of which at least three must be during the last ten years. For the other two cases (IPT and IPTC), eligibility depends on age. For persons aged twenty-six or younger, the requirement is half of the number of years between the age of sixteen and the age when disability began. For persons older than twenty-six, the requirement is either five years or one-fourth of Table 9A.2

Fraction of New Disability Pensions Awarded to Individuals Aged 55+, by SS Program and Level of Disability (1994) Program

IPT

IPA

GI

RGSS RETA REA REMC RETM REEH

4.0 53.4 58.5 0.3 14.9 25.0

43.5 64.4 63.7 48.6 32.1 75.0

39.3 49.3 68.9 60.0 32.0 80.6

Notes: Inability to do the usual job (IPT); inability to do any kind of job (IPA); complete inability (GI). See table 9.A1 for definitions of program variables.

Micro-Modeling of Retirement Behavior in Spain

573

the number of years between the age of twenty and the age when disability began, whichever is largest. Furthermore, at least one-fifth of the required contributory years must have occurred during the last ten years. The benefit base depends on the source of disability. In case of ordinary illness, it is computed the same as for old age pensions. For a non-workrelated accident, it is the average annual wage over a period of twenty-four consecutive months chosen by the person from the last seven years of work. For a work-related accident or professional illness, it is the average wage in the last year of work. The pension equals 55 percent of the benefit base under IPT and increases to 75 percent under IPTC. In case of IPA, it is equal to 100 of the benefit base, whereas for GI it is equal to 100 percent of the benefit base plus another 50 percent covering the person taking care of the disabled. Disability pensions are indexed to inflation, like the other pensions of the RGSS. Unlike the latter, however, disability pensions may be kept while earning income from a job different from the one for which the disability (even a complete one) was determined. Data and Variables In this section, we define the variables that have been employed in the specification of the reduced-form probit (see table 9A.3 for descriptive statistics). The data source is the HLSS (see Martínez 1999 for description), unless we otherwise stated. Variables from HLSS: Experience, Education, and Occupation

• Spell: length of the current spell in the data set • History: history in the data set (i.e., length of participation to the labor market)

• Part Time: indicator variable that takes the value of 1 if the individual does not work full time

• Fraction Working: history divided by potential experience (time elapsed since first time observed in the data set)

• Temporary illness: length of history spent in temporary illness • Sector: 1-digit SIC industry classification • Contributive group: ten groups, from college degree to unskilled bluecollars

• Education: proxy for the level of education, constructed as follows— all individuals in contributive group 1 (i.e., college) are assigned to the college level of the educational variable, People belonging to contributive groups 2, 3, and 4 are assigned to the high school (diploma) category, and people in the remaining contributive groups are assigned to a general class labeled “less than high school” • Years of contributions: number of years contributed • Eligibility indicator: a dummy variable that takes the value of 1 if the

Descriptive Statistics

Retirement Age SSW Accrual Peak value Option value Eligibility Average-LFI id squared Expected wage id squared Expected-wage peak value indicator Expected-wage option value indicator Pension below MP indicator Pension id squared Spell in the firm (years) id squared Years of contributions id squared Potential experience (days) Part time Contributions censored below

Table 9A.3

0.1130 58.89 118.31 2.9720 17.5364 90.0784 0.9499 1035.973 1,459,921 12.4631 203.014 9.2904 14.7937 0.0593 5.7539 46.4389 7.9719 134.736 30.3409 972.835 10,601.4 0.0078 0.1636

Male 0.3166 2.948 36.4737 7.9557 24.7313 89.7464 0.2181 621.855 181,245 6.9057 235.350 8.2419 12.0910 0.2362 3.6514 72.5773 8.4374 248.1310 7.2298 396.92 2,859.6 0.0882 0.3669

RGSS

0.1098 59.29 90.3704 0.4338 13.3365 68.8603 0.8063 663.854 735,039 9.5710 117.873 7.4796 11.3172 0.1449 5.5854 46.4762 7.9754 121.7048 23.7570 638.737 7,708.3 0.0802 0.2233

Female 0.3127 3.179 28.0817 8.6825 22.3410 69.7920 0.3952 542.598 128,180 5.1261 137.711 7.0318 8.4369 0.3520 3.9094 83.3227 7.6232 209.830 8.6233 406.99 3,471.4 0.2717 0.4165

0.3038 3.262 10.7495 4.1614 10.4062 24.4423 0.1498 119.85 145,097

1.9530 2.4331 0.4549 2.9124 31.2876 8.7681 343.308 7.1568 422.886 3,337.0

0.1029 59.54 87.1748 –0.9075 1.3826 25.0808 0.9770 337.578 128,320

2.9242 5.5842 0.2923 2.6031 15.2561 18.5711 421.745 31.2213 1025.98 10,090.1

Male

RETA

3.2752 5.8361 0.4393 2.8775 21.234 12.740 206.84 22.414 550.49 5,940.5

0.1180 60.18 67.9971 –1.5855 1.3056 27.3171 0.8581 237.63 62,491.9

Female

2.0349 2.4006 0.4964 3.6001 61.2252 6.6738 213.08 6.9346 329.20 3,112.2

0.3227 3.644 10.9800 6.5413 9.8729 19.0409 0.3490 77.647 47,122.

0.3508 0.2869 0.1745 0.2546 0.1665 0.2377 0.3358 0.3121 0.3604 0.1329 0.1983 0.2109 0.2654 0.2807 0.2919 0.2134 0.2284 0.4012 0.3734 0.3509 0.1543

0.1147 0.0312 0.2227

0.1437 0.0905 0.0314 0.0697 0.0285 0.0601 0.1296 0.1094 0.1534 0.0180 0.0410 0.0466 0.0763 0.0862 0.0941 0.0478 0.0552 0.1870 0.1674 0.1438 0.0288

0.0154 0.0039 0.2249

Note: Blank cells indicate that data are not available.

Contributions censored above College Diploma Engineers Technical engineers Supervisor and foremen Aux. and clerks assistant Janitors and clerks assistant Unskilled Energy Mineral, chemicals Mechanic & engineering Other manufacturing Construction Retail Transportation Communications and financial Administration and other Code 9130 Temporary illness Other codes Collective bargaining clauses favoring retirement at 65 64 Early retirement 0.1242 0.0314 0.2130

0.0922 0.0908 0.1984 0.1121 0.3327 0.1086 0.2214 0.5035 0.3549 0.3252 0.1197

0.0086 0.0083 0.0410 0.0127 0.1267 0.0119 0.0517 0.4457 0.1477 0.1202 0.0145

0.0194 0.0045 0.2222

0.2007 0.1972 0.2464 0.1695 0.2408 0.1629 0.3181 0.3984 0.4828

0.0421 0.0405 0.0649 0.0296 0.0618 0.0273 0.1142 0.1978 0.3697

576

• • • • • • •

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

individual meets the contributive threshold (fifteen years of contributions) and zero otherwise Pension amount: see section 9.2 for a detailed description Average lifecycle earnings: constructed on the basis of a fixed-effect model for each contributive group Expected earnings: see section 9.4.1 for description Expected-earnings peak indicator: discounted sum of the expected earning from the present to the year the peak is reached Expected-earnings OV indicator: discounted sum of the expected earning from the present to the year the option value (OV) is maximized Minimum pension indicator: a dummy variable that takes value of 1 if the individual’s expected retirement pension falls below the minimum retirement pension Censoring earnings indicators: two dummy variables—the first takes value of 1 if the individual’s level of contributions falls below the minimum (mandatory) level of contributions, and the second takes value of 1 if the individual’s level of contribution is greater than the maximum level of contributions

Variables from the Collective Settlements Register (Estadística de Convenios Colectivos; ECC) Since we do not have direct information about regulations affecting specific workers, we use the Spanish register of collective settlements in order to construct proxies for such regulations. In particular, using the ECC [see Jiménez-Martín (1998) for a brief description of the source], we have constructed three indicators of the coverage of early and mandatory retirement provisions for each (2-digit SIC) industry.

• Early retirement indicator: fraction (weighted by employment) of collective settlements including a prevision favoring early retirement

• Retirement at sixty-four: fraction (weighted by employment) of collective settlements including a provision to facilitate retirement of workers aged sixty-four without incurring any age penalty (this variable only applies to people aged sixty-four enrolled in RGSS) • Mandatory retirement at sixty-five: fraction (weighted by employment) of collective settlements including a provision promoting mandatory retirement at sixty-five (this variable only applies to people aged sixty-five enrolled in RGSS) We refer to Martínez (1999) for a detailed description of the variables and for summary statistics of the histories, covered earnings, and benefits files. Complementary Data Sources At various stages of this work, we have made use of the following complementary data sets: Encuesta de Presupuestos Familiares (EPF; INE

Micro-Modeling of Retirement Behavior in Spain

577

1995a), 1973–1974, 1980–1981, and 1990–1991; Encuesta de Estructura Salarial (EES; INE 1997) 1995; Encuesta Continua de Presupuestos Familiares (ECPF; INE 1998) 1985–1995; EPA 1987–1998. All these data sets are collected by the Spanish National Statistical Institute (INE). A brief description follows.

• EPF: A cross-sectional household budget survey carried out in 1974, 1981, and 1991, with reference to income and expenditure in the previous calendar year (the 1990–1991 sample covers 21,155 households and 72,123 individuals) • EES: A cross-sectional survey of the Spanish wage structure carried out in 1995 with reference to wages paid that year. It collects detailed information about gross wages, SS contributions, working hours, and personal characteristics of about 175,000 workers in 19,000 establishments. The EES is useful for understanding the relation between covered earnings and actual earnings for those cases in which the latter exceed the former, since it simultaneously reports both gross wages and contributions, together with relevant professional characteristics of the individual. • ECPF: A rotating household survey carried out quarterly since 1985. It collects data on income, consumption, net quarterly income (broken down by source: wages, self-employment income, capital income, transfers, and subsidies), and personal characteristics (especially those of the household head and their spouse) for about 3,000 households. • EPA: A quarterly CPS-like survey of roughly 60,000 Spanish households. It contains detailed information on labor force status and education and family background variables but no information on wages and income. Publicly released cross-sectional files are available from 1976 onwards. Starting with 1987, INE also releases the so-called Encuesta de Poblacion Activa Enlazada, or EPAL, which is the panel version of EPA obtained by exploiting the rotating cross-sectional nature of the original survey. It contains fewer variables than EPOA, but it permits to follow individuals for up to six consecutive months.

References Boldrin, M., S. Jiménez-Martín, and F. Peracchi. 1999. Social security and retirement in Spain. In Social security programs and retirement around the world, ed. J. Gruber and D. Wise, 305–353. Chicago: University of Chicago Press. ———. 2001. Sistema de pensiones y mercado de trabajo en España (Pension system and labor market in Spain). Madrid: Fundación BBV. Cordon, F. 1999. Proyección de la población española 1991–2026 (Spanish population projections: 1991–2026). http://www.fedea.es.

578

Michele Boldrin, Sergi Jiménez-Martín, and Franco Peracchi

Instituto Nacional de Estadística (INE). 1995a. Encuesta de presupuesto familiares 1990–1991 Metodología. (EPF; Household budget survey 1990–1991. Methodology) http://www.ine.es. ———. 1995b. Proyecciones de la población de España (Spanish population predictions). Madrid: INE. ———. 1997. Encuesta de estructura salarial (EES; Wage structure survey). http:// www.ine.es. ———. 1998. Encuesta continua de presupuesto s familiares metodología (ECPF; Continuous household budget survey. Methodology). http://www.ine.es/pralyser/catalogo/ecptmeto.htm. ———. 2002. Encuesta de población active. Descripción del cuestinario (EPA; Labor force survey. Survey description, definitions, and instructions for the questionnaire). http://www.ine.es/prodyser/catalogo/epa.htm. Jiménez-Martín, S. 1998. Indexation and wage change settlements: Evidence from Spanish manufacturing firms. Oxford Bulletin of Economics and Statistics 60: 449–84. ———. 1999. El impacto de la reforma de 1997 en los incentivos y la deuda implícita de la Seguridad Social (The impact of the 1997 reform in the incentives and implicit social security debt). Economístas 84:200–08. Jiménez-Martín, S., and A. Sánchez. 2000. Incentivos y reglas de jubilación en España (Incentives and pension rules in Spain). Cuadernos Económicos de ICE 65: 45–88. Martínez, P. 1999. Historiales laborales de la Seguridad Social (Social security labor histories). Universitat Carlos III, Departmento de Economia. Working Paper. Peracchi, F. 2002. The European community household panel survey: A review. Empirical Economics 27:63–90.

10 Income Security Programs and Retirement in Sweden Mårten Palme and Ingemar Svensson

10.1 Introduction As compared to most other industrialized economies, Sweden has high labor force participation among older workers. In the group of countries studied in Gruber and Wise (1999),1 only France, Germany, and Japan have higher male labor force participation among fifty-five year olds, while only Japan has higher participation rates among sixty year olds. By age sixty-five, labor force participation is higher in the United States, Canada, and Japan than in Sweden. Despite these relatively high participation rates, Sweden shares the trend of declining labor force participation among elderly workers experienced in recent decades in other Western industrialized economies: Labor force participation among men aged between sixty and sixty-four has declined from about 85 percent in the early 1960s to about 55 percent in the mid1990s. Since the general health status of the Swedish population has improved over this period of time, the decline in labor force participation has to be explained by other factors. These may include changes in the availability of Mårten Palme is associate professor of economics at Stockholm University. Ingemar Svensson is senior researcher at the Swedish National Social Insurance Board. We are grateful for comments from the editors of this volume and from the seminar participants of the National Bureau of Economic Research (NBER) International Social Security project. The work has also benefited from comments by Anders Björklund, Ed Palmer, and Ann-Charlotte Ståhlberg, as well as from discussions on seminars at Swedish National Institute for Economic Research, Stockholm School of Economics, Stockholm University and University of Umeå. Mårten Palme acknowledges financial support from the Swedish Council for Social Research. 1. These countries are Belgium, Canada, France, Italy, Germany, Japan, the Netherlands, Spain, the United Kingdom, and the United States.

579

580

Mårten Palme and Ingemar Svensson

programs for financing early exit from the labor market, improved economic conditions in the population, changes in economic incentives in general, and changes in collective agreements on retirement age between trade unions and employers’ confederations. In this study, we estimate how economic incentives inherent in the income security system and compulsory old age pensions affect retirement behavior. Social security policy may have a dual effect on economic incentives for labor force participation. First, it may affect the level of an individual’s social security wealth. An increase in social security wealth will increase an individual’s demand for all goods, including leisure, and will, therefore, increase their propensity to retire early. Second, it may affect the accrual in the social security wealth from additional work. High accrual from working one additional year implies a substitution effect to delay retirement. In our retirement probit regressions, net social security wealth is included to measure the income effect. We use three different measures of accrual: benefit accrual, peak value, and option value. The measure of benefit accrual is myopic in the sense that it refers only to the immediate gain or loss to a worker from remaining in the labor force one additional year. The measure of peak value encompasses future possible gains from remaining in the labor force. In addition to considering future possible income gains, the measure of option value also allows for different valuations of leisure time when retired. A large share of those who permanently leave the Swedish labor market receive their main income from disability, sickness, or unemployment insurance. The replacement levels in these programs are higher in general than in the old age pension scheme, which has to be taken into account when measuring the economic incentive variables. However, since labor market insurance programs have requirements on health or employment status, support from these programs is not an available option for all workers. To avoid potential endogeneity problems, we use the eligibility probabilities as weights when calculating net social security wealth and the accrual measures for all workers in the sample. In addition to estimating the econometric model, we simulate the effect of two hypothetical policy reforms. The first delays the age of eligibility for all retirement schemes by three years. In the second experiment, existing retirement schemes are superseded by a program that replaces 60 percent of predicted earnings at age sixty. The normal retirement age is sixty-five, but the pension can be claimed from age sixty to age seventy with an actuarial reduction or increase of 6 percent for every year in advance or delay from age sixty-five. We use a panel data set containing individual characteristics (such as education and sector of employment), detailed information on income components between 1983 and 1997, and contributions to the public pension scheme extending back to 1960. The data set was obtained by merging in-

Income Security Programs and Retirement in Sweden

581

formation from censuses along with tax, social insurance, and education registers. We restricted the sample to men and women born between 1927 and 1940, which resulted in a sample size of about 30,000 individuals, and we study their retirement behavior between 1984 and 1996. This is the first time this comparatively rich data set has been used in a study on retirement behavior. Therefore, we provide detailed descriptions of how different variables, date of retirement, in particular are measured. Our results support the view that economic incentives matter for retirement behavior on the Swedish labor market. Estimates of the econometric model reveal that measures of economic incentives are, in general, significant and with the expected signs. However, the results from the simulations emphasize the importance of collective agreements on normal retirement age, which are supported by Swedish labor legislation. The paper is organized as follows. In section 10.2 we describe the institutions that affect the economic incentives for retirement decisions—such as rules for different pension schemes, labor market insurance, housing allowances, and income taxes—during the period under study, that is, 1983– 1997. Parts of this section are very detailed, mainly to facilitate further research on economic incentives and retirement behavior. (Most readers can skip large parts of this section without losing the ability to follow the rest of the paper.) Section 10.2 also provides a descriptive analysis of the frequency of different pathways to permanent exit from the labor market. Section 10.3 reviews previous research on retirement behavior in Sweden. We present the data set in section 10.4 and show how the different measures of economic incentives are obtained in section 10.5. The empirical models and estimation results are outlined in section 10.6. Section 10.7 reports results from simulations of the estimated models. Section 10.8 concludes. 10.2 Institutional Background We begin by describing the general policy environment for income security, old age pensions, and income taxes in Sweden pertinent to the cohorts in our sample, born between 1927 and 1940, and the time period under study (i.e., 1983–1997). We then describe the frequency of different pathways to retirement. 10.2.1 The Social Security System The public old age pension system in Sweden consists of three parts: a basic pension, a supplementary pension (known as allmän tilläggspension; ATP), and the part-time retirement pension. These are financed through proportional payroll taxes (employers’ contributions) levied on wages. All Swedish citizens and all persons residing in Sweden are entitled to a basic pension. In principle, everyone receives the same amount regardless of previous earnings. The amount is reduced if the duration of residence in

582

Mårten Palme and Ingemar Svensson

Sweden is less than forty years and the number of years with labor income in Sweden is less than thirty years. Like all social-insurance schemes, the basic pension is related to a basic amount (BA). Although the BA is linked to the consumer price index (CPI), it is decided each year by the government. A majority in the Swedish Parliament can thus make discretionary changes that deviate from the development of the CPI. During the 1990s, pensions were not fully aligned with price indexing, due mainly to several measures aimed at cutting the government budget deficit. In 1995, the BA was SKr34,986, and the annual wage of an average production worker was SKr189,488. The basic pension for a single old age pensioner is 96 percent of the BA; it is reduced to 78.5 percent if the person is married. Before 1995, it was not reduced unless the individual was married to someone who also received the basic pension. Individuals with no or low ATP are entitled to a special supplement. This supplement is independent of marital status and has grown from 42 percent of the BA in 1983 to 55.5 percent as of 1993. The special supplement is reduced on a one-to-one basis against the supplementary pension. Thus a single old age pensioner with only a basic pension and a special supplement receives 151.5 percent of the BA. In 1995, this amounted to SKr53,004 in annual pension or 28.0 percent of the annual earnings of an average production worker. The basic pension also contains a survivor’s pension. Widows receive 90 percent of a BA until they reach the age of sixty-five. If a woman is younger than fifty when her husband dies, the amount is reduced. The basic pension for widows has been income tested since 1997. Children normally receive 25 percent of a BA, but the amount may be higher if there is no ATP. A new, gender-neutral transitional pension for men and women born in 1945 and later was implemented in 1990. The transitional pension is paid for six months after the decease of a spouse and amounts to 90 percent of a BA. The transitional pension can be prolonged for a survivor who has the custody of children. The benefit level of ATP is related to an individual’s earnings history and is determined in three steps. The first step involves determining pensionrights income for each year from the age of sixteen. Pension-rights income is calculated on the basis of income from labor reported in an individual’s annual tax return and is the share of the income exceeding 1 BA and below the social security ceiling at 7.5 BA.2 It is set to zero if annual income from labor does not exceed 1 BA. In addition to earnings and income from selfemployment, the pension-rights income includes transfer payments from social insurance (such as income from sickness and unemployment insurance), the parental cash benefit, and the partial retirement pension. Three years of 2. The proportional payroll tax used to finance the ATP scheme is paid also on the share of income exceeding 7.5 BA.

Income Security Programs and Retirement in Sweden

583

pension-rights income greater than zero between the ages of sixteen and sixty-four are required to receive an old age pension from the ATP scheme. In the second step, average pension points are calculated by dividing pension-rights income by the corresponding year’s BA to obtain the pension points for each year. Thus, due to the social security ceiling at 7.5 BAs, the maximum number of pension points an individual may receive in any given year is 6.5. Average pension points comprise the average of an individual’s fifteen best years of earnings. The final step is to calculate an individual’s ATP benefit (Yi ) by applying the formula (1)




Ni Yi
0.6  APi  min , 1  BA, 30

where APi denotes individual average pension points, BA is the basic amount, and Ni is the number of years the individual has reported a pension-rights income greater than zero. Thirty years with pension points are required for full ATP for individuals born in 1924 and thereafter. Inserting the amount of the BA in 1995 into the ATP formula reveals that the maximum pension amount from the Swedish national pension system in 1995 was SKr170,032. There are no dependent’s benefits within the ATP scheme—that is, the amount of the pension is independent of marital status, and there is no splitting of future ATP benefits in the event of a divorce. For women born before 1945, the survivor’s benefit in the ATP system is 35–40 percent of the deceased husband’s ATP pension for a surviving wife and 10–15 percent (20–30 percent after the 1990 reform) for a surviving child, depending on the number of children. The widow’s pension is 35 percent if there are children in the household who are eligible for a children’s pension and 40 percent otherwise. Before the 1990 reform, the widow’s pension from ATP was lifelong. As of 1991, extensive transition rules apply for new survivors. Women born before 1930 still receive a lifelong widow’s pension. For a widow born between 1930 and 1944, her survivor’s ATP is reduced after age sixty-five, taking into account her own ATP. The rules vary somewhat for different birth cohorts. As for the basic pension, for women born after 1 January 1945, the widow’s pension is replaced by a gender-neutral transition pension. The transition pension is paid for six months after the decease of a spouse. However, women born in 1945 and thereafter may also receive a widow’s pension according to special rules and based on the deceased husband’s pension points up until 1990. The basic pension and ATP can be claimed in advance at age sixty3 or postponed until age seventy. If an individual chooses to withdraw from the 3. In 1998, the early retirement age was raised to sixty-one.

584

Mårten Palme and Ingemar Svensson

labor market in advance of reaching sixty-five, the amount of the monthly benefit is permanently reduced by 0.5 percent for each month of early withdrawal: For example, if an individual retires at sixty, the permanent reduction is 30 percent. If an individual decides to claim a pension later than at age sixty-five, the pension income is permanently increased by 0.7 percent for each month of postponement. A partial retirement pension allows workers aged sixty-one and older to reduce their hours of work and receive a benefit to replace lost earnings. To be eligible for part-time retirement, a worker must have accumulated ten years of pension-rights earnings after age forty-five and must work at least seventeen hours per week after the reduction. As of 1 July 1994, the benefit is 55 percent of the difference in earnings before and after part-time retirement. The principal rules of a new pension system intended to replace the basic pension, ATP, and partial retirement pension were decided in 1994. The main changes are that earnings from an individual’s entire life cycle are counted when pension income is determined, rather than only the fifteen best years; pensions are related to the real growth rate in the entire economy rather than price indexes; and changes in life expectancy also affect annual pension income (that is, increased life expectancy and lower economic growth rates reduce individual pension income at a given retirement age). The first birth cohort affected by the new system comprises those born in 1938, who will have four-twentieths of their pension determined according to the new rules and the remainder according to the old rules. The share in the new system is then increased by one-twentieth for each successive birth cohort. 10.2.2 Occupational Pensions Almost all of the Swedish labor market is covered by central agreements between the unions and employers’ confederations. These central agreements include occupational pension schemes financed through employers’ contributions. There are basically four occupational pension plans covering different groups on the labor market: (a) blue-collar workers in the private sector; (b) white-collar workers in the private sector; (c) central government employees; and (d) local government employees.4 Pension rights are transferable among these four main schemes. Each of these pension schemes is briefly described below concentrating on the time period covered by the panel data set used in this study (i.e., 1983–1997). Occupational Pensions for Private-Sector Blue-Collar Workers In 1996, the earlier pay-as-you-go pension scheme (STP) was replaced by a fully funded pension plan. The blue-collar workers in our sample are 4. See Kangas and Palme (1989) for a more detailed description of different occupational pension schemes in Sweden.

Income Security Programs and Retirement in Sweden

585

thus covered by three different occupational pension regimes: those born between 1927 and 1931 who are covered entirely by STP; those born between 1932 and 1938 who may choose between the STP rules and a “transitional pension;” and those born in 1939 and later are covered by the transition pension. As a monthly payment starting the month of a worker’s sixty-fifth birthday, STP could not be claimed in advance or postponed. The size of the pension was determined as 10 percent of the average of the worker’s monthly earnings during his three best years between age fifty-five and fifty-nine. If the worker contributed to the scheme for less than thirty years after age twenty-eight, the pension was reduced proportionally. To receive any pension at all, a worker was required to contribute at least three years between ages fifty-five and fifty-nine. The STP is indexed by the BA, and the social security ceiling at 7.5 BAs applies here as well. In the new, fully funded pension scheme for blue-collar workers, a share of gross earnings is paid into a personal account in a pension fund. Between 1996 and 2000, the contribution rate was 2.0 percent of gross earnings and, according to the new agreement in effect from 2000, the share is 3.5 percent. Each worker can choose among about a dozen insurance companies to manage their pension fund. The first cohort affected by the new system was born in 1932. However, since this cohort, as well as the later cohorts in our sample, worked under the STP system, they have not made any payments to pension funds and are, therefore, subject to so-called transition rules. Pensions under these transition rules are determined by the sum of two parts. The first part is 10 percent of average earnings, deflated by the BA, that are below the social security ceiling from age thirty. The second part is the amount a worker receives from the funded pension. Since the STP scheme allows a worker to choose the average of their best three years between ages fifty-five and fifty-nine, and the pension from the funded system is very low, pensions under the transition rules are, in general, lower than STP. However, the birth cohorts between 1932 and 1938 may opt for the old STP scheme if it turns out to be more favorable. Occupational Pensions for Private-Sector White-Collar Workers White-collar workers in the private sector are covered by ITP, ITPK, and ITPG. The ITP is a defined-benefit scheme, ITPK is fully funded, and ITPG guarantees that a worker covered by ITP receives at least what he would have been entitled to if he had been covered by the STP scheme. The ITP is determined by a worker’s earnings the year before they retire: I is 10 percent of that year’s salary up to 7.5 BAs, 65 percent of the salary between 7.5 and 20 BAs, and 32.5 percent between 20 and 30 BAs. As in the STP scheme, the pension is reduced proportionally if a worker has contributed for less than thirty years since age twenty-eight. Contributions to ITP have been around 4.5 percent of gross earnings in the 1980s and 1990s.

586

Mårten Palme and Ingemar Svensson

Table 10.1

Reduction/Enhancement if ITP is Claimed in Advance/Postponed Retirement Age

Reduction/Enhancement

60 61 62 63 64 65 66 67 68 69 70

0.739 0.783 0.831 0.884 0.942 1.000 1.076 1.154 1.241 1.338 1.448

The normal retirement age for ITP is sixty-five. Table 10.1 shows the reduction (or enhancement) if a worker starts to claim (or postpone) retirement benefits between ages sixty and seventy. Also, ITP can also be claimed before age sixty, in which case the amount of the pension is determined by an individual actuarial adjustment. The ITPK was introduced in 1977 and is a fully funded system. During the 1980s and 1990s, contributions amounted to approximately 2 percent of each worker’s labor earnings up to 30 BAs. Contributions to the ITPK scheme start when a worker is aged thirty. They are free to choose a company to manage their ITPK pension. ITPK is normally claimed as monthly payments over a five-year period after retirement. As ITPK was introduced in 1977, it is maturing during the period covered by our data (the 1927 cohort of workers were aged fifty in 1977 and the 1940 cohort was aged thirty-seven). This implies that the ITPK pensions are, on average, larger for the younger cohorts. Pensions for Central Government Employees Pensions for central government employees are regulated in central agreements between the trade unions and the state. Prior to 1992, the occupational pension scheme for employees in the central government provided a gross pension in the sense that it totally replaced the state pension for workers covered by the scheme. The size of the pension was calculated as 65 percent of earnings for the year before retirement. A full pension required thirty years of earnings, and the pension was reduced proportionally if the worker did not fulfill that requirement. Most people employed by central government have a mandatory retirement age of sixty-five. There are several exceptions, such as military personnel, whose mandatory retirement age is fifty-five. Before 1992, central government employees could not claim their occupational pension prior to their mandatory retirement age. If they wanted to retire earlier, they could

Income Security Programs and Retirement in Sweden

587

claim their state pension with actuarial adjustment and their occupational pension as a lifelong annuity. This annuity was calculated as 65 percent of 95 percent of an individual’s earnings the year before they retired. This amount was not indexed and not paid out until they reached the mandatory retirement age. After that, it was indexed by the BA and paid as a lifelong annuity. After 1992, two supplementary occupational pension schemes, one fully funded and one pay-as-you-go, replaced the former gross pension. In the fully funded system, 1.7 percent of a worker’s annual salary, starting from 1991, is paid to a pension fund. The pay-as-you-go is very similar to the ITP, but it is determined by average earnings during the five years preceding retirement rather than by an individual’s earnings the year before retirement. It is 10 percent of this five-year average up to 7.5 BAs, 65 percent between 7.5 and 20 BAs, and 32.5 percent between 20 and 30 BAs. The pension is reduced proportionally if the requirement of thirty years of contributions to the scheme since age twenty-eight is not fulfilled. In contrast to the pre-1992 occupational pension for central government employees, the post-1992 pension can be claimed five years before the mandatory retirement age with an actuarial adjustment. This adjustment is a 0.4 percent lifelong reduction for each month the pension is received prior to an individual’s sixty-fifth birthday. However, if someone retires before age sixty, the pre-1992 rules apply (i.e., no benefit prior to the mandatory retirement age). This pension can also be postponed with a 0.4 percent lifelong increase for each month it is delayed up to five years after the mandatory retirement age. Pensions for Local Government Employees The pension plan for employees in local governments (or municipalities) is regulated by a central agreement between the union and a confederation for Sweden’s municipalities. Two agreements affect pensions for the time period covered by the data in this study: the first was made in 1978 and the second in 1985. According to the 1985 agreement, the size of the pension is determined by the average of the employee’s five best years of earnings during the seven-year period prior to the year of retirement. The pension is then calculated as 96 percent of this amount below 1 BAs, 78.5 percent between 1 and 2.5 BAs, 60 percent between 2.5 and 3.5 BAs, 65 percent between 7.5 and 20 BAs, and 32.5 percent between 20 and 30 BAs. A full pension requires thirty years of employment in the local government sector between ages eighteen and sixty-five; otherwise the pension is reduced proportionally. This pension scheme is fully coordinated with the state pension. This means that only the amount exceeding the state pension is paid. The normal retirement age is sixty-five for most local government employees, but an individual can enter retirement at age sixty or postpone it

588

Mårten Palme and Ingemar Svensson

until sixty-seven. If they retire before age sixty-five, their pension is reduced for the rest of their life by 0.3 percent per month between ages sixtythree and sixty-five, by 0.4 percent between ages sixty-two and sixty-three, and by 0.5 percent per month between ages sixty and sixty-two. The pension is increased by 0.1 percent for each month the individual decides to continue to work after age sixty-five. The rules for claiming before age sixty are very similar to those in the pension scheme for central government employees: The pension is transformed into a lifelong annuity that is paid out starting at age sixty-five. 10.2.3 Disability, Sickness, and Unemployment Insurance Disability Insurance The disability insurance (DI) scheme is very similar to the state old age pension during the period covered by the study.5 It consists of a basic pension, an income-related ATP supplement, and a special supplement. Pension income is determined in much the same way as the old age pension benefit but without any actuarial reduction for early retirement. An “assumed” pension point is calculated for each year between the year of retirement with DI and age sixty-four. The formula for old age ATP is then applied to actual as well as assumed points between ages sixteen and sixty-four. A disability pension can be received from age sixteen. Eligibility requires certification from a physician that an individual’s capacity to work is permanently reduced by at least 25 percent due to illness, physical or mental incapacity, or so forth. To receive a full disability pension, working ability must be completely lost, although an individual may also be awarded 25, 50, or 75 percent DI, corresponding to different degrees of lost ability to work.6 Between 1972 and 1991, disability pensions were also granted for labor market reasons. The requirements then were that the individual was sixty years old or more and had exhausted his right to unemployment insurance. In practice, the strictness with which medical screening is applied varies over time. When analyzing granting rates of different local social insurance offices, it is also evident that it varies between different parts of the country. The rules regarding eligibility for DI have been tightened considerably through successive changes in legislation in July 1993, October 1995, and January 1997. Sickness Insurance Sweden has universal sickness insurance covering all employees and self-employed that is financed through payroll taxes. This insurance pro5. In addition to the public labor market insurance, which we will consider here, there are negotiated occupational insurances for disability and long term sickness. 6. Before 1993, the levels were 50, 67, and 100 percent.

Income Security Programs and Retirement in Sweden

589

vides compensation for foregone earnings to workers who are not able to carry out their regular work due to temporary health problems. It has undergone several changes over the time period covered by our data set. Prior to the first major reform in 1987, compensation was calculated on the basis of annual earnings, but during the first two weeks of illness, it only covered foregone earnings during scheduled work time from the second day in a sickness spell. After the reform, 90 percent of foregone earnings up to the social security ceiling were compensated from the first day of a sickness spell. The second major reform took place in 1992 when employers had to take responsibility for sickness insurance during the first two weeks of a spell. The replacement level—the share of foregone earnings replaced by the insurance—has been changed on several occasions between 1983 and 1997. In 1993, the replacement level for long sickness spells, which is most relevant for the purpose of this study, was reduced from 90 to 80 percent of foregone earnings between days 91 and 365 in a spell, and reduced to 70 percent after one year. In 1996, it was changed to 75 percent for all long-term spells, and as of 1998, it is 80 percent for all spells. Eligibility for compensation after seven days of a sickness spell requires a certificate from a physician. The certificate then has to be renewed at least every third month for continued compensation. A physician has to certify that temporary illness does not permit the insured individual to perform his regular work and that he will be able to return to the labor force after recovery. Otherwise, the worker should be granted DI. The compensation level of sickness insurance is higher than that of DI for most workers. This implies that a worker has economic incentives to remain on sick leave, rather than DI, even if the probability of returning to the labor force is very low. The law does not stipulate an upper limit on the length of a sicknessbenefit spell. Unemployment Insurance Unemployment insurance (UI) is twofold: One part consists of the same amount for all unemployed workers, and the second depends on an insured worker’s income level before he became unemployed. A worker is not eligible for the second part unless he belongs to an unemployment benefit fund. All members of labor unions automatically belong to an unemployment benefit fund. It is also possible to be a member of an unemployment benefit fund without being a union member, if the worker has the occupation covered. Unemployed workers who actively search for a new job are eligible for UI. Refusal to accept a “suitable” job offer from the public employment office might lead to exclusion from compensation. In general, a worker can reject two, but must accept the third suitable job offer. An unemployed worker is entitled to UI compensation for 300 days up to age fifty-five and

590

Mårten Palme and Ingemar Svensson

for 450 days thereafter.7 However, if a worker undergoes one or more training programs, the compensation period can be renewed several times. The compensation level is very similar to sickness insurance with one important exception: The income ceiling of UI is lower. The UI ceiling is not indexed: Changes in the ceiling are made on a discretionary basis by the government. By the end of the period covered by our data (1997), the income ceiling for UI was SKr199,650 compared to SKr272,250 for the priceindexed social security ceiling used for sickness insurance. The changes in the compensation level of sickness insurance, reported in the preceding subsection, also apply to UI. 10.2.4 Income Taxes and Housing Allowances Besides the effect of the social security system, retirement incentives are also affected by income taxes.8 Sweden has an integrated income tax system. Individuals pay local and national income taxes. The national government determines the tax base for national and local taxes. After a major tax reform in 1991, the tax base is now divided into earned income and capital income. All income from the social insurance system is included in earned income along with wages and salaries. As of 1991, there is a national proportional tax of 30 percent on taxable income from capital. Earned income is taxed nationally and locally. The local tax rate is determined independently by each of Sweden’s 283 municipalities. Local tax rates are clustered around 31 percent. Prior to 1991, the marginal tax rate on pension income was affected by capital income, since there was no division of the tax base into earned income and capital income. Local income taxes are proportional, while the national income tax is progressive. After the 1991 tax reform, national income tax was set at (almost) zero below a certain point of earnings, and at 20 percent on all income above that level. In 1995, the latter tax was temporarily increased to 25 percent. These rules may give the false impression that there are only two possible marginal tax rates on earned income, but there is a basic deduction that varies among different earned-income brackets. There are also special rules for the basic deduction for old age pensioners, which largely determine their marginal tax rates. Old age, disability, and survivor’s pensioners with low income are entitled to a housing allowance. In 1995, this allowance was, at most, 85 percent of the housing cost up to a certain ceiling and above a certain floor. It is reduced by 40 percent (45 percent at high-income levels) of income in excess of a basic pension and special supplement and by 2 percent of wealth. In 1994, about 30 percent of all old age pensioners received housing allowances, and the average amount was about SKr17,673—that is, 33 7. This was changed to age fifty-seven in December 1997. 8. See Aronsson and Walker (1997) for a detailed description of the Swedish tax system.

Income Security Programs and Retirement in Sweden

591

percent of the amount of the lowest pension from the national pension system. 10.2.5 Mandatory Retirement Rules on the Swedish Labor Market Sweden has a normal retirement age of sixty-five.9 Older workers are not covered by employment security legislation,10 that is, workers older than sixty-five are not covered by seniority rules and therefore are protected the least if a firm wants to scale down. Furthermore, workers over sixty-five are not entitled to UI. On the other hand, the wage cost for employers is lower because they do not have to pay payroll-taxes to national or occupational pension schemes for employees over sixty-five. Central and local government employees automatically lose their jobs at age sixty-five. Exceptions from this rule are permitted for one year. In the private sector, collective agreements between the trade unions and the employers’ confederations as a rule also prescribe strict rules for mandatory retirement at age sixty-five. As the number of these agreements is very large, it is hard to get an overview of the overall strictness of the rules for mandatory retirement. 10.2.6 Sources of Income after Retirement As already indicated, the Swedish welfare system provides several options for early exit from the labor market. In order to gain an understanding of to what extent these different options are used, let us consider the cohorts in our data set for persons born between 1927 and 1932. These are the birth cohorts that had reached the normal retirement age of sixty-five in 1997 (the end of the period under study). Table 10.2 shows the percentage share of workers in this subsample who receive their main income (more than 50 percent of their total nonlabor income) from one of ten different sources of income after retirement. The last row in the table indicates that none of the sources of income accounts for more than 50 percent of the retired worker’s nonlabor income. The sources of income listed in table 10.2 can be divided into three groups. The first group consists of schemes designed to serve as old age pension programs: the state old-age pension (pathway 1), occupational pensions (pathway 2), pensions provided by the employer or severance payments (pathway 6), private pensions (pathway 7), and partial retirement benefits provided by social security (pathway 10). The second group comprises insurance programs against income loss from poor health or unemployment: DI (pathway 3), wife’s supplement (pathway 5), sickness insurance (pathway 8) and UI (pathway 9). In contrast to the first group, 9. Wadensjö (1989) discusses the implications of sixty-five as a normal retirement age. 10. Less than 5 percent of employees in the Swedish labor market are not covered by a central agreement, and in which case, they are protected by employment security legislation until age sixty-seven.

592

Mårten Palme and Ingemar Svensson

Table 10.2

Percentage Share of the Pathways to Permanent Exit From the Labor Market Showing Main Source of Income (more than 50% from the indicated source) for Cohorts Born 1927–1932, by Gender

Pathway

Men

Women

1. State old age pension 2. Occupational pension 3. Disability pension 4. Survivor’s pension 5. Wife’s supplement 6. Severance payments from employer 7. Private pension 8. Sickness insurance 9. Unemployment insurance 10. Partial retirement benefit 11. No income source 50%

33.70 13.68 6.55 0.00 0.02 0.60 0.86 20.53 8.35 10.04 5.67

26.99 14.21 6.59 3.99 2.00 0.69 0.76 26.88 6.42 6.83 4.64

Notes: The 10.02% of the male and the 6.11% of the female subsample not yet retired by the end of the panel are included in pathway 1. Pathway 5 also includes some other minor benefits in addition to wife’s supplement.

claiming support from the sources in the second group is not a viable option for everyone, due to the health or unemployment requirements of the programs. The third group contains only one source: survivor’s pension (pathway 4). According to table 10.2, the second insurance group accounts for about 35 percent of the male and about 40 percent of the female subsample. Within this group, sickness insurance is the dominant initial source of income: More than 20 percent of the men and 27 percent of the women. In the first (old age retirement) group, private pensions and employerprovided pensions are relatively unimportant as the main source of income of the newly retired. To study whether or not the initial path to retirement varies among workers assigned to different occupational pension schemes (i.e., different socioeconomic groups) we repeated the analysis above, but divided the subsample into groups corresponding to assignment to different occupational pension schemes. These results are shown in table 10.3. It is evident from table 10.3 that there are large differences between workers in different occupational pension schemes. Blue-collar workers in the private sector, covered by the STP scheme, are much more likely than all other groups to receive their main initial income from sickness insurance or UI. A further distinct result is that employees in the public sector, both state and local government, are less likely to have their main income from UI when they exit the labor force. Since workers are able to switch between different sources of income after permanent exit from the labor force, it may be misleading to describe

36.24 19.45 4.36 0.00 0.00 1.65 1.00 12.55 7.66 8.84 8.25

(2) 28.34 31.98 8.87 0.29 0.00 0.00 0.00 13.95 4.8 7.27 4.51

(3) 41.12 16.89 5.71 0.00 0.00 0.75 0.37 17.02 2.48 9.32 6.34

(4) 48.93 4.41 8.96 0.00 0.00 0.14 3.13 17.78 5.69 7.68 3.27

(5) 24.26 4.17 5.96 2.70 2.78 0.25 0.33 34.80 14.05 6.21 4.49

(1) 31.00 12.56 5.32 3.28 0.23 3.05 0.79 18.78 8.60 10.29 6.11

(2)

25.86 16.44 10.76 3.74 1.94 0.30 0.00 22.42 5.38 8.97 4.19

(3)

Women

25.41 20.54 6.37 4.82 1.94 0.17 0.60 28.34 1.81 5.68 4.31

(4)

41.58 5.84 5.15 5.50 4.81 1.03 4.81 16.84 6.87 3.09 4.47

(5)

Notes: The 10.02% of the male and the 6.11% of the female subsample who are not yet retired by the end of the panel are included in pathway 1. Column (1) shows blue-collar workers in the private sector; column (2) shows white-collar workers in the private sector; column (3) shows central government employees; column (4) shows local government employees; and column (5) shows self-employed.

25.65 5.11 7.06 0.00 0.05 0.05 0.37 31.13 13.10 12.92 4.55

(1)

Men

Percentage Share of the Pathways to Permanent Exit from the Labor Market Showing Main Source of Income (more than 50% from the indicated source) for Cohorts Born 1927–1932, by Gender and Socioeconomic Group

1. State old age pension 2. Occupational pension 3. Disability pension 4. Survivor’s pension 5. Wife’s supplement 6. Severance payments from employer 7. Private pension 8. Sickness insurance 9. Unemployment insurance 10. Partial retirement benefit 11. No income source 50%

Pathway

Table 10.3

594 Table 10.4

Mårten Palme and Ingemar Svensson Percentage Shares of Main Source of Income After a Spell with Sickness or Unemployment Insurance After Permanent Exit From the Labor Market, with Sickness and Unemployment Insurance, Respectively, as Main Source of Income Before Transition (cohorts born 1927–1932) Number of Years Living on First Main Income Source

Pathway

1

State old age pension 12.00 Occupational pension 7.07 Disability insurance 61.03 Survivor’s pension 0.62 Wife’s supplement 0.50 Severance payments 0.54 Private pension 0.23 Unemployment insurance 5.11 Partial retirement benefit 0.04 Mixed sources 12.76 No. of observations 2,601 State old age pension Occupational pension Disability insurance Survivor’s pension Wife’s supplement Severance payments Private pension Partial retirement benefit Mixed sources No. of observations

47.64 3.48 20.27 0.12 1.12 0.37 14.18 0.25 12.56 804

2

3

4

5+

Average

Sickness Insurance to: 62.18 22.44 10.90 97.28 2.72 0 30.50 44.72 18.87

3.85 0 4.09

0.64 0 1.83

1.59 1.02 2.03

93.98

3.01

2.26

0

0.75

1.50

42.17

39.76

11.45

5.12

1.50

1.84

Unemployment Insurance to: 27.94 39.69 18.54 10.71 85.71 3.57 7.98 57.06 31.29

7.05 0 3.07

6.79 0 0.61

2.26 1.93 2.31

45.61

44.74

7.89

1.75

0

1.66

30.69

54.46

8.91

3.96

1.98

1.92

Note: Blank cells indicate that data is not available.

only the first main source.11 Table 10.4 shows the percentage distributions of the second main source of income for those who initially left the labor market with sickness insurance or UI, the number of years they retain their first source, as well as the average number of years on their first main source of nonwork income. Since those who start to receive old age pension benefits at retirement are most likely to continue to do so, and those who leave the labor force with DI as their main source of income will automatically begin receiving old age pensions at age sixty-five; these groups are excluded from the transitions listed in table 10.4. According to table 10.4, most of those who initially had sicknessinsurance benefits as their main income source receive a disability pension 11. For example, we found 677 different permutations of the main source of income after retirement in our sample.

Income Security Programs and Retirement in Sweden Table 10.5

595

Percentage Share of Main Source of Income after Sickness or Unemployment Insurance, and Mixed Sources of Income as Second Main Source after Permanent Exit from the Labor Market (cohorts born 1927–1932) Number of Years on Second Main Income Source 1

State old age pension Occupational pension Disability insurance Wife’s supplement Mixed sources No. of observations State old age pension Occupational pension Disability insurance Survivor’s pension Wife’s supplement Unemployment insurance No. of observations

2

3

4

Sickness Insurance to Unemployment Insurance to: 29.96 89.47 2.63 7.89 0 1.56 55.47 94.37 4.23 0 0 2.34 10.94 100 0 0 0 128 Sickness Insurance to Mixed Sources to: 28.57 45.74 35.11 12.77 1.82 65.05 38.79 43.46 10.28 3.04 0.30 1.22 328

5+

Average

0

1.18

1.41

1.10

0

1.00

6.38

0

1.80

5.14

2.34

1.89

0

1.70

0

1.63

0

1.77

0

1.95

State old age pension Occupational pension Disability insurance Wife’s supplement Mixed sources No. of observations

Unemployment Insurance to Sickness Insurance to: 27.27 36.67 56.67 6.67 0 0.91 60.91 49.25 41.79 5.97 2.99 0.91 10.00 110

State old age pension Occupational pension Disability insurance Sickness insurance No. of observations

Unemployment Insurance to Mixed Sources to: 73.78 37.10 51.61 8.06 3.23 1.19 23.81 20.00 70.00 5.00 5.00 1.19 84

Note: Blank cells indicate that data is not available.

as their second main source. More than 70 percent of this group receive sickness insurance only one or two years before the transition to DI. The picture is somewhat more diverse for those who initially receive UI benefits as their main source of income. More than 45 percent switch to an old age pension. Almost 70 percent of UI-benefit recipients have a UI benefit prior to the transition to some other benefit for one or two years. About 20 percent switch to a DI pension, and a considerable fraction, 14.18 percent, switch to sickness-insurance benefits as their next main source of income. Table 10.5 goes one step further and reports what happens after the

596 Table 10.6

Mårten Palme and Ingemar Svensson Percentage Distribution of the Number of Years after Permanent Exit from the Labor Force before Disability Insurance Becomes the Main Income Source (retirees with initial income from sickness or unemployment insurance only) Number of Years 1

2

3

4

5+

Average

Age 50–55 37.60 33.47 7.69 30.77 35.66 3.33

9.09 15.38 9.30

10.74 46.15 12.79

2.75 4.00 2.81

Sickness insurance Unemployment insurance All

9.09 0.00 8.91

Sickness insurance Unemployment insurance No income source 50% All

20.10 3.92 51.02 18.91

Age 55–60 42.44 40.20 18.37 41.21

27.17 38.24 24.49 28.70

6.67 13.73 6.12 7.66

3.62 3.92 0.00 3.53

2.31 2.74 1.86 2.36

Sickness insurance Unemployment insurance No income source 50% All

49.31 10.42 75.56 49.74

Age 60–65 40.79 64.58 18.89 39.48

9.95 25.00 5.56 10.78

0.00 0.00 0.00 0.00

0.00 0.00 0.00 0.00

1.60 2.15 1.30 1.61

states considered in table 10.4. Most of those who switch from sickness insurance (SI) to UI or from UI to SI, ended up with DI as their main source of income. Table 10.5 also reveals that most of the second transitions took place within one or two years. The percentage distribution of the number of years during which retirees received their main income from other sources and then started to receive their main income from DI is reported in table 10.6. It is evident that those who retire at relatively older ages make a faster transition to DI. Table 10.6 may also serve as a summary of the results previously obtained on transitions between sources of income after permanent exit from the labor force. It shows that a majority in most groups make the transition to DI within two years after they retired. Finally, figure 10.1 shows the relation between retirement age and the average number of years before a worker receives DI as their main source of income, provided that their initial main source was from one of the labor market insurance programs. In particular, there is a very clear relationship between age of exit from the labor force and the average number of years with UI- or SI-benefits. In summary, this section showed that there is a great deal of heterogeneity in the way Swedish workers finance their retirement. Two important conclusions emerge. First, pathways to retirement vary considerably between different groups of workers. Blue-collar workers in the private sector, in particular, get their income from insurance against poor health or unemployment after having permanently left the labor force to a much

Income Security Programs and Retirement in Sweden

597

Fig. 10.1 Average number of years after retirement before DI becomes the main source of income; only workers with main initial source of income from insurance

larger extent than other groups. Second, although a large share of workers rely on SI and UI as their main source of income in the initial state after a permanent exit from the labor force, most of them switch to DI after one to two years. This period decreases with the age of retirement. 10.3 Research Background Despite the importance of early retirement from the labor market, the empirical research on retirement behavior is very meager in Swedish data. The most ambitious attempt to formally model retirement choice is found in Hansson-Brusewitz (1992). In the empirical part of his study, HanssonBrusewitz estimates a labor-supply model with joint decisions on the number of hours of work and labor force participation. Among other things, he simulates the effects on total labor supply of introducing a partial pension scheme and replacing the ATP system with a scheme in which pensions amount to 60 percent of average lifetime earnings. He found that the partial retirement scheme has a positive effect on total number of hours of work. As regards the hypothetical reform of the ATP system, he found a small positive effect on hours of work but a small negative effect on desired retirement age. Sundén (1994) studies to what extent changes in rules, in general, and the introduction of the partial retirement benefit, in particular, could account for the changes in retirement behavior between 1974 and 1981, or to what extent these changes rather could be attributed to changes in individual

598

Mårten Palme and Ingemar Svensson

preferences. She estimates a multinomial-logit model with four different retirement options and then decomposes the overall change in retirement behavior between 1974 and 1981. She finds that most of the observed changes could be attributed to preferences, that is, an estimated coefficient in the model. Changes in rules, reflected in variables in the model, have a very small effect. Skogman Thoursie (1999) investigates whether or not economic incentives affect the probability that a disability pension will be granted. He used a sample from the Swedish Level of Living Survey and estimated a reducedform conditional-logit model. The difference between the predicted income from DI and the predicted income from labor was used as a measure of economic incentives. The results showed that a gain in predicted income from DI, relative to income from labor, increases the probability that a worker will exit the labor market with a disability pension. The interpretation of the result is that economic incentives do in fact affect the number of new disability pensions. Wadensjö and Palmer (1996) compare disability policies in Sweden and the Netherlands. Both countries have generous disability programs, which provide major pathways to early exit from the labor market. Despite the similarities, there is a higher labor force participation rate among older workers in Sweden than in the Netherlands. The authors point to some peculiarities in the Swedish labor market and disability policies that might account for the different outcomes. Among these distinctive features are the emphasis on the “work principle” in Swedish social and labor market policy, the low unemployment rates (until the recession in the 1990s), the possibilities (specific to Sweden) of combining work and pensions through partial benefits, and the vocational rather than medical focus of rehabilitation policy.

10.4 Data 10.4.1 The Data Set We use the Longitudinal Individual Data panel data set (LINDA) recently constructed by Statistics Sweden, the Department of Economics at the University of Uppsala, and the National Social Insurance Board. The LINDA is a pure-register sample, that is, no interviews were made when the data were collected. LINDA contains data from three main registers. 1. Income and Wealth Register (Inkomst- och Förmögenhetsstatistiken; IoF): This income tax register consists of tax-return data on all people registered as taxpayers in Sweden. The LINDA contains data from this register for each year between 1968 and 1997. For the years 1983 to 1997, the IoF includes detailed data on taxable and nontaxable transfers based on

Income Security Programs and Retirement in Sweden

599

registers from the National Social Insurance Board, the National Board for Educational Assistance, and some other authorities. 2. Population Census (Folk- och Bostadsräkningen; FoB): The FoB exists for every fifth year between 1960 and 1990 and is obtained from mailed questionnaires. Everyone living in Sweden is included in the FoB, and participation in the census is compulsory. 3. The National Social Insurance Board Registers for pension points (based on earnings): The LINDA contains data from this register for each year between 1960 and 1997. The LINDA also contains data on education from the National Education Register and on employment from the National Labor Market Board register. The original sample for the LINDA panel is a random draw of 300,000 individuals from the 1995 IoF. The sampling procedure used to update the panel backwards and forwards from 1995 is designed so that each yearly cross section of LINDA is also a representative sample of the whole Swedish population. The LINDA panel also contains information on the spouse of each individual originally included in the sample. In general, the same variables as for the original individuals are also available for their spouses. There are two, somewhat different, definitions of “spouse” in LINDA. The first is the tax-authority definition of spouses (samtaxerad) as either formally married or as cohabiting and having common children. Information on spouses according to this definition is available for each year between 1968 and 1997. The second definition refers to all spouses that have reported in the mailed questionnaire that they are living together (i.e., share housing). This information is only available for the years of the FoB. When calculating incentive variables for this analysis, we used the first definition because it is available for all years under study. 10.4.2 Sample Selection For purpose of our study, we have restricted the population in several dimensions. First, the period of analysis is restricted to the years 1983 to 1997, primarily because the LINDA panel contains much more detailed information on individual sources of income for this period compared to the period preceding 1983. Second, the population is restricted to individuals born between 1927 and 1940, that is, those who were born in 1927 were age fifty-six in 1983 and seventy in 1997; those born in 1940 were age forty-three in 1983 and fiftyseven in 1997. Third, we have restricted the sample to those who had not already permanently exited from the labor force at age fifty. Table 10.7 shows the number of individuals remaining in the sample at different stages of the selection process.

600

Mårten Palme and Ingemar Svensson

Table 10.7

Number of Individuals Remaining After Each Step in the Sample Selection

Individuals born 1927–1940 Neither emigrated nor dead in 1983 Usable earnings histories Not retired at age 50 Not retired in 1983 Employed in 1983

Table 10.8

Men

Women

Total

22,375 22,055 22,046 20,364 18,163 15,619

21,948 21,798 21,781 19,576 15,916 14,820

44,323 43,853 43,827 39,940 34,079 30,439

Number of Individuals Remaining in the Sample During the Period Under Study

1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997

Men

Women

Total

15,619 15,578 15,535 15,479 15,390 15,325 15,237 15,144 15,043 14,914 14,789 14,664 14,518 14,370 14,194

14,820 14,812 14,794 14,775 14,731 14,698 14,654 14,612 14,550 14,495 14,438 14,363 14,282 14,194 14,103

30,439 30,390 30,329 30,254 30,121 30,023 29,891 29,756 29,593 29,409 29,227 29,027 28,800 28,564 28,297

Since LINDA is a national register sample, the attrition of the panel differ from survey panels. There are two main sources of attrition: mortality and permanent emigration. Table 10.8 shows the number of individuals remaining in the sample in different years covered by the panel. 10.4.3 Measurement of Variables Measuring Date of Retirement As the data set only includes register information, there is no selfassessed information on date of retirement. It does, however, contain detailed information on sources of income for each individual in the sample. The sources of income that enable workers to remain out of the labor force, listed in table 10.2, enable us to indirectly measure the date of permanent exit from the labor market (i.e., date of retirement). We investigated two definitions for measuring full-time retirement.

Income Security Programs and Retirement in Sweden Table 10.9

601

Difference in Years Between Age of Permanent Exit From the Labor Force Using the Earnings-From-Labor (definition 2) Measure of Retirement and Source-of-Income Definition (definition 1), by Gender and Age Group of Retirement (source-of-income definition) Difference (years)

Gender and Age Men 50–55 55–60 60–65 65–70 Women 50–55 55–60 60–65 65–70

–3

–2

–1

0

1

2

3

10.67 7.89 2.75 1.24

4.94 2.12 1.24 0.13

37.58 34.66 29.26 14.76

44.27 51.56 61.73 74.06

2.23 2.46 3.87 5.66

0.32 0.81 0.60 2.02

0.00 0.50 0.55 2.15

16.47 11.34 4.47 2.84

7.72 4.96 2.26 0.69

37.83 38.46 38.76 21.86

36.80 43.54 52.98 72.03

1.19 1.10 1.14 1.81

0.00 0.53 0.14 0.60

0.00 0.08 0.26 0.17

1. Out of the labor force full time by source of income: An individual is considered to be out of the labor force full time if, in a particular year, they receive more than 80 percent of their income from the sources listed in table 10.2. 2. Out of the labor force full time by earnings from labor: An individual is considered to be out of the labor force full time if, in a particular year, he has labor earnings of less than one BA. This leads us to two different definitions of date of retirement: (a) the year preceding the first year an individual is out of the labor force full time, according to the source-of-income definition, and remains so for the rest of the period covered by the panel or (b) the year preceding the first year an individual is out of the labor force full time, according to the earningsfrom-labor definition, and remains so for the rest of the period covered by the panel.12 These two definitions of retirement are compared in table 10.9, which shows the distribution in percentage shares of the difference between the age of retirement resulting from two measures. The results are shown for four subsamples by age group for retirement according to the earningsfrom-labor definition. Table 10.9 shows that there are differences between the two measures. First, there is a thick clustering of observations in the 0 and –1 columns. A relatively simple explanation as to why the earnings-from-labor measure 12. An obvious problem with this way of measuring date of retirement is that workers who are regarded as retired could in fact have returned to the labor market after 1997 (the last year included in the panel)—that is, on average, we will underestimate the date of retirement and the degree of underestimation is positively correlated to the date of retirement.

602

Mårten Palme and Ingemar Svensson

gives a retirement age one year before the source-of-income measure is that a worker is likely to earn more than one BA the year he retires—unless he does not retire until the end of the year. Therefore, as indicated above, we have set the year of retirement at the year before the worker starts to permanently earn less than 1 BA from labor. Likewise, it is very likely that labor income exceeds 20 percent of total income the year a worker retires. Consequently, we set the year of retirement at the year before the worker starts to permanently receive less than 20 percent of his income from labor. For the majority of individuals in our sample, one BA is larger than 20 percent of income. If, due to the timing of retirement during the year, earnings fell between 20 percent and one BA, a one-year difference between the two measures is recorded. Second, there are relatively many observations in the –3 column. These workers reach the earnings-from-labor criterion three or more years before the source-of-income definition, that is, they have several years with earnings from labor below one BA but exceeding 20 percent of total income. There are several explanations for such observations. They can refer to partially retired low-income workers (those who live on their own savings) or on the income of other members of the household, which is probably more common. Another possibility is that workers exit from the regular labor market and enter the informal sector of the economy. Such individuals appear more frequently in the female subsample and, more importantly, in the age group that, according to the earnings-from-labor measure, retired early. These groups represent very few observations: the figure of 16.47 percent for women who retired between ages fifty and fifty-five in the –3 column corresponds to only sixty-three observations. Moreover, table 10.9 shows that a considerable share of the individuals who retire after age sixty-five, according to the earnings-from-labor criterion, had retired according to the source-of-income criterion two or more years earlier. In other words, they continued to work part time after retirement while simultaneously receiving their main income from old age pension benefits. There is no earnings test in Sweden’s old age pension schemes (i.e., it is possible to receive full pension benefits and continue to work). The decision to retire (i.e., leave the labor force) and the decision to claim a benefit are separate. Table 10.9 shows that almost 10 percent of the men who retired between ages sixty-five and seventy according to the earningsfrom-labor definition claimed a benefit at an earlier age. Women did this to a considerably smaller extent. In the case of high-income workers, the source-of-income definition might be more appropriate, since earnings above one BA correspond to relatively few hours of work. To conclude the comparison between the two definitions of full-time retirement, let us first note that the resemblance between the two measures of retirement seems to be good for most individuals in the sample. However, the source-of-income definition missed that some individuals, pri-

Income Security Programs and Retirement in Sweden

603

marily women, leave the labor force without immediately claiming full benefits from any of the programs considered in table 10.2. Also, some individuals, primarily men, remain in the labor market part time at a relatively old age, but are still measured as retired full time by the source-ofincome definition of retirement. These two disparities imply that the earnings-from-labor definition of retirement is more useful, and we restrict ourselves to using this definition when describing the transition to full-time retirement. Measuring Other Included Variables We used the extensive earnings histories as well as information on the incomes of spouses included in LINDA to calculate the economic incentive for remaining in the labor force. Here, there are two problems associated with using the earnings-from-labor variable. First, some of the observations are missing. This could be due to the fact that a worker is temporarily out of the labor force or out of the country. In such cases, we simply imputed the missing earnings observation by taking the average of the surrounding observations or, if the missing observations are from the beginning of the observed period of time, we imputed the missing observation by taking the average of the first three earnings observations. Second, and more importantly, when a worker retires, the contrafactual earnings from labor cannot be observed. Nevertheless, this earnings level is, of course, important for the retirement decision, and (as discussed in section 10.5.5) it is required in order to calculate the incentive measures. To predict future earnings, we simply take the average of earnings over the last three years of a worker’s observed earnings records indexed by the CPI. Lifetime earnings are measured as the sum of the constant and the fixed effect, from a fixed-effects regression on labor earnings between 1983 and 1997 on age, age squared, and dummy variables for each year included. The same strategy is used for measuring lifetime earnings for the spouse. Our data set does not contain any direct information that specifies to which occupational pension scheme each individual belongs. Given the importance of occupational pension schemes, this is, of course, essential information. However, as described in section 10.2, occupational pension schemes are associated with the different trade unions, which, in turn, are associated with different personnel categories and sectors of the economy. The FoBs contain information on the sector in which each individual works as well as socioeconomic group. This information can then be used to predict to which occupational pension scheme each individual belongs. We use information from the FoBs (censuses) in 1980, 1985, and 1990. If an individual has retired by the date of a census, it does not contain any information on either their socioeconomic group or sector of employment. This means that there is less information missing from the 1980 census compared to the other two censuses. Therefore, we used the 1980 census to

604

Mårten Palme and Ingemar Svensson

Table 10.10

Classification of Individuals into Pension Schemes Men

Private sector Blue-collar White-collar Government employees Central Local All

Table 10.11 Level 1 2 3 4 5 6

Women

Total

Number

%

Number

%

Number

%

6,188 4,972

39.62 31.83

3,137 2,614

21.17 17.64

9,325 7,586

30.64 24.92

2,110 2,349 15,619

13.51 15.04 100.00

2,348 6,721 15,165

15.84 45.35 100.00

4,458 9,070 30,439

14.65 29.80 100.00

Number of Observations at Different Education Levels (%) Description

Men

Women

Compulsory school only (7 or 8 years) Junior secondary school (9 or 10 years) Vocational school 2 years Upper secondary school 3 years; sixth form of comprehensive school (U.K.); senior high school (U.S.) Post-upper secondary school 2 years; junior college (U.S.), e.g., nursing school Post-upper secondary school 3 years, e.g., business administration, engineering or medicine, and PhDs

42.58 4.44 20.06

35.63 8.24 33.06

14.61

4.71

7.35

9.03

10.96

9.33

predict the occupational pension scheme. However, for missing values in this census, we used information from the 1985 census and, if necessary, also from the 1990 census. The resulting distribution among occupational schemes is shown in table 10.10. We used the same strategy to measure individual education level. In the first place, we used information from the 1993 education register. For missing observations, we used data from the 1994, 1995, and 1995 registers, respectively. Table 10.11 gives a short description of each education level along with the percentage share of observations in each category. Finally, we used controls for place of residence. Sweden is divided into twenty-five counties, and LINDA contains annual information on in which county an individual is registered for local taxes; this is the measure used for place of residence. 10.5 Construction of Incentive Measures 10.5.1 Definitions of the Incentive Measures We use four different measures of economic incentives for retirement decisions:

Income Security Programs and Retirement in Sweden

1. 2. 3. 4.

605

Benefit accrual, Effective tax or subsidy rate, Peak value, and Option value.

Benefit accrual measures the increase in pension wealth that a worker gains by postponing retirement and the claiming of benefits for one year. The present value of a worker’s pension wealth at year t if he retires at age r is defined as max age

SSW(r, t)


(2)

∑

stEt B(s, r),

s
r

where  is the discount factor and Et B(s, r) is the expected benefit at age s if the worker retires at age r. The expected benefit is defined as (3)

Et B(s, r)
p(st)q(st)BM(s, r) p(st)[1  q(st)]BS(s, r) [1  p(st)]q(st)S(s, r, t),

where BM(s, r) is the worker’s pension benefit at age s if they are married and retire at age r; BS(s, r) is the worker’s pension benefit at age s if they are not married and retire at age r; S(s, r) is the survivor’s benefit when the worker would have been aged s and retired at age r; p (st) is the probability of survival at time s conditional on survival at time t; and q (st) is the probability of the spouse surviving at age s conditional on survival at age t. The value S(s, r, t) depends on the spouse at time t as well as the retirement age r, while BM(s, r) and BS(s, r) are not dependent on t, since we assume perfect foresight about wages. The benefit accrual at age t is defined as man age

(4)

ACCR(t)


∑

max age

stEt B(s, t 2) 

s
t 2

∑

stEt B(s, t 1).

s
t 1

The effective-tax-or-subsidy-rate measure relates benefit accrual to the net wage if the worker stays in the labor market one additional year, that is, ACCR(t) TS(t)
, W(t 1)

(5)

where W(t 1) denotes labor earnings at age t 1. Peak value is defined as social security wealth (SSW) at its maximum value minus SSW at time t, that is,



max age

(6)

PEAK(t)


max

r
t 2,t 3, .... .71

∑ s
r

max age

stEt B(s, r) 

∑

s
t 1



stEt B(s, t 1) .

This measure is forward looking in the sense that it not only takes into account the immediate accrual in SSW of working an additional year, but also the accruals in future years.

606

Mårten Palme and Ingemar Svensson

Option value is related to the notion that an individual’s retirement decision also depends on how they value consumption and leisure at different ages. At any given age (t), it is assumed that the worker compares the expected present value of retiring at that age with the value of retiring at each age (r) in the future. The expected utility at age t of retiring at age r is defined as r1

max age

s
t

s
r

V(t, r)
∑ st [Y(s)] p(st)

∑

st [kB(s, r)] p(st),

where is the subjective discount rate, k reflects the marginal utility of leisure, and  measures marginal utility of consumption. The option value of retiring at age t is (7)

OPT(t)
V(t, r ∗)  V (t, t 1),

where r ∗ is the optimal retirement age; that is, the option value can be interpreted as the loss in utility of retiring today rather than preserving the option to retire at the preferred age. All of these incentive measures abstract from the possibility of retiring without claiming a benefit and of claiming a benefit without retiring. In an expanded model, an individual who is not retired and does not claim a benefit in one year could choose between four options the next year. 1. 2. 3. 4.

Continue to work and not claim a benefit Retire and start claiming a benefit Retire without claiming a benefit Claim a benefit without retiring

Here we continue to abstract from the numerous possibilities of partial retirement in the Swedish system. In a utility-maximizing framework, it is conceptually straightforward to take all four options into account. However, this approach complicates the retirement model considerably and, considering the extent to which our data are dominated by options 1 and 2, we do not think it is justified. The fact that we have relatively few observations on options 3 and 4 may be explained by the progressive income tax, which creates an incentive to smooth income over time. To the extent that options 3 and 4 are dominated by 1 and 2, they can be disregarded, as we do in our model. 10.5.2 Sources of Income after Retirement As pointed out in section 10.2, workers may use several different sources of income provided by the Swedish social insurance system after having permanently left the labor force. Moreover, different sources of income also implied varying income levels after retirement. In general, the replacement levels from DI and from UI and SI, in particular, are signifi-

Income Security Programs and Retirement in Sweden

607

cantly higher compared to the old age pension alternative. However, as explained in section 10.2, these sources of income are not available to all workers. It is only possible to observe ex post that an individual who actually receives support from a particular insurance is eligible for it. Whether or not an individual who continues to work one additional year is, in fact, qualified for benefits from a particular program cannot be determined. This complicates the construction of the incentive measures, since they are based on expected income after retirement. An extreme way of handling this problem is to assume that each worker, at each point of time, is eligible for support from the program that provides the most generous support. However, since this does not apply to some of the workers in the sample, such a measure would overestimate the true income after retirement for some of the workers, thereby underestimating the effect of economic incentives. Another extreme would be to assume that the old age pension is the only alternative available. But this would certainly not apply to those workers who are affected by the economic incentives inherent in labor market insurance, also thereby underestimating the effect of the economic incentives. A third alternative would be to assume that insurance is an alternative available only to those who, in fact, claim some kind of insurance when retiring. This procedure, however, would give rise to an endogeneity problem. If the retirement income from an insurance program, which is considerably higher than that from an old age pension, were assigned only at the point in time when a worker actually retired, and not in the preceding time period, it would be recorded as an increase in retirement income for the year retirement actually take place. This, in turn, would imply that the effect of economic incentives is overestimated. To avoid the problems involved in these approaches, we used a “probabilistic” or instrumental variable (IV) approach. To explain how the incentive measures are calculated using this approach, let us take SSW as an example; SSW from the old age pension system, which is available without any requirements regarding health status or unemployment, is denoted SSWOAP. A worker’s SSW, if they are eligible for labor market insurance, is denoted SSWLI. If the worker’s probability of access at a particular point in time is p, then their SSW can be written (8)

SSW
SSWOAP p(SSWLI  SSWOAP ).

Calculating this measure involves two problems. The first concerns calculation of SSWLI . Not only does Sweden’s welfare system offer several different labor market insurance programs, but workers are also able to shift between different programs. Ideally, SSWLI should be divided to account for different systems with a probability assigned to eligibility for each of them. However, as noted in section 10.2, considering all permutations of the main source of income over time resulted in 677 different

608

Mårten Palme and Ingemar Svensson

Fig. 10.2 Fitted values from regressing the average number of years with sickness or unemployment insurance as the main source of income before DI becomes the main source on a quadratic function in retirement age along with actual sample averages

combinations in the sample. In practice, it is obviously not feasible to calculate the economic incentive measures for hundreds of pathways. Several simplifications can be made, however. For example, a behavioral model could be applied to predict how workers choose among different insurance programs. On the other hand, according to section 10.2, most workers who retire by claiming labor market insurance follow a similar pattern. So, rather than applying a behavioral model, we used a common, synthetic “insurance path” to approximate the shifts between different insurance programs over time. It was noted in section 10.2 that the replacement rates for SI and UI are quite similar, particularly for workers who have an insured income below both the social security and the UI ceilings. This applies to most of the blue-collar workers in our sample for most of the time period under study. These are also the workers that are the most likely to initially finance their retirement from insurance. Thus, the accuracy in predicting income after retirement it is not likely to be impaired if, when choosing between SI and UI, the “right” insurance program is not used. As shown in figure 10.1, the length of the time before the transition to DI is highly dependent on a worker’s age when they permanently leave the labor force. In constructing the synthetic insurance path to retirement, we therefore use retirement age as a predictor of the length of the period with UI or SI before the transition to DI. In figure 10.2, predicted values from a regression of the average number of years with SI or DI before DI on a quadratic function of retirement age are added to the data shown in figure

Income Security Programs and Retirement in Sweden Table 10.12

Probit Estimates of Probability of Getting Disability, Sickness, or Unemployment Insurance Benefits Men

Age Age2 Occ2 Occ3 Occ4 Occ5 Occ6 Elev2 Elev3 Elev4 Elev5 Elev6 Intercept Controls for counties Pseudo R 2

609

Women



/s



/s

0.444 –0.003 –0.253 –0.148 –0.214 –0.214 –0.290 0.119 0.075 –0.035 –0.151 –0.281 –17.03

7.71 –6.20 –13.37 –6.05 –8.68 –8.50 –7.56 3.60 4.26 –1.05 –4.60 –8.88 –10.28

0.088 0.00 –0.255 –0.140 –0.189 –0.237 –0.404 –0.066 –0.051 –0.077 –0.201 –0.339 –6.853

45.69 0.00 –0.03 –5.89 –9.89 –6.09 –7.63 –2.45 –2.99 –2.16 –6.80 –10.64 –59.70

Yes 0.0761

Yes 0.0717

Notes: Occ
socioeconomic group; Elev
education level.

10.1. As is apparent from the figure, the function gives a very good fit to the observed averages.13 When calculating the incentives measures, we assume—for each age— that a worker receives SI or UI with a replacement rate of 80 percent during the number of years predicted by the quadratic function and, after that point, shifts to DI as their main source of income. Beginning at age sixtyfive, no workers are eligible for any type of labor market insurances, and consequently, all incentive measures are calculated using the old age pension alternative only. The second main problem with using the IV, or probability approach, concerns assigning the probability of being eligible for labor market insurance. Ideally, we would like to know, for each point in time, every worker’s probability of being eligible for labor market insurance. Since this information is not available, we estimated a probit equation where the dependent variable is the observed take-up rate of the labor market insurance programs. The specification of the probit equation is a polynomial in age, indicators for six education levels, indicators for four socioeconomic groups, marital status, and indicators for the twenty-five different counties in Sweden. The results from the probit regression are shown in table 10.12. 13. As there are very few observations on retirement before age fifty-four and as the estimated function actually increases between retirement ages fifty-two and fifty-three, we used 2.75 years up to age fifty-four. After that age, we applied the quadratic function.

610

Mårten Palme and Ingemar Svensson

10.5.3 Calculating Retirement Income Components and Income Taxes Calculations of the incentive measures for the individuals in the sample require calculating the old age pension, DI, SI, and UI benefits for each individual at every possible retirement age. It is also imperative that income taxes and housing allowances be taken into account. The frequent changes that have taken place in the Swedish system for housing allowances and income taxes not only require some approximations, but also raise issues about expectations. The most straightforward component to calculate in the economic incentive measures is income from the defined-benefit pension schemes— that is, the state basic and supplementary pension (ATP) and the four main occupational-pension schemes. Compared to income taxes, for example, the rules for these pension schemes do not change very often, and given the detailed earnings histories in our data set, we were able to calculate these incomes accurately. For the defined-contribution schemes—ITPK, the central government employees’ supplementary pension, and the post-reform pension scheme for blue-collar workers in the private sector—the size of the pension depends on the return of the particular fund that each may choose to manage their pension. Calculating the hypothetical outcome for these pensions therefore involves approximations. As regards the ITPK scheme, we used an algorithm for calculating the size of the pension provided by the insurance company (called SPP), which administers the largest share of the ITPK pensions. As suggested by SPP, we used an annual interest rate of 2.25 percent (net after taxes and administrative costs). We used the same algorithm for the supplementary pension scheme for central government employees, while taking into account that a lower share of the wage sum that is paid into the pension scheme compared to the ITPK, as well as the fact that this scheme went into effect after ITPK. In the case of the postreform pension scheme for blue-collar workers in the private sector, we used an algorithm provided by the company, which manages the largest share of these pensions. The calculations with respect to income taxes are more complicated. Although there has only been one major reform of the Swedish income tax system during the period covered by our data, several year-to-year changes have taken place. Since the number of years included is quite large, considering all changes would be unrealistic. To simplify matters, we chose an approximate strategy. We began by regressing the amount of taxes on taxable income. Since income tax rates are different for people still in the labor force compared to retirees, we have estimated separate functions for retired and nonretired. We used a third degree polynomial to model the marginal tax rates in the prereform income tax system and three linear segments for the postreform marginal tax rates. We then used the estimated

Income Security Programs and Retirement in Sweden

611

functions for each year to calculate individual taxes. A similar procedure was applied for housing allowances. The forward-looking measures—peak value and option value—require the worker to compare expected income among all possible retirement ages. This implies predicting hypothetical individual labor earnings after the workers had actually retired. For these calculations, we chose a latest possible retirement age of seventy-one. This, in turn, requires predicting labor earnings up until year 2010. For these out-of-sample predictions, we used the same strategy as outlined in section 10.4, that is, we took the threeyear averages for the years preceding the year with missing labor earnings. All three incentive measures used in this study involve individual expectations on future net income streams, which to a large extent are affected by future changes in benefits and income taxes. For example, the economic incentives for a retirement decision in the late 1980s are affected by whether or not an individual anticipated the 1991 income tax reform, and we assume that it was anticipated. Another example is the occupational pension scheme for blue-collar workers in the private sector. The trade union and the employers confederation agreed on a new pension scheme in 1996. However, another new agreement went into effect on 1 January 2000, and the pensions of some workers were affected retroactively. Needless to say, it is impossible to know which changes the workers anticipated. We assume that they anticipated all changes until 1 January 2000, but none thereafter. 10.5.4 Sample Estimates of Different Incentives Measures Tables 10.13–10.16 report the sample distribution of the incentive measures by age. Table 10.13 shows SSW, benefit accrual, and the tax or subsidy rate, that is, benefit accrual as a share of labor earnings minus payroll and income taxes in the individual’s last year of work. Table 10.14 reports the peak value and option value distributions for men. Tables 10.15 and 10.16 list the corresponding distributions for women. In addition to the median, the tenth and ninetieth percentiles of the distribution are given for the benefit accrual, peak value, and option value measures. Benefit accrual exhibits a marked increase at age fifty-seven. This is due to the rule in the STP scheme that at least three years of work between ages fifty-five and fifty-nine are required in order to be eligible for the STP pension. The increase is also more marked regarding the median in the male subsample because a larger share of the male labor force comprises bluecollar workers. The next, noticeable increase is at age fifty-nine. This may be explained by the way the pension schemes for central and local government employees are constructed (see section 10.2). This spike is substantially more marked in the female subsample, mainly because the largest share of the female labor force in Sweden works in the local government service sector.

Table 10.13

Social Security Wealth, Benefit Accrual, and Tax or Subsidy Rates for Men, by Age (1995 SKr; CPI used as deflator) Accrual

Last Age of Work 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

SSW Median 1,067,750 1,103,079 1,145,999 1,230,600 1,278,554 1,332,801 1,369,422 1,402,465 1,427,187 1,447,386 1,471,473 1,500,672 1,525,269 1,545,787 1,561,635 1,571,791

Median

10th Percentile

90th Percentile

SD

Tax Subsidy Rate Median

Previous Volume

14,863 15,260 38,432 10,210 11,004 –2,452 –11,171 –19,918 –28,814 –24,106 –23,631 –31,293 –39,412 –47,679 –56,298

–14,914 –15,693 –8,627 –18,052 –20,526 –31,416 –37,200 –46,657 –63,317 –59,611 –90,149 –72,891 –74,802 –83,300 –100,704

47,384 53,690 127,339 45,075 109,772 39,160 26,246 13,460 –5,090 –4,044 –7,890 –19,904 –29,946 –38,275 –46,628

71,558 72,946 86,832 76,746 100,362 77,248 72,601 59,917 58,197 53,811 58,009 39,608 27,395 25,987 24,605

0.225 0.220 0.072 0.250 0.249 0.330 0.392 0.457 0.520 0.478 0.177 0.232 0.291 0.359 0.440

0.231 0.221 0.056 0.153 0.146 0.350 0.358 0.253 0.290 0.313 0.036 0.085 0.128 0.169 0.193

Note: SD
standard deviation.

Table 10.14

Forward-Looking Incentive Measures (peak value and option value) for Men, by Age (1995 SKr; CPI used as deflator) Option Value

Peak Value

Last Age of Work

Median

10th Percentile

90th Percentile

SD

Median

10th Percentile

90th Percentile

SD

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

45,950 39,521 33,227 19,618 12,822 4,688 1,768 –38 –1,316 –727 –1,033 –2,149 –3,394 –4,609 –5,862

15,156 12,921 11,155 3,877 1,656 –1,798 –2,810 –3,895 –5,441 –5,080 –8,973 –6,211 –6,652 –8,248 –9,800

79,966 69,369 60,132 46,890 36,711 19,721 12,560 7,866 4,480 2,968 1,457 –413 –1,874 –3,113 –4,380

29,275 26,296 23,560 21,475 18,752 13,949 11,557 9,380 7,992 6,501 6,194 4,362 2,748 2,443 2,276

111,968 102,906 99,200 34,187 20,638 –727 –10,308 –19,595 –28,504 –23,711 –23,624 –31,287 –39,412 –47,679 –56,298

2,283 4,108 3,822 –14,539 –18,041 –30,358 –36,314 –45,173 –61,698 –55,125 –90,065 –72,446 –74,802 –83,300 –100,704

310,487 280,804 254,886 210,786 176,940 83,890 46,448 21,454 –1,070 –2,489 –7,876 –19,893 –29,946 –38,275 –46,628

167,153 156,075 146,097 138,783 126,933 98,434 84,042 69,578 61,668 51,264 52,730 39,478 27,395 25,987 24,605

Notes: SD
standard deviation. Parameter values for option value measure
0.97, 
0.75, and k
4.7.

Table 10.15

Social Security Wealth, Benefit Accrual, and Tax or Subsidy Rate for Women, by Age (1995 SKr; CPI used as deflator) Accrual

Last Age of Work 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

SSW Median 827,876 861,445 898,823 945,692 981,653 1,039,325 1,071,957 1,095,325 1,113,421 1,126,691 1,143,643 1,143,224 1,144,681 1,137,244 1,137,490 1,158,467

Median

10th Percentile

90th Percentile

SD

Tax Subsidy Rate Median

13,359 13,758 19,745 12,815 16,637 5,127 –2,507 –11,511 –16,910 –14,378 –12,733 –16,525 –20,105 –25,737 –32,298

–4,465 –5,362 –2,612 –7,463 –19,034 –21,227 –26,163 –34,581 –49,134 –45,515 –76,788 –54,609 –43,002 –45,050 –47,004

33,588 34,854 71,591 35,116 115,672 39,044 31,383 23,540 15,471 23,469 7,313 –4,154 –12,120 –18,568 –21,987

27,906 29,790 36,585 34,400 65,566 35,988 32,331 30,403 32,297 33,250 43,044 26,593 18,995 16,415 17,032

0.174 0.170 0.111 0.171 0.132 0.253 0.328 0.420 0.495 0.443 0.186 0.204 0.231 0.275 0.324

Note: SD
standard deviation.

Table 10.16

Forward-Looking Incentive Measures (peak value and option value) for Women, by Age (1995 SKr; CPI used as deflator) Option Value

Peak Value

Last Age of Work

Median

10th Percentile

90th Percentile

SD

Median

10th Percentile

90th Percentile

SD

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

55,472 50,667 45,767 39,903 34,936 28,565 24,400 20,704 17,598 14,823 11,896 9,574 7,146 4,693 2,316

31,397 28,052 24,544 20,487 17,085 12,420 9,976 7,937 5,996 4,935 3,265 2,502 1,622 893 454

78,714 72,457 66,102 58,841 52,382 44,776 39,244 33,919 29,139 24,545 20,120 15,913 11,744 7,663 4,040

19,403 18,227 17,016 15,699 14,368 12,835 11,537 10,177 8,934 7,541 6,390 5,053 3,857 2,689 1,611

107,531 97,138 88,609 56,027 43,210 10,656 –275 –9,347 –15,506 –12,558 –12,290 –16,523 –20,090 –25,723 –32,298

3,356 2,522 2,547 –2,275 –11,355 –19,482 –25,322 –34,069 –47,250 –42,778 –75,789 –53,451 –43,002 –45,015 –47,004

280,500 258,400 239,264 205,810 182,714 107,011 74,868 51,408 34,902 28,088 7,746 –4,115 –12,098 –18,539 –21,987

125,036 116,511 108,763 100,545 92,363 66,102 53,948 47,330 42,157 37,060 43,266 26,639 19,004 16,407 17,061

Notes: SD
standard deviation. Parameter values for option value measure
0.97, 
0.75, and k
1.25.

614

Mårten Palme and Ingemar Svensson

The forward-looking incentive measures, peak value and option value, show—although on different levels—a similar pattern as they both decrease over the observed ages. However, the quantiles of the option value measure evolve more smoothly over the ages, which is not surprising given the way in which it is calculated. The ninetieth percentiles of the peak value fall considerably between ages fifty-nine and sixty, which, again, reflects the makeup of the pension schemes for central and local governments employees. The last column in table 10.13 gives the tax or subsidy rates obtained in Palme and Svensson (1999) for a representative worker born in 1930, assigned to the STP occupational pension scheme, and a median income earner throughout their working career. As can be seen in the table, the tax rates are somewhat higher in the data set used here. This may be explained by differences in the way these two sets of tax rates were obtained. First, in Palme and Svensson (1999), labor market insurance was not considered in the base case. This may explain the different general levels up to age sixty-three since this type of insurances entails more generous replacement rates compared to old age pension. Second, the representative worker was assumed to be assigned to the STP pension scheme. The dip in the tax rate at age fifty-six is definitely related to the way this pension scheme is constructed, which is less marked in the present data set that comprises individuals from all occupational pension schemes. Third, the data set now used encompasses the prereform income tax system up until 1991 with substantially higher marginal tax rates. This difference may explain the somewhat higher tax rates after age sixty-three. Finally, the results from the present data set are medians from the distribution of tax rates, rather than tax rates from the median income earner. The direction in which this difference works is not clear.

10.6 Empirical Model and Results 10.6.1 Empirical Specification We use the following empirical specification for the retirement decision model: (9) Rit
0 1ACCit 2 SSWit 3 AGEit 4 PREARNit 5 EARNit 6 PREARNit  EARNit 7SPEARNit Xit vit , where ACCit is the measure of accrual at time t; SSWit is the net present value of SSW discounted back to time t; AGEit represents the individual’s age either by a linear variable or by indicators for each age; PREARN is the individual’s predicted earnings at time t and the square of this measure; EARN is a measure of the individual’s lifetime earnings and its square; SPEARN is lifetime earnings of the spouse, its square, and the spouse’s net

Income Security Programs and Retirement in Sweden

615

SSW discounted back to time t; and X is a set of personal characteristic variables, including marital status, education level (1–6),14 socioeconomic group (1–4),15 and indicators for each of Sweden’s twenty-five counties (cf. section 10.4 for the construction of these variables). The focus of our interest is on the dual effect of economic incentives for retirement created by the social security system. Higher SSW will increase an individual’s demand for all goods—including leisure time (i.e., retirement). This effect is measured by SSW, and we expect a positive sign on this variable. However, if the accrual from working one additional year is sufficiently large, then the substitution effect, induced by the accrual, will dominate the income effect, and the worker will choose to continue to work. Therefore we expect the sign on ACC to be negative. As explained in section 10.5, we use three different measures of accrual: benefit accrual, peak value, and option value. Among the changes in the Swedish income tax and security systems outlined in section 10.2, the following are particularly helpful for identification of the empirical model: the income tax reform (1991); the reform of the occupational pension scheme for central government employees (1992); the maturing of the fully funded supplementary pension scheme for white-collar workers in the private sector (ITPK, in 1977) and for central government employees (1991); and the transition from a STP to a fully funded pension scheme for blue-collar workers in the private sector (1996). Although the data set has a panel structure, identification of the empirical model prevents us from using, for example, a fixed-effect approach to control for unobserved heterogeneity. We use the data set as a cross section in the estimation and use observable demographic characteristics to control for heterogeneity. This means that most individuals are included in the data several times. To correct the standard errors for dependence between different observations on the same individual, we use the Huber-White sandwich estimator, which allows for general dependence within clusters of observations. 10.6.2 Sample Characteristics Since only workers older than age fifty are included in the sample, the first observations on the cohorts born between 1934 and 1940 were not used in the estimation. The final sample consists of 127,390 observations (from 15,619 individuals) in the male subsample and 123,979 observations (from 14,820) individuals in the female subsample. Table 10.17 reports means and standard deviations of most of the variables included. To save space, we exclude descriptive statistics of the twenty-five county dummies. 14. The education levels are described in table 10.11. 15. The socioeconomic groups are explained in table 10.10.

616

Mårten Palme and Ingemar Svensson

Table 10.17

Means and Standard Deviations of the Variables Included in the Sample Used in Estimations (monetary values in SKr; 1997 prices deflated by the CPI) Men

Retired Benefit accrual Peak value Option value SSW Log lifetime earnings Predicted log earnings Education 1 Education 2 Education 3 Education 4 Education 5 Education 6 Occupation 1 Occupation 2 Occupation 3 Occupation 4 Married Log lifetime earnings (spouse) SSW (spouse)

Women

Mean

SD

Mean

SD

0.055 15,196 107,746 37,044 1,265,103 11.73 12.30 0.446 0.039 0.176 0.143 0.077 0.119 0.382 0.331 0.129 0.158 0.754 24.01 940,866

0.228 83,856 170,984 35,893 437,801 0.39 0.43 0.497 0.194 0.381 0.350 0.266 0.324 0.486 0.471 0.336 0.354 0.431 15.08 762,528

0.060 13,769 104,804 48,365 933,227 9.81 11.89 0.351 0.079 0.321 0.406 0.095 0.103 0.200 0.180 0.154 0.466 0.723 23.58 1,050,550

0.238 37,361 128,149 24,210 325,278 0.42 0.45 0.477 0.269 0.467 0.209 0.293 0.304 0.400 0.384 0.361 0.499 0.448 15.43 805,860

Notes: SD
standard deviation. For the option value measure, we use
0.97, 
0.75, and k
4.7 for the male subsample and
0.97, 
0.75, and k
1.25 for the female subsample.

10.6.3 Estimation Results The results from the probit regression on retirement decisions are shown in table 10.18 for men and table 10.19 for women. Each of the tables shows six different specifications. The three accrual measures described in section 10.5.1 are used in alternative specifications (i.e., the one-year benefit accrual, the peak value, and the option value accrual measures, respectively). For each measure of accrual, there is one specification with a linear age variable (M1) and one with indicator variables for each one-year age group (M2). To evaluate the model specification, we tested for joint significance of the main groups of variables included—that is, the incentive variables and the additional variables for the workers’ economic situation, education level, socioeconomic group, family income, and county of residence. The results show that all these groups are significantly different (on the 5 percent level) from zero in all specifications. We estimated the parameters in the option value measure by a grid search in which the maximum of the log likelihood from the M2 model was

Table 10.18

Results from Probit Regressions on Individual Retirement Decision (men) Accrual M1

ACCR/106 1 million change SSW/106 1 million change Lifetime earnings Lifetime earnings2 Predicted earnings Predicted earnings2 Lifetime  Predicted (Lifetime  Predicted) 2 Education2 Education3 Education4 Education5 Education6 Occupation2 Occupation3 Occupation4 Age Married Lifetime earnings (spouse) Lifetime earnings (spouse) 2 SSW (spouse)/106 Indicators for age Indicators for counties Pseudo R 2 Log-likelihood

–0.21 (–2.08) –0.02 0.55 (10.81) 0.04 –2.60 (–1.81) 0.02 (0.30) –0.41 (–0.26) –0.08 (–1.79) 0.35 (2.34) –0.01 (–2.90) 0.21 (6.63) 0.18 (10.70) 0.12 (6.09) 0.10 (3.90) 0.04 (1.56) –0.19 (–10.83) 0.02 (1.01) –0.18 (–8.84) 0.12 (38.26) –0.11 (–2.37) 0.02 (1.67) –0.01 (–1.70) 0.04 (3.26) No Yes 0.1593 –24,654

Peak Value M2

–0.09 (–0.81) –0.01 0.57 (10.99) 0.04 –2.51 (–1.73) 0.01 (0.14) –0.57 (–0.36) –0.07 (–1.76) 0.37 (2.39) –0.01 (–2.87) 0.21 (6.55) 0.18 (10.47) 0.12 (5.92) 0.09 (3.59) 0.04 (1.42) –0.19 (–10.60) 0.02 (1.11) –0.19 (–8.90)

–0.11 (–2.51) 0.02 (1.95) –0.01 (–1.97) 0.04 (3.37) Yes Yes 0.1813 –24,011

Option Value

M1

M2

M1

–0.93 (–10.12) –0.07 0.34 (6.41) 0.02 –2.76 (–1.92) 0.11 (1.45) 1.47 (0.93) –0.09 (–2.14) 0.13 (0.91) –0.01 (–1.78) 0.22 (6.91) 0.18 (11.19) 0.13 (6.87) 0.12 (4.71) 0.07 (2.50) –0.17 (–9.77) 0.03 (1.38) –0.18 (–8.68) 0.11 (38.39) –0.05 (–1.21) 0.03 (2.45) –0.01 (–2.51) 0.04 (3.01) No Yes 0.1621 –24,571

–0.92 (–9.94) –0.07 0.35 (6.43) 0.03 –2.80 (–1.92) 0.011 (1.39) 1.40 (0.87) –0.09 (–2.16) 0.15 (0.96) –0.01 (–1.77) 0.22 (6.85) 0.18 (10.99) 0.14 (6.75) 0.12 (4.44) 0.07 (2.41) –0.17 (–9.53) 0.03 (1.46) –0.19 (–8.78)

–5.11 (–9.39) –0.36 0.31 (5.50) 0.02 –2.43 (–1.71) 0.10 (1.31) 1.26 (0.80) –0.08 (–1.91) 0.12 (0.83) –0.01 (–1.62) 0.22 (6.92) 0.18 (11.23) 0.14 (6.88) 0.12 (4.75) 0.07 (2.46) –0.17 (–9.68) 0.03 (1.42) –0.18 (–8.82) 0.11 (33.28) –0.04 (–0.93) 0.02 (2.11) –0.01 (–2.16) 0.03 (2.98) No Yes 0.1612 –24,599

–0.06 (–1.29) 0.04 (2.77) –0.01 (–2.83) 0.04 (3.11) Yes Yes 0.1841 –23,928

M2 –6.74 (11.42) –0.49 0.24 (4.16) 0.02 –2.55 (–1.76) 0.12 (1.58) 1.59 (1.00) –0.10 (–1.94) 0.08 (0.53) –0.01 (–1.33) 0.22 (6.91) 0.19 (11.24) 0.15 (7.04) 0.12 (4.73) 0.07 (2.59) –0.17 (–9.16) 0.03 (1.62) –0.19 (–8.90)

–0.02 (–0.47) 0.03 (2.63) –0.01 (–2.68) 0.03 (3.01) Yes Yes 0.1844 –23,920

Notes: Results based on 15,619 individuals and 127,390 observations. Numbers in parentheses are t-values.

Table 10.19

Results From Probit Regressions on Individual Retirement Decision (women) Accrual

ACCR/106 1 million change SSW/106 1 million change Lifetime earnings Lifetime earnings2 Predicted earnings Predicted earnings2 Lifetime • Predicted (Lifetime • Predicted) 2 Education2 Education3 Education4 Education5 Education6 Occupation2 Occupation3 Occupation4 Age Married Lifetime earnings (spouse) Lifetime earnings (spouse) 2 SSW (spouse)/106 Indicators for age Indicators for counties Pseudo R 2 Log-likelihood

Peak Value

M1

M2

M1

–1.00 (–4.06) –0.08 0.35 (7.28) 0.03 –5.87 (–3.18) 0.33 (4.22) 4.21 (2.52) –0.20 (–3.05) 0.14 (0.79) –0.00 (–1.80) 0.05 (2.14) 0.06 (3.88) 0.06 (1.99) –0.01 (–0.45) –0.09 (–3.21) –0.12 (–5.22) –0.04 (–1.57) –0.13 (–7.14) 0.15 (51.92) 0.27 (4.37) 0.01 (0.54) –0.01 (–0.93) –0.01 (–0.60) No Yes 0.1736 –23,615

–0.81 (–0.34) –0.01 0.44 (8.93) 0.04 –5.67 (–2.94) 0.28 (3.32) 4.27 (2.44) –0.23 (–3.24) 0.21 (1.12) –0.00 (–1.76) 0.04 (1.63) 0.06 (3.53) 0.05 (1.43) –0.01 (–0.54) –0.11 (–3.61) –0.12 (–4.97) –0.12 (–4.97) –0.02 (–1.10)

–1.42 (–10.39) –0.10 0.07 (1.27) 0.01 –4.68 (–2.47) 0.40 (5.07) 5.94 (3.50) –0.19 (–3.05) –0.19 (–1.06) –0.00 (–0.37) 0.06 (2.27) 0.07 (4.15) 0.07 (2.23) –0.00 (–0.07) –0.08 (–2.75) –0.11 (–5.12) –0.04 (–1.81) –0.13 (–7.27) 0.14 (50.07) 0.29 (4.61) 0.01 (0.30) –0.00 (–0.66) –0.01 (–0.58) No Yes 0.1762 –23,540

0.29 (4.51) 0.00 (0.33) –0.00 (–0.71) –0.01 (–0.71) Yes Yes 0.1977 –22,926

M2 –1.29 (–9.69) –1.70 0.13 (2.16) 0.01 –4.60 (–2.34) 0.36 (4.34) 6.02 (3.38) –0.22 (–3.20) –0.12 (–0.66) –0.00 (–0.38) 0.05 (1.80) 0.06 (3.87) 0.06 (1.70) –0.00 (–0.05) –0.09 (–3.04) –0.11 (–4.93) –0.04 (–1.52) –0.13 (–6.82)

0.32 (4.87) 0.00 (0.12) –0.01 (–0.48) –0.01 (–0.68) Yes Yes 0.2004 –22,850

Option Value M1 –23.4 (–20.43) –1.87 –0.47 (–7.13) –0.03 –6.31 (–3.39) 0.66 (7.32) 5.53 (3.56) –0.07 (–1.26) –0.46 (–2.85) –0.00 (0.04) 0.09 (3.44) 0.09 (5.72) 0.11 (3.32) 0.04 (1.74) –0.02 (–0.82) –0.03 (–1.48) –0.02 (–0.67) –0.21 (–11.34) 0.09 (24.64) 0.32 (4.96) 0.00 (0.27) –0.02 (–0.67) –0.02 (–1.12) No Yes 0.1828 –23,351

M2 –24.0 (–21.67) –1.87 –0.48 (–7.14) –0.04 –6.22 (–3.25) 0.65 (6.80) 8.82 (3.56) –0.08 (–1.47) –0.46 (–2.68) –0.00 (0.19) 0.08 (2.96) 0.09 (5.52) 0.09 (2.78) 0.05 (1.80) –0.04 (–1.10) –0.03 (–1.29) –0.01 (–0.42) –0.21 (–10.95)

0.35 (5.28) 0.00 (0.04) –0.02 (–0.43) –0.02 (–1.23) Yes Yes 0.2083 –22,624

Notes: Results based on 14,820 individuals and 123,979 observations. Numbers in parentheses are t-values.

Income Security Programs and Retirement in Sweden

619

used as a value function. Although the log-likelihood function for men was shown to be very flat with respect to  and k, a maximum was found at
0.97, 
0.75, and k
4.7. These parameter values were then used to obtain the estimates presented in table 10.18. However, we also estimated the model using approximately the parameter values obtained by Stock and Wise (1990), i.e.,
0.97, 
0.75, and k
1.25. These values gave smaller coefficient estimates for the option value variable (coefficient 106 at –4.28 [–4.56] for the M1 and –4.98 [–5.36] for the M2 specification, respectively), but slightly larger estimates for the SSW variable (coefficient 106 at 0.43 [7.28] for the M1 and 0.41 [6.86] for the M2 specification, respectively). For the female subsample, the grid search did not result in any maximum since the log-likelihood function was decreasing in k in the permitted region of values for k. For the estimates reported in table 10.19 we used
0.97, 
0.75, and k
1.25. The most important result from the estimates is that economic incentives seem to matter for retirement behavior: The coefficients for the SSW variable are, in general, significantly positive, and those for the different accrual measures are, as expected, significantly negative. There are, however, some exceptions to this pattern. In the male sample, the benefit accrual measure is not significantly different from zero in the M2 specification. In the female sample, the variable for SSW of the spouse is insignificant in all specifications, and the variable for the individual’s own SSW is significant with an unexpected sign in both specifications including the option value measure. The magnitude of the effect from the economic incentive variables is hard to quantify from the parameter estimates alone. According to the results of the implied probability effect of changing the incentive measure by SKr1 million, the effect appears to be very small. It should be kept in mind, however, that (as is evident from table 10.17) the average probability of retirement is fairly low in the sample (0.055 for men and 0.060 for women)— that is, an implied probability effect of 0.03 corresponds to about a 50 percent increase in retirement. To gain a better understanding of the implications for the magnitude of the effects of economic incentives from the estimates, we simulated the effects of two hypothetical reforms (cf. section 10.7). It is evident from the results that the forward-looking incentive measures, peak value and option value, work better than the benefit accrual measure. The benefit accrual coefficient is only significantly different from zero with the linear specification in age. However, also for this specification, the log-likelihood values are larger for the models with the peak value and option value measures. Considering the design of many of the pension schemes, for example, the STP scheme in which three years of earnings between ages fifty-five and fifty-nine are required to be eligible for any pension at all, this outcome was expected.

620

Mårten Palme and Ingemar Svensson

There is no unambiguous ranking between the peak value and the option value measures: The log likelihood for the option value specification is lower in the male subsample for the M1 specification, but higher for the M2 specification and in the female subsample. In the female subsample, the measure of SSW takes an unexpected negative sign when the option value incentive measure is used. Another issue that could not be resolved on the basis of the results reported in tables 10.18 and 10.19 alone is the extent to which the economic incentive measures capture the observed pattern of retirement behavior. Figure 10.3 shows the implied probability effect of the age indicator variables, along with the actual hazard rate out of the labor force, by age for the male and female subsamples. Our interpretation of this result is that the economic incentive measures do not fully capture the age pattern of retirement. The spike at age sixty-five tells us that collective agreements on retirement ages, as described in section 10.2.5, have considerable influence on retirement behavior. 10.7 Simulations To evaluate the implications of the estimates, we simulated the effects of two hypothetical policy reforms on labor force participation. For the male subsample, all three measures of accrual were used in the simulations, whereas we only use one measure for the female subsample. Since we were not able to estimate the parameters in the option value measure for females, we used the peak value measure. In the first policy reform, the age of eligibility for all programs is delayed by three years. That is, the state old age pension, as well as the occupational pension programs, could be claimed beginning at age sixty-three rather than age sixty. The actuarial adjustments of pension levels, within both the public and occupational pension systems, start from age sixty-eight rather than sixty-five. Moreover, the probabilities for access to labor market insurance (DI, SI, and UI) is delayed by three years. The probit regression model for probability of insurance eligibility was used to predict eligibility probabilities under this policy reform. In the second reform, the entire income security program (the state old age pension, occupational pensions, and labor market insurance) is replaced by a hypothetical common pension scheme. This scheme replaces 60 percent of predicted earnings16 at age sixty if it is claimed at the normal retirement age of sixty-five. The pension could be claimed beginning at age sixty or delayed until age seventy. There is a 6 percent actuarial reduction for each year of retirement before age sixty-five and a 6 percent increase for each year retirement is delayed after age sixty-five. Since there is no labor 16. We use the strategy for predicting future earnings described in section 10.5.

Income Security Programs and Retirement in Sweden

621

A

B

Fig. 10.3 The implied probability effect of the age dummies from the M2 specification with different incentive measures along with the actual hazard rate out of the labor force by age: A, men; B, women

market insurance, a worker who decides to retire before age sixty receives no income until age sixty. Figure 10.4 compares the sample median SSW by age, under the three policy alternatives: the actual and the two hypothetical policies for males and females. For males, it can be seen that there is a substantial difference between the actual system and the two hypothetical schemes. At age sixty,

A

B

Fig. 10.4 Median social security wealth; actual and simulated under policy alternatives 1 and 2, respectively: A, men; B, women

Income Security Programs and Retirement in Sweden

623

Fig. 10.5 Median benefit accrual; actual and simulated under policy alternatives 1 and 2, respectively, men

the SSW is about 20 percent lower under both policy alternatives. As regards the first policy alternative, this difference is due to delaying of all benefits, and in the case of the second policy alternative, it is due to the abolition of labor market insurance and a reduction in replacement rates. For females, panel B shows that the median SSW is very similar under the actual and the second policy alternative for all ages. The most likely explanation to this outcome is that, since the pension income is determined by predicted earnings at age sixty, there is no reduction for being temporarily out of the labor force before age sixty, as there is in the actual system. For both males and females, all differences are counteracted by progressive income taxes and housing allowances. Figures 10.5 to 10.7 show the three alternative measures of accrual by age under the different policy alternatives. Figure 10.5 shows that there are two spikes in the benefit accrual of working: the first one is for working during one’s fifty-seventh year under the actual system (due to the rules in the occupational pension scheme for blue-collar workers) and the second one is for working during one’s fifty-ninth year (due to the rule that pension benefits could not be claimed until age sixty). Under the first policy alternative, these spikes are delayed by three years under the first policy alternative and are entirely removed under the second policy alternative. It can also be seen that benefit accrual is zero up to age sixty under the second policy alternative. Finally, it can be seen that benefit accrual is zero up to

A

B

Fig. 10.6 Median peak value; actual and simulated under policy alternatives 1 and 2, respectively: A, men; B, women

Income Security Programs and Retirement in Sweden

625

Fig. 10.7 Median option value; actual and simulated under policy alternatives 1 and 2, respectively, men

age sixty under the second policy alternative, since earnings before age sixty do not affect the benefit after retirement under this policy alternative. The 6 percent actuarial adjustment under the second policy alternative is not enough to generate positive benefit accrual after age fifty-nine. The peak value measures under the different policy alternatives, shown in figure 10.6 for men and women, also reflect the particulars of each policy: The fall in peak value, due to the rules in the blue-collar worker pension scheme, is delayed by three years under the first policy alternative, while the peak value under the second policy alternative is constant up to age sixty. The median option value measure in figure 10.7 shows a marked difference between the first policy alternative on one hand, and the actual system and the second policy alternative on the other. The option value of not retiring is, of course, higher if the benefits are delayed by three years. We carried out three different simulations for each policy change. The first simulation, S1, used the model with the linear specification in age (M1). The second, S2, used the model with age dummies (M2) without changing anything except the measures of economic incentives according to the two proposed policy changes. The specification with indicator variables for each age group is likely to be overparameterized in the sense that the estimated age pattern of retirement reflects some features of the pension system, in addition to variations in preferences for leisure by age and

626 Table 10.20

Mårten Palme and Ingemar Svensson Average Retirement Rates and Retirement Ages in Simulations Baseline

Model Men S1 and benefit accrual S2 and benefit accrual S3 and benefit accrual S1 and peak value S2 and peak value S3 and peak value S1 and option value S2 and option value S3 and option value Women S1 and peak value S2 and peak value S3 and peak value

Policy 1

Policy 2

Retirement Rate

Retirement Age

Retirement Rate

Retirement Age

Retirement Rate

Retirement Age

0.0577 0.0577 0.0577 0.0576 0.0577 0.0577 0.0577 0.0577 0.0577

62.60 62.57 62.57 62.60 62.57 62.57 62.57 62.57 62.57

0.0461 0.0463 0.0246 0.0480 0.0483 0.0264 0.0457 0.0461 0.0269

63.40 63.37 65.33 63.24 63.20 65.12 63.44 63.37 65.07

0.0447 0.0447 0.0447 0.0526 0.0529 0.0529 0.0507 0.0529 0.0529

63.52 63.50 63.50 62.91 62.88 62.88 63.10 62.92 62.92

0.0626 0.0625 0.0625

61.97 61.95 61.95

0.0593 0.0590 0.0288

62.15 62.14 64.36

0.0730 0.0724 0.0724

61.37 61.38 61.38

Note: Since workers older than age 70 are not included in the data, we set the hazard rate at age 70 to one (100%).

institutions on the labor market. This, in turn, implies that the predicted effect of a change in the social security system is underestimated. We therefore use the outcome from this simulation as a lower bound. In the third simulation, S3, we again used the M2 model, but here, for the first policy alternative, each dummy variable is incremented by three years in addition to the changes done in S2. This procedure corresponds to the (unrealistic) assumption that the entire age pattern of retirement behavior estimated by the age indicator variables is determined by the social security system. We use this simulation as an upper bound on the predicted outcome. For the second policy alternative, the ages for early and normal retirement coincide with those in the actual Swedish system. This means that the S2 and S3 simulations coincide. Since we used three different measures of accrual for each of the three simulations in the male subsample, and only peak value in the female subsample, there are nine simulations for men and three for women. We present three different outcomes from the simulations. First, table 10.20 shows the predicted average retirement rate and age for each policy alternative and each simulation. Figures 10.8–10.19 show predicted hazard rates out of the labor force and cumulative distribution functions of retirement by age. Each figure shows three graphs: one for the model prediction of actual policy and one for each of the policy alternatives. It is evident from table 10.20 and the figures that most models predict the largest effect on retirement behavior from the first policy alternative. The

A

B

Fig. 10.8 S1 using benefit accrual estimates, men: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.9 S2 using benefit accrual estimates, men: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.10 S3 using benefit accrual estimates, men: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.11 S1 using peak value estimates, men: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.12 S2 using peak value estimates, men: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.13 S3 using peak value estimates, men: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.14 S1 using option value estimates, men: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.15 S2 using option value estimates, men: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.16 S3 using option value estimates, men: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.17 S1 using peak value estimates, women: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.18 S2 using peak value estimates, women: A, predicted hazard rates; B, predicted CDF

A

B

Fig. 10.19 S3 using peak value estimates, women: A, predicted hazard rates; B, predicted CDF

Income Security Programs and Retirement in Sweden

639

exception from this pattern is the models using the benefit accrual measure. For the peak and option value measures, the predicted delay in average retirement age is between 0.63 and 0.87 years for first policy alternative for the S1 and S2 simulations, compared to between 0.31 and 0.53 for the second policy alternative. The predictions of average delay in retirement for the S3 simulation are between 2.50 and 2.76 for the first policy alternative. The difference of these predictions vis-à-vis those of S1 and S2 shows that there is a large element of uncertainty in the predictions from all our models. There are two different sources of this uncertainty. The first source regards the extent to which spikes at ages sixty and sixtyfive in the hazard rates for retirement can be attributed to economic incentives or institutional arrangements on the labor market. For the spike at age sixty, it is likely that it can be attributed largely to economic incentives. In most pension plans, pensions can be claimed starting from age sixty. It is also the case that most workers have fulfilled the requirements for number of years of contributions to the plan at age sixty. For the second spike at age sixty-five, however, it is likely that it can be attributed largely to the collective agreements on retirement age between the unions and the employers’ confederation on the labor market. Thus, to some degree, determining the extent to which the reform proposal on delayed eligibility age also includes delayed retirement age in the collective agreements is a matter of interpretation of the reform proposal. The second source of uncertainty is the extent to which the dummy variable specification for the way age affects retirement reflects changes in preferences for being retired due to aging (primarily through changes in health status) or the extent to which it reflects unmeasured economic incentives from the income security system. If it is primarily due to changes in preferences due to aging, the lower bound is a better prediction of the outcome. However, without detailed data on changes in health status, it is not possible to disentangle these effects. The results in the female sample can be seen in figures 10.29–10.34 and in table 10.20. The simulations of the first policy reform show that there is a smaller predicted behavioral effect of the reform for women than for men. In S1 and S2, there is only, on average, a 0.18 or 0.19 years delay of retirement due to the reform. The big effect in S3 shows that the uncertainty of the predictions is even greater for females compared to males. For the second policy alternative, there is an effect toward earlier retirement in the female sample. The background to this result contains several different elements. First of all, most pension plans in Sweden have a linear reduction in the replacement rates if the worker has contributed less than thirty years to the plan. This is binding for a large share of the women studied here, but a much smaller share of the men. Since the pension under the second policy alternative is determined as a fraction of earnings at age

640

Mårten Palme and Ingemar Svensson

sixty, the SSW is actually lower for a large share of the female sample, compared to the baseline policy. This can be seen for the median SSW in figure 10.6. The effect through the SSW of the spouse is reduced, or actually reversed, since the coefficient estimates of the spouse’s SSW were insignificant (with the unexpected sign) in the female sample. Finally, since there are no actuarial adjustments before age sixty under the second policy alternative, the effect through the peak value accrual measure under the baseline policy is thus reduced under the second policy alternative. This explains the higher hazard rates to retirement under the second policy alternative before age sixty. 10.8 Conclusions The results of the econometric analysis in our study support the notion that economic incentives matter for the retirement decision. The parameter estimates for the economic incentive variables were, in general, significantly different from zero with the expected signs. This applies in particular to the male sample. The results also show that the forward-looking accrual measures, the peak and option value measures, work somewhat better than the one-year benefit accrual measure, since they give a better fit to the data. Our simulations of two hypothetical policy reforms showed that there would be a substantial effect on labor force participation from changing the economic incentives of retirement. However, there is a large element of uncertainty in such predictions in the sense that the lower bound on the predictions, where the one-year age group dummies were maintained, predicted much lower labor force participation rates compared to the upper bound on the predictions, where the age dummies were shifted by three years. The extent to which the age indicator variables capture features of income security programs, which are unmeasured by the economic incentive variables, changes in preferences for leisure by age, collective agreements on normal retirement ages in the labor market, or social norms regarding retirement, is an open question to be explored in further research.

References Aronsson, Thomas, and James R. Walker. 1997. The effects of Sweden’s welfare state on labor supply incentives. In The welfare state in transition, ed. Richard B. Freeman, Birgitta Swedenborg, and Robert Topel, 203–65. Chicago: University of Chicago Press. Gruber, Jonathan, and David A. Wise. 1999. Social security and retirement around the world. Chicago: University of Chicago Press.

Income Security Programs and Retirement in Sweden

641

Hansson-Brusewitz, Urban. 1992. Labor supply of elderly men: Do taxes and pension systems matter? Ph.D. diss. Uppsala University, Department of Economics. Uppsala, Sweden. Kangas, Olli, and Joakim Palme. 1989. Public and private pensions: The Scandinavian countries in a comparative perspective. Stockholm University, Institute for Social Research. Working Paper no. 3. Palme, Mårten, and Ingemar Svensson. 1999. Social security, occupational pensions and retirement in Sweden. In Social security and retirement around the world, ed. Jonathan Gruber and David A. Wise, 355–402. Chicago: University of Chicago Press. Skogman Thoursie, Peter. 1999. Disability and work in Sweden. Ph.D. diss. Stockholm University, Swedish Institute for Social Research. Stock, James, and David A. Wise. 1990. Pensions, the option value of work, and retirement. Econometrica 58 (5): 1151–80. Sundén, Annika. 1994. Early retirement in the Swedish pension system. Ph.D. diss. Cornell University, Department of Economics. Wadensjö, Eskil. 1989. Varför har vi normal pensionsålder (Why do we have a normal retirement age)? In Vingarnas trygghet. Arbetsmarknad, ekonomi och politik (The confidence of the wings), ed. Eskil Wadensjö, Åke Dahlberg, and Bertil Holmlund, 173–85. Lund: Dialogos. Wadensjö, Eskil, and Edward E. Palmer. 1996. Curing the Dutch disease from a Swedish perspective. In Curing the Dutch disease: An international perspective on disability policy reform, ed. Leo J. M. Aarts, Richard V. Burkhauser, and Philip R. de Jong, 133–56. Aldershot: Avebury.

11 Pension Incentives and the Pattern of Retirement in the United Kingdom Richard Blundell, Costas Meghir, and Sarah Smith

11.1 Introduction Like many other Organization for Economic Cooperation and Development (OECD) countries, the United Kingdom has been experiencing a trend towards earlier labor market exits among older, particularly male, workers. The proportion of men aged sixty to sixty-four in employment halved from 1968, when 80 percent were employed, to only a little over 40 percent in 1996. The fall in the proportion of older men who were in full-time employment was even greater than the fall in the proportion in any form of employment, with a relative shift within the employed to self-employment and part-time employment (see figure 11.1). Female employment has not experienced the same downward trend—but this contrasts with rising participation among most other age groups across the same period. This paper looks at the extent to which these trends might be explained by the financial incentives in the pension system that people faced when making their retirement decisions. In doing so, we focus not only on the Richard Blundell is Leverhulme Research Professor of Economics at University College, London, and research director at the Institute for Fiscal Studies (IFS). Costas Meghir is professor of economics at University College, London, and deputy director of the Economic and Social Research Council (ESRC) Research Centre at the IFS. Sarah Smith is a lecturer at the London School of Economics and a research associate at the IFS. She worked on the project while at the IFS. This paper forms part of the International Social Security project at the National Bureau of Economic Research (NBER). The authors are grateful to Jon Gruber, David Wise, and the participants of that project, and to Richard Disney and Paul Johnson for comments. The Department of Social Security is thanked for financing the primary analysis of the second wave of the Retirement Survey and for making the data available. Both waves of the Retirement Survey are now deposited at the ESRC Data Archive at the University of Essex. This research of part of the program of research by the ESRC Centre for the Micro-Economic Analysis of Public Policy at IFS, and we are grateful to the ESRC for funding.

643

Fig. 11.1 Trends in employment Source: Family Expenditure Survey 1968–1996.

Pension Incentives and Pattern of Retirement in the United Kingdom

645

pensions provided by the state, but also on employer-provided pensions and on other state benefits, such as invalidity benefit, all of which have played a crucial role in the United Kingdom. Compared to many other European countries, the United Kingdom stands out as having a high level of coverage by private pensions and, at least in recent years, a trend towards less generous state pension provision. This has not always been the case. In the 1970s, the trend was going the other way toward more generous state provision. The main element of the state pension system, the basic state pension, was increased each year in line with earnings or prices, whichever was the greatest. In 1978, a new second-tier earnings-related pension—the State Earnings Related Pension Scheme (SERPS)—was introduced that was originally intended to pay a pension worth 25 percent of an individual’s best twenty years of earnings. However, SERPS was never a universal scheme for all employees. When it was introduced, workers who already belonged to a (defined-benefit) occupational pension could opt out of the state scheme, as well as pay reduced National Insurance contributions, so long as their occupational scheme guaranteed at least the same pension as SERPS. This applied to more than half of all employees and to more than two-thirds of male employees. Since the early 1980s, successive reforms have cut back the generosity of state pension provision. The indexation of the basic state pension to earnings lasted only until 1982, since when it has been formally indexed to prices and has fallen relative to average earnings. Reforms to SERPS in 1986 and 1995 have reduced its generosity for anyone retiring after 2000. Also, the state pension age for women, currently sixty, is set to increase to sixty-five by 2020. These reforms were coupled with further encouragement for individuals to make a private pension provision. In 1988, the right to opt out of SERPS was extended to those with a defined-contribution scheme. In practice, this meant a growth in individual retirement accounts (personal pensions) and the development of defined-contribution occupational pensions, although these are still a minority of all employer schemes. The growth in personal pensions was rapid. By the early 1990s, they covered nearly one-quarter of employees and an even higher proportion of younger workers. The trend towards less generous state pension provision means that, in spite of an aging population, the future cost of the state pensions is set to fall as a proportion of the gross domestic product (GDP) by 2050 (see table 11.1), making the situation in the United Kingdom different to most other OECD countries.1 However, it is worth bearing in mind that spending on pensions represents only part of total government spending on benefits for older 1. See Johnson (1999) for a discussion.

646

Richard Blundell, Costas Meghir, and Sarah Smith

Table 11.1

% of GDP

Projected State Spending on Pensions in the United Kingdom 2000

2010

2020

2030

2040

2050

4.5

5.2

5.1

5.5

4.0

4.1

Source: Banks and Emmerson (2000).

nonworkers. In the 1980s, there was a very large increase in the number of older nonworkers on disability benefits2 (see figure 11.2) and spending on these benefits has more than doubled in real terms since 1990. Also, as the level of the basic state pension has now fallen below the level of meanstested benefits for pensioners, many pensioners are eligible for means-tested benefits on top of their state pension. By 2000–2001 more than one-third of pensioners were receiving means-tested benefits. Means testing is becoming an increasingly important element in state provision for pensioners with the introduction of an earnings-indexed means-tested Minimum Income Guarantee for pensioners from April 1999. In this paper, we consider a cohort of workers retiring at the beginning of the 1990s and study the impact of the incentives in public and private pension schemes on their retirement. This cohort was in employment when coverage of defined-benefit occupational pensions was at its peak. Most men in the cohort belonged to an occupational pension scheme, and this is likely to be the key financial determinant of when they retire. Previous analysis has shown clear differences in the retirement behavior of people with and without occupational pensions—see Disney, Meghir, and Whitehouse (1994) and Blundell and Johnson (1998, 1999). Those with occupational pensions are more likely to remain in employment up to age sixty than those without, but are more likely to leave after this age (see figure 11.3). This difference in behavior has been attributed to the incentive structure of occupational pensions, but this has never been modeled explicitly. This paper therefore represents an important contribution to increasing understanding of the incentive effects of occupational pensions on retirement. The state pension scheme is likely to have a smaller incentive effect on retirement behavior in the United Kingdom than that in other countries. The earnings-related element (SERPS) was adopted only in 1978 and is of relatively smaller magnitude than in other European countries. It will also be irrelevant to those people who opted out into occupational pension or personal pension schemes (and to many married women who opted out of the state pension system altogether). Only a minority of people in our sample of retirees remained in SERPS, although they form an interesting group to look at since SERPS was nearing its peak in terms of generosity at the time they were retiring. 2. The main benefit was invalidity benefit, which was replaced by incapacity benefit in 1995.

Fig. 11.2

Recipients of invalidity benefit

Fig. 11.3

Survival probabilities, by pension status

Pension Incentives and Pattern of Retirement in the United Kingdom

649

This paper models retirement incentives for the cohort of individuals in the U.K. Retirement Survey (Department of Social Security and Office for Population and Census Surveys, various years). This is a two-wave panel survey of a sample of individuals born between 1919 and 1933. The first wave, conducted between November 1988 and January 1989, collected information on 3,543 key respondents then aged fifty-five to sixty-nine. About two-thirds of the original sample were reinterviewed in 1994. The Retirement Survey has a larger sample of individuals in the relevant age range than any general household or individual surveys in the United Kingdom and is therefore the best currently available data for analyzing retirement behavior. However, it is considerably smaller than the administrative data sets used in other countries’ studies. It also lacks complete earnings histories and full information on the rules of individuals’ occupational pension schemes. Instead, we match earnings profiles from crosssectional surveys on the basis of cohort, education, and industry. We also model the individual’s occupational pension entitlement according to the rules of the most common scheme in the sector that the individual works in. The plan of the paper is as follows. The next section describes the U.K. pension system and the key elements that are likely to affect retirement behavior. Section 11.3 provides further information on the Retirement Survey and the selection criteria that we use for choosing a sample of individuals for analyzing retirement behavior. Section 11.4 describes the construction of earnings- and pension-incentive measures. Section 11.5 contains the results from estimating probit models of retirement that include these incentive measures and discusses their implications for retirement behavior by means of alternative scenarios for reforms to the pension system. In section 11.6, we present simulation results from two policy reforms designed to reduce the incentives for early retirement in the current pension system. Section 11.7 concludes. 11.2 Policy Environment The U.K. pension system is two-tiered. The first tier, provided by the state, consists of the basic state pension and a significant level of meanstested benefits (made more significant by the introduction of the Minimum Income Guarantee for pensioners in April 1999). The second tier, compulsory for all employees with earnings above a certain floor, is made up of the SERPS3 and a large and growing level of private provision (see figure 11.4). 3. The SERPS will be replaced by the state second pension from 2002. This will effectively be a flat-rate top-up to the basic state pension and more generous than SERPS to low earners. Most workers will be encouraged to opt out into private provision.

650

Richard Blundell, Costas Meghir, and Sarah Smith

Fig. 11.4

U.K. pension system, 1990

11.2.1 The Basic State Pension The basic state pension is a flat-rate contributory benefit payable to people aged over the state pension age (sixty-five for men and sixty for women4) who have made sufficient contributions throughout their working lives.5 From April 2000, the basic state pension is worth £72.50 a week for a single pensioner. Prior to 1978, married women could opt to pay a reduced rate of National Insurance, which meant they did not qualify for a basic state pension in their own right. Couples in which one partner does not qualify for the basic state pension receive a dependent addition, irrespective of whether they have ever worked or not. Since 1989, there has been no earnings test for receipt of the basic state pension,6 although individuals who choose to defer will increase the value of their pension by 10 percent for each year of deferral.7 11.2.2 The State Earnings Related Pension Scheme (SERPS) The first part of the second tier of pension provision is the SERPS. Introduced in 1978, this pays a pension equal to a fraction of an individual’s 4. The retirement age for women will be raised by six months each year from 2010 to 2020 so that equalization is achieved in 2020. 5. To qualify for the basic state pension, individuals need to have made or be credited with National Insurance contributions for 90 percent of their working lives. Credits are available for periods of illness, disability, or unemployment. 6. See Disney and Smith (2002) for a discussion of the effects of the abolition of the earnings test on labor supply. 7. Increased from 7.5 percent in 1995.

Pension Incentives and Pattern of Retirement in the United Kingdom

651

qualifying annual earnings (above a specified lower-earnings limit) each year since 1978. When it was introduced, SERPS was intended to pay a pension worth one-quarter of an individual’s best twenty years’ earnings (up to a specified upper-earnings limit). Subsequent reductions in the generosity of SERPS mean that it is worth only 20 percent of average lifetime earnings to anyone retiring after 2000. Married women who opted to pay reduced-rate National Insurance contributions do not qualify for SERPS. Currently widows can claim their husbands’ SERPS pensions in full if they receive no additional pension in their own right.8 After retirement the SERPS pension is uprated each year in line with prices. 11.2.3 Income Support and Invalidity Benefit In addition to the basic state pension and SERPS, there are two other state benefits that are taken up widely by older nonworkers—income support and incapacity benefit (formerly invalidity benefit). Income support is a flat-rate, noncontributory, means-tested benefit. It is paid automatically to people aged sixty or more who do not work. Unlike people in younger age groups, the over-sixties do not have to show that they are actively seeking work in order to qualify. From April 1999, income support for pensioners was renamed the Minimum Income Guarantee and made more generous with an increase in the level and a commitment to uprate in line with earnings, at least for the short to medium term. Incapacity benefit (formerly invalidity benefit) is a contributory benefit paid to the long-term sick and disabled. In the case of invalidity benefit, an individual qualified on the basis of medical certificates from their general practitioner (GP) showing them to be incapable of the work that was “reasonable” to expect them to do (given their age, qualifications, and so forth). With the introduction of incapacity benefit in 1995, this was changed to a stricter “all work test” carried out by a doctor employed by the Benefits Agency Medical Service. The change from invalidity benefit to incapacity benefit was a response to very rapid growth in receipt during the 1980s. A key feature of incapacity benefit (and invalidity benefit) is that, before April 2001, it was not means tested and could be received in conjunction with private pension income (unlike income support). From April 2001, it will be means tested against occupational pension income. 11.2.4 Occupational and Personal Pensions Compared to most other European countries, the United Kingdom has a high level of coverage of private pensions, including both occupational pensions and individual retirement accounts, known in the United King8. This was due to be reduced to half from April 2000. However the failure of the government to properly inform individuals of the change in entitlement led to the reform being delayed.

652

Richard Blundell, Costas Meghir, and Sarah Smith

Table 11.2

DB plans DC plans Hybrid

Occupational Schemes: Defined Benefit (DB) versus Defined Contribution (DC) % Private-Sector Schemes

% Public-Sector Schemes

% All Schemes

78 16 6

98 2 0

80 14 6

Source: National Association of Pension Funds (NAPF) 1998.

dom as personal pensions. Any individual can choose to contract out of SERPS, into one of these two types of secondary private pensions (and from April 2001, people are also able to choose to opt out into a stakeholder pension, which is effectively a benchmarked individual retirement account). Members of defined-benefit occupational schemes pay a reduced rate of National Insurance, while those with defined-contribution occupational pensions or personal pensions receive a National Insurance rebate paid directly into their fund. Occupational pensions currently cover around 45 percent of employees, down from a peak of over 50 percent in the early 1980s. They are typically defined-benefit schemes (DB; see table 11.2), although since 1988, employees have also been allowed to opt out into defined contribution (DC) occupational schemes, and there has been a gradual shift from DB to DC schemes since then (see Disney and Stears 1996). The decline in coverage of occupational pension schemes is due to a number of factors. It reflects changing employment patterns and a shift to smaller employers. Also, it reflects increasing pension choice among individuals working for employers offering occupational pensions who, since 1988, can no longer be compelled to join the scheme. Since 1988, individuals have been able to contract out of SERPS (and leave their occupational scheme) and take out a personal pension. To kick start these schemes when they were introduced, a bonus National Insurance contribution of 2 percent was paid by the government, in addition to the contracted-out rebate. By the mid-1990s, around 6 million people (more than one-quarter of all employees) had taken out a personal pension. Takeup was higher among younger workers, as would be expected. However, there is a serious issue over the number of older workers who were “missold” personal pensions by financial advisers who wrongly advised them that they would be better off leaving their occupational pension scheme. Table 11.3 summarizes labor market participation and income receipt by age using data from the Family Expenditure Survey 1994–1995 (corresponding to the second wave of the Retirement Survey). It shows relatively high rates of labor market withdrawal among men before the state pension age. The two most important sources of income before state pension age are income from private (predominantly occupational) pensions and disability benefit. It is important to stress that these two sources of income are not al-

Pension Incentives and Pattern of Retirement in the United Kingdom Table 11.3

Gender and Age Men 50–54 55–59 60-64 65–69 Women 50–54 55–59 60–64 65–69

653

Labor Market Participation and Benefit Receipt

Full-Time Work

Part-Time Work

Not Working

Public Pension

Private Pension

Disability Benefits

Disability Benefits + Private

Other Benefits

0.6447 0.4620 0.2680 0.0213

0.2053 0.1881 0.1787 0.0816

0.1500 0.3598 0.5533 0.8972

0.0000 0.0000 0.0000 0.8121

0.0947 0.3432 0.5395 0.7411

0.0737 0.1386 0.2096 0.1667

0.0237 0.0825 0.1478 0.1312

0.0658 0.0728 0.1237 0.0532

0.4667 0.2936 0.0909 0.0156

0.2427 0.2385 0.1394 0.0688

0.2907 0.4679 0.7697 0.9156

0.0507 0.0975 0.7970 0.9594

0.1040 0.1988 0.3606 0.4125

0.0400 0.0398 0.0242 0.0000

0.0133 0.0061 0.0152 0.0000

0.0480 0.0520 0.0485 0.0469

Source: Family Expenditure Survey 1994–1995 (U.K. Data Archives, 1996).

ways alternative preretirement income sources, but are typically received together by the same people. The fact that disability benefit was not means tested meant that it could be received in conjunction with other forms of income. Three-quarters of people in receipt of disability benefit income also received some money from a private pension. 11.3 Data Overview 11.3.1 The Retirement Survey The main data used for analyzing retirement behavior are drawn from the U.K. Retirement Survey (RS), a household panel survey collected by the Office for Population and Census Surveys on behalf of the Department for Social Security. This is the first large-scale panel data set in the United Kingdom to focus on individuals around the time of retirement (see Bone et al. 1992). Two waves of data were collected on a national random sample of individuals born between 1919 and 1933. The first wave of the survey was conducted between November 1988 and January 1989 and collected information on 3,543 key respondents (who were aged fifty-five to sixtynine). The key respondents include spouses if they were in the relevant age range. In addition, information was also collected on 609 spouses outside this age range. About two-thirds of the original sample were reinterviewed in 1994, and 11 percent of respondents disappeared in this interval due to mortality; the residual attrition is a combination of nonresponse and (perhaps) unreported mortality.9 9. The high attrition rate is largely due to the fact that the survey was not originally intended to be a panel survey. Hence, little attempt was made to keep in touch with respondents after the first wave.

654

Richard Blundell, Costas Meghir, and Sarah Smith

The RS offers a relatively large sample of people in the relevant age range, compared to more general panel surveys, such as the British Household Panel Survey. It also offers very rich demographic, economic, and health information on individuals—and their spouses—in both waves. And it has employment history information and private pension history information dating all the way back to individuals’ first jobs.10 However, compared to the administrative datasets available in other countries, the sample in the RS is relatively small (and is reduced by the high attrition rate between the two waves). Also, the survey does not collect earnings-history information, which is needed to calculate exact pension entitlements for each individual. Instead, as we describe in the next section, we have to impute earnings histories on the basis of employment-history information. The analysis of retirement behavior in this paper is based on a subsample of people in the RS. The group we look at comprises those who were

• below the state pension age in wave 1, that is, men aged fifty-five to sixty-four or women aged fifty-five to fifty-nine in 1988–1989;

• working in wave 1 with nonmissing earnings information and no income from occupational pension schemes, unemployment benefit, or income support; • and interviewed in both waves. Excluding people who fail to meet any one of these criteria leaves 456 individuals—283 men and 173 women. Each of these individuals remains in the sample from 1989 until they leave employment, leaving a total sample of 1,998 person observations. Summary sample characteristics based on all person observations are given in table 11.4. 11.3.2 Earnings Histories and Projections To calculate state pension entitlements, we need individual earnings profiles going back to 1978, when SERPS was introduced. These are absent in the RS, but the survey does provide detailed work histories documenting spells in employment, whether the employment was part time or full time, and in which industry the individual worked, which, together with information on age and education, allow us to match earnings profiles from cross-sectional data. There is no single dataset with consistent information on these variables going back to 1978. Instead, we combine information from two datasets to get consecutive cross-sectional waves of data from 1978 to 1989—the Family Expenditure Survey (1978–1986) and the General Household Survey (1987–1989). Projecting forward from 1989, we assume constant real wages. 10. For a good overview of information in the RS (Department of Social Security and Office for Population and Census Surveys, various years), see Disney, Grundy, and Johnson (1998).

Table 11.4

Sample Characteristics

No. of observations Mean age Proportion currently married Age difference between individual and spouse (years) Net earnings ($) Proportion with an occupational pension Proportion of women paying reduced-rate NI Length of time in current job (years) Proportion of time since leaving education in full-time employment Industry, energy Industry, engineering Industry, manufacturing Industry, distribution Industry, services Industry, government Zero financial wealth £1–£3,000 financial wealth £3,000–£10,000 financial wealth
£10,000 financial wealth Missing financial wealth School dropout High school education College Health in 1988 (severity score) Variable High school dropout High school graduate

College Severity score

Men

Women

1,276 61.50 0.8659

722 59.87 0.7659

2.80 18,157 0.6857 0.0000 12.16

–1.17 9,064 0.3850 0.7410 9.85

0.6143 0.0940 0.0030 0.2014 0.1951 0.2429 0.0635 0.1897 0.4036 0.2045 0.1575 0.0447 0.4397 0.4287 0.1317 0.3017

0.2341 0.0000 0.0457 0.1191 0.1551 0.6053 0.0748 0.1856 0.4460 0.1717 0.1399 0.0568 0.6108 0.3047 0.0545 0.3670

Definition No qualification O levels; A levels; school certificate; certificate of sixth-form studies; clerical and commercial qualifications (e.g., typing, shorthand, bookkeeping, and commerce); city and guilds; nursing qualifications; other qualification; recognized trade apprenticeship University degree or diploma; teaching qualification; membership of professional institution Measure of self-assessed health status based on the international classification of impairments, disabilities, and handicaps (ICDIH). Separate scales are constructed for areas of locomotion, reaching and stretching, dexterity, seeing, hearing, continence, communication, personal care, behavior, intellectual functioning, consciousness, digestion, and disfigurement. The severity score is constructed as a weighted average of the three highest-severity scores from the 13 areas: Highest + 0.4 (second highest) + 0.3 (third highest).

Note: NI = National Insurance; FT = Full-time

656

Richard Blundell, Costas Meghir, and Sarah Smith

We also exploit the earnings information that is available in the first wave of the RS to construct an individual fixed effect, which we use to adjust the individual’s entire earnings profile. We assume that the wage of individual i in cohort, education, and industry subgroup g in period t can be expressed as Wigt  iWgt where i is a constant individual fixed effect, Wig88 /Wg88 , where Wig88 is taken from the RS and Wg88 is calculated from the cross-sectional data. Our underlying assumption is that macro-shocks affect everyone in the cohort, education, and industry subgroup in the same way. 11.4 Construction of Incentive Measures Each individual’s total pension wealth and pension accrual measures are built up from combining four separate elements of the pension system—the basic state pension, the state earnings-related scheme (SERPS), occupational pensions, and invalidity benefit.11 In this section, we discuss how each of these individual elements is constructed. We also discuss potential sources of variation in total pension wealth and accrual rates by which we might identify the impact of pension incentives on retirement behavior. 11.4.1 The Basic State Pension Calculation of basic state pension entitlement is straightforward. It depends on the total number of years of contribution and, for a married woman, on whether or not she opted to pay reduced-rate National Insurance contributions. This latter piece of information is known directly from the RS. Although the basic state pension is flat rate, total wealth will vary across individuals because of the dependent’s allowance and because of the fact that widows not entitled to a pension in their own right can claim their former spouse’s pension in full when their spouse dies. In these cases, we need to compute husbands’ total pension wealth over the life of the couple, based on the age difference between the spouses. Obviously, the larger the age difference between husband and wife, the greater the husband’s total pension wealth. 11.4.2 State Earnings Related Pension Scheme The precise formula for calculating an individual’s SERPS pension is given by R

SERPS 

∑

t1978


W˜ Y  LEL  , YR

t

R1

Rt

˜t  min(Wt , UEL). where W

t

11. We ignore income support since it is a universal benefit.

Pension Incentives and Pattern of Retirement in the United Kingdom

657

Earnings up to the annual upper-earnings limit (UEL) are revalued to the year of reaching state pension age (R) using an index of economy-wide average earnings (YR /Yt ). The lower-earnings limit (LEL) in the year prior to the individual reaching state pension age is deducted from each year’s revalued earnings figure and the net of LEL earnings are multiplied by an accrual factor ( Rt ).12 For people retiring before 2000, the accrual rate was 1.25 percent a year. Details of earnings factors, upper- and lower-earnings limits, and accrual rates are given by table 11A.1 in the appendix. Having calculated earnings profiles for each individual in the RS, their SERPS entitlements are fairly straightforward to calculate. We assume zero SERPS pension for people who are in occupational pension schemes and for married women who have opted to pay reduced-rate National Insurance contributions. There are several potential sources of variation in SERPS pension wealth across individuals. Total wealth, but not accrual, will be affected by an individual’s employment history since 1978—both the number of years they have been in employment and their earnings—while projected earnings in the future will have an impact on expected total wealth and accrual. Another important factor for determining total wealth (but not accrual) will be the individual’s age in 1978. This was when SERPS was introduced and an individual’s age in that year will determine the period over which they are able to accrue rights to a SERPS pension before reaching state pension age. The maximum SERPS pension to which an individual could be entitled, for each year of retirement since 1978 is shown in figure 11.5 (and also the SERPS entitlement based on average earnings). For example, someone reaching state pension age in 1979 would receive practically no SERPS pension since they would only have been building up entitlement for one year.13 Someone retiring in 1998 could have accrued rights to a SERPS pension of up to £5,000 a year by earning the upper-earnings limit for twenty years. As shown in table 11A.1, accrual rates will change after 2000, but this reform will not affect the cohort of individuals in the RS, all of whom will have reached the state pension age before then. Finally, the fact that widows can claim their former husbands’ SERPS pensions if they receive no pension in their own right means that, as with the basic state pension, a man’s marital status, and the age difference between them and their spouse also affects their total pension wealth and accrual. Table 11.5 compares our estimates of individuals’ SERPS pension with the actual SERPS pension they received where this information is available 12. Starting April 2000 this formula changes. Instead of uprating annual earnings and then subtracting the LEL from the year prior to retirement, the LEL from the year worked is subtracted from earnings first, and then the difference is uprated in line with earnings growth. Since the LEL is annually uprated in line with the basic state pension (i.e., with prices), this has the effect of reducing the generosity of SERPS. 13. Individuals cease to build up entitlements once they pass the state pension age.

658

Richard Blundell, Costas Meghir, and Sarah Smith

Fig. 11.5

Table 11.5

SERPS entitlement

Predicted and Actual SERPS Pensions Actual SERPS pension received in 1994 ($) Imputed SERPS pension in retirement year ($) Correlation coefficient No. of observations

4,315 3,849 0.3334 102

(i.e., for individuals who had retired by the second wave of the RS and reported pension receipt). On average, we underpredict individuals’ SERPS pensions, and while the correlation coefficient is positive and significant, it is fairly low, compared to that for occupational pensions (see following discussion). One possible explanation is that individuals who are in SERPS— and therefore not in occupational pension schemes—are likely to have had more variable employment histories than those who are in occupational schemes. Our method for estimating earnings profiles may be missing a lot of variation in their previous earnings, which would also matter more for SERPS than for occupational pensions that are typically determined only according to recent years’ earnings. 11.4.3 Invalidity Benefit One possible way to treat entitlement to invalidity benefit would be to assume that only individuals who received the benefit were eligible and that all those who satisfied the eligibility conditions received the benefit. However, given the potential for subjective evaluations of “incapacity for work”

Pension Incentives and Pattern of Retirement in the United Kingdom

659

and “reasonable work” and in the light of significant variation in the number of people receiving the benefit over time (as well as anecdotal evidence of differences between doctors in their willingness to certify individuals as being incapable of work), this assumption is inappropriate. Instead, we calculate an individual’s invalidity-benefit wealth on the basis of an assigned probability that they will receive the benefit. These probabilities are derived from a probit model for receipt of invalidity benefit as a function of characteristics such as age, education, region, tenure, marital status, and spouse’s employment status, which we estimate using data drawn from the Family Expenditure Survey from April 1988 to March 1994. We impute probabilities for individuals in the RS on the basis of matched characteristics. The probit results are reported in appendix table 11A.2. 11.4.4 Occupational Pensions The pension received in a DB occupational pension scheme is typically determined by a formula of the type: P  (PER  LELR1 )N, where P is the annual occupational pension,  is the scheme-specific accrual rate, PER is pensionable earnings at the time of retirement (which are typically the individual’s average earnings in the last year or last few years before retirement),  is the integration factor, and N is the number of years that the individual has belonged to the scheme. From information in the RS, we know N, the number of years the individual has belonged to the scheme. However, we have to make reasonable assumptions about  Rt , PER , and . The key distinction that we make is between individuals who work in the public sector versus those in the private sector. We assume that different typical schemes apply in the two sectors with different accrual rates, definitions of pensionable earnings, and integration factors. This assumption, and the choice of parameter values that we adopt, are based on information from the 1997 (NAPF) Survey of Occupational Pension Funds (1998), which shows a clear difference between public and private sector schemes (see table 11.6). We assume an accrual rate of one-sixtieth for private sector and oneeightieth for public sector. For pensionable earnings, we take the best three out of last ten years’ earnings for individuals working in the private sector and the best single year’s earnings out of the last ten years for individuals working in the public sector. We assume an integration factor of 1 for private-sector schemes and 0 for public-sector schemes. By construction, total occupational pension wealth—and accrual rates—will vary across individuals according to whether they work in the public or private sector. But there are other sources of variation in both total wealth and accrual rates. Total wealth will vary according to the num-

660

Richard Blundell, Costas Meghir, and Sarah Smith

Table 11.6

Private versus Public Schemes Private Schemes (%)

Public Schemes (%)

Accrual Rates 1/80th 1/60th Other

15 65 20

Definition of Pensionable Earnings Actual earnings at retirement 11 Actual earnings at fixed date 4 Average earnings over the last 12 months 23 Best year’s earnings within 3–10 years 25 Best 3 years’ earnings within 10–13 years 30 Other 7

92 8 0 2 3 9 86

Integration With State Schemes Integration No Yes Adjustment based on: Basic state pension Lower earnings limit Other

44 56

92 8

43 55 2

50 50

Source: NAPF (1998).

ber of years that the individual has belonged to the scheme, while projected earnings in the future will have an impact on expected total wealth and accrual. Further variation in accrual rates comes from differences across occupational schemes in the age at which individuals are entitled to start drawing their pension, which is also asked in the RS.14 We assume that people can continue to accrue rights to occupational pensions beyond this age (up to a maximum of forty years), but for each year that they continue to work beyond this age, they lose a year’s pension. This is clearly a simplification of the actual rules of occupational pension schemes, and not least because, around this time, many firms implemented early retirement schemes to encourage exits. With no information about the availability of these schemes in the RS, we are almost certain not to capture the actual set of retirement incentives facing some individuals. Even so, we do appear to estimate fairly well the level of occupational pension income received in retirement. Table 11.7 compares our estimates of individuals’ occupational pension with the actual occupational pension they received, where this information is available (i.e., for individuals who had retired by the second wave of the RS and 14. The survey asks “at what age will you start to receive the pension,” and then asks “is that the usual age for drawing a pension,” which is true for 90 percent of respondents. Where information on usual pension age is missing, we assume that it is sixty-five (the modal age).

Pension Incentives and Pattern of Retirement in the United Kingdom Table 11.7

Predicted and Actual Occupational Pensions Actual occupational pension received in 1994 ($) Imputed occupational pension in retirement year ($) Correlation coefficient No. of observations

Table 11.8

661

8,140 7,762 0.7868 172

Sources of Variation in Pension Incentives Across Individuals

Marital status and age of spouse (survivor’s benefits) Whether paid reduced-rate NI (married women) Whether spouse paid reduced-rate NI (married men) Past earnings Future earnings Date of birth Number of years with current employer Accrual rate—SERPS, public sector, private sector Pensionable earnings—public sector, private sector Normal retirement age Region, tenure, spouse’s employment, education, age

BSP, SERPS, OP BSP, SERPS BSP, SERPS SERPS, OP SERPS, OP SERPS OP OP OP OP IVB

Wealth

Accrual

✓ ✓ ✓ ✓

✓ ✓ ✓

✓ ✓ ✓ ✓ ✓ ✓

Note: BSP = basic state pension; OP = occupational pension; IVB = invalidity benefit.

reported pension receipt). As with SERPS, we underpredict individuals’ total level of occupational pension income, but the correlation coefficient is positive, significant, and high. 11.4.5 Total Pension Wealth and Accrual Measures Identifying the effects of incentive measures on retirement behavior requires these measures to vary across individuals, over time, or both, conditional on the other sociodemographic covariates that would be included in a model of retirement. As the previous discussion of the construction of the pension incentive measures has shown, there are a number of potential sources of variation in total wealth and in the forward-looking accrual measures for each of the four separate elements of the pension system. Table 11.8 summarizes these sources of variation, indicating which of the four elements of the system—the basic state pension, SERPS, occupational pension, and invalidity benefit—is affected and whether the source drives variation in total pension wealth or forward-looking accrual measures (or both). Almost all of the sources of variation affect both total pension wealth and accrual. However, future earnings will affect forwardlooking accrual measures but not current total pension wealth, while total wealth (but not accrual) varies with past earnings and with the individual’s date of birth (in the case of individuals with SERPS).

✓ ✓ ✓ ✓ ✓ ✓

662

Richard Blundell, Costas Meghir, and Sarah Smith

In our analysis of the incentive effects of pensions on retirement, three different forward-looking measures of accrual are used. The first is simply the one-period accrual—that is, how much an individual can add to their total pension wealth by working this period. The second is peak value. This represents the difference between total pension wealth accumulated by the start of the period and the maximum total pension wealth an individual could accumulate looking forward across all future years. This is a more appropriate measure if it is assumed that labor market exits by older workers are irreversible. In this case, when someone leaves the labor market, they are giving up all possible future additions to their pension and will therefore consider how much they could increase their pension by staying in the labor market not just for this period, but in all future periods. By not retiring now, individuals retain an option to retire in the future and, thereby, to increase their pension. This is very similar in spirit to the option value (Stock and Wise 1990), which is the third measure used. In the option value model, individuals are assumed to compare the value of retiring now to the maximum of the expected values of retiring at all future ages, where the value of retiring at future ages includes both possible pension additions and future earnings, that is, OV  Vt (r ∗)  Vt (t)

r1

T

st

sr

where Vt (r)  ∑ stY s ∑ st [kBs (r)] ,

where Ys is earnings and Bs retirement benefits. The option value differs from the peak value by incorporating the future value of earnings until retirement and by incorporating utility parameters k, the differential value of income in leisure compared to earned income, and , the coefficient of relative risk aversion. In our calculation of option values we assume k  1.50 and   0.75. We assume a discount factor () of 0.97 throughout. Tables 11.9–11.11 summarizes the distribution of pension incentive measures for men and women by age. These are calculated across all men and women of the same age who remain in our sample (i.e., those who have not yet exited the labor force) and will therefore be affected by differential selection into the sample at each age. All the figures are expressed in 1998 prices and in dollars.15 Table 11.9 summarizes pension incentive measures for men, pooling those with and without an occupational pension. There is a clear effect of the state pension age—sixty-five for men—on the incentive measures. For men over sixty-five, median accruals are negative and total pension wealth starts to fall.16 It is worth pointing out that the selection effects will tend to increase average accrual measures—and reduce average total wealth— 15. Assuming an exchange rate of $1.50 to £1. 16. Individuals can choose to defer their pension after the state pension ages. However, since deferral is actuarially unfair for an average male and with no earnings test, we assume that all men start to draw their state pension at age sixty-five.

Pension Incentives and Pattern of Retirement in the United Kingdom Table 11.9

663

Incentive Measures for Men ($1998 prices) One-Year-Ahead Accrual

Age

Wealth Median

Median

10th Percentile

90th Percentile

SD

N

56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

89,821 93,850 97,320 103,990 108,244 113,266 117,994 123,886 127,333 128,514 111,329 84,831 68,720 43,102 27,458

3,017 2,544 3,137 3,146 2,932 2,829 2,781 3,080 3,326 –6,038 –5,976 –6,857 –7,162 –7,892 –7,864

1,164 970 1,301 1,115 1,026 791 692 182 801 –10,570 –10,210 –9,859 –9,564 –9,277 –9,024

7,796 13,862 13,819 13,248 9,099 7,947 8,886 9,560 9,254 –1,914 –3,525 –4,975 –4,826 –4,976 –4,826

4,402 5,335 6,537 5,142 3,927 3,577 5,095 4,645 4,545 4,068 2,546 2,062 1,695 1,540 1,799

31 64 104 133 155 170 162 144 124 96 36 24 17 12 4

Peak Value

56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

Option Value

Median

10th Percentile

90th Percentile

SD

Median

10th Percentile

90th Percentile

15,936 12,766 12,764 10,916 8,824 7,234 5,118 2,993 3,269 –6,038 –5,976 –6,857 –7,162 –7,892 –7,864

3,209 2,377 2,650 1,666 1,190 884 653 182 770 –10,570 –10,210 –9,859 –9,564 –9,277 –9,024

37,966 37,675 31,027 24,728 23,424 19,690 14,313 8,541 9,091 –1,914 –3,525 –4,975 –4,826 –4,976 –4,826

13,088 25,228 18,650 10,250 9,974 8,447 6,227 4,355 4,471 4,015 2,546 2,062 1,695 1,540 1,799

10,476 8,857 7,449 6,168 5,034 4,060 2,745 1,615 1,072 681 312 126 322 1,480 2,129

5,375 4,237 3,524 2,920 2,332 1,675 1,214 534 268 –298 –207 –283 –418 –219 –585

13,813 12,711 11,162 10,581 9,083 7,938 7,165 6,316 4,864 5,089 4,190 3,616 4,629 6,041 3,466

SD 3,268 3,270 2,938 3,047 2,975 2,821 2,652 2,298 2,085 2,380 1,851 1,831 1,875 2,369

Notes: SD = standard deviation; N = number of observations.

since those with lower accrual rates and higher total wealth will tend to exit the labor market earlier and so drop out of the sample. The peak values and option values yield more pronounced incentives for people to stay in work at younger ages than the single period accruals. The median option values remain positive up to age seventy, reflecting relatively low replacement rates in the United Kingdom. With the assumption that real earnings remain constant indefinitely, this appears to create an in-

664

Richard Blundell, Costas Meghir, and Sarah Smith

Table 11.10

Incentive Measures for Men With or Without Occupational Pension ($, 1998 price Peak Value

Age

Median Wealth

Median

Option Value SD

Median

SD

N

56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

89,813 94,818 100,479 106,350 111,982 115,244 119,875 129,065 129,262 134,894 120,887 88,656 73,094 45,510 28,010

With Occupational Pension 20,617 12,318 16,313 28,814 15,433 20,429 12,622 10,697 10,424 10,793 8,406 9,287 6,407 6,956 3,932 4,940 4,906 5,033 –4,710 4,264 –5,248 1,630 –6,300 1,947 –7,162 1,622 –7,785 435 –9,024

11,060 9,509 8,128 7,041 5,858 4,898 3,742 2,443 1,301 1,147 1,795 682 340 1,566 2,441

2,429 2,873 2,745 3,042 3,098 2,984 2,858 2,466 2,326 2,526 1,941 2,012 1,760 2,636

22 44 77 95 112 123 113 102 86 62 18 10 7 3 1

56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

94,040 81,424 89,181 92,931 98,296 105,402 110,036 116,857 122,692 126,471 103,687 82,364 50,315 41,864 26,906

Without Occupational Pension 3,636 7,777 4,374 6,166 5,658 4,944 4,403 4,931 4,267 4,116 3,949 3,277 2,575 2,185 1,373 1,145 1,576 1,174 –9,286 2,314 –8,614 2,254 –8,355 1,915 –7,550 1,832 –8,800 1,772 –7,760 1,757

5,585 4,976 4,776 4,002 3,358 2,543 1,727 863 926 –45 –29 43 305 1,394 1,817

2,474 2,488 2,133 1,756 1,342 1,129 964 573 1,010 1,104 1,334 1,708 2,045 2,387 2,037

9 20 27 38 43 47 49 42 38 34 18 14 10 9 3

Notes: SD = standard deviation; N = number of observations.

centive for some individuals to carry on working even at older ages. This will be reinforced by increasing selection of high-wage individuals into the sample with age. Table 11.10 compares the incentive measures for men with and without occupational pensions. Figure 11.3 showed a clear difference in the labor market exit behavior of these two groups; men with an occupational pension are more likely to stay in work at younger ages. Table 11.10 shows that men with occupational pensions tend to have higher median peak values and option values up to the state pension age, as well as higher wealth. These incentives could work toward either earlier or later retirement. The

Pension Incentives and Pattern of Retirement in the United Kingdom Table 11.11

665

Incentive Measures for Women ($1998 prices) Accrual

Age

Wealth Median

Median

10th Percentile

90th Percentile

SD

N

56 57 58 59 60 61 62 63 64 65 66

3,018 4,633 2,324 4,124 1,231 0 604 0 303 0 19,916

0 0 0 0 0 0 0 0 0 0 154

0 0 0 0 –5,039 –4,888 –4,741 –4,599 –4,473 –4,413 0

7,200 9,841 8,081 6,739 2,231 2,089 1,899 1,409 3,809 631 307

5,905 6,809 5,203 6,068 3,183 3,316 3,447 3,703 4,994 2,135 217

38 68 98 114 142 107 68 43 25 17 2

Peak Value

Median 56 57 58 59 60 61 62 63 64 65 66

0 0 0 0 0 0 0 0 0 0 39

Option Value

10th Percentile

90th Percentile

SD

Median

10th Percentile

90th Percentile

SD

0 0 0 0 –5,039 –4,888 –4,741 –4,599 –4,473 –4,413 0

35,228 29,180 20,693 17,108 14,269 13,120 8,724 4,886 10,106 3,644 78

18,676 15,825 10,208 10,558 10,236 9,199 8,177 7,051 5,268 3,319 55

6,191 5,646 4,910 4,277 3,877 3,445 3,833 3,375 2,291 2,593 3,427

2,028 647 421 322 –308 –346 –391 –295 –329 –377 3,005

12,753 11,485 10,868 9,254 8,858 8,208 7,901 6,149 5,119 4,472 3,848

3,855 3,720 3,343 3,070 3,609 3,349 3,219 2,707 2,456 1,840 596

Notes: SD = standard deviation; N = number of observations.

observed pattern of exits suggests that the effect of the higher option values is likely to dominate, at least at younger ages, encouraging men with occupational pensions to stay in employment. It is worth pointing out that although the typical annual occupational pension is considerably higher than the typical SERPS pension (comparing tables 11.5 and 11.7), the difference between total pension wealth for people with occupational pensions and those without is reduced by the more generous survivors’ benefit provisions of SERPS. In the case of SERPS, the surviving spouse inherits the pension in full; in the case of occupational pensions, they inherit only half.17 Table 11.12 summarizes the incentive measures for women. The large 17. The survivors’ benefit was due to be cut to half in SERPS from April 2001. However, in the build-up to the preannounced reform many people were issued the wrong information in the form of leaflets that did not refer to the reform. The change has been delayed to October 2002, and those who can show that they were misinformed will keep their original entitlement.

666 Table 11.12

Richard Blundell, Costas Meghir, and Sarah Smith The Probability of Male Retirement (1,276 observations) Linear Age Marginal Effect

Cohort Dummies

SE

Marginal Effect

SE

Age Dummies Marginal Effect

SE

Single Period Accruals Excluding age first eligible Pension wealth Single-period accrual Spouse wealth Pseudo R2 Log-likelihood F test (PW, Accrual)

.0999 –.4975 .0324 .1961 –404.11 34.59

.0183 .1574 .0154

.1158 –.8890 .0386 .1885 –407.93 60.44

.0197 .1513 .0164

.0915 –.1251 .0305 .2352 –384.43 21.43

.0196 .1761 .0159

Including age first eligible Pension wealth Single-period accrual Spouse wealth Penage dummy Pseudo R 2 Log-likelihood F test (PW, Accrual)

.0699 –.1954 .0224 .1787 .2235 –390.33 13.50

.0188 .1623 .0155 .0453

.0861 –.6365 .0281 .1586 .2087 –397.77 26.67

.0206 .1580 .0166 .0455

.0825 –.1144 .0278

.0201 .1769 .0158

.2386 –382.72 16.68

Peak Value Excluding age first eligible Pension wealth Peak accrual Spouse wealth Pseudo R 2 Log-likelihood F test (PW, Accrual)

.0883 –.0946 .0296 . .1866 –408.88 25.81

.0173 .0795 .0152

.0999 –.3170 .0297 .1670 –418.71 40.54

.0192 .0780 .0166

.0892 –.0257 .0293 .2348 –384.62 21.10

.0192 .0574 .0157

Including age first eligible Pension wealth Peak accrual Spouse wealth Penage dummy Pseudo R 2 Log-likelihood F test (PW, Accrual)

.0629 .0084 .0194 .1991 .2220 –391.07 12.09

.0178 .0765 .0153 .0452

.0696 –.2025 .0199 .2012 .1987 –402.77 16.75

.0198 .0786 .0167 .0471

.0802 –.0115 .0266 .0962 .2382 –382.92 16.33

.0197 .0760 .0157 .0636

number of zeros arises as a result of the number of married women who are not eligible for a pension in their own right. This means that the median single-period accruals and median peak accruals are all equal to zero. As with men, the effect of the state pension age is clear, with the tenth percentile single-period accruals and peak values turning negative at age sixty. The ninetieth percentile peak values and option values remain positive after this age because of occupational pension schemes, which may have normal pension ages for women that are actually higher for women than their state pension age.

Pension Incentives and Pattern of Retirement in the United Kingdom Table 11.12

667

(continued) Linear Age Marginal Effect

Cohort Dummies

SE

Marginal Effect

SE

Age Dummies Marginal Effect

SE

Option Value Excluding age first eligible Pension wealth Option value Spouse wealth Pseudo R 2 Log-likelihood F test (PW, Accrual)

0.7706 –.3619 .0244 .1858 –409.25 25.06

.0205 .4196 .0153

.0509 –1.7598 .0188 .1731 –415.68 44.58

.0206 .3675 .0165

.0750 –.6140 .0267 .2366 –383.72 22.71

.0213 .4462 .0158

Including age first eligible Pension wealth Option value Spouse wealth Penage dummy Pseudo R 2 Log-likelihood F test (PW, Accrual)

.0531 –.3739 .0175 .1977 .2228 –390.68 12.78

.0209 .4201 .0154 .0440

.0246 –1.5893 .0105 .2142 .2111 –396.54 27.43

.0209 .3653 .0164 .0465

.0654 –.6397 .0239 .0995 .2403 –381.88 18.19

.0218 .4460 .0157 .0637

Notes: The full set of demographic controls include earnings (and earnings squared), education, health, job tenure, industry, proportion of time spent in full-time employment, whether individual has an occupational pension, housing tenure, financial wealth, age difference within couples, spouse’s earnings, spouse’s health, and whether spouse is retired. PW = pension wealth; SE = standard error.

11.5 Estimated Pension Incentives and the Probability of Retirement 11.5.1 The Statistical Model We estimate the impact of the incentive and wealth variables on retirement decisions by modeling the conditional probability of exit from employment for individuals in the RS. For each individual i, we write Dit  1 if the individual has left the labor market in period t (conditional on being in the labor market in period t – 1). The probability of this event is then modeled as a function of observable household and individual characteristics, as well as the pension incentive variables. The pension incentive variables, defined in the previous section, are discounted wealth, option value (or single-period accrual), spouse’s pension wealth, and the pension age. The latter measures the earliest age at which someone can draw their pension. This varies not only across gender, but also across type of pension plan. Denoting the observable characteristics as Zit , and the pension incentive variables as Iit , our conditional probability model may be expressed as Pr(Dit  1)  G (a Zit b Iit ),

668

Richard Blundell, Costas Meghir, and Sarah Smith

where G is the cumulative distribution function of unobservables in the conditional exit model and a and b are unknown response coefficients. In estimation, we assume G is a cumulative standard normal and consequently estimate a and b using a probit model for the conditional exit probability, pooled over all five years of retirement information in the RS. In constructing the standard errors, we need to allow for dependence over time in the unobservables for the same individual in the panel who survives more than one period before retiring. This is implemented using the block bootstrap method. 11.5.2 The Impact on Retirement The discussion in section 11.4 has highlighted the sources of variation in the pension incentive variables. We argue that there is sufficient variability in the pension variables, conditional on the full set of other variables included in the regressions. Generally it is difficult to gauge how much variation one needs for a credible estimate—after all, this crucially depends on the amount of variance in the errors. However, we note that, for our most general specification, 40 percent of total pension wealth in the case of men and 43 percent of total pension wealth in the case of women remains unexplained by all the other included regressors, including option values (see table 11A.3 in appendix). For the option value, 24 percent of that for men and 32 percent of that for women remains unexplained by the other regressors, including total pension wealth. Overall, the pension variables, conditional on our functional form assumptions and exclusion restrictions, seem to display sufficient variability. Turning to the conditional exit probability estimates, table 11.12 presents the marginal effects and standard errors from a probit regression for a variety of specifications estimated using data on our sample of men in the RS. Table 11.13 provides the equivalent estimates for women. The results are separated into three panels (single accrual, peak value, and option value) according to the specification of the incentive variable. These are precisely as defined in the previous section and, in particular, allow for the basic pension, SERPS, and occupational schemes where the individual is eligible. They also allow for eligibility to invalidity benefit according to the assigned probability model described in section 11.4.3. The columns in each panel differ according to the specification of age effects. In the first column, a linear age term is allowed. It may be that all other age effects are simply due to the wealth and pension incentives, in which case this specification will be adequate. However, given that we are mixing different date of birth cohorts in this survey and that age effects may represent preferences as well as incentives, the next two columns allow for alternative age specifications. The second column includes a date-of-birthcohort dummy and the final column includes a full set of age dummies. Each panel is further separated according to whether a dummy for the

Pension Incentives and Pattern of Retirement in the United Kingdom Table 11.13

669

The Probability of Female Retirement (722 observations) Linear Age Marginal Effect

Cohort Dummies

SE

Marginal Effect

SE

Age Dummies Marginal Effect

SE

Single Period Accruals Excluding age first eligible Pension wealth Single-period accrual Spouse wealth Pseudo R 2 Log-likelihood F test (PW, Accrual)

.0677 –.3856 .0343 .1085 –285.17 6.23

.0274 .3243 .0172

.0726 –.5986 .0344 .1025 –287.09 7.60

.0283 .3300 .0177

.0568 –.0744 .0346 .1407 –274.86 4.92

.0265 .3142 .0167

Including age first eligible Pension wealth Single-period accrual Spouse wealth Penage dummy Pseudo R 2 Log-likelihood F test (PW, Accrual)

.0536 –.1176 .0308 .1309 .1291 –278.58 4.12

.0269 .3212 .0168 .0405

.0590 –.3857 –.0317 .1077 .1165 –282.62 4.57

.0282 .3307 .0175 .0400

.0536 –.0391 .0334 .0546 .1420 –274.44 4.48

.0266 .3164 .0167 .0644

Peak Value Excluding age first eligible Pension wealth Peak accrual Spouse wealth Pseudo R 2 Log-likelihood F test (PW, Accrual) Including age first eligible Pension wealth Peak accrual Spouse wealth Penage dummy Pseudo R 2 Log-likelihood F test (PW, Accrual) (continued )

.0564 –.2848 .0313 .1111 –284.33 7.72

.0255 .1683 .0171

.0557 –.3934 .0303 .1065 –285.80 9.70

.0263 .1706 .0176

.0551 –.2212 .0329 .1440 –273.83 6.76

.0250 .1572 .0167

.0503 –.1930 .0292 .1268 .1312 –277.90 5.35

.0252 .1638 .0168 .0397

.0482 –.3277 .0286 .1070 .1209 –281.21 7.04

.0262 .1682 .0175 .0392

.0534 –.2046 .0322 .0430 .1448 –273.56 6.05

.0251 .1588 .0167 .0631

age at which individuals become eligible to receive a pension is included— the “pension age.” For recipients of the basic state pension and SERPS, this is the normal state pension age (sixty-five for men and sixty for women). For individuals with an occupational pension, we use the age at which they are entitled to start drawing their occupational pension.18 This 18. The results presented in the tables in this section focus only on the pension measures. A full result for a limited number of specifications are presented in table 11A.3. For example, table 11A.5 shows the effect of other demographic and economic characteristics on the probability of retirement.

670 Table 11.13

Richard Blundell, Costas Meghir, and Sarah Smith (continued) Linear Age Marginal Effect

Cohort Dummies

SE

Marginal Effect

SE

Age Dummies Marginal Effect

SE

Option Value Excluding age first eligible Pension wealth Option value Spouse wealth Pseudo R 2 Log-likelihood F test (PW, Accrual)

.0470 –.4044 .0322 .1069 –285.69 5.38

.0292 .6307 .0172

.0277 –1.2782 .0298 .1042 –286.55 8.78

.0291 .6113 .0176

.0463 –.3528 .0366 .1412 –274.72 5.22

.0284 .6029 .0167

Including age first eligible Pension wealth Option value Spouse wealth Penage dummy Pseudo R 2 Log-likelihood F test (PW, Accrual)

.0429 –.2992 .0297 .1334 .1293 –278.53 4.26

.0289 .6195 .0168 .0397

.0205 –1.2237 .0168 .1150 .1208 –281.23 7.28

.0290 .6073 .0275 .0393

.0452 –.3112 .0326 .0531 .1424 –274.31 4.73

.0284 .6045 .0167 .0640

Note: The full set of demographic controls include earnings (and earnings squared), education, health, job tenure, industry, proportion of time spent in full-time employment, whether individual has an occupational pension, housing tenure, financial wealth, age difference within couples, spouse’s earnings, spouse’s health, and whether spouse is retired. PW = pension wealth; SE = standard error.

varies across individuals in occupational pension schemes so that it has potential explanatory power even when added to the specification with the full set of age dummies in the final column. A broad look across the results in Table 11.12 is quite encouraging for the retirement model. In all cases, the pension wealth and incentive variables are jointly significant. In all but one of the eighteen specifications, the signs are as we would expect—a positive wealth effect and a negative accrual effect. These results are consistent with the presence of both income and substitution effects in retirement decisions.19 The positive coefficient on the total pension wealth variable points to an income effect, whereby individuals who accumulate a lot in earlier years retire earlier. The impact of the option value reflects future opportunities foregone by stopping working now; the negative coefficient on this term indicates that the greater those foregone opportunities, the less likely individuals are to retire. Since the incentive variables are measured in $100,000, the coefficient of –0.6397 on the option value in the final column, for example, implies that a $10,000 rise in the option values (leaving pension wealth unaffected) reduces the probability of retirement by a little over 6 percentage points. The counter19. The option value and total pension wealth measures are in hundred-thousands while net earnings are in thousands.

Pension Incentives and Pattern of Retirement in the United Kingdom

671

factual simulations in the next section are intended to shed more light on what these magnitudes are likely to mean in reality. The significance of these coefficients requires some discussion. The panel nature of the survey means that the standard errors calculated from the standard formula for the probit model will not account for the dependence across time periods. In table 11A.9, we present bootstrap confidence intervals that do allow for this dependence. Interestingly, these intervals maintain the significance found in the wealth and incentive variables in table 11.10. A more detailed examination of table 11.12 reveals some further interesting features. On pure likelihood grounds, the specifications that include the option value dominate specifications with the more ad hoc incentive variables. The dummy for the age at which the pension is first eligible is typically significant, although slightly less so for the option value specifications. The inclusion of date-of-birth-cohort effects, in the second column of each panel, tends to reduce the impact of the wealth variables. This seems consistent with the strong differences in wealth across cohorts. At the same time, it leads to a strong increase in the incentive variable. Indeed, for the peak value specifications, it is the only case where the incentive variable remains significant. Including a completely unrestricted set of age effects reduces the magnitude of the substitution effect, although the wealth effect remains positive and significant. As we saw in the previous section, the option value has quite a lot of variation, even after including a full set of age effects. The estimates for the option value that also include the pension age dummy are the preferred (on the grounds of likelihood) and yield a marginally significant option value coefficient, albeit much reduced from the cohort dummy specification. Interestingly, the block bootstrap standard errors reported in table 11A.6 show a 95 percent interval that remains negative, suggesting a significant negative effect even in this specification with age effects and the pension age dummy. Figure 11.6 compares the within-sample predictive performance of these model specifications. A number of immediate features stand out. First, even without including a full set of age dummies, we manage to predict a large amount of retirement before state pension age (sixty-five). This is due to the impact of invalidity benefit and early retirement incentives in the occupational systems. Second, the linear-age and cohort-effects specifications completely fail to capture the spike at sixty-five. Note that these specifications do not include the age first eligible to pension variable—we discuss this specification in more detail in our simulation of pension reforms. The linear-age effects specification does not manage to capture the downturn in retirement hazards that occurs after sixty-five. The retirement model results for women, presented in table 11.13, are similar to those for men, although the magnitude of the coefficients is typically smaller. For the majority of women, their decision to continue working—

672

Richard Blundell, Costas Meghir, and Sarah Smith

Fig. 11.6

Actual and predicted retirement hazards (option value)

and their decision to retire—is completely unaffected by the incentives in their own pension scheme, since they do not receive a pension in their own right. A lot of the identification of incentive effects is likely to come from exploiting variation between the set of women who do have their own pension and those who do not. In practice, however, these two groups of women are likely to differ in several other key respects, which makes it difficult to measure the pension incentive effects with a high degree of accuracy. Very few of the other included demographic and economic variables are individually significant (see table 11A.5). Among those that are significant—for both men and women—are self-reported health status at the time of the first interview, the retirement status of the spouse (someone is more likely to retire if their spouse is already retired), and whether or not someone has an outstanding mortgage, which tends to reduce the probability of retirement. This is consistent with the idea that people might carry on working in order to pay off their mortgage. Typically, the occupational pension dummy is positive and significant. This might reflect unmeasured incentives arising from occupational pension rules or the fact that people tend to select jobs with occupational pension schemes according to their underlying preferences for early retirement. 11.5.3 Evidence for Liquidity Constraints An interesting feature of the results in tables 11.12 and 11.13 is the significance of the pension age dummy. One possibility is that, prior to the age

Pension Incentives and Pattern of Retirement in the United Kingdom Table 11.14

Evidence of Liquidity Constraints (men) Marginal Effect

Pension wealth Option value LiqPenW LiqOV Log-likelihood

673

SE

Marginal Effect

All Men (N = 1,276) .0195 .0371 .3650 –.3572 –.0083 –1.4586 –397.33 –387.79 .0282 –1.5592

SE

.0202 .4390 .0195 .3975

Marginal Effect

SE

.0039 –1.6173 –390.63

.0190 .3443

Pension wealth Option value LiqPenW LiqOV Log-likelihood

Financial Wealth £3,000 (N = 757) –.0186 .0255 –.0096 .0307 –2.3928 .4741 –1.7453 .6182 –.0012 .0266 –.6366 .5145 –208.54 –207.07

–.0096 –1.5393 –211.72

.0241 .4244

Pension wealth Option value LiqPenW LiqOV Log-likelihood

With Occupational Pension (N = 875) .0509 .0239 .0587 .0250 –1.5178 .4270 –.1866 .5295 –.0102 .0228 –1.7021 .5054 –293.12 –285.33

.0060 –1.8349 –288.86

.0255 .4317

With No Educational Qualifications (N = 561) –.0338 .0597 .0388 .0655 .7419 –.1174 .9399 –.0558 .0595 –2.9933 .9198 –184.33 –176.77

.0470 –3.0554 –177.10

.0568 .7341

Pension wealth Option value–2.4158 LiqPenW LiqOV Log-likelihood

Notes: Controls included for demographics, earnings, cohort dummies, and age of first entitlement to pension. SE = standard error; N = number of observations.

at which individuals become entitled to start drawing their pension, they are liquidity constrained and unable to borrow against their future pension wealth, even if it is quite large.20 Reaching pension age and being able to start drawing their pension therefore may have a significant effect on the probability of retirement in addition to the incentive effects. Table 11.14 is an attempt to examine this. The first panel considers the complete sample of men used in table 11.12 and includes two new variables—LiqPenW and LiqOV. These variables calculate wealth and incentives assuming that pension wealth only matters at the time the individual becomes entitled to start drawing on the pension income. Because individuals cannot directly draw on their wealth before this age, it is maybe assumed not to matter for retirement decisions. 20. It may well also be the case that they are uncertain about the amount they will receive. Although this should not be the case for state pension incomes.

674

Richard Blundell, Costas Meghir, and Sarah Smith

At first sight, this hypothesis seems quite plausible. However, the results presented in table 11.14 are mixed. The LiqPenW variable, which is where one might think the dominant impact of such an effect would occur, is never significant, and the original pension wealth variable always dominates. Perhaps the impact would be more important for those with smaller amounts of financial wealth. The second panel does not lend support to this. Here we just select those with financial wealth holdings less than £3,000. There is no noticeable impact of the liquidity measures of pension wealth. Why should this be? One fact that we have pointed out is the low value of the state pension for most of those eligible for state pension. For many they will have their incomes in retirement topped up by welfare benefits. Moreover, if they retire before the state pension age, they will often be eligible for disability benefits and will receive an income much like the state pension. Their net incomes from employment as they age will stay quite stable and they have little reason to save or borrow. Moreover, since those eligible for state pension are typically lower skilled, their earnings in work are also quite low. The consequence is that their net replacement rates are little under 100 percent. Although they may implicitly face liquidity constraints, these are never binding and therefore have no impact on retirement decisions. The remaining panels of table 11.14 further investigate the evidence for liquidity effects among those with occupational pensions and also those with lower educational qualifications. Again, in neither case is there much evidence that such constraints are binding on the retirement decision. 11.6 Counterfactual Simulations To illustrate the size of the estimated incentive effects on retirement behavior, we consider the effects of reforms to the pension system on the predicted probabilities of retirement at different ages. Two alternative scenarios are considered. The first counterfactual is to increase the pension age for everyone by three years. This means that the state pension age is sixty-eight for men and sixty-three for women. We also augment the normal occupational pension retirement ages by three years. There is clearly a correspondence in practice between the state pension ages and the normal retirement ages in occupational pension schemes, so increasing the state pension is likely to have a knock-on effect on occupational pension schemes. Moreover, the underlying demographic pressures that are likely to cause the government to reduce the generosity of the state pension system will have a similar effect on occupational schemes. The second counterfactual assumes a pension system of the following form.

Pension Incentives and Pattern of Retirement in the United Kingdom

• • • • •

675

An early entitlement age of sixty A normal retirement age of sixty-five A 60 percent replacement rate at age sixty-five A 6 percent actuarial adjustment from sixty to seventy no other pathways to retirement

The effects of each of these alternative scenarios on the distribution of total pension wealth and option values by age are presented in table 11.15. We report results for men only since the majority of women, who have no pension in their own right, will be unaffected. The effect of raising the retirement age is to reduce the median level of total pension wealth and to increase option values, compared to the existing pension system. The income and substitution effects work in the same direction, and the combined effect is to reduce the conditional probability of retirement at younger ages. The effects can most clearly be seen by plotting and comparing the predicted retirement probabilities under the base case of the existing pension system and under the reform. This is done in Figure 11.7, corresponding to the one-period accruals and option values respectively. The precise magnitude of the effects of reforming the pension system depend on which specification is used. When a full set of age dummies is in-

Table 11.15

Incentive Measures: The Impact of Reform (men, 1,272 observations) Base

Reform 1

Reform 2

Age

Median TW

Median OV

Median TW

Median OV

Median TW

Median OV

56 57 58 59 60 61 62 63 64 65 66 67 68 69 70

89,821 93,850 97,320 103,990 108,244 113,266 117,994 123,886 127,333 128,514 111,329 84,831 68,720 43,102 27,458

10,476 8,857 7,449 6,168 5,034 4,060 2,745 1,615 1,072 681 312 126 322 1,480 2,129

76,862 80,639 83,564 88,016 90,293 93,686 97,107 101,717 103,070 105,048 114,466 111,553 109,707 82,729 67,553

13,353 11,785 10,320 9,430 8,084 7,113 6,046 5,252 4,229 3,021 1,544 1,034 –119 –138 –217

102,841 101,072 105,531 113,593 116,247 119,261 118,727 127,663 125,523 121,037 131,675 115,668 115,164 105,251 71,404

13,660 13,019 12,880 13,200 13,384 11,962 11,112 10,182 9,768 9,151 8,332 7,007 6,211 4,574 2,990

Notes: OV = option value; TW = total wealth. Reform 1 raises the pension age by three years; reform 2 introduced a pension system with a 60% replacement rate at the normal pension age of 65, an early retirement age of 60, and a 6% deferral rate.

Fig. 11.7

A

Predicted retirement hazards: A, one-period accruals; B, option value

Fig. 11.7

B

(cont.)

678

Richard Blundell, Costas Meghir, and Sarah Smith

cluded these tend to dominate any of the pension wealth and accrual incentives and the effect of reforming the pension system appears to be very small. To the extent that the age dummies pick up the incentive effects, these would need to be adjusted to reflect the pivotal ages in the new system. The bottom-right-hand graph in figure 11.7 shows the effect of changing the “pension age.” When a set of cohort dummies is included, the effect of the forwardlooking accrual and option value measures is much stronger—as was seen from the regression results in the previous section, and this is reflected in bigger predicted responses from increasing the retirement age by three years. Looking at figure 7, panel A, which shows the retirement probabilities based on the one-period-ahead accrual, the effect is to halve the predicted probability of retirement at sixty-five. Including option values tends to smooth the effects over a longer period, as would be expected from a more forward-looking model. The probability of retirement is reduced by between 3–7 percentage points between the ages of sixty and sixty-six. The effect of the second simulated reform is to increase the level of pension wealth and to increase the option values, compared to the existing system. Under the simulated reform, the level of annual pension income that the simulated pension system produces is relatively generous compared to the existing U.K. system. The option value effect is reinforced by the absence of any nonpension benefits (such as disability benefits) before retirement age under the simulated reform which increases the incentive to stay in work. The effect on one-period accruals is slightly different and worth commenting on. Under the simulated reform, everyone is granted a full entitlement at age sixty and the level of pension is determined on the basis of earnings at age sixty-five. Therefore, the decision to continue working, before age sixty, has no effect on pension entitlement, and up until age sixty, the one-period-ahead accrual is zero. Only after age sixty is the one-period accrual positive (and higher than under the existing system). If the decision to leave work were reversible, then the optimal thing to do in terms of maximizing pension benefits might be to leave work until age sixty and then reenter to increase pension entitlements. The option value approach implicitly rules this out, and before age sixty gives a positive value to the option to increase pension value by working between sixty-one and sixty-two. The effects of this can be seen from the simulated retirement probabilities in figure 11.7. Looking at the simulations with cohort dummies and the one-period-ahead accruals, the probability of retirement before age sixty is higher under the case of the second reform than it is in the base case. This reflects the higher level of pension wealth and the lower (effectively zero) accrual rate. After age sixty, however, the one-period accrual is positive and higher than under the existing system. The income and substitution effects now work in the opposite direction, with the substitution effect

Pension Incentives and Pattern of Retirement in the United Kingdom

679

being more powerful; the probability of retirement is lower under the reform. The option value makes the higher substitution effect count at younger ages, and the probability of retirement is reduced at all ages. 11.7 Summary and Conclusions The United Kingdom experienced a serious decline in labor market attachment among older workers in the 1980s and 1990s. This was especially acute among men aged fifty-five or older. The analysis we present shows that during the two recessions—the first in the early 1980s and the second in the early 1990s, the fraction of such men in employment declined by more than 30 percentage points to record-low levels of little over 50 percent and has shown no sign of recovery. For older women this decline was less evident, reflecting the growing participation rate among younger cohorts that offsets the decline in employment. To what extent can these low levels of labor market attachment be attributed to the workings of the U.K. pension system and to what extent can these trends be reversed by reforms to this system? These questions formed the motivation for this study. We began the paper with a comprehensive evaluation of the economic incentives for retirement underlying the U.K. pension system. This accounted for the changing impact of the SERPS, introduced in 1978 and of growing importance for those retiring in the 1990s. It also accounted for the complex set of private defined-benefit occupation pension schemes, which provided coverage for nearly 70 percent of those approaching retirement in the 1990s. We highlighted the importance of invalidity benefit as a mechanism for income support in early retirement whose take-up approached nearly 1.5 million among individuals below state retirement age in the 1990s. To examine the impact of these factors on retirement, we used a sample of individuals aged fifty-five or older from the U.K. RS. Their retirement probability was modeled in terms of the incentives underlying their own pension plans and other socioeconomic factors. Our analysis followed an option value approach and allows a separate role for pension wealth. We also allowed for the spouse’s economic and demographic characteristics. The estimation results pointed to significant incentive and wealth effects through the pension system. The magnitude of these turned out to be quite sensitive to the specification of age effects. To allow for this, we considered three basic specifications. The first restricted age effects to enter linearly. This is clearly rejected by the data in favor of more general specifications but we retain it as a baseline specification. The second allowed for cohort effects, and the third allowed for a complete set of age dummies. Not surprisingly, the full set of age dummies was found to provide the best fit; but even in this case, the wealth and incentive variables remained correctly signed and significant. Overall the option value model performed better than models that used simpler and more ad hoc incentive measures.

680

Richard Blundell, Costas Meghir, and Sarah Smith

On their own, the incentive and pension wealth measures were unable to explain the large increase in the retirement at the normal retirement age (sixty-five for men and sixty for women). This explained the much improved fit of the age dummy specification. Nonetheless, the incentive and wealth variables alone managed to explain the most part of large amount of early retirement that occurs prior to the normal retirement age. This appeared to be due to a combination of the ability for invalidity benefit to act as an early retirement incentive and the significant incentives for early retirement that occur in occupational schemes in the United Kingdom. To explain the spike in the exit probability at the normal retirement age, we included a dummy for the age at which the individual could first draw down their pension. This variable was shown to contain variation over and above the age dummies because this age varied across occupational schemes. Even with this variable included, the option value incentive and pension wealth variables remained significant. We also investigated whether or not the significance of this variable could be attributed to a liquidity effect— that is, a wealth effect that only became important at the point at which the pension could be drawn. We found little success in this explanation. Each of these specifications was used to simulate two policy reforms. The first—reform 1—involved an increase by three years in the pension age. The second—reform 2—was more complex. This later reform had an early retirement age at sixty with a normal retirement at age sixty-five. This was matched by a 60 percent replacement rate at sixty-five and a 6 percent actuarial adjustment from sixty to seventy. Finally, all other pathways to retirement were eliminated. Reform 1 had a clear impact on retirement in all specifications—moving the retirement peak to a later age and significantly cutting the incidence of early retirement. Reform 2 had an even more dramatic impact on early retirement, resulting in a smooth and lower rate of exit into retirement at all ages. As a more cautious final note, it should be pointed out that the data source we used had a number of drawbacks. Most notably, the high attrition between waves and the resulting small sample size used in our analysis. In addition, many of the features of the occupational plans that we would like to include are missing from the data. More optimistically, the new English Longitudinal Survey of Ageing, which will produce the first wave in December 2003, will remedy both of these defects and will also provide a comprehensive and detailed data source on health and retirement.

Pension Incentives and Pattern of Retirement in the United Kingdom

681

Appendix Table 11A.1

Calculating SERPS using Accrual Rates, Earnings Factors, and Upper- and Lower-Earnings Limits (UEL, LEL, respectively) Accrual Rate on Earnings

Year of Retirement

Between 1978–79 and 1987–88

1988–89 Onwards

25-20 = 1.25 25-21= 1.19 25-26 = 0.96 25-31 = 0.81 25-36 = 0.69 25-41 = 0.61 25-46 = 0.54 25-49 = 0.51

25-20 = 1.25 25-21 = 1.19 22.5-26 = 0.87 20-31 = 0.65 20-36 = 0.56 20-41 = 0.49 20-46 = 0.431 20-49 = 0.41

Before 1998–99 2000–01 2005–06 2010–11 2015–16 2020–21 2025–26 2027–28 onwards Year of Earnings

Factor (%)

1978–79 1979–80 1980–81 1981–82 1982–83 1983–84 1984–85 1985–86 1986–87 1987–88

377 320.7 251.6 194.8 167.6 148 129.6 115.5 97.9 84.4

Year

LEL

UEL

1978–79 1979–80 1980–81 1981–82 1982–83 1983–84 1984–85 1985–86 1986–87 1987–88

17.5 19.5 23.0 27 29.5 32.5 34 35.5 38 39

120 135 165 200 220 235 250 265 285 295

Year of Earnings

Factor (%)

1988–89 1989–90 1990–91 1991–92 1992–93 1993–94 1994–95 1995–96 1996–97 Year 1988–89 1989–90 1990–91 1991–92 1992–93 1993–94 1994–95 1995–96 1996–97 1997–98

69.5 53.5 43.1 30 22.1 16.3 11.7 8 5 LEL

UEL

41 43 46 52 54 56 57 58 61 62

305 325 350 390 405 420 430 440 455 465

Table 11A.2

Probit Results, Invalidity Benefit Men Coefficient

Women SE

Coefficient

SE

Yorks and Humberside Northwest East Midlands West Midlands East Anglia South East Greater London South West Wales Scotland

–.0958 .0217 –.4242 –.3005 –.6051 –.5109 –.5336 –.4568 .2494 .0179

.0748 .0713 .0857 .0773 .1189 .0835 .0731 .0852 .0809 .0743

–.1601 .2176 –.2936 –.3128 –.4210 –.2288 –.4302 –.3078 .3606 .2237

.1129 .0989 .1256 .1155 .1700 .1165 .1101 .1253 .1088 .0998

Age Age2 College education Spouse employed Homeowner with mortgage Outright owner

.5859 –.0054 –.6891 .1476 .0109 –.0850

.0246 .0002 .0962 .0416 .0484 .0399

.4968 –.0048 –.0452 –.1923 –.0529 –.2829

.0354 .0003 .1019 .0467 .0614 .0595

Constant No. of observations Pseudo R 2

–15.7892 9,363 0.2047

.6933

–13.5498 14,192 0.2013

.9405

Source: Data from the Family Expenditure Survey (U.K. Data Archive 1996) for the period from April 1988 to March 1994. Note: SE = standard error.

Table 11A.3

Regression Results Dependent Variable Total Pension Wealth

Option value Total pension wealth Spouse pension wealth Difference in ages Job tenure % of full-time employment Education Health score Spouse health score Renter Mortgage Industry, engineering Industry, manufacturing Industry, distribution Industry, services Industry, government Spouse retired Occupational pension £1–£3,000 wealth £3,000–£10,000 wealth
£10,000 wealth Missing wealth Earnings Spouse earnings Pension age Age dummies R2

Accrual

Peak Value

A. Wealth and Accrual, Men (1,276 observations) –8.971 .5268 .0326 .0027 .0382 –.2272 .0189 .0114 .0021 .0107 .0299 .0026 .0001 .0002 .0036 .0072 .0018 .0000 .0002 .0014 .0765 .0464 .0041 .0049 –.0189 .1296 .0229 .0016 .0024 .0032 .0089 .0126 –.0001 .0013 .0033 .0094 .0075 –.0006 .0008 –.0018 –.0205 .0279 .0034 .0029 –.0012 .2254 .0269 .0044 .0029 .0137 –.1137 .0415 .0066 .0044 –.0145 –.0927 .0425 –.0039 .0045 –.0400 –.0401 .0413 .0031 .0044 –.0226 .0205 .0404 .0012 .0043 –.0265 .3681 .0541 –.0016 .0058 –.0476 –.0146 .0365 –.0072 .0039 –.0061 .1923 .0271 .0177 .0028 .0355 –.0727 .0309 .0060 .0033 .0136 –.0039 .0350 .0136 .0037 .0297 .3155 .0404 .0236 .0042 .0642 –.0010 .0557 .0111 .0059 .0133 .0272 .0019 .0001 .000 .0003 .0025 .0019 –.0008 .0002 –.0015 .4908 .0623 –.0166 .0068 –.0704 Yes .6054

Yes .5976

.0064 .0050 .0006 .0004 .0117 .0058 .0031 .0018 .0070 .0069 .0104 .0107 .0103 .0101 .0138 .0091 .0066 .0077 .0087 .0099 .0140 .0004 .0004 .0160

Yes .4890

B. Wealth and Accrual, Women (722 observations) Option value –10.84 .6996 Total pension wealth .0329 .0027 .0183 Spouse pension wealth –.1515 .0211 .0031 .0018 –.0022 Difference in ages .0132 .0063 –.0007 .0005 –.0023 Job tenure –.0033 .0023 –.0003 .0001 –.0010 % of Full-Time employment .2429 .0631 .0110 .0054 .0468 Education .1602 .0396 .0087 .0033 .0324 Health score –.0228 .0170 –.0019 .0014 –.0033 Spouse health score –.0029 .0164 .0010 .0013 –.0010 Renter .0246 .0485 –.0055 .0041 –.0160 Mortgage .0890 .0402 .0026 .0034 .0118 Industry, engineering –.3168 .1006 –.0133 .0085 –.0065 Industry, manufacturing –.1573 .0771 –.0348 .0065 –.0486 Industry, distribution –.2116 .0747 –.0259 .0063 –.0341 Industry, services –.0207 .0648 –.0195 .0054 –.0107 Spouse retired –.0413 .0572 –.0106 .0048 –.0188 (continued )

.0064 .0043 .0012 .0004 .0126 .0078 .0033 .0032 .0095 .0079 .0199 .0152 .0148 .0128 .0112

Option Value

–.0211 –.0042 .0016 .0000 .0016 .0021 –.0008 –.0011 .0017 .0045 .0046 –.0002 .0005 .0001 –.0030 –.0011 .0149 .0058 .0060 .0202 .0046 .0020 –.0003 –.0012

.0012 .0009 .0001 .0000 .0022 .0011 .0006 .0003 .0013 .0013 .0020 .0020 .0020 .0019 .0026 .0017 .0012 .0014 .0016 .0019 .0027 .0000 .0000 .0031

Yes .7597

–.0238 –.0013 –.0005 –.0001 –.0007 .0044 –.0014 .0005 –.0015 .0015 .0076 –.0065 –.0029 .0008 –.0060

.0015 .0010 .0003 .0001 .0029 .0018 .0007 .0007 .0022 .0018 .0047 .0036 .0035 .0030 .0026

Table 11A.3

(continued) Dependent Variable Total Pension Wealth

Occupational pension £1–£10,000 wealth £3,000–£10,000 wealth
£10,000 wealth Missing wealth Earnings Spouse earnings Pension age

.4871 .0245 –.1071 .2858 –.1200 .0343 –.0097 .0759

Age dummies R2

Table 11A.4

Peak Value

.0041 .0040 .0048 .0055 .0070 .0003 .0001 .0070

.0865 –.0250 –.0200 –.0019 –.0243 –.0003 –.0010 –.0560

Yes .4833

Option Value

.0095 .0094 .0113 .0129 .0164 .0007 .0003 .0163

.0160 –.0054 –.0048 .0006 –.0063 .0036 –.0002 –.0056

Yes .4546

.0022 .0022 .0026 .0030 .0039 .0001 .0000 .0038

Yes .6795

Retirement Probabilities: One-Year Accrual, Men (1,276 observations)

Marginal Effect

SE

Demographics Marginal Effect

SE

Demographics, Earnings, and Cohort Dummies Marginal Effect

SE

A. Not Including Age of First Entitlement to Pension .1024 .0158 .1152 .0189 .1128 .0185 –1.1481 .1514 –1.1869 .1512 –.8903 .1519 .0176 .0167 .0367 .0167 .0380 .0163 .0001 .0013

Pseudo R 2 Log-likelihood F test (PW, Accrual) Pension wealth Accrual Spouse wealth Penage dummy Net earnings

.0159 –.0122 –.0106 –.0101 –.0139 .0004 –.0001 –.0183

Yes .5683

No Controls

Pension wealth Accrual Spouse wealth Net earnings

.0466 .0482 .0574 .0644 .0835 .0045 .0019 .0827

Accrual

.0786 –463.2 76.71

.1435 –430.5 83.25

.1869 –408.7 61.87

B. Including Age of First Entitlement to Pension (Penage) .0638 .0166 .0812 .0193 .0853 .0192 –.6461 .1632 –.7627 .1604 –.6342 .1588 .0062 .0166 .0254 .0167 .0283 .0165 .2649 .0498 .2205 .0484 .1594 .0455 .0004 .0013

Pseudo R 2 Log-likelihood F test (PW, Accrual) Note: SE = standard error.

.1229 –440.9 21.33

.1831 –410.1 29.50

.2074 –390.4 27.76

Demographics, Earnings, and Age Dummies Marginal Effect

SE

.0888 –.1153 .0302 .0003

.0180 .1768 .0157 .0012

.2337 385.2 24.29 .0802 –.1049 .0276 .0955 .0004

.0185 .1777 .0156 .0634 .0012

.2370 –383.5 18.71

Table 11A.5

Retirement Probabilities: Peak Values, Men (1,276 observations)

No Controls Marginal Effect

Pension wealth Peak value Spouse wealth Net earnings

Marginal Effect

SE

Marginal Effect

SE

A. Not Including Age of First Entitlement to Pension .0851 .0150 .0923 .0181 .0982 .0179 –.5018 .0766 –.5054 .0759 –.3211 .0779 .0114 .0167 .0246 .0168 .0296 .0165 .0004 .0013

Pseudo R 2 Log-likelihood F test (PW, Accrual) Pension wealth Peak value Spouse wealth Penage dummy Net earnings

SE

Demographics

Demographics, Earnings, and Cohort Dummies

.0625 –471.3 59.96

.1225 –441.1 62.50

.1190 –442.8 16.90

.1706 –416.9 22.71

Marginal Effect

SE

.0868 –.0283 .0290 .0002

.0175 .0764 .0155 .0012

.1662 –419.1 42.61

B. Including Age of First Entitlement to Pension (Penage) .0514 .0155 .0628 .0183 .0702 .0185 –.2689 .0793 –.3041 .0777 –.2045 .0787 .0022 .0165 .0167 .0166 .0206 .0165 .2935 .0489 .2602 .0485 .2007 .0470 . .0006 .0013

Pseudo R 2 Log-likelihood F test (PW, Accrual)

Demographics, Earnings, and Age Dummies

.2334 –385.4 24.08 .0780 –.0141 .0264 .0953 .0004

.1979 –403.2 18.21

.0181 .0763 .0155 .0635 .0012

.2367 –383.7 18.45

Note: SE = standard error. Table 11A.6

Retirement Probabilities: One-Year Accruals, Women (722 observations)

No Controls Marginal Effect

Pension wealth Accrual Spouse wealth Net earnings

Marginal Effect

SE

Marginal Effect

SE

A. Not Including Age of First Entitlement to Pension .0862 .0256 .0827 .0279 .0821 .0276 –.6064 .3176 –.6549 .3398 –.5645 .3296 .0247 .0158 .0226 .0170 .0314 .0174 –.0007 .0030

Pseudo R 2 Log-likelihood F test (PW, Accrual) Pension wealth Accrual Spouse wealth Penage dummy Net earnings

SE

Demographics

Demographics, Earnings, and Cohort Dummies

.0188 –313.9 11.29

.0702 –297.4 9.38

.0973 –288.8 9.29

B. Including Age of First Entitlement to Pension (Penage) .0730 .0256 .0694 .0278 .0694 .0275 –.3590 .3206 –.4076 .3409 –.3523 .3307 .0223 .0157 .0199 .0169 .0290 .0172 .1312 .0414 .1228 .0411 .1063 .0399 –.0009 .0030

Pseudo R 2 Log-likelihood F test (PW, Accrual) Note: SE = standard error.

.0378 –307.8 8.58

.0874 –291.9 6.27

.1109 –274.4 6.36

Demographics, Earnings, and Age Dummies Marginal Effect

SE

.0710 –.0521 .0309 –.0025

.0262 .3181 .0166 .0028

.1306 –278.1 8.23 .0678 –.0170 .0298 .0543 –.0025

.0264 .3202 .0166 .0655 .0028

.1319 –277.7 7.69

Table 11A.7

Retirement Probabilities: Peak Values, Women (722 observations)

No Controls Marginal Effect

Pension wealth Peak value Spouse wealth Net earnings

Marginal Effect

SE

Marginal Effect

SE

A. Not Including Age of First Entitlement to Pension .0735 .0225 .0657 .0251 .0673 .0253 –.2388 .1407 –.3990 .1781 –.3608 .1707 .0231 .0158 .0185 .0170 .0278 .0173 –.0011 .0031

Pseudo R 2 Log-likelihood F test (PW, Accrual) Pension wealth Peak value Spouse wealth Penage dummy Net earnings

SE

Demographics

Demographics, Earnings, and Cohort Dummies

.0711 –314.2 10.94

.0730 –296.5 11.04

.1004 –287.8 11.04

B. Including Age of First Entitlement to Pension (Penage) .0668 .0224 .0597 .0250 .0607 .0252 –.1634 .1392 –.3151 .1752 –.2930 .1684 .0216 .0156 .0169 .0168 .0264 .0172 .1340 .0409 .1223 .0404 .1055 .0392 –.0012 .0030

Pseudo R 2 Log-likelihood F test (PW, Accrual) Note: SE = standard error.

.0381 –307.7 8.89

.0907 –290.9 8.18

.1143 –283.3 8.35

Demographics, Earnings, and Age Dummies Marginal Effect

SE

.0712 –.1791 .0296 –.0027

.0244 .1592 .0166 .0028

.1327 –277.4 9.37 .0693 –.1627 .0289 .0454 –.0027

.0245 .1607 .0166 .0646 .0028

.1336 –277.2 8.62

Table 11A.8

Retirement Probabilities: Option Values Men Demographics, earnings, and cohort dummies

Total wealth Option value Spouse wealth Net earnings Spouse net earnings Pension age Difference in ages Job tenure % of Full-Time employment Education Health score Spouse health score Renter Mortgage Industry, engineering Industry, manufacturing Industry, distribution Industry, services Industry, government Spouse retired Occupational pension £1–£3,000 wealth £3,000–£10,000 wealth
£10,000 wealth Missing wealth Cohort born 1934 Cohort born 1935 Cohort born 1936 Cohort born 1937 Cohort born 1938 Cohort born 1939 Cohort born 1940 Cohort born 1941 Cohort born 1942 (continued )

Women Demographics, earnings, and age dummies

Demographics, earnings, and cohort dummies

Demographics, earnings, and age dummies

Marginal Effect

SE

Marginal Effect

SE

Marginal Effect

SE

Marginal Effect

SE

.0282 –1.559 .0123 .0032 –.0031 .2135 –.0010 .0005

.0195 .3650 .0162 .0014 .0016 .0464 .0021 .0013

.0650 –.6130 .0243 .0014 –.0027 .0985 –.0033 –.0001

.0198 .4431 .0154 .0014 .0015 .0636 .0021 .0012

.0302 –1.243 .0251 .0032 –.0017 .1126 –.0067 .0028

.0282 .6109 .0172 .0036 .0016 .0392 .0049 .0018

.0594 –.3382 .0291 –.0014 –.0021 .0523 –.0065 .0029

.0280 .6152 .0166 .0034 .0015 .0650 .0048 .0018

.0213 –.0147 .0158 –.0078 –.0021 –.0254 .0280

.0356 .0191 .0086 .0062 .0224 .0216 .0354

.0419 –.0195 .0205 –.0086 –.0155 –.0388 .0511

.0338 .0187 .0083 .0059 .0202 .0197 .0384

.0361 –.0143 .0229 –.0210 –.0154 –.0346 .0119

.0505 .0330 .0131 .0155 .0369 .0320 .0847

.0307 .0035 .0265 –.0236 –.0035 –.0313 –.0107

.0487 .0319 .0127 .0147 .0374 .0307 .0730

–.0218 –.0187 –.0547 –.0285 .0769

.0306 .0286 .0240 .0327 .0378

.0022 –.0038 –.0469 –.0197 .0604

.0335 .0306 .0241 .0340 .0352

–.0397 .0525 –.0179

.0522 .0689 .0527

–.0441 .0535 –.0283

.0492 .0674 .0515

.1496

.0603

.1325

.0585

.0535 .0237

.0184 .0270

.0490 .0260

.0182 .0260

.0428 –.0030

.0428 .0405

.0381 –.0052

.0412 .0395

.0473 .0607 .0330 –.0168 –.0350 .0094 –.0364 –.0406 –.0632 –.0668 –.0729 –.0411

.0351 .0423 .0552 .0368 .0327 .0474 .0320 .0302 .0265 .0265 .0249 .0367

.0398 .0255 .0405

.0335 .0363 .0565

.0295 –.0097 –.0287

.0509 .0522 .0658

.0434 –.0321 –.0534

.0523 .0457 .0518

.0834 .0260 .0440 .0843 .1098

.0590 .0440 .0425 .0355 .0346

Table 11A.8

(continued) Men Demographics, earnings, and cohort dummies Marginal Effect

SE

Age = 57 Age = 58 Age = 59 Age = 60 Age = 61 Age = 62 Age = 63 Age = 64 Age = 65 Age = 66 Age = 67 Age = 68 Age = 69 Age = 70

Women Demographics, earnings, and age dummies

Demographics, earnings, and cohort dummies

Marginal Effect

SE

Marginal Effect

.0268 –.0040 .0072 –.0268 –.0008 –.0006 .0615 .0955 .2944 .2087 .3098 .2218 .4065 .7167

.1069 .0830 .0885 .0698 .0848 .0856 .1198 .1395 .2241 .2083 .2415 .2401 .2744 .2439

SE

Demographics, earnings, and age dummies Marginal Effect

SE

–.0239 .0051 –.0034 .1531 .1438 .0905 .1516 .1266 .4152 .5461

.0748 .0782 .0782 .1193 .1067 .1070 .1308 .1435 .1830 .3560

Note: SE = standard error. Table 11A.9

Bootstrap Standard Error (male retirement model) Linear Age (Marginal Effect)

Sample size Pension wealth Mean estimate 5% 95% Option value Mean estimate 5% 95% Sample size Pension wealth Mean estimate 5% 95% Option value Mean estimate 5% 95%

Cohort Dummies (Marginal Effect)

A. Excluding Age First Eligible 1,202–1,375 1,183–1,347

Age Dummies (Marginal Effect)

1,158–1,345

.0749 .0383 .1083

.0464 .0089 .0808

.0692 .0372 .1014

–.3861 –1.2363 .4071

–1.9043 –2.6422 –1.2743

–.6032 –1.3079 .0747

B. Including Age First Eligible 1,194–1,345 1,179–1,359

1,106–1,350

.0510 .0156 .0815

.0192 –.0235 .0555

.0629 .0283 .1068

–.3706 –1.1093 .2624

–1.7371 –2.4786 –1.1214

–.6833 –1.4395 –.0250

Pension Incentives and Pattern of Retirement in the United Kingdom

689

References Banks, J., and C. Emmerson. 2000. Public and private pension spending: principles, practice and the need for reform. Fiscal Studies 21 (1): 1–63. Blundell, R., and P. Johnson. 1998. Pensions and labor force participation in the UK. American Economic Review 88 (2): 173–78. ———. 1999. Pensions and retirement in the UK. In Social security and retirement around the world, ed. J. Gruber and D. Wise, 403–36. Chicago: University of Chicago Press. Bone, M., J. Gregory, B. Gill, and D. Lader. 1992. Retirement and retirement plans. London: Her Majesty’s Stationary Office (HMSO) for Office of Population Census and Surveys. Department of Social Security and Office for Population and Census Surveys. Various years. Retirement Survey. London: Department of Social Security. Disney, R. 1996. Can We Afford to Grow Older? A Perspective on the Economics of Aging. Cambridge, Mass.: MIT Press. Disney, R., C. Meghir, and E. Whitehouse. 1994. Retirement behaviour in Britain. Fiscal Studies 15 (1): 24–43. Disney, R., and S. Smith. 2002. The labour supply effects of the abolition of the earnings rule. Economic Journal 112 (478): C136–C152. Disney, R., and G. Stears. 1996. Why is there a decline in defined benefit plan membership? Working Paper no. 96/4. London: Institute for Fiscal Studies. Johnson, P. 1999. Older Getting Wiser. Sydney: Institute of Chartered Accountants in Australia. National Association of Pension Funds (NAPF). 1998. Annual survey of occupational pension schemes 1997. London: NAPF. Stock, J., and D. Wise. 1990. Pensions, the option value of work and retirement. Econometrica 58 (5): 1151–80. U.K. Data Archive. 1996. Family expenditures survey 1994–1995. http://www.dataarchive.ac.uk/finding Data/snDescription.asp?sn3478.

12 The Effect of Social Security on Retirement in the United States Courtney Coile and Jonathan Gruber

12.1 Introduction One of the most striking labor force phenomena of the second half of the twentieth century in the United States has been the rapid decline in the labor force participation rate of older men. In 1950, for example, 81 percent of sixty-two-year-old men were in the labor force; by 1995, this figure had fallen to 51 percent, although it has rebounded slightly in the past few years (Quinn 1999). Over the same period, the labor force participation rate of older women has risen dramatically, as shown in figure 12.1, due in large part to changing roles and opportunities for women during the period. Much has been written about the proximate causes of the decline in older men’s labor force participation and, in particular, about the role of the Social Security (SS) program. A large number of articles have documented pronounced “spikes” in retirement at ages sixty-two and sixty-five, which correspond to the early and normal retirement ages for SS, respectively. While there are some other explanations for a spike at age sixty-five, such as entitlement for health insurance under the Medicare program or rounding error in surveys, there is little reason to see a spike at sixty-two as Courtney Coile is assistant professor of economics at Wellesley College and a faculty research fellow of the National Bureau of Economic Research (NBER). Jonathan Gruber is professor of economics at MIT, director of the research program on children at NBER, and research associate of NBER. We are grateful to Dean Karlan for research assistance and to Peter Diamond, Alan Gustman, Jim Poterba, Andrew Samwick, and seminar participants at MIT, Harvard, NBER, and the Social Security Administration for helpful comments, and especially to David Wise and other participants in the International Social Security Comparisons project for their insights. Coile gratefully acknowledges support from the National Institute on Aging through NBER. Gruber acknowledges support from the National Institute on Aging.

691

Fig. 12.1

LFPR of men and women fifty-five to sixty-four, 1948–1998

The Effect of Social Security on Retirement in the United States

693

attributable to anything other than the SS program. Indeed, as Burtless and Moffitt (1984) document, this spike at age sixty-two only emerged after the early retirement eligibility age for men was introduced in 1961. The presence of these strong patterns in retirement data suggest that the underlying structure of SS plays a critical role in determining retirement decisions, but the impact of increases in SS generosity on retirement decisions is less obvious. A large body of literature dating from the mid-1970s has investigated this relationship, and the broad conclusion of that literature is that the level of SS benefits has a significant, but modest, effect on retirement dates. However, much of this literature either relies on data that are now decades old or otherwise flawed or it suffers from methodological problems. The purpose of our paper is to revisit the impact of SS on retirement, taking advantage of newly available data on retirement behavior and methodological advances in retirement modeling over the past decade. Our data set, the Health and Retirement Study (HRS), follows a sample of nearretirement-age individuals starting in 1992 and contains detailed information on demographic and job characteristics, labor force attachment, earnings histories, health, and private pensions. Our empirical analysis relies on the important observation of Stock and Wise (1990a,b) that it is not simply the level of retirement wealth or the increment with one additional year of work that matters, but the entire evolution of future wealth with further work. Their “option value” model posited retirement decisions as a function of the difference between the utility of retirement at the current date and at the date that maximizes one’s utility. We use this model in a reduced-form context, as well as an alternative forward-looking measure called “peak value,” introduced in Coile and Gruber (2001) and described in more detail below. We have two major findings. First, retirement appears to respond much more to SS incentive variables defined with reference to the entire future stream of retirement incentives than to the accrual in retirement wealth over the next year alone, indicating that it is important to include forwardlooking measures such as peak value or option value in retirement models. These forward-looking measures have a significant impact on retirement decisions for men, although for women only the option value model generates a significant result. Second, we conduct simulations of the effect of two possible policy changes—raising the early and normal retirement ages by three years or moving to a system with a flat benefit of 60 percent of earnings—and find that these policy changes could have significant effects on retirement behavior. Our paper proceeds as follows. In section 12.2, we briefly discuss the relevant institutional features of the SS system in the United States and provide an overview of the previous literature in this area. In section 12.3, we describe our data and incentive variable calculations. In section 12.4, we

694

Courtney Coile and Jonathan Gruber

describe the empirical framework for our regression analysis and present the results of our estimation. In section 12.5, we conduct a series of simulation exercises to assess the impact of SS reform using our model, and we present our conclusions in section 12.6. 12.2 Background 12.2.1 Institutional Features of Social Security The SS system is financed by a payroll tax that is levied equally on workers and firms. The total payroll tax paid by each party is 7.65 percentage points; 5.3 percentage points are devoted to the Old Age and Survivors Insurance (OASI) program, with 0.9 percentage points funding the Disability Insurance (DI) system and 1.45 percentage points funding Medicare’s Hospital Insurance (HI) program. The payroll tax that funds OASI and DI is levied on earnings up to the taxable maximum, $76,200 in 2000; the HI tax is uncapped. Individuals qualify for an OASI pension by working for forty quarters in covered employment, which now encompasses most sectors of the economy. Benefits are determined in several steps. The first step is computation of the worker’s averaged indexed monthly earnings (AIME), which is onetwelfth of the average of the worker’s annual earnings in covered employment, indexed by a national wage index. Importantly, additional higherearnings years can replace earlier lower-earnings years, since only the highest thirty-five years of earnings are used in the calculation.1 The next step is to convert the AIME into the primary insurance amount (PIA). This is done by applying a three-piece linear progressive schedule to an individual’s average earnings, whereby 90 cents of the first dollar of earnings is converted to benefits, while only 15 cents of the last dollar of earnings (up to the taxable maximum) is so converted. As a result, the rate at which SS replaces past earnings (the “replacement rate”) falls with the level of lifetime earnings. The final step is to adjust the PIA based on the age at which benefits are first claimed. For workers commencing benefit receipt at the normal retirement age (NRA; legislated to rise slowly from age sixty-five to sixtyseven over the next twenty years), the monthly benefit is the PIA. For workers claiming before the NRA, benefits are decreased by an actuarial reduction factor of five-ninths of one percent per month; thus, a worker 1. While earnings through age fifty-nine are converted to real dollars for averaging, earnings after age sixty are treated nominally. There is a two-year lag in availability of the wage index, calling for a base in the year in which the worker turns sixty in order to be able to compute benefits for workers retiring at their sixty-second birthdays. This implies particularly large effects of this dropout-year provision for earnings near the age of retirement, particularly in high-inflation environments.

The Effect of Social Security on Retirement in the United States

695

with an NRA of age sixty-five claiming on his sixty-second birthday receives 80 percent of the PIA.2 Individuals can also delay the receipt of benefits beyond the NRA and receive a delayed retirement credit (DRC). For workers reaching age sixty-five in 2000, an additional 6 percent is paid for each year of delay; this amount will steadily increase until it reaches 8 percent per year in 2008. While a worker may claim as early as age sixty-two, receipt of SS benefits is conditioned on the earnings test until the worker reaches age sixtyfive.3 A worker age sixty-two to sixty-five may earn up to $9,600 in 1999 without the loss of benefits, then benefits are reduced $1 for each $2 of earnings above this amount. Months of benefits lost through the earnings test are treated as delayed receipt, entitling the worker to a DRC on the lost benefits when he resumes full-benefit receipt. One of the most important features of SS is that it also provides benefits to dependents of covered workers. Spouses receive a benefit equal to 50 percent of the worker’s PIA, which is available once the worker has claimed benefits and the spouse has reached age sixty-two; however, the spouse only receives the larger of this and their own entitlement as a worker.4 Dependent children are also each eligible for 50 percent of the PIA, but the total family benefit cannot exceed a maximum that is roughly 175 percent of the PIA. Surviving spouses receive 100 percent of the PIA, beginning at age sixty, although there is an actuarial reduction for claiming benefits before age sixty-five or if the worker had an actuarial reduction. Finally, benefit payments are adjusted for increases in the consumer price index (CPI) after the worker has reached age sixty-two; thus, SS provides a real annuity. 12.2.2 Labor Market Participation and Program Participation Table 12.1 documents the transition of men and women out of the labor force and into receipt of SS and other benefits. At ages fifty to fifty-four, 81 percent of men are working full time, 4 percent are working part time, and 15 percent are not working. The fraction of men in this age group receiving some type of benefit is about equal to the fraction not working and is divided roughly equally among those receiving DI benefits (6 percent), Unemployment Insurance (UI) benefits (5 percent), and private pensions (5 percent).5 At ages fifty-five to fifty-nine, an additional 11 percent of men 2. The reduction factor will be only five-twelfths of one percent for months beyond thirtysix months before the NRA, which is relevant for workers with an NRA past age sixty-five. 3. Until 2000, workers aged sixty-five to sixty-nine were subject to an earnings test with a higher earnings floor and lower tax rate than that for workers aged sixty-two to sixty-five. However, the Senior Citizens’ Freedom to Work Act of 2000 eliminated the earnings test for persons aged sixty-five to sixty-nine as of January 2000. 4. Spousal benefits can begin earlier if there is a dependent child in the household; spousal benefits are also subject to actuarial reduction if receipt commences before the spouse’s NRA. 5. In addition, 2 percent of men are receiving supplemental security income (SSI), a meanstested benefit for people who are poor and either disabled or aged sixty-five or older.

696

Courtney Coile and Jonathan Gruber

Table 12.1

Labor Market Participation and Program Participation, 1997 (%) Age 50–54

55–59

60–64

65–69

70–74

Labor Market Participation Men Working full time Working part time Not working Women Working full time Working part time Not working

81.2 3.5 15.2

69.8 5.2 25.0

44.6 8.3 47.1

14.8 10.5 74.7

7.1 8.3 84.6

57.4 13.6 29.0

48.2 12.8 39.0

27.5 12.1 60.4

8.6 10.0 81.5

2.6 6.5 90.9

Program Participation Men who received SS retired worker benefits SS dependent spouse benefit SS survivor’s benefits DI benefits SSI benefits UI benefits Private pension benefits Women who received SS retired worker benefits SS dependent spouse benefit SS survivor’s benefits DI benefits SSI benefits UI benefits Private pension benefits

0.0 0.0 0.0 5.7 2.1 5.1 4.8

0.0 0.0 0.0 9.0 2.7 3.8 13.2

26.9 0.1 0.4 12.9 3.1 3.4 27.4

85.9 0.1 0.1 0.0 4.0 1.0 45.2

95.3 0.2 0.1 0.0 4.1 0.2 48.5

0.0 0.3 0.6 4.0 3.4 2.7 2.2

0.0 0.3 1.4 5.9 4.3 2.5 5.3

21.3 7.0 10.6 7.4 5.2 1.6 12.4

57.5 15.9 14.2 0.0 6.9 0.7 20.9

61.3 15.8 18.3 0.0 7.6 0.3 24.6

Sources: Population figures are from table 14 of the 1998 Statistical Abstract of the United States (U.S. Bureau of the Census 1998) and from the Bureau of Census website (available at http://www.census.gov/population/estimates/nation/intfile2-1.txt). The SS, DI, and SSI benefit figures are from tables 5.A1 and 7.E3 of the 1998 Annual Statistical Supplement to the Social Security Bulletin (Social Security Administration 1998). Labor force participation, UI benefits, and private-pension benefits are authors’ calculations from the March 1998 Current Population Survey.

leave the labor force, and there is a concurrent rise in the fraction receiving DI (to 9 percent) and private pensions (to 13 percent). At ages sixty to sixtyfour, there is a large movement out of the full-time labor market (down to 45 percent) and on to SS (27 percent) and private pensions (27 percent), and to a lesser extent, DI (13 percent). By ages sixty-five to sixty-nine, the vast majority of workers have moved out of the full-time labor market (down to 15 percent) and on to SS (86 percent) and often private pensions (45 percent).6 In short, the U.S. system features one main pathway to retirement, 6. Receipt of DI benefits goes to zero at age sixty-five, as DI recipients are automatically transferred to SS benefit receipt at age sixty-five.

The Effect of Social Security on Retirement in the United States

697

from full-time work to receipt of SS (and frequently private pension benefits), a move that typically occurs between ages sixty-two and sixty-five. This is in contrast to many other developed countries, where many people exit the labor force at earlier ages and receive UI or DI benefits prior to becoming eligible for retirement benefits. As use of these other paths to retirement is minimal in the United States, they will not factor into our analysis. For women, the patterns are similar but with a few notable differences. First, a lower fraction of women are initially working full time at ages fifty to fifty-four (57 percent); this reflects both a higher fraction of women out of the labor force entirely (29 percent) and a higher fraction working part time (14 percent). Second, while many women receive SS benefits based on their own work record (58 percent of women at ages sixty-five to sixtynine), a significant fraction receive benefits only as a result of being a dependent spouse (16 percent) or widow (14 percent). Third, fewer women receive private pension benefits (21 percent at ages sixty-five to sixty-nine, versus 45 percent of men). 12.2.3 Previous Related Literature A number of studies have used aggregate information on the labor force behavior of workers at different ages to infer the role played by SS. Hurd (1990) and Ruhm (1995) emphasize the spike in the age pattern of retirement at age sixty-two; as Hurd states, “there are no other institutional or economic reasons for the peak” (597). Using quarterly data, Blau (1994) finds that almost one-quarter of the men in the labor force at their sixtyfifth birthday retire in the next three months; this hazard rate is over 2.5 times as large as the rate in surrounding quarters. Lumsdaine and Wise (1994) examine this excess retirement at sixty-five and conclude that it cannot be explained by the change in the actuarial adjustment at this age, by the incentives in private pension plans, or by the availability of retirement health insurance through Medicare. However, SS may still play an important role by setting up the focal point of a normal retirement age. The main body of the retirement incentives literature attempts to specifically model the role that potential SS benefits play in determining retirement. The earliest work in this area considered reduced-form models of the retirement decision as a function of SS wealth (SSW) and pension levels. Much of this literature is reviewed in Mitchell and Fields (1982); more recent cites include Diamond and Hausman (1984) and Blau (1994). While these articles differ in the estimation strategies, with the more recent work using richer models, such as nonlinear 2SLS or hazard modeling, their results generally suggest that SS’s role is significant but small relative to the time trends in retirement behavior. A key limitation of these studies is that they consider SS effects at a point in time, but not any impacts on the retirement decision arising from the time pattern of SSW accruals. This was remedied in three different ways by subsequent literatures. The first was to use structural models of retirement

698

Courtney Coile and Jonathan Gruber

decisions by workers facing a lifetime budget constraint; for example, see Burtless (1986), Burtless and Moffitt (1984), Gustman and Steinmeier (1985, 1986), and Rust and Phelan (1997). The second was to estimate reduced-form models, but incorporate the accrual of SSW with a year of additional work; for example, see Fields and Mitchell (1984), Hausman and Wise (1985), and Sueyoshi (1989). Both of these types of studies continued to find an important, but modest, role for SS, and some indicated a larger role for private pensions. The final type of literature is the option value work of Stock and Wise noted previously.7 A final article that deserves particular mention is that of Krueger and Pischke (1992). They note that the key regressor in many of these articles, SS benefits, is a nonlinear function of past earnings and that retirement propensities are clearly correlated with past earnings. They solve this problem by using a unique natural experiment provided by the end of doubleindexing for the “notch generation” that retired in the late 1970s and early 1980s. For this cohort, SS benefits were greatly reduced relative to what they would have expected, yet the dramatic fall in labor force participation continued unabated in this era. This raises important questions about the identification of the cross-sectional literature. However, Krueger and Pischke still find significant and sizeable impacts of SS accruals on retirement, which highlights the value of the dynamic approach and suggest that the additional nonlinearities that govern the evolution of SSW (as opposed to its level) may be a fruitful source of identification for retirement models. Each of these dynamic literatures has important limitations. The first suffers from the perhaps untenable assumptions that are required to identify these very complicated structural models.8 The second suffers from the limited way in which dynamic retirement incentives are specified. Some of these problems are remedied by the option value literature, but this literature has not separated the impact of SS incentives, as distinct from pension incentives, on retirement.9 If all dollars of retirement wealth are not weighed equally by potential retirees, either because individuals understand their firm’s pension incentives better than SS incentives or because the real annuity provided by SS is valued differently than the nominal annuity provided by most defined-benefit pensions, then it is important to separately estimate SS and private pension impacts.10 In addition, all of these studies suffer from important data deficiencies, as they use data from the 1970s (when the structure of the SS system was fairly different), data from only a handful of firms, or data without com7. See also Samwick (1998), who uses the option value model in a reduced-form context. 8. For a criticism of this type in the context of this type of estimation of general labor supply responses, see MaCurdy (1981). 9. Stock and Wise did not attempt this decomposition, and Samwick’s (1998) attempt to do so with a reduced-form version of the option value model was unsuccessful, perhaps due to the measurement error in SS incentives arising from a lack of earnings-history data. 10. The latter is suggested by Diamond and Hausman (1984), who find much smaller effects of pensions on retirement than those of SS.

The Effect of Social Security on Retirement in the United States

699

plete information on SS incentives. Finally, all of the literature suffer from a lack of careful attention to the sources of identification of the retirement incentive effects that they estimate. As highlighted by Krueger and Pischke (1992), SS benefits are a nonlinear function of earnings, making it difficult to disentangle their impact from the separate impact of earnings on the work decision. This problem is not necessarily surmounted and is potentially compounded, by the option value literature, as this measure is largely determined by wage differences across individuals and only secondarily influenced by the structure of retirement incentives. In principle, this problem can be surmounted by structural estimation of the option value model, which will identify the difference in the impacts of wages and retirement income on retirement decisions through the value of leisure parameter. But, in practice, this is only true if the particular utility structure is correct; for example, if the additional leisure of utility enters the model only as a multiplier on postretirement income and not in some other way. To address these concerns, Coile and Gruber (2000, 2001) introduce a new measure, peak value, which incorporates the insights of the option value measure but focuses solely on variation in SS incentives. This is comparable to the accrual, but looks forward more than just one year: It calculates the difference between SSW at its maximum expected value and SSW at today’s value in order to measure the incentive to continued work. The peak value appropriately considers the trade-off between retiring today and working to a period with much higher SSW, thereby capturing the option value of continued work even before SS entitlement ages are reached. Since wage is not included specifically into the peak value calculation, there is much more variation from the structure of the SS entitlement.11 In the empirical analysis below, both peak value and option value are used in a reduced-form context. 12.3 Data and Empirical Strategy 12.3.1 Data Our data for this analysis comes from the HRS.12 The HRS is a survey of 12,652 individuals aged fifty-one to sixty-one in 1992 with reinterviews every two years; the first four waves of the survey (1992, 1994, 1996, and 1998) are used in this analysis.13 Spouses of respondents are also interviewed, so the total age range covered by the survey is much wider. 11. In our sample, an earnings quartic and age dummies explain only 33 percent of the variation in peak value versus 74 percent of the variation in option value. 12. The HRS is conducted by the Survey Research Center at the University of Michigan in Ann Arbor, Michigan. The data is available at http://www.umich.edu/~hrswww/. Most of the data is publicly available, although the SS and firm-level private-pension data is restricted to approved users. 13. The 1998 wave 4 data are preliminary.

700

Courtney Coile and Jonathan Gruber

A key feature of the HRS is that it includes SS earnings histories back to 1951 for most respondents. This provides two advantages for our empirical work. First, it allows us to appropriately calculate benefit entitlements, which depend on the entire history of earnings.14 Second, it allows us to construct a large sample of person-year15 observations by using the earnings histories to compute SS retirement incentives and labor force participation at each age. We use all person-year observations age fifty-five to sixty-nine for our analysis, subject to the exclusions detailed below. Our sample is selected conditional on working, so that we examine the incentives for retirement conditional on being in the labor force. Work is defined in one of two ways. For those person-years before 1992, when we are using earnings histories, we define work as positive earnings in two consecutive years; if earnings are positive this year, but zero the next (and if the year of zero earnings occurs at or after age fifty-five), we consider the person to have retired this year.16 For person-years from 1992 onwards, when we have the actual survey responses, we cannot use this earnings-based definition since we only have earnings at two year intervals. For this era, we use information on self-reported retirement status and dates of retirement to construct retirement measures.17 We only consider individuals before their first retirement; if a person who is categorized as retired reenters the labor force, the later observations are not used. Our sample selection criteria are as follows for men. There are 5,886 men who appear in waves 1, 2, or 3 of the HRS.18 We first exclude 1,533 men who are missing SS earnings history data. These data, fortunately, appear to be missing essentially randomly, as noted by Haider and Solon (1999). We then exclude 99 observations where the respondent or spouse was born prior to 1922, as these individuals are subject to different SS benefit rules. We also exclude 240 observations where the wife is missing SS earnings history data (necessary due to the family structure of benefits) and 67 observations with an ambiguous work history.19 Next, we exclude 730 men 14. Only earnings since 1950 are required to compute SS benefits for our sample’s age range; the benefit rules specify that a shorter averaging period is used for persons born prior to 1929. 15. “Person-years” means that an observation represents a given person in a given year, so that there are multiple observations for each person. 16. One potential problem with using earnings histories to define retirement is that an individual may move from the private sector to the state and local government sector, in which case they would be classified as retired when, in fact, they are still working. We find that results are similar when individuals who list their industry as public administration are dropped. 17. If an individual simultaneously reports their labor supply status as working and retired, we treat them as working. 18. Observations that enter the sample at wave 4 will not be used in the analysis, as multiple observations on the same person are required to establish work and retirement status. 19. Observations with missing spouse data are those for which we know that the spouse worked at least half as many years as their partner, but for which we don’t have their SS earnings records. Observations with an ambiguous work history are those who have zero covered

The Effect of Social Security on Retirement in the United States

701

who retired prior to age fifty-five. The remaining 3,217 men are converted into 18,733 person-year observations by creating one observation for each year from 1980 through 1997 in which the individual is between the ages of fifty-five and sixty-nine and working at the beginning of the year. Finally, we exclude 988 person-year observations that represent labor force reentry after a previous retirement. The final sample size is 17,745 male observations. A similar process generates a sample size of 11,419 female observations. The means of our key variables are shown in table 12.2 for men and women, respectively. In any given year, a similar percentage of the male and female sample retire, 5.7 percent for men and 5.6 percent for women. The average age of our sample is 58.5 for men and 58 for women. Some 91 percent of our male sample is married, and the typical man is 4.3 years older than his wife, while only 64 percent of our female sample is married, and the typical woman is 1.3 years younger than her husband.20 Roughly 80 percent of both samples are white. Among the male sample, 24 percent are high school dropouts, 36 percent have only a high school degree, 14 percent have some college, and 25 percent are college graduates; for women, the figures are 19 percent, 42 percent, 17 percent, and 22 percent, respectively. The average projected earnings for the next year of work are $36,152 for men and $20,984 for women (in 1998 dollars), and the average monthly earnings over the working life are $2,470 for men and $1,102 for women. The typical spouse’s earnings (averaging over single people, people with nonworking spouses, and people with working spouses) are $10,813 for an additional year of work and $612 per month on average over the spouse’s lifetime for the male sample, and $16,489 and $1,475 for the female sample. The typical man in our sample has forty years of labor market experience and seventeen years of tenure on their current job, and 5.4 percent of our sample is missing tenure information (indicating a short-term job); equivalent figures for women are thirty-nine and thirteen years and 6.6 percent. 12.3.2 Incentive Variable Calculation—Accrual Our goal is to measure the retirement incentives inherent in the SS system. The first step in this calculation uses a simulation model we have developed to compute the PIA for any individual at all possible future retirement dates. This process is based on a careful modeling of SS benefits rules and has been cross-checked against the SS Administration’s ANYPIA earnings in the administrative data from age fifty-four through 1991, have positive selfreported earnings in 1991, and report that they have changed jobs between age fifty-four and 1991; they are excluded because it is impossible to know whether they have retired prior to 1991 and reentered the labor force. 20. The fact that a lower fraction of the female sample is married is a result of the fact that when the sample selection criteria is applied, women who are still working at age fifty-five are less likely to be married than men who are still working at age fifty-five.

702

Courtney Coile and Jonathan Gruber

Table 12.2

Summary Statistics Variable Male Sample Retired Age Education; less than high school Education; high school Education; some college Married Age different with spouse Race; black Race; other nonwhite Earnings AIME Spouse’s earnings Spouse’s AIME Experience Job tenure Job tenure missing Have pension No. of observations No. of individuals

Mean

SD

0.057 58.5 0.241 0.363 0.136 0.914 4.3 0.101 0.081 36,152 2,470 10,813 612 40 17 0.054 0.348

0.232 3.0 0.428 0.481 0.343 0.281 4.9 0.301 0.273 18,926 945 14,117 654 4 12 0.227 0.476 17,745 3,217

Female Sample Retired 0.056 0.230 Age 58.0 2.6 Education; less than high school 0.193 0.395 Education; high school 0.424 0.494 Education; some college 0.168 0.374 Married 0.640 0.480 Age different with spouse –1.3 3.6 Race; black 0.163 0.369 Race; other nonwhite 0.071 0.257 Earnings 20,984 14,887 AIME 1,102 735 Spouse’s earnings 16,489 21,168 Spouse’s AIME 1,475 1,413 Experience 39 4 Job tenure 13 10 Job tenure missing 0.066 0.248 Have pension 0.325 0.468 No. of observations 11,419 No. of individuals 2,526 Note: SD = standard deviation.

model for accuracy. The appropriate actuarial adjustment is applied to the PIA to obtain the monthly benefit entitlement. The next step is to compute the expected net present discounted value of SSW associated with each retirement date. Our methodology for doing so is described in Coile and Gruber (2000, 2001). For single workers, this is

The Effect of Social Security on Retirement in the United States

703

simply a sum of future benefits discounted by time preference rates and survival probabilities. For married workers, it is more complicated since we must include dependent spouse and survivor benefits and account for the joint likelihood of survival of the worker and dependent. We use a real discount rate of 3 percent and survival probabilities from the age- and sexspecific U.S. life tables from the U.S. Department of Health and Human Services (National Center for Health Statistics 1990, sec. 6). We next compute the other SS incentive variables. We first calculate the accrual, the change in SSW resulting from an additional year of work. There are two routes through which an additional year of work affects SSW. First, the additional year of earnings will be used in the recomputation of SS benefits. For workers who have not yet worked thirty-five years, this replaces a zero in the benefits computation; for workers who have worked thirty-five years, it may replace a previous low-earnings year. So the recomputation raises SSW (or leaves it unchanged). Second, at ages sixty-two and beyond, the additional year of work implies a delay in claiming; this raises future benefits through the actuarial adjustment, but reduces the number of years of benefit receipt, so the net effect is uncertain. Both of these factors will affect workers differently, depending on their potential earnings next year, earnings history, mortality prospects (which will vary over time and cohort in our data), family structure, and spouse’s earnings. Thus, the net effect of an additional year of work on SSW is theoretically ambiguous and will vary significantly across people. Computing the accrual and other incentive variables requires projecting the worker’s potential earnings next year (or in all future years). We considered a number of different projection methodologies and found that the best predictive performance was from a model which simply grew real earnings from the last observation by 1 percent per year, so we use this assumption in our simulations.21 Our SS incentive variables incorporate dependent spouse and survivor benefits, since these are important components of SSW. For a worker with a nonworking spouse, these benefits are based solely on the worker’s earnings record. For a worker whose spouse is entitled to benefits on their own, the spouse’s benefits are based (partially or fully) on the spouse’s record, but are also included in SSW. Since a full modeling of the joint retirement decision is beyond the scope of this paper, we simply assume that spouses who are working will retire at age sixty-two; this seems reasonable, given that the median retirement age is sixty-two for women in the sample and sixty-three for men. For more evidence on joint retirement decisions, see Coile (1999). 21. Projected earnings always represent potential earnings for one full year. For example, in the case where an individual earns $2X in year t and $X in year t
1 because they retire halfway through the year, the year t
1 observation has projected real earnings of $2X  (1.01) and there is no t
2 observation (since the individual retires in year t
1).

704

Courtney Coile and Jonathan Gruber

For the simulations below, we assume that workers claim SS benefits at retirement or when they become eligible (age sixty-two), if they have retired before then. In fact, this is not necessarily true; retirement and claiming are two distinct events, and for certain values of mortality prospects and discount rates, it is optimal to delay claiming until some time after retirement, due to the actuarial adjustment of benefits. Coile et al. (2002) investigate this issue in some detail, and they find that a relatively small share of those retiring before age sixty-two delay claiming until age sixty-three (about 10 percent) and that virtually none of those retiring at age sixty-two or later delay claiming. Given these findings, we choose not to jointly model delayed claiming here. Our incentive measures will therefore slightly overstate any subsidies to continued work, since part of this subsidy will come from delayed claiming that could be obtained without delaying retirement. We do not incorporate private pension incentives into our analysis. Coile and Gruber (2000) estimate retirement models that include both SS and pension incentives, and they find that the results differ significantly from those for SS alone. This suggests that changes in private pension provisions may have different impacts on retirement than changes in public pensions, so that one should not extrapolate the effect of public pension reform from private pension responses. Thus, since our primary goal is to discuss the impacts of public pensions on retirement, we exclude private pensions here. Table 12.3 shows the medians of the retirement incentive variables for our male sample by age. The median present discounted value (PDV) rises from $179,316 at age fifty-five to a peak of $205,584 at age sixty-five, then falls to $194,555 at age sixty-nine.22 The age pattern of accruals demonstrates how the various effects of working an additional year enter in at different ages. From ages fifty-five to sixty-one, accruals are positive, but small, reflecting the value of the dropout year provision. From ages sixtytwo to sixty-four, accruals are two to three times larger; this is the delayedclaiming effect, whereby an additional year of work increases the actuarial adjustment and raises future benefits.23 After age sixty-five, accruals become negative and rise rapidly, as the delayed retirement credit is insufficient to compensate for the value of lost benefits. Most importantly for our analysis, there is enormous heterogeneity in accruals, as is also shown in table 12.3. The standard deviation in accru22. The SSW median displayed in table 12.2 is the median SSW at age fifty-five increased or decreased each year by the median accrual. The median SSW at each age in the sample rises much more rapidly with age due to a sample selection effect (those working at later ages have higher SSW). 23. This large subsidy to work at age sixty-two is at odds with the common wisdom that the actuarial reduction at age sixty-two is approximately fair. This point is developed much further in Coile and Gruber (2000).

The Effect of Social Security on Retirement in the United States Table 12.3

705

The Distribution of the One-Year Accrual, Male Sample Accrual

Age 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

Number of Obs.

Median SSW

Median

10th Percentile

90th Percentile

SD

Median Tax Rate

Median Tax Rate 2

2,809 2,747 2,444 2,143 1,823 1,546 1,255 1,021 716 483 344 191 110 71 42

179,316 181,870 184,239 186,367 188,315 190,108 191,770 193,440 197,789 202,300 205,584 204,461 202,160 198,858 194,555

2,554 2,369 2,128 1,948 1,793 1,662 1,670 4,349 4,511 3,284 (1,123) (2,301) (3,302) (4,303) (4,758)

322 243 167 66 12 0 0 978 430 0 (4,741) (5,591) (6,773) (7,530) (7,825)

5,492 5,032 4,555 4,233 4,136 4,190 4,170 7,032 7,394 6,152 1,785 600 0 0 0

2,334 3,859 3,382 2,497 3,191 2,007 5,113 2,691 3,532 2,631 4,470 2,996 8,027 2,693 2,735

–0.072 –0.066 –0.061 –0.054 –0.048 –0.042 –0.043 –0.118 –0.122 –0.093 0.027 0.059 0.114 0.134 0.119

–0.022 0.046 0.060 0.069 0.072 0.071 0.064 –0.028 –0.005 0.031 0.118 0.225 0.269 0.439 0.455

Notes: Tax rate 2 is from Diamond and Gruber (1999). Definitions of other variables are provided in the text. Median SSW is the age-55 median SSW incremented by the median accrual. All figures are 1998 dollars. One source of difference between the median tax rate and tax rate 2 is that tax rate 2 includes SS contributions. Adding the 12.4% payroll tax to the median tax rate results in figures very similar to tax rate 2. SD = standard deviation. Numbers in parentheses are negative.

als is substantial, averaging roughly $3,000 per year. At sixty-two, for example, while there is a sizeable positive median accrual, the tenth percentile person has an accrual of only $978 and the ninetieth percentile person has an accrual of $7,032; the standard deviation at that age is $2,691. It is this sizeable variation that identifies our models. Table 12.4 shows the median retirement incentives for women by age. For women, median accruals are much smaller at all ages than they are for men, and there are no large accruals at ages sixty-two to sixty-four. In large part, this is because women have typically been less attached to the labor force over their working lives and are often dually entitled to benefits as both retired workers and as a dependent (or divorced) spouses or widows. If a woman’s retired-worker benefits are less than what she is entitled to as a dependent spouse, then additional work will typically not result in a higher benefit. It is also worth noting that in the HRS, earnings histories are not available for divorced or deceased spouses; since more women than men are likely to be receiving benefits based on the record of a divorced or deceased spouse, estimates of women’s incentives are likely to be subject to greater measurement error. Table 12.4 also shows that there is significant heterogeneity in women’s accruals.

706

Courtney Coile and Jonathan Gruber

Table 12.4

The Distribution of the One-Year Accrual, Female Sample Accrual

Age

Number of Obs.

Median SSW

Median

10th Percentile

90th Percentile

SD

Median Tax Rate

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

2,124 1,954 1,711 1,455 1,201 944 726 547 352 217 119 48 12 5 4

163,139 164,239 165,415 166,601 167,801 169,004 170,286 171,603 172,963 173,882 173,986 172,452 170,803 168,111 164,045

1,100 1,176 1,186 1,200 1,203 1,282 1,317 1,360 919 104 (1,534) (1,649) (2,692) (4,066) (4,178)

0 0 0 0 0 0 0 (586) (1,264) (1,927) (4,711) (5,031) (6,940) (7,998) (5,280)

3,200 3,185 3,168 3,138 3,122 3,119 3,365 5,126 4,647 4,126 1,319 1,251 332 (443) (745)

2,506 2,503 2,319 2,975 2,576 2,355 4,334 3,732 3,094 2,597 7,173 5,572 5,831 2,780 2,081

–0.055 –0.057 –0.059 –0.061 –0.064 –0.065 –0.069 –0.639 –0.045 –0.012 0.052 0.062 0.087 0.186 0.151

Note: See table 12.3.

12.3.3 Incentive Variable Calculation—Forward-Looking Measures The recent work on retirement has highlighted an important weakness of the accrual measure. For any given year from age fifty-five to sixty-one, as we show in table 12.3, a typical worker sees a small positive accrual from additional work through the recomputation of the AIME. But, by working, that worker is also buying an option on the more-than-fair actuarial adjustment that exists from age sixty-two to sixty-four. Incorporating this option dramatically changes the nature of SS incentives, particularly at ages before age sixty-two, as documented in Coile and Gruber (2001). For a sizeable minority of workers, accrual patterns are nonmonotonic, so that forward-looking measures can deliver very different incentives than oneyear accruals. As noted above, Stock and Wise (1990a) propose to account for these option values by contrasting the utility of retiring today versus at the optimal point in the future. Their option value model is based on the individual’s indirect utility function over work and leisure R1

(1)

Vt (R) 

∑p

st

st

T

d st( ys ) g


∑p

st

d st [k  Bs (R)]g,

sR

where R is the retirement date, d is the discount rate, p is the probability of being alive at some future date conditional on being alive today, y is income while working, B is retirement benefits, g is a parameter of risk aversion,

The Effect of Social Security on Retirement in the United States

707

k is a parameter to account for disutility of labor (k  1), and T is maximum life length. In this model, additional work has three effects. First, it raises total wage earnings, increasing utility. Second, it reduces the number of years over which benefits are received, lowering utility. Third, it may raise or lower the benefit amount, depending on the shape of the benefit function, B(R). The latter two effects are weighted more heavily because of the disutility of labor, which acts as a devaluation of wage income relative to retirement income. The optimal date of retirement is the date at which the utility gained from the increase in earnings resulting from additional work is outweighed by the utility lost from the decrease in retirement income. The option value is the difference between the indirect utility from retirement at the optimal date, R∗, and the indirect utility from retirement today. As a structural estimation of the option value model is beyond the scope of this paper, we instead calculate the option value using reasonable utility parameters and include it as a regressor in a retirement model.24 As mentioned above, one possible weakness of the option value model is that much of the variation in this measure arises from differences in wages, which may not be a legitimate source of identification of retirement effects. We take two approaches to addressing this potential shortcoming. First, we include rich controls for earnings in the retirement model to capture the heterogeneity that may bias these estimates. However, since wages enter highly nonlinearly in the option value and the form of heterogeneity is unknown, even rich wage controls may not fully capture the underlying correspondence between option value and tastes for work. Therefore, we also estimate retirement models utilizing the peak value measure. As described above, peak value is the difference between SSW at its maximum expected value and SSW at today’s value.25 In this way, the peak value incorporates the insights of the option value measure and appropriately considers the trade-off between retiring today and working to a period with much higher SSW, but focuses solely on variation in SS incentives. Table 12.5 shows the age pattern and heterogeneity for peak and option value for the male sample. The important differences between peak value and accrual, particularly at younger ages, are immediately apparent; peak values are quite large, at ages fifty-five to sixty-one, a range where accruals 24. We follow Stock and Wise in assuming values of 1.5 and 0.75 for k and g, respectively, but we found that the fit of our model was much better with a more reasonable assumption for d of 0.97, relative to the very high discount rate of 0.75 obtained from their model. We also tested the robustness of our model to the choice of k and g and the results are not sensitive to this choice. 25. If the individual is at an age that is beyond the SSW optimum, then the peak value is the difference between retirement this year and next year, which is exactly the accrual rate.

708

Courtney Coile and Jonathan Gruber

Table 12.5

The Distribution of the Peak Value and Option Value, Male Sample Peak Value

Option Value

Age

Median

10th Percentile

90th Percentile

SD

Median

10th Percentile

90th Percentile

SD

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

23,579 21,266 18,548 16,804 15,456 14,083 12,925 11,886 8,102 3,508 (1,042) (2,250) (3,302) (4,303) (4,758)

4,785 4,276 3,929 3,630 3,105 2,479 2,059 1,769 762 0 (4,692) (5,591) (6,773) (7,530) (7,825)

42,312 39,893 37,021 34,450 31,339 28,931 24,597 21,665 15,267 8,805 4,908 1,520 0 0 0

16,253 15,790 14,386 13,865 13,398 12,559 12,320 10,363 9,707 7,954 9,099 7,556 9,033 2,871 2,735

23,755 22,030 20,207 18,390 16,894 15,225 13,623 11,877 10,143 8,057 6,187 4,342 2,690 1,962 893

4,956 4,728 4,134 3,258 2,879 2,041 1,775 1,442 1,257 452 0 0 0 0 0

37,331 35,302 32,641 30,507 28,240 25,728 23,146 20,767 18,164 15,343 12,606 9,850 7,662 5,151 2,616

11,461 10,780 10,119 9,512 8,861 8,217 7,477 6,649 5,913 5,113 4,396 3,612 3,048 1,937 963

Notes: Peak value is in 1998 dollars; option value is in utility units. SD = standard deviations. Numbers in parentheses are negative.

are small.26 The peak value declines sharply with age, as people move closer to or reach their optimal retirement date; the declines occur at a fairly constant rate up until about age sixty-two, then become very large. The peak value is positive for the median person until they reach age sixty-five, and then it becomes negative. As with the accrual, there is an enormous amount of heterogeneity in all of these measures that can be used to identify our models. Part of this variance arises from heterogeneity in the peak year. For 38 percent of our sample, age sixty-five is the peak; for 11 percent, it is age seventy, and there are substantial masses at ages sixty-six, sixtyseven, sixty-eight, and sixty-nine. Partly, this reflects the evolving generosity of the DRC over time; the peak occurs after age sixty-five for 28 percent of the workers in the oldest cohorts in our sample versus 73 percent of the workers in the youngest cohorts. Although option value is measured in utility units and cannot be directly compared to peak value, option value follows the same declining pattern 26. Note that we take the median of each variable, so that all the numbers in any given row do not necessarily represent the incentives facing a single person. This explains a seeming inconsistency between tables 12.3 and 12.5, which is that the accruals from age fifty-five through age sixty-four add up to more than the peak value at age fifty-five, despite the fact that age sixty-five is often the peak for SSW. As we show in Coile and Gruber (2001), this is a fallacy of composition, and for any given individual the peak value is just the sum of accruals to the peak SSW age.

The Effect of Social Security on Retirement in the United States Table 12.6

709

The Distribution of the Peak Value and Option Value, Female Sample Peak Value

Option Value

Age

Median

10th Percentile

90th Percentile

SD

Median

10th Percentile

90th Percentile

SD

55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

9,359 8,981 8,097 6,799 5,724 4,889 3,685 2,121 1,424 280 (1,534) (1,601) (2,692) (4,066) (4,178)

166 134 120 90 78 70 39 (553) (1,264) (1,927) (4,711) (5,031) (6,940) (7,998) (5,280)

33,369 32,237 30,074 28,006 24,602 21,578 19,680 17,115 12,171 7,806 3,880 3,651 332 (443) (745)

14,639 14,144 13,791 13,325 12,323 11,991 12,077 9,432 8,539 7,574 9,901 7,387 6,362 2,780 2,081

12,506 12,195 11,309 10,235 9,334 8,705 7,584 6,447 5,514 4,382 3,496 2,504 2,281 1,417 274

1,174 953 794 574 369 338 131 130 88 0 0 0 0 130 0

27,876 26,443 24,471 22,223 20,350 18,780 16,594 14,218 12,292 10,169 8,995 7,761 5,800 4,796 706

10,113 9,599 8,952 8,311 7,649 7,046 6,352 5,460 4,778 4,040 3,535 2,761 2,194 1,804 340

Note: See table 12.5.

as peak value. The median option value falls monotonically with age, but remains positive even beyond age sixty-five, as additional earnings offset losses in SSW. There is also substantial heterogeneity in the option value measure. Table 12.6 shows the distribution of the peak value and option value measures by age for the female sample. The age trends are largely similar to those for men, although the dollar amounts are smaller since women typically benefit less from additional work for reasons described previously and also have lower earnings, which lowers the option value. 12.4 Empirical Framework and Results 12.4.1 Regression Framework In a standard retirement model, SS will play two roles in the decision regarding whether to retire this year or to continue working. The first is through wealth effects: higher SSW will induce individuals to consume more of all goods, including leisure, and to retire earlier. The second is through accrual effects: the individual’s decision to continue to work is a function of the increase in retirement consumption resulting from additional work. Following this discussion, we use the incentive variables described above to run regressions of the form

710

(2)

Courtney Coile and Jonathan Gruber

Rit  b0
b1 SSWit
b2 INCENTit
b3 Xit
b4 AGE it
b5 EARNit
b6 AIMEit
b7 MARit
b8 AGEDIFFi
b9 SPEARNit
b10 SPAIMEit
b11Yt
e,

where SSW is the expected PDV of SS benefits that is available to the person if he retires that year (t); INCENT is one of the incentive measures noted above (accrual, option value, and peak value); X is a vector of control variables that may importantly influence the retirement decision, but do not enter directly into the calculation of SSW (education, race, veteran status, born in the United States, region of residence, experience in the labor market and its square,27 tenure at the firm and its square, thirteen major industry dummies, and seventeen major occupation dummies); AGE is either entered linearly or as a set of dummies for each age fifty-five to sixtynine; EARN is a control for potential earnings in the next year; AIME is a control for average monthly lifetime earnings as of period t;28 MAR is a dummy for marital status; AGEDIFF controls for the age difference with the spouse; SPEARN and SPAIME are the spouse’s next year and average lifetime earnings; and Y is a series of year dummies. Since our dependent variable is dichotomous, we estimate the model as a probit. We have also estimated these models as Cox proportional hazard models and the results were very similar; this is not surprising, given that the models all include a full set of age dummies that pick up the same factors captured by the baseline in the hazard model. This model parallels the types of models used in the first round of research on SS and retirement, with one important exception: the earnings controls. Most articles in this literature did not control for earnings, and no articles controlled for both earnings around time of retirement and average lifetime earnings. Yet both of these variables are clearly important determinants of both SS incentives and retirement decisions, so excluding them from the model imparts a potential omitted variables bias. Moreover, there is no reason to suspect that heterogeneity is a purely linear function of earnings. Thus, for each of the earnings controls previously listed, we include squared, cubed, and quartic terms as well. Moreover, it is possible that heterogeneity in retirement is also related to the relationship between current and average lifetime earnings; we therefore also include a full set of interactions between the EARN and AIME quartics in order to reflect this. Finally, it is important to highlight that our work is focused on the im27. Experience is defined as age minus years of education minus six, since the HRS selfreported earnings histories may have gaps and administrative data do not include employment in noncovered sectors. 28. Note that AIME is time varying because additional years of work change average lifetime earnings through the dropout-year provision.

The Effect of Social Security on Retirement in the United States

711

pact of SS on the labor force participation decision. A separate and interesting issue is the impact of SS on the marginal labor supply decision among those participating in the labor force. This is more complicated for those around retirement age, since it involves incorporating the role of the earnings test, which we avoid with our analysis of participation. This, in turn, would involve modeling expectations about the earnings test, since individuals appear not to understand that this is just a benefits delay instead of a benefits cut. This is clearly a fruitful avenue for further research. 12.4.2 Social Security Incentives and Retirement Table 12.7 shows the results of estimating equation (2) for men for the three incentive measures and the two possible sets of age controls. Peak value, accrual, and SSW are expressed in $100,000; option value is expressed in units of 10,000. The magnitude of the coefficients is illustrated by the term in square brackets, which gives the implied percentage-point impact of a $1,000 increase in the accrual/peak value and a $10,000 increase in SSW. In all the models, we estimate a positive impact of SSW levels, as expected; however, the coefficient is significant at the 5 percent level in only two of the six models. The coefficient implies that each $10,000 increase in SSW increases the probability of retirement by about 0.2 percent, or about 3.5 percent of the sample average retirement rate; evaluated at the mean, this corresponds to an elasticity of nonparticipation with respect to benefits of 0.60. The coefficients are about 50 percent greater in the models with linear age than in those with age dummies. The coefficient on the accrual is the wrong sign (positive) and is highly insignificant once age dummies are included in the model. This suggests that there is little impact of one-year-forward incentives on retirement decisions. This could reflect the fact that individuals are not at all forwardlooking in their decisions. Alternatively, given nonlinearities in future accruals, it could represent the fact that individuals are not considering solely the accrual to the next year but the entire future path of incentives. This possibility is addressed in the next two sets of columns, which show the estimates from the peak value and option value models. In both cases, we now estimate significant negative impacts of the forward-looking incentive measures for retirement decisions. We find that each $1,000 in peak value lowers the odds of retirement by 0.05 percent, or about 1 percent of the sample average retirement rate; this corresponds to an elasticity of nonparticipation with respect to benefits of 0.15. For option value, it is not possible to calculate the impact of a simple $1,000 increment since this is a utility-based metric; we will return to comparisons of these two models in the simulation section below. The coefficients on peak value and option value are similar whether age

Table 12.7

Retirement Probits, Male Sample Specification Accrual

Variable SSW $10,000 changea Incentive measure $1,000 changeb Age

Option Value

(1)

(2)

(3)

(4)

(5)

(6)

0.3581 (0.1163) [0.0030] 1.8360 (0.3648) [0.0015] 0.0922 (0.0154)

0.2190 (0.1162) [0.0018] 0.4560 (0.4137) [0.0004]

0.2519 (0.1157) [0.0021] –0.4289 (0.2238) [–0.0004] 0.0877 (0.0155)

0.1718 (0.1170) [0.0014] –0.5697 (0.2367) [–0.0005]

0.1730 (0.1146) [0.0015] –0.2368 (0.0539)

0.1075 (0.1173) [0.0009] –0.2106 (0.0522)

Age56 Age57 Age58 Age59 Age60 Age61 Age62 Age63 Age64 Age65 Age66 Age67 Age68 Pseudo R 2

Peak Value

0.1215

–0.0311 (0.0706) –0.0693 (0.0808) 0.0587 (0.0877) 0.0696 (0.0986) 0.2079 (0.1088) 0.3032 (0.1178) 0.8999 (0.1279) 0.8122 (0.1446) 0.6908 (0.1622) 1.0796 (0.1779) 0.5924 (0.2062) 0.4542 (0.2539) 0.7229 (0.2387) 0.1379

0.1198

0.0691 (0.0162) –0.0332 (0.0707) –0.0759 (0.0810) 0.0479 (0.0878) 0.0579 (0.0989) 0.1932 (0.1092) 0.2847 (0.1179) 0.8897 (0.1268) 0.7852 (0.1439) 0.6424 (0.1628) 0.9881 (0.1819) 0.4928 (0.2097) 0.3467 (0.2578) 0.5974 (0.2429) 0.1386

0.1223

–0.0470 (0.0711) –0.1075 (0.0816) 0.0009 (0.0890) –0.0065 (0.1006) 0.1082 (0.1117) 0.1794 (0.1215) 0.7650 (0.1321) 0.6512 (0.1495) 0.5050 (0.1682) 0.8534 (0.1870) 0.3540 (0.2146) 0.1953 (0.2615) 0.4231 (0.2483) 0.1402

Notes: All regressions include controls for education, race, experience, marital status, industry, occupation, region, year, as well as a quartic in earnings, a quartic in lifetime earnings, and the interactions of these quartics (plus same earnings variables for the spouse). Standard deviations in parentheses. Magnitude of coefficients is in square brackets. a Implied percentage-point impact of a $10,000 increase in SSW. b Implied percentage-point impact of a $1,000 increase in SSW.

The Effect of Social Security on Retirement in the United States

713

is controlled for using a linear variable or age dummies. The goodness of fit of all six models is similar, with a pseudo R 2 of about 12 percent in models without age dummies and 14 percent in models with age dummies. These findings suggest that the forward-looking models of the type advocated by Stock and Wise are very important for explaining retirement behavior. Individuals do appear to recognize the future path of SSW accumulation, and they take this into account in making their retirement decisions. The other variables in the regression have their expected impacts.29 There is a rising pattern of retirement propensities with age, with particularly large effects at ages sixty-two, sixty-three, sixty-five, and sixty-nine. Figure 12.2 displays the empirical retirement hazard for the sample and the age dummies estimated in the three models. The age dummies in the accrual and peak value models are nearly identical to the empirical hazard, indicating that these models explain little of the variation across ages in retirement propensities; on the other hand, the age dummies in the option value model are significantly below the empirical hazard. Being married and having a larger age difference with one’s wife decrease the probability of retirement, although only the former is significant. More experience lowers the odds of retirement, conditional on age, but this relationship is decreasing in absolute value. There is no distinct relationship with tenure, although there is a very significant positive impact of being in the 6 percent of the sample with missing tenure data; this is consistent with lower labor force attachment among those in jobs of short duration. The industry and occupation dummies do not show a particularly strong pattern, with the exception of higher retirement rates in the armed forces and the cleaning- and building-services occupations. There is no significant time pattern to retirement behavior, which is consistent with Quinn (1999) who shows that the strong time series trend towards earlier retirement was arrested beginning in the mid-1980s. There is no strong regional pattern, other than a higher retirement rate in the western region and a lower rate in New England.30 The retirement probits for the female sample are shown in table 12.8. The SSW coefficients are roughly the same size as in the men’s probits and are significant in the accrual and peak value models. As in the men’s probits, the coefficients on accrual are positive and highly insignificant. Surprisingly, the results using the forward-looking incentive variables are mixed: The coefficients on peak value are negative, but small and insignificant, while the coefficients on option value are negative and significant. One possible explanation for the insignificant coefficients is the measurement error in women’s incentive variables due to a lack of earnings histories for 29. Only coefficients on age or age dummies are shown in table 12.7. 30. See Coile and Gruber (2000) for a discussion of results incorporating pensions in the retirement incentive variables and including health status and health insurance as regressors.

Fig. 12.2

A

Retirement hazard and age dummies: A, males; B, females

Fig. 12.2

B

(cont.)

716

Courtney Coile and Jonathan Gruber

Table 12.8

Retirement Probits, Female Sample Specification Accrual

Variable SSW $10,000 changea Incentive measure $1,000 changeb Age

Option Value

(1)

(2)

(3)

(4)

(5)

(6)

0.2619 (0.1132) [0.0020] 0.7727 (0.7291) [0.0006] 0.1219 (0.0219)

0.2265 (0.1130) [0.0017] 0.4773 (0.8169) [0.0004]

0.2430 (0.1132) [0.0019] –0.0618 (0.0253) –[0.00005] 0.1217 (0.0219)

0.2169 (0.1129) [0.0017] –0.0132 (0.2856) [–0.00001]

0.2080 (0.1142) [0.0016] –0.2695 (0.0772)

0.1838 (0.1135) [0.0014] –0.2414 (0.0752)

Age56 Age57 Age58 Age59 Age60 Age61 Age62 Age63 Age64 Age65 Age66 Age67 Age68 Pseudo R 2

Peak Value

0.1418

0.0223 (0.0825) 0.0474 (0.0960) 0.1254 (0.1104) 0.1186 (0.1295) 0.4364 (0.1415) 0.4587 (0.1585) 1.0015 (0.1702) 0.9872 (0.1939) 0.8106 (0.2229) 1.2619 (0.2464) 0.8181 (0.3377) 0.2807 (0.5925) 0.8993 (0.5079) 0.1530

0.1416

0.1040 (0.0223) 0.0217 (0.0824) 0.0480 (0.0959) 0.1268 (0.1103) 0.1212 (0.1294) 0.4389 (0.1414) 0.4632 (0.1584) 1.0075 (0.1699) 0.9918 (0.1937) 0.8119 (0.2223) 1.2533 (0.2456) 0.8086 (0.3371) 0.2714 (0.5941) 0.8815 (0.5084) 0.1530

0.1441

0.0101 (0.0826) 0.0312 (0.0961) 0.0964 (0.1104) 0.0784 (0.1298) 0.3811 (0.1422) 0.3928 (0.1599) 0.9102 (0.1720) 0.8721 (0.1966) 0.6718 (0.2262) 1.0901 (0.2497) 0.6318 (0.3406) 0.0774 (0.6078) 0.6609 (0.5215) 0.1549

Note: See table 12.7.

divorced and deceased spouses; however, this would not explain why option value is significant.31 The linear age variable and age dummies are sim31. Coile (2003) estimates similar models and finds that women respond to both peak value and option value measures. Her sample differs in two ways from the female sample here: First, she looks only at married women (who may have less measurement error in their incentive variables than unmarried women) and second, she conditions on working at age fifty or later (versus fifty-five here).

The Effect of Social Security on Retirement in the United States

717

ilar to those in the men’s model, and again the age dummies from the option value model are below the empirical hazard (figure 12.2, panel B). The pseudo R 2 is about 15 percent in all six models. In summary, SSW has a positive and marginally significant effect on retirement behavior for both men and women. The one-year accrual has the wrong sign and an insignificant effect, while the forward-looking incentive measures, peak value and option value, have a significant negative effect (although peak value is not significant for women). However, the implications of the estimates that we have presented thus far are difficult to interpret in a vacuum; are $1,000 changes in peak value considered large or small? To provide some more context for the magnitudes of our results, we conduct simulations of changes to the SS system in the following section. 12.5 Policy Simulations In this section, we consider two potential major reforms to the SS system. The first policy change examined is to raise both the early retirement age (ERA) and the normal retirement age (NRA) by three years, to sixtyfive and sixty-eight, respectively. The second policy change is to move from the current SS system to a common system simulated by all chapters in this volume: an ERA of sixty and NRA of sixty-five, a replacement rate of 60 percent of AIME at age sixty-five, and a 6 percent annual actuarialadjustment factor between ages sixty and seventy. Our basic procedure is to reestimate the incentive variables under the new policy, then use the probit estimates discussed above to predict changes in retirement behavior. But executing these simulations raises the difficult question of how to translate the earlier models into policy responses. In particular, we face the difficulty that our models are largely unable to explain the age pattern of retirement, an age pattern that is certainly at least partly due to SS incentives (in particular the spike at age sixty-two). We therefore consider three possible simulation approaches. In the first simulation (S1), we use the model with linear age. This simulation does not allow for any age-specific deviations from a linear baseline, therefore increasing the explanatory power of our financial incentive variables. In the second simulation (S2), we use the model with age dummies, but we only consider the impact of changing the financial incentive variables; that is, when retirement ages change, we only consider the impact that this has through changing peak or option value, and not through any other structural shifts. In contrast, the third simulated approach (S3) is also based on the model in which age dummies are used, but imposes a shift in the spikes of the retirement hazard when the policy is changed; that is, when retirement ages change, we assume that there is a corresponding change in the

718

Courtney Coile and Jonathan Gruber

Table 12.9

Average Retirement Rates in Simulations Simulated Reform Case

Plus 3 Years

Common

Males Base retirement rate Peak value S1 S2 S3 Option value S1 S2 S3 Base retirement rate Option value S1 S2 S3

0.057

0.057

0.048 0.051 0.037

0.080 0.064 0.076

0.048 0.051 0.039

0.081 0.071 0.083

Females 0.066

0.066

0.057 0.056 0.039

0.094 0.088 0.100

underlying hazard of retirement by age.32 Simply put, S2 corresponds to the assumption that any age pattern not captured by our financial measures is not due to retirement programs; S3 assumes that the entire age pattern is driven by retirement programs. Since it is unknown whether or not the policy changes would move the spikes in the retirement hazard, S2 and S3 can be thought of as bounding the true effect of the policy change, with S1 somewhere in between. We present our findings in two formats. Table 12.9 shows the baseline average retirement rate as well as the average retirement rates in the various policy simulations.33 The second format is graphical. Each of the figures 12.3–12.11 shows the impact on the hazard of retirement and the cumulative probability of being in the labor force for the baseline and for each of the two reforms. The different figures correspond to different models and simulations: Figures 12.3–12.5 are for the peak value model for males for 32. For the first policy, all age dummies are incremented by three years, so that the retirement hazard at age sixty-two is moved to age sixty-five, and so forth. For the second policy, the age-sixty-two dummy is moved to age sixty, the dummies before age sixty and at ages sixtyfive and older are unaffected, and the age dummies at ages sixty-one to sixty-four are replaced with an average of the age-sixty-three and -sixty-four dummies. Admittedly, these are ad hoc adjustments, but it is difficult to predict how these policy changes would affect the underlying propensity to retire at various ages. 33. Accrual is not used in the simulations, as the coefficients are the wrong sign and insignificant. Peak value is not used in the simulations for women, as the coefficients are highly insignificant.

The Effect of Social Security on Retirement in the United States

719

A

B

Fig. 12.3 S1 for males, peak value model: A, simulated hazard; B, cumulative probability

simulations S1, S2, and S3; figures 12.6–12.8 are for the option value model for males; and figures 12.9–12.11 are for the option value model for females. 12.5.1 Raising the ERA and NRA by Three Years The first policy change, raising the ERA and NRA, would have the effect of lowering the average retirement rate for both men and women. The reduction is 1 percentage point or less in S1 and S2, but 2–3 percentage

720

Courtney Coile and Jonathan Gruber

A

B

Fig. 12.4 S2 for males, peak value model: A, simulated hazard; B, cumulative probability

points under S3. The larger reduction in average retirement rate under S3 is not surprising, as this case moves the spikes in the retirement hazard back by three years. We can assess the wealth and accrual effects underlying these results. This change will have a negative wealth effect on retirement, since this amounts to a benefit cut for any retirement age, which will encourage work. The accrual effects are more complicated: For ages sixty-two to sixty-four, this change will decrease work incentives, as work in these years now only benefits the individual through the dropout year provision and no longer

The Effect of Social Security on Retirement in the United States

721

A

B

Fig. 12.5 S3 for males, peak value model: A, simulated hazard; B, cumulative probability

through a more than fair actuarial adjustment for delayed claiming; and for ages sixty-five to sixty-seven, there will be an increase in work incentives, as the less fair DRC is replaced by the 6.67 percent per year actuarial adjustment. Due to offsetting wealth and accrual effects, there are only modest effects of this change on labor supply when there is no change in retirement norms; however, when a change in retirement norms is applied, the labor supply effects will be substantially larger. The results of these simulations for the first policy change are clearly visible in figures 12.3–12.11. In each case, for S1, there is relatively little im-

722

Courtney Coile and Jonathan Gruber

A

B

Fig. 12.6 S1 for males, option value model: A, simulated hazard; B, cumulative probability

pact on retirement; the simulated pattern of retirement or labor force participation closely follows the linear baseline model. Similarly, in S2, there is little deviation from the baseline; here, the baseline has a nonlinear shape, as it is allowed to reflect variation in the age pattern according to the model with age dummies. But in S3 there are more significant impacts. Indeed, at age sixty-five, this policy raises the odds of participating in the labor force by about one-half from the baseline. For example, for the option value model, the odds of participating at age sixty-five rise from 0.46 to 0.68. This is an enormous effect. This effect peaks at age sixty-five and

The Effect of Social Security on Retirement in the United States

723

A

B

Fig. 12.7 S2 for males, option value model: A, simulated hazard; B, cumulative probability

then fades over time, as retirement is very high under either model at older ages. 12.5.2 Common Retirement System The second policy change, moving to a flat 60 percent replacementrate benefit with an age sixty early retirement age, has a somewhat different pattern of effects. First of all, it significantly raises, rather than lowering, retirement rates. The policy has the effect of raising both SSW (from $177,000 to $269,000 for the median man in the sample) and the

724

Courtney Coile and Jonathan Gruber

A

B

Fig. 12.8 S3 for males, option value model: A, simulated hazard; B, cumulative probability

incentive variables (peak value rises from $13,000 to $35,000 for the median man). Again, there are offsetting wealth and accrual effects, but here the wealth effects are much larger and the result is a much higher retirement rate. The rise in retirement rates for men ranges from 0.7 percentage points in the peak value model with S2 to 2.6 percentage points for the option value model with S3. For females, the range is from 2.2 percentage points, with S2, to 3.4 percentage points, with S3.

The Effect of Social Security on Retirement in the United States

725

A

B

Fig. 12.9 S1 for females, option value model: A, simulated hazard; B, cumulative probability

Once again, we show the implications for retirement and cumulative labor force participation at each age in figures 12.3–12.11. In this case, there are fewer differences across our modeling and simulation methods: There is a general finding of a small rise in the hazard rate at all ages. Unlike the effects of the first reform, which fade over time, these impacts are either constant or grow at all ages, reflecting the fact that this policy does not so much shift incentives toward earlier retirement as it does raise the wealth level of retirees at all ages.

726

Courtney Coile and Jonathan Gruber

A

B

Fig. 12.10 S2 for females, option value model: A, simulated hazard; B, cumulative probability

12.6 Conclusion The SS program is the most important source of retirement income support for older Americans. As such, it is possible that the incentives embodied in this system for continued work or retirement at various ages are a critical determinant of retirement decisions. Understanding the influence that SS has on retirement decisions is particularly important now, as any reforms to the SS system will change the structure of the program in a manner that has important impacts on retirement incentives. Our paper has used the richest available current data, the HRS, to pro-

The Effect of Social Security on Retirement in the United States

727

A

B

Fig. 12.11 S3 for females, option value model: A, simulated hazard; B, cumulative probability

vide new evidence on the impact of SS on retirement. We find that retirement decisions appear to be made with reference to the entire stream of future SSW accruals, rather than just the level of wealth or the accrual over the next year, so that forward-looking measures, such as our peak value measure, are important variables to include in retirement models. These forward-looking measures have a significant impact on retirement decisions for men, although for women only the option value model generates significant results. Simulations of policy changes indicate that these changes could have significant impacts on retirement decisions. An in-

728

Courtney Coile and Jonathan Gruber

crease in the ERA and NRA could result in a 2 percentage point decrease in the average annual retirement rate if the increase has the effect of changing retirement norms, although the effect would be much smaller if norms are unchanged, due to offsetting wealth and accrual effects. A move to a policy with a 60 percent replacement rate at age sixty-five, a much more generous policy than the current SS system, would have very large wealth effects, raising the average annual retirement rate by 2–3 percentage points.

References Blau, David M. 1994. Labor force dynamics of older men. Econometrica 62 (1): 117–56. Burtless, Gary. 1986. Social security, unanticipated benefit increases, and the timing of retirement. Review of Economic Studies 53 (5): 781–805. Burtless, Gary, and Robert Moffitt. 1984. The effect of social security benefits on the labor supply of the aged. In Retirement and economic behavior, ed. Henry Aaron and Gary Burtless, 135–75. Washington, D.C.: Brookings Institution. Coile, Courtney. 2003. Retirement incentives and couples’ retirement decisions. NBER Working Paper no. 9496. Cambridge, Mass.: National Bureau of Economic Research. Coile, Courtney, Peter Diamond, Jonathan Gruber, and Alain Jousten. 2002. Delays in claiming social security benefits. Journal of Public Economics 84 (3): 357–86. ———. 2000. Social security and retirement. NBER Working Paper no. 7830. Cambridge, Mass.: National Bureau of Economic Research, August. Coile, Courtney, and Jonathan Gruber. 2001. Social security incentives for retirement. In Themes in the Economics of Aging, ed. David A. Wise. Chicago: University of Chicago Press. Diamond, Peter, and Jonathan Gruber. 1999. Social security and retirement in the United States. In Social Security and Retirement around the World, ed. Jonathan Gruber and David A. Wise, 437–73. Chicago: University of Chicago Press. Diamond, Peter, and Jerry Hausman. 1984. Retirement and unemployment behavior of older men. In Retirement and economic behavior, ed. Henry Aaron and Gary Burtless, 97–135. Washington, D.C.: Brookings Institution. Fields, Gary S., and Olivia S. Mitchell. 1984. Economic determinants of the optimal retirement age: An empirical investigation. Journal of Human Resources 19 (2): 245–62. Gustman, Alan L., and Thomas L. Steinmeier. 1985. The 1983 social security reforms and labor supply adjustments of older individuals in the long run. Journal of Labor Economics 3 (2): 237–53. ———. 1986. A structural retirement model. Econometrica 54 (3): 555–84. Haider, Steven, and Gary Solon. 1999. Nonrandom selection in the HRS social security earnings sample. University of Michigan, Department of Economics. Unpublished Manuscript. Hausman, Jerry A., and David A. Wise. 1985. Social security, health status, and retirement. In Pensions, labor, and individual choice, ed. David A. Wise, 159–91. Chicago: University of Chicago Press.

The Effect of Social Security on Retirement in the United States

729

Hurd, Michael D. 1990. Research on the elderly: Economic status, retirement, and consumption and saving. Journal of Economic Literature 28 (2): 565–637. Krueger, Alan B., and Jorn-Steffan Pischke. 1992. The effect of social security on labor supply: A cohort analysis of the notch generation. Journal of Labor Economics 10 (4): 412–37. Lumsdaine, Robin, and David Wise. 1994. Aging and labor force participation: A review of trends and explanations. In Aging in the United States and Japan: Economic trends, ed. Yukio Noguchi and David Wise, 7–42. Chicago: University of Chicago Press. MaCurdy, Thomas E. 1981. An empirical model of labor supply in a life-cycle setting. Journal of Political Economy 89 (6): 1059–85. Mitchell, Olivia S., and Gary S. Fields. 1982. The effects of pensions and earnings on retirement. In Research in labor economics, Vol. 5, ed. Ronald Ehrenberg, 115–55. Greenwich, Conn.: JAI Press. National Center for Health Statistics. 1990. Vital statistics of the United States, 1987. Vol. 2. Washington, D.C.: Washington Public Health Service. Quinn, Joseph. 1999. Retirement patterns and bridge jobs in the 1990s. Policy brief for the employee benefit research institute. Washington, D.C.: Employee Benefit Research Institute (EBRI). Ruhm, Christopher. 1995. Secular changes in the work and retirement patterns of older men. Journal of Human Resources 30 (2): 362–85. Samwick, Andrew A. 1998. New evidence on pensions, social security, and the timing of retirement. Journal of Public Economics 70 (2): 207–36. Stock, James H., and David A. Wise. 1990a. Pensions, the option value of work, and retirement. Econometrica 58 (5): 1151–80. ———. 1990b. The pension inducement to retire: An option value analysis. In Issues in the economics of aging, ed. David A. Wise, 205–29. Chicago: University of Chicago Press. Social Security Administration. 1998. Annual statistical supplement, 1998, to the Social Security bulletin. Baltimore, Mary.: U.S. Department of Health and Human Services. Sueyoshi, Glenn T. 1989. Social security and the determinants of full and partial retirement: A competing risks analysis. NBER Working Paper no. 3113. Cambridge, Mass.: National Bureau of Economic Research, September. U.S. Bureau of the Census. 1998. No. 1 population and area: 1790 to 1990. In Statistical abstract of the United States. 118th ed. Washington, D.C.: Bureau of the Census.

Contributors

Michael Baker Department of Economics University of Toronto 150 St. George Street Toronto, Ontario M5S 3G7 Canada Paul Bingley National Centre for Register-Based Research University of Aarhus Tåsingegade 1 DK-8000 Aarhus C Denmark Didier Blanchet Institut National de la Statistique et des Études Économiques (INSEE), France 18 Boulevard Adolphe-Pinard 75675 Paris Cedex 14 France Richard Blundell Department of Economics University College London Gower Street London WC1E 6BT England

Michele Boldrin Department of Economics University of Minnesota 908 Walter W. Heller Hall Minneapolis, MN 55455 Axel Börsch-Supan Department of Economics University of Mannheim Building L13, 17D-68131 D-68131 Mannheim Germany Agar Brugiavini Department of Economics University of Venice, Ca’Foscari Cannaregio 873 30121 Venice Italy Courtney Coile Department of Economics Wellesley College 106 Central Street Wellesley, MA 02481 Nabanita Datta Gupta Aarhus School of Business Department of Economics Prismet, Silkeborgvej 2 DK 8000 Aarhus C Denmark

731

732

Contributors

Arnaud Dellis Department of Economics Cornell University 404 Uris Hall Ithaca, NY 14853 Raphaël Desmet Economics Department University of Liège Boulevard du Rectorat, 7-Bâtiment B31 4000 Liège Belgium Jonathan Gruber Department of Economics Massachusetts Institute of Technology 50 Memorial Drive Cambridge, MA 02142-1347 Sergi Jiménez-Martín Department of Economics Universidad Carlos III de Madrid C./Madrid, 12628903 Getafe (Madrid) Spain Alain Jousten Economics Department University of Liège Boulevard du Rectorat, 7-Bâtiment B31 4000 Liège Belgium Arie Kapteyn RAND 1700 Main Street P.O. Box 2138 Santa Monica, CA 90407-2138 Simone Kohnz Graduate School of Economics University of Munich 80539 Munich Germany

Ronan Mahieu Caisse Nationale des Allocations Familales Direction des Statistiques, des Études et de la Recherche 23 Rue Daviel 75634 PARIS Cedex 13 France Giovanni Mastrobuoni Princeton University Department of Economics Fisher Hall Princeton, NJ 08544-1021 Costas Meghir Department of Economics University College London Gower Street London WC1E 6BT England Kevin Milligan Department of Economics University of British Columbia 997-1873 East Mall Vancouver BC V6T 1Z1 Canada Akiko Sato Oishi National Institute of Population and Social Security Research Hibiya Kokusai Building, 6th Floor 2-2-3 Uchisaiwai-cho Chiyoda-ku, Tokyo 100-0011 Japan Takashi Oshio Department of Education Tokyo Gakugei University 4-1-1-Nukui Kitamachi Koganei, Tokyo 184-8501 Japan Mårten Palme Department of Economics Stockholm University Universitetsvägen 10A SE-106 91 Stockholm Sweden

Contributors Peder J. Pedersen Department of Economics University of Aarhus DK-8000 Aarhus Denmark Franco Peracchi Faculty of Economics University of Rome, Tor Vergata Via Columbia, 2 I-00133 Rome Italy Sergio Perelman Economics Department University of Liège Boulevard du Rectorat, 7-Bâtiment B31 4000 Liège Belgium Reinhold Schnabel Department of Economics University of Essen 45177 Essen Germany

Sarah Smith The Financial Services Authority 25 The North Colonnade Canary Wharf London E145HS England Ingemar Svensson National Social Insurance Board SE-103 61 Stockholm Sweden Klaas de Vos CentER Applied Research Tilburg University 5000 LE Tilburg The Netherlands David A. Wise John F. Kennedy School of Government, Harvard University and NBER 1050 Massachusetts Avenue Cambridge, MA 02138-5398

733

Author Index

Aarts, L., 467 Abe, Y., 412 Baker, Michael, 99, 100, 102, 111n19, 115, 128n36 Benjamin, Dwayne, 99, 100, 115 Blanchet, D., 245 Blau, David M., 697 Blöndal, S., 42n1 Boldrin, M., 500, 502, 502n1, 503, 537 Bommier, A., 235 Bone, M. J., 653 Börsch-Supan, A., 286n2, 287, 287n3, 304, 310, 310n22 Bouillot, L., 43n3 Brugiavini, A., 364n24 Burbridge, John, 100n2 Burtless, Gary, 693, 698 Coile, Courtney, 11n1, 12n3, 117, 287n4, 699, 702, 706 Colombino, U., 357, 358 Compton, Janice, 100 Deboosere, P., 53 Diamond, Peter, 697, 698n10 Disney, R., 646, 650n6, 652 Falkinger, J., 358 Fields, Gary S., 697, 698 Franco, D., 356n12

Gadeyne, S., 53 Gruber, Jonathan, 1, 7, 11n1, 12n3, 100, 102, 111n19, 115, 117, 128n36, 156, 285n1, 472, 579, 697, 699, 702, 706 Gustman, Alan L., 698 Hansson-Brusewitz, Urban, 597 Hausman, Jerry A., 698, 698n10 Henkens, K., 467 Heyma, A., 468 Hurd, Michael D., 697 Iwamoto, Y., 412 Jiménez-Martin, S., 500, 502, 502n1, 503, 537, 539 Johnson, P., 645n1 Jong, P. R. de, 467 Jousten, A., 42 Kapteyn, A., 472, 475 Kerkhofs, M., 467, 468, 469 Kotlikoff, Laurence J., 11n1 Krueger, Alan B., 699, 729 Lazear, Edward, 11n1 Lindeboom, M., 468, 469, 567 Lumsdaine, Robin, 697 MaCurdy, Thomas E., 698 Magnac, T., 235

735

736

Author Index

Martinéz, P., 515 Meghir, C., 310n22, 646 Milligan, Kevin, 102, 111n19, 128n36 Miniaci, R., 357 Mitchell, Olivia S., 697, 698 Moffit, Robert, 467, 693, 698 Nikami, K., 403 Ogawa, H., 412 Oshio, T., 414, 422 Palmer, E., 598 Pedersen, P. J., 158 Pelé, L. P., 245 Peracchi, F., 500, 502, 502n1, 503, 537 Perelman, S., 43n3 Pesando, James E., 100n2 Pestieau, P., 42, 58 Pischke, Jorn-Steffan, 698, 699 Ralle, P., 245 Rea, Samuel A., 100n2 Riphahn, Regina T., 287n3 Roger, M., 284 Rossi, N., 363n20 Ruhm, Christopher, 697 Samwick, Andrew A., 698n7, 698n9 Sánchez, A., 539 Scarpetta, S., 42n1 Schmidt, P., 287, 287n3, 310 Schnabel, R., 286n2, 287, 303n21 Seike, A., 405, 412, 414

Siddiqui, S., 287 Siegers, J., 467 Skogman Thoursi, Peter, 598 Smith, S., 650n6 Sorgato, A., 363n20 Spataro, L., 357 Stears, G., 652 Steinmeier, Thomas L., 698 Stijns, J. P., 58 Stock, James H., 11n2, 13–14, 58, 118, 175, 253, 287, 357, 662, 698n9, 706, 707n24 Sueyoshi, G. T., 310, 310n22 Sundén, Annika, 597 Takayama, N., 401n2, 412, 414 Theeuwes, J. J. M., 467, 468, 469 Thio, V., 467, 568 Tompa, Emile, 100 Toniolo, G., 363n20 Vos, K. de, 472, 475 Wadensjö, Eskil, 598, 641n9 Whitehouse, E., 310, 646 Winter-Ebmer, R., 358 Wise, David A., 1, 7, 11n1, 11n2, 13–14, 58, 118, 156, 175, 253, 285n1, 287, 357, 472, 579, 662, 697, 698, 706, 707n24 Woittiez, I., 467, 468 Yashiro, N., 403, 414, 421–22 Zweimüller, J. R., 358

Subject Index

Age, retirement rates and, 13. See also Eligibility ages; Retirement ages Belgium: earnings histories and projections for, 52–53; effect of labor mobility on social security systems, 41–42; guaranteed-minimum-income pensions in, 49; labor force participation rates in, 5; labor market participation rates and retirement in, 49; option value incentive measures and, 58–60; overview of data used in study of, 51–52; pathways to retirement in, 49–51; peak value incentive measures and, 58–60; pillars of social security schemes in, 43–44; regression results for, 60–69; retirement scheme for public-sector employees in, 47–48; retirement scheme for self-employed in, 48–49; retirement scheme for wage earners in, 44–47; simulations results, 69–96; social security wealth concept and, 53–58; uncertain future of social security systems in, 41 Benefit eligibility ages, 8 Britain. See United Kingdom Canada: calculation of benefit entitlements in, 114–15; Canada Pension Plan in, 103–6; data sets for, 109–12; earnings and nonlabor income projections for study of, 112–14; empirical framework

for study of ISW for, 122–26; Guaranteed Income Supplement in, 106–7; labor force participation rates in, 100, 101; Old Age Security (OAS) system in, 103; one-year accrual calculation of ISW in, 115–16; option value calculation of ISW in, 117–18; overview of data for study of, 102–3; overview of transfer programs to older Canadians and, 99–103; paths to retirement in, 107–9; peak value calculation of ISW in, 117; policy simulations for, 133–46; private pension coverage in, 107; Quebec Pension Plan in, 100, 103–6; recent studies of Income Security programs, 100–102; retirement regression results for, 128–33; sample characteristics of study of, 126–28; sample estimates of difference incentive measures for, 118– 22; spousal behavior for collecting entitlements in, 115; Spouse Allowance benefit, 106–7; study findings for, 146– 49 Canada Pension Plan (CPP), 100, 103–6 Common reform: comparing three-year eligibility delay and, 34–35; effect of, by country, 23–25; the Netherlands simulation of, 483–94; simulation for France, 266, 272–79; simulation methodology for, 23–25; simulation results for, 30–34. See also Simulations; specific country

737

738

Subject Index

Denmark: constructing measures of social security wealth for, 169–75; defined benefits in, 155; estimating optionvalue model of retirement for, 175–77; exit routes from labor force in, 160–64; labor market pensions in, 155–56; labor participation rates in, 156–59; overview of data for study of, 167–68; overview of retirement in, 153–54; paths to early retirement in, 156; pension system in, 154; policy simulations for, 177–81; pressure for policy reforms in, 164–67; sample descriptives and earnings projections for, 168–69; social disability pensions in, 154–55; structure of working and nonworking populations in, 159–60; study conclusions for, 181–82; transitional benefits program in, 156 Disability insurance, in Sweden, 588 Earnings histories: for Belgium, 52–53; for Canada, 112–14; for Germany, 300– 303; for Japan, 415–18; for the Netherlands, 471–72; for Spain, 526–31; for United Kingdom, 654–56 Eligibility ages, effect of, on retirement, 17– 23. See also Retirement ages France: analysis of incentives for retirement for civil servants in, 258–60; analysis of incentives for retirement in private sector in, 255–58; common reform simulation for, 266, 272–79; complementary schemes for wage earners in private sector, 241–42; computation of pensions for civil servants in, 242–43; computation of pensions for wage earners in private sector in, 238–41; data set for study of, 246–48; demographic data for studying pension benefits for, 249–51; early retirement in, 243–44; econometric analysis of retirement in, 260–65; general structure of pension system in, 237–38; incentives linked to unemployment benefits and early retirement in, 254–55; labor force participation rates in, 5, 244–45; mandatory retirement in, 243; option value incentive measures for, 253–54; overview of retirement in, 235–37; pathways to retirement in, 251–52; peak value incentive measures

for, 253; reconstructing wages and pension levels for, 248–49; social security wealth for, 252–54; study’s conclusions for, 279–81; three-year delay simulation for, 265–72 German Socio-Economic Panel (GSOEP), 298–300 Germany: constructing earning histories for, 300–303; constructing social security wealth for, 303–7; data sources for study of, 298–300; defining retirement status in, 301; effect of eligibility ages on retirement in, 19–22; labor force participation rates in, 5; option value incentive measures in, 309–10; overview of retirement in, 285–87; peak value incentive measures in, 307– 9; public retirement insurance (GRV) system in, 287–95; public sector pensions, 295–98; regression results for, 310–17; retirement options in, 14–15; selecting retirement programs in, 301– 3; self-employed workers in, 288; simulation results for, 317–23 Great Britain. See United Kingdom Guaranteed Annual Income Supplement for the Aged (GAINS-A), 107 Guaranteed Income Supplement (GIS), 106–7 Hazard rates: for men in Germany, 19–22; for men in United states, 17–18 Identification issue, 12–13 Incentive measures. See Option value incentive measures; Peak value incentive measures; Social security wealth (SSW) Incentives. See Social security incentives Income security wealth (ISW), in Canada, 102–3; one-year accrual calculation for, 115–16; present discounted value of, 115. See also Social security wealth (SSW) Italy: data set for study of, 358–62; earnings projections for, 362–63; labor force participation rates in, 5; literature review of retirement decisions in, 357–58; modeling retirement choices in, 372– 85; overview of retirement in, 345–46; pre-1993 social security system rules

Subject Index in, 348–51; recent expenditure trends for social security system, 347–48; reforms of 1990s in, 351–54; simulating retirement choices in, 385–90; social security wealth in, 367–72; study’s conclusions for, 390–93; types of transitions into retirement in, 363–67 Japan: calculations of incentive measures for, 418–23; disability pension in, 404; earning histories and projections for, 415–18; empirical framework for regression analysis for retirement in, 423–25; employer-provided pension in, 404–5; estimation results for, 425–31; Kosei Nenkin Hoken in, 401–3, 404–5; labor force participation rates in, 5; labor market participation of elderly in, 405–10; 1999 Pension Reform Act of, 405; option value incentive measures for, 419–23; overview of data for study of, 414–15; overview of retirement in, 399–401; pathways to retirement in, 410–12; peak value incentive measures for, 419–23; research on retirement incentives for, 412–14; simulations results for, 431–59; social security wealth for, 418–19; types of pensions in, 401–5; unemployment insurance in, 403; wage subsidy for elderly workers in, 403–4; wage subsidy in, 407–8; Zaishoku Pension in, 402, 403 Kosei Nenkin Hoken (KNH, Japan), 401–3, 404 Kyosai Kumiai (Japan), 402 Labor force participation rates: in Canada, 100, 101table; decline in, 4–5; in Denmark, 156–57; disability insurance in, 588; in France, 244–45; in the Netherlands, 465–66; occupational pensions in, 554–88; sickness insurance in, 588– 89; social security system in, 581–84; in Spain, 502, 504; in Sweden, 579; in United States, 5, 691–93, 695–97 Labor mobility. See Mobility, labor Micro-estimation approach, advantages of, 2 Mobility, labor, in Belgium, and social security systems, 41–42

739

Netherlands, the: common reform simulation for, 483–94; data overview for, 469–71; early retirement schemes in, 464; earnings histories and projections for, 471–72; estimation results for, 475– 82; labor force participation rates in, 5, 465–66; literature on retirement effects of social security in, 466–69; option value incentive measures for, 473–75; overview of retirement in, 461–62; simulations for, 482–94; social security system in, 462–66; social security wealth in, 472–73; three-year delay in eligibility simulation for, 482–94 Option value incentive measures, 606; for Belgium, 58–60; calculating, in Canada, 117–18; for France, 253–54; for Italy, 367–68; for Japan, 419–23; for the Netherlands, 473–75, 577; for United Kingdom, 15n4, 662–67; for United States, 706–7. See also Peak value incentive measures Option value model, 11–12, 13–14 Pathways to retirement: in Belgium, 49–51; in France, 251–52; in Japan, 410–12; in Spain, 514 Peak value incentive measures, 12–13, 605; for Belgium, 58–60; calculating, in Canada, 117; for France, 253; for Germany, 307–9; for Italy, 367–68; for Japan, 419–23; for the Netherlands, 474–75, 477; for United Kingdom, 15n4, 662–66; for United States, 706–9. See also Option value incentive measures Pension Reform Act of 1999 (Japan), 405 Personal pensions. See Private pensions Private pensions: in Canada, 107; in United Kingdom, 651–53 Quebec Pension Plan (QPP), 100, 103–6. See also Canada Regímenes Especiales de la Seguridad Social (RESS, Spain), 506–7, 531–32 Regimén General de la Seguridad Social (RGSS, Spain), 499–500; benefit computation for, 507–9; early retirement and, 509–10; earning distributions, earnings histories, and projections for,

740

Subject Index

526–31; family considerations and, 510–11; financing and eligibility of, 507; maximum and minimum pensions for, 510 Retirement: definitions of, 408; effect of eligibility ages on, 17–23; effect of social security incentives on, 2 Retirement ages: retirement rates and, 13; and social security provisions, 8. See also Eligibility ages Retirement incentives. See Social security incentives Retirement rates, age and, 13 Self-employed workers: in Belgium, 48–49; in Germany, 288; in Spain, 511–12 Simulations, 2; for delaying eligibility ages by three years, 23–25; results for threeyear delay, 25–30. See also Common reform; specific country Social security implicit tax, on earnings, 6–7 Social security incentives: effect of, on retirement, 2; and withdrawal of workers at older ages, 7–8 Social security wealth (SSW), 6, 16; Belgium study and, 53–58; for France, 252–54; for Germany, 303–7; in Italy, 367–72; for Japan, 418–19; in the Netherlands, 472–73; in Sweden, 605. See also Income security wealth (ISW) Spain: early retirement in, 502; labor force participation rates in, 5, 502, 504; life expectancy in, 503–4; overview of data set for, 514–26; overview of social security system in, 499–505; pathways to retirement in, 514; policy simulations for, 547–67; public programs for old-age workers in, 505–6; regression results for, 542–47; RESS in, 506–7; RGSS in, 499–500, 507–11; sample evidence for RGSS, 539–42; schemes for farmers in, 512; schemes for public servants in, 512–14; schemes for self-employed in, 511–12; social security incentives in, 532–39; social security regimes and rules in, 506–7; structure of population in, 503–4; summary of results for, 567– 68 Spouse Allowance (SPA) benefit, 106–7; provincial programs in, 107 State Earnings Related Pension Scheme (SERPS), 645, 646, 650–51

Sweden: data set used for study of, 598–600; empirical model for retirement behavior in, 614–15; housing allowances in, 590–91; incentive measures used for study of, 604–14; income tax system in, 590; labor force participation rates in, 5, 579; mandatory retirement rules in, 591; overview of research on retirement behavior in, 597–98; results of study for, 615–20; review of methodology for study of, 600–604; simulations for, 620–40; social security wealth and, 605; sources of income after retirement in, 591–97; unemployment insurance in, 589–90 Tax force to retire, 6–7 Tax rates, 5–6 Unemployment insurance: in France, 254– 55; in Japan, 403; in Sweden, 589–90 United Kingdom, 15n4; basic state pension in, 650, 656; construction of incentive measures for, 656–61; earnings histories and projections for, 654–56; effect of eligibility ages on retirement in, 22– 23; impact of incentive and wealth variables on retirement decision in, 667–74; income support in, 651; invalidity benefit in, 651, 658–59; labor participation rates in, 643–45; occupational pensions in, 659–61; option value incentive measures for, 662–66; overview of data used for, 653–54; peak value and option value incentive measures in, 15n4; peak value incentive measures for, 662–66; pension system in, 649–53; private pensions in, 651–53; SERPS and, 645, 646, 650–51, 656–58; simulations for, 674–79; summary and conclusions of impact on incentives on retirement decisions in, 679–80; total pension wealth and accrual measures for, 661–68; trends in state pension provision in, 645–46 United States: conclusions of study of, 726– 28; effect of eligibility ages on retirement in, 17–18; impact of Social Security on retirement decisions in, 693; labor force participation rates in, 5, 691– 93; labor market participation rates in, 695–97; measuring retirement incentives inherent in Social Security sys-

Subject Index tem, 701–5; overview of data for, 699– 701; peak value incentive measures for, 706–9; regression model for, 709–11; regression results for, 711–17; retirement options in, 14; review of retire-

741

ment-incentives literature for, 697–99; simulations for, 717–26; Social Security system in, 694–95 Zaishoku Pension (Japan), 402