Social Security Programs and Retirement around the World: Disability Insurance Programs and Retirement 9780226262604

Even as life expectancy in many countries has continued to increase, social security and similar government programs can

181 4 7MB

English Pages 544 [559] Year 2016

Report DMCA / Copyright

DOWNLOAD FILE

Polecaj historie

Social Security Programs and Retirement around the World: Disability Insurance Programs and Retirement
 9780226262604

Citation preview

Social Security Programs and Retirement around the World

National Bureau of Economic Research Conference Report

Social Security Programs and Retirement around the World Disability Insurance Programs and Retirement

Edited by

David A. Wise

The University of Chicago Press Chicago and London

David A. Wise is the John F. Stambaugh Professor of Political Economy at the Kennedy School of Government at Harvard University. He is the area director of Health and Retirement Programs and director of the Program on the Economics of Aging at the National Bureau of Economic Research.

The University of Chicago Press, Chicago 60637 The University of Chicago Press, Ltd., London © 2016 by the National Bureau of Economic Research All rights reserved. Published 2016. Printed in the United States of America 24 23 22 21 20 19 18 17 16 1 2 3 4 5 ISBN-13: 978-0-226-26257-4 (cloth) ISBN-13: 978-0-226-26260-4 (e-book) DOI: 10.7208/chicago/9780226262604.001.0001 Library of Congress Cataloging-in-Publication Data Social security programs and retirement around the world : disability insurance programs and retirement / edited by David A. Wise. pages cm — (National Bureau of Economic Research conference report) ISBN 978-0-226-26257-4 (cloth : alk. paper) — ISBN 978-0-226-26260-4 (e-book) 1. Social security. 2. Postemployment benefits. 3. Disability retirement. I. Wise, David A. II. Series: National Bureau of Economic Research conference report. HD7091.S6244 2015 362—dc23 2015011454

♾ This paper meets the requirements of ANSI/NISO Z39.48–1992 (Permanence of Paper).

National Bureau of Economic Research Officers Martin B. Zimmerman, chairman Karen N. Horn, vice chairman James M. Poterba, president and chief executive officer Robert Mednick, treasurer

Kelly Horak, controller and assistant corporate secretary Alterra Milone, corporate secretary Denis Healy, assistant corporate secretary

Directors at Large Peter C. Aldrich Elizabeth E. Bailey John H. Biggs John S. Clarkeson Don R. Conlan Kathleen B. Cooper Charles H. Dallara George C. Eads Jessica P. Einhorn

Mohamed El-Erian Linda Ewing Jacob A. Frenkel Judith M. Gueron Robert S. Hamada Peter Blair Henry Karen N. Horn John Lipsky Laurence H. Meyer

Michael H. Moskow Alicia H. Munnell Robert T. Parry James M. Poterba John S. Reed Marina v. N. Whitman Martin B. Zimmerman

Directors by University Appointment Jagdish Bhagwati, Columbia Timothy Bresnahan, Stanford Alan V. Deardorff, Michigan Ray C. Fair, Yale Edward Foster, Minnesota John P. Gould, Chicago Mark Grinblatt, California, Los Angeles Bruce Hansen, Wisconsin–Madison

Benjamin Hermalin, California, Berkeley Marjorie B. McElroy, Duke Joel Mokyr, Northwestern Andrew Postlewaite, Pennsylvania Cecilia Rouse, Princeton Richard L. Schmalensee, Massachusetts Institute of Technology David B. Yoffie, Harvard

Directors by Appointment of Other Organizations Jean-Paul Chavas, Agricultural and Applied Economics Association Martin Gruber, American Finance Association Ellen L. Hughes-Cromwick, National Association for Business Economics Arthur Kennickell, American Statistical Association William W. Lewis, Committee for Economic Development Robert Mednick, American Institute of Certified Public Accountants

Alan L. Olmstead, Economic History Association Peter L. Rousseau, American Economic Association Gregor W. Smith, Canadian Economics Association William Spriggs, American Federation of Labor and Congress of Industrial Organizations Bart van Ark, The Conference Board

Directors Emeriti George Akerlof Carl F. Christ Franklin Fisher George Hatsopoulos

Saul H. Hymans Rudolph A. Oswald Peter G. Peterson Nathan Rosenberg

John J. Siegfried Craig Swan

Relation of the Directors to the Work and Publications of the National Bureau of Economic Research 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 co-authors 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

1.

2.

Acknowledgments

ix

Introduction Courtney Coile, Kevin Milligan, and David A. Wise

1

Disability Insurance Incentives and the Retirement Decision: Evidence from the United States Courtney Coile Effect of Pensions and Disability Benefits on Retirement in the United Kingdom James Banks, Carl Emmerson, and Gemma Tetlow

3.

Option Value of Disability Insurance in Canada Kevin Milligan and Tammy Schirle

4.

Health Status, Disability, and Retirement Incentives in Belgium Alain Jousten, Mathieu Lefebvre, and Sergio Perelman

5.

6.

Health, Disability Insurance, and Labor Force Exit of Older Workers in the Netherlands Adriaan Kalwij, Klaas de Vos, and Arie Kapteyn Retirement, Early Retirement, and Disability: Explaining Labor Force Participation after Fifty-Five in France Luc Behaghel, Didier Blanchet, and Muriel Roger

45

81

137

179

211

251

vii

viii

7.

8.

9.

10.

11.

12.

Contents

Health, Financial Incentives, and Early Retirement: Microsimulation Evidence for Germany Hendrik Jürges, Lars Thiel, Tabea Bucher-Koenen, Johannes Rausch, Morten Schuth, and Axel Börsch-Supan Health, Disability Insurance, and Retirement in Denmark Paul Bingley, Nabanita Datta Gupta, Michael Jørgensen, and Peder J. Pedersen Pathways to Retirement and the Role of Financial Incentives in Sweden Per Johansson, Lisa Laun, and Mårten Palme Health Status, Disability Insurance, and Incentives to Exit the Labor Force in Italy: Evidence from SHARE Agar Brugiavini and Franco Peracchi Financial Incentives, Health, and Retirement in Spain Pilar García-Gómez, Sergi Jiménez-Martín, and Judit Vall Castelló Option Value of Work, Health Status, and Retirement Decisions in Japan: Evidence from the Japanese Study on Aging and Retirement (JSTAR) Satoshi Shimizutani, Takashi Oshio, and Mayu Fujii Contributors Author Index Subject Index

285

331

369

411

455

497

537 541 545

Acknowledgments

Funding for this project was provided by the National Institute on Aging, grant numbers P01-AG005842 and P30-AG012810 to the National Bureau of Economic Research. We thank two anonymous reviewers for detailed and thoughtful comments. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Institute on Aging, the National Institutes of Health, or the National Bureau of Economic Research.

ix

Introduction Courtney Coile, Kevin Milligan, and David A. Wise

Through the coordination of work by a team of analysts in twelve countries for over fifteen years, the International Social Security (ISS) project has used the vast differences in social security programs across countries as a natural laboratory to study the effects of retirement program provisions on the labor force participation of older persons. A central finding of the project is that in many countries the provisions of social security and related government programs provide strong incentives for workers to leave the labor force at relatively young ages and that reducing the inducement to leave the labor force can lead workers to delay retirement and yield large improvements in the financial position of government budgets. The work to date has also made clear that disability insurance (DI) programs can play a large role in the departure of older persons from the labor force, as many workers pass through DI on their path from employment to retirement. This is the sixth phase of the ongoing ISS project. This phase is particularly related to the fifth phase (Wise 2012) and the second phase (Gruber and Wise 2004) of the project. This volume continues the focus of the previous volume on DI programs while extending the methodology to study retirement behavior used in the second phase to focus in particular on the effects of the DI programs. The key question this volume seeks to address is: Given Courtney Coile is professor of economics at Wellesley College and a research associate of the National Bureau of Economic Research. Kevin Milligan is associate professor at the Vancouver School of Economics, University of British Columbia, and a research associate of the National Bureau of Economic Research. David A. Wise is the John F. Stambaugh Professor of Political Economy at the Kennedy School of Government at Harvard University. He is the area director of Health and Retirement Programs and director of the Program on the Economics of Aging at the National Bureau of Economic Research. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http://​www​.nber​.org​/chapters​/c13323​.ack.

1

2     Courtney Coile, Kevin Milligan, and David A. Wise

health status, to what extent are differences in labor force participation across countries determined by the provisions of disability insurance programs? The fifth phase presented an analysis of historical trends in our group of countries to set the stage for the more formal analysis of disability insurance programs in the current volume. In that phase, the countries summarized DI program reforms and considered how DI reforms were related to changes in health, in particular as measured by changes in mortality. We also treated DI reforms as natural experiments—­not prompted by changes in the health or employment circumstances of older persons—­and showed that these “exogenous” reforms often had a very large effect on the labor force participation of older workers. The second phase, which was based on microeconomic analysis of the relationship between a person’s decision to retire and the social security and other program incentives faced by that person, documented the large effects that changing plan provisions would have on the labor force participation of older workers. In that phase the country teams considered the employment implications of increasing retirement program eligibility ages, including the eligibility age for DI, and showed that these changes would have very large effects on employment at older ages. As described in more detail below, the current phase of the project differs from the second in incorporating a more careful modeling of the incentives arising from the DI program and simulating how changes in access to DI might affect labor force participation. To summarize the findings of the remaining phases: The first phase of the project described the retirement incentives inherent in plan provisions and documented the strong relationship across countries between social security incentives to retire and the proportion of older persons out of the labor force (Gruber and Wise 1999). The third phase (Gruber and Wise 2007) demonstrated the consequent fiscal implications that extending labor force participation would have on net program costs—­reducing government social security benefit payments and increasing government tax revenues. The analyses in the first two phases, as well as the analysis in the third phase, are summarized in the introduction to the third phase. In the fourth phase (Gruber and Wise 2010) we directed attention to the oft-­claimed proposition that incentives to induce older persons to retire—­ inherent in the provisions of social security systems—­were prompted by youth unemployment. Many have worried that if the incentives to retire were removed and older persons stayed longer in the labor force, the job opportunities of youth would be reduced. We found no evidence to support this “boxed economy” proposition. In short, we concluded: “the overwhelming weight of the evidence, as well as the evidence from each of the several different methods of estimation, is contrary to the boxed economy proposition. We find no evidence that increasing the employment of older persons will reduce the employment opportunities of youth and no evidence that

Introduction    3

increasing the employment of older persons will increase the unemployment of youth.” The results of the ongoing project are the product of analyses conducted for each country by analysts in that country. Researchers who have participated in the project are listed below: Belgium

Alain Jousten, Mathieu Lefebvre, Sergio Perelman, Pierre Pestieau, Raphaël Desmet, Arnaud Dellis, and Jean-­Philippe Stijns Canada Kevin Milligan, Tammy Schirle, Michael Baker, and Jonathan Gruber Denmark Paul Bingley, Nabanita Datta Gupta, Michael Jørgensen, and Peder J. Pedersen France Luc Behaghel, Didier Blanchet, Muriel Roger, Thierry Debrand, Melika Ben Salem, Antoine Bozio, Ronan Mahieu, Louis-­Paul Pelé, and Emmanuelle Walraet Germany Axel Börsch-­Supan, Tabea Bucher-­Koenen, Hendrik Jürges, Johannes Rausch, Morten Schuth, Lars Thiel, Reinhold Schnabel, Simone Kohnz, and Giovanni Mastrobuoni Italy Agar Brugiavini and Franco Peracchi Japan Mayu Fujii, Takashi Oshio, Satoshi Shimizutani, Akiko Sato Oishi, and Naohiro Yashiro Netherlands Adriaan Kalwij, Arie Kapteyn, and Klaas de Vos Spain Pilar García Gómez, Sergi Jiménez-­Martín, Judit Vall Castelló, Michele Boldrín, and Franco Peracchi Sweden Per Johansson, Lisa Laun, Mårten Palme, and Ingemar Svensson United Kingdom James Banks, Carl Emmerson, Gemma Tetlow, Richard Blundell, Antonio Bozio, Paul Johnson, Costas Meghir, and Sarah Smith United States Courtney Coile, Kevin Milligan, Jonathan Gruber, and Peter Diamond An important goal of the project has been to present results that were as comparable as possible across countries. Thus the chapters for each phase were prepared according to a detailed template that we developed in consultation with country participants. In this introduction, we summarize the collective results of the country analyses and borrow freely from the country chapters. In large part, however, the results presented in the introduction could only be conveyed by combined analysis of the data from each of the countries. The country chapters themselves present much more detail for each country and, in addition to the common analyses performed by all countries, often present country-­specific analysis relevant to a particular

4     Courtney Coile, Kevin Milligan, and David A. Wise

Fig. I.1  Proportion of men age sixty to sixty-­four receiving DI benefits in 2009, by country Note: The data for Belgium and Italy pertain to the number of DI participants divided by the number of active wage earners plus the number of DI participants (rather than the population ages sixty to sixty-­four). Data for Germany are for ages fifty-­five to fifty-­nine. Data for France are for 2007 and for ages fifty-­five to fifty-­nine and pertain to inactivity due to health reasons. Data for Italy are for 2004. The value for Japan is an estimate.

country. In addition, the country chapters typically present results separately for both men and women. As we have noted in our past work, the share of the population receiving disability benefits at older ages varies substantially across countries. ­Figure I.1 shows the share of men ages sixty to sixty-­four collecting DI benefits by country in 2009. This value varies by a factor of eight within the participant countries, from 17 percent in Belgium to 16 percent in the United Kingdom, 14 percent in the United States, 6 percent in Italy and France, and 2 percent in Japan. (It is important to note that the data for Belgium and Italy pertain to the number of DI participants divided by the number of active wage earners plus the number of DI participants, rather than the population age sixty to sixty-­four. This same caveat applies to figures I.2, I.6, I.7, I.8, and I.9.) It seems unlikely that differences of this magnitude would be driven exclusively, or even primarily, by differences in the health status of the population across countries. In the introduction to the prior phase of the project (Milligan and Wise 2012), we grouped countries according to the share of men collecting disability benefits at age forty-­five, which was 2 to 3 percent in one set of countries and 5 to 6 percent in another. By age sixty-­ four, both groups of countries were exhibiting large differences in the share of men collecting DI (or similar) benefits—­among countries with the lower rates of DI usage at age forty-­five, for example, participation at age sixty-­four ranged from less than 10 percent to over 35 percent. The emergence of these

Introduction    5

vast differences in the use of DI at older ages among countries with similar rates of disability in middle age strongly suggests that DI usage depends on factors other than health. These statistics also indicate that the DI program serves as a source of retirement income before the social security eligibility age for a sizable share of the population in some countries. It is these observations that lead us to seek a better understanding of how financial incentives from DI programs affect labor supply. This introduction is organized in several sections. The first section presents background information on DI participation, including changes over time, participation gradients by education and health status, and other relevant statistics. The second section explains the Poterba, Venti, and Wise (PVW) index of health that is used throughout the analysis. The third section explains the estimation procedure that is followed. The last section discusses the simulations based on the estimation results. While the simulations in the second phase of the project emphasized the implications of increasing program eligibility ages, the simulations here emphasize employment (­ retirement) effects of incentives inherent in the provisions of the country retirement plans, particularly of changing the accessibility of the DI program. Background Trends in DI Participation: We begin by documenting changes in DI participation over time. Figure I.2 shows the DI participation rate for men ages sixty to sixty-­four by country for selected years from 1970 through 2012 (years of data available for each country vary; data for France and Germany is for ages fifty-­five to fifty-­nine). Disability insurance participation is not shown for Japan, where DI participation has been extremely low. Similar figures in the individual country chapters show results for men ages fifty to fifty-­four and fifty-­five to fifty-­nine; for women trends in these other groups are often similar to those shown here, though participation levels are lower at younger ages. Perhaps the most striking feature of these data is the sharp decline in the DI participation rate for older men in many European countries beginning between the late 1980s and the mid-­1990s. In five countries—­most striking in Sweden, Canada, and the United Kingdom, but also in Italy and ­Germany—­an inverted U-­shaped pattern is evident, with DI participation rising until the mid-­1990s and falling sharply thereafter. The DI participation rate reached 36 percent in Sweden and 27 percent in the United Kingdom before dropping by 53 and 50 percent respectively over the next fifteen to twenty years. The drop was 50 percent from the peak in Canada, 41 percent in Germany, and 15 percent in Italy. In the Netherlands, Denmark, and Belgium there was also a large decline after the late 1980s, ranging from 32 to 45 percent. In these three countries the time series begins too late to see the rise, but the fall in DI participation is quite evident.

Fig. I.2  Share of men age sixty to sixty-­four on DI (fifty-­five to fifty-­nine in Germany and France), for selected years

Introduction    7

Fig. I.2  (cont.)

In the remaining countries, the pattern is different. In the United States, the DI participation rate for men ages sixty to sixty-­four rose from 4.7 to 13.6 percent between 1960 and 1980 and then fell by 3 percentage points during the 1980s from 13.6 to 10.4 percent. Since that time, while DI participation in many European countries has fallen dramatically, the DI participation rate in the United States increased by 30 percent in a trend that shows no signs of stopping. Spain, too, has experienced an increase in the DI participation rate over the past two decades. In France the trend in DI participation between 1990 and 2007 is unclear, although there was a decline in DI participation in the last years of available data. The changes are summarized in table I.1. The countries are ordered by the decline in the percent on DI with the greatest decline in Sweden and the greatest increase in the United States. As we discuss subsequently, the dramatic changes in the DI participation rate over time experienced by many countries cannot be explained by changes in health. This feature of the data is documented in substantial detail in the previous phase of the project—­the individual country chapters in that volume (Wise 2012) and the introduction to that volume (Milligan and Wise 2012). The rapid changes in the level of DI participation that can be seen in figure I.2 are often associated with reforms in the DI program or in other government programs and are also documented in the prior phase of the project. In addition to looking at the DI participation rate in isolation, it is instruc-

8     Courtney Coile, Kevin Milligan, and David A. Wise Table I.1

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

Change in percent of men on DI from most recent maximum or minimum to year of most recent data (by country) Year of most recent minimum (or maximum)

Year of most recent data

1993 1995 1996 1994 1993 1996 1987 2004 2000 1988 1990

2012 2009 2012 2010 2008 2009 2010 2007 2004 2012 2012

DI percent in these years 0.360 0.139 0.272 0.219 0.212 0.196 0.255 0.074 0.069 0.102 0.109

0.170 0.070 0.137 0.121 0.123 0.115 0.174 0.059 0.059 0.120 0.142

Percent change between years –52.8 –49.6 –49.6 –44.7 –42.0 –41.3 –31.8 –20.3 –14.5 17.6 30.3

Fig. I.3  Pathways to retirement for men in Germany

tive to consider how the use of different benefit programs as pathways from employment to retirement has changed over time. Figure I.3 provides this information for German men. As the figure makes evident, the proportion of men retiring by way of DI fluctuated widely between 1960 and 2012. For example, the proportion retiring through the two DI programs (for work-

Introduction    9

Fig. I.4  Decline in mortality at age sixty-­five, by country

ers under and over age sixty, respectively) ranged from a high of 68 percent in 1981 to a low of 28 percent in 2005—­a decline of over 58 percent—­and then increased by over 14 percent by 2012. This figure also shows that the decline in retirement through DI coincided with an increase in retirement through a special unemployment insurance program for older workers. The decline in the sum of DI plus unemployment insurance (UI) programs between 1981 and 1999 was a more modest 33 percent. This example suggests that government programs may substitute for one another—­a decline in participation in one program may be offset by an increase in participation in another program and may not necessarily be associated with an equal increase in labor supply. Therefore it is important to take a holistic view and model the incentives arising from all programs that are potential sources of (early) retirement income, as we aim to do in the analysis that follows. Trends in DI Participation versus Trends in Health: In the prior phase of the project (Wise 2012), we emphasized the absence of a relationship between DI participation and health, as measured by mortality. Figure I.4, taken from the introduction to this earlier study (Milligan and Wise 2012), shows the decline in mortality at age sixty-­five between 1970 and the early twenty-­ first century for our twelve participating countries. Mortality declined in all of the countries over this period, generally in a similar way. Yet as shown above in figure I.2, DI participation fluctuated widely over the same time period. The juxtaposition of these trends casts doubt on the possibility that

10     Courtney Coile, Kevin Milligan, and David A. Wise

Fig. I.5  Percent change in DI participation versus percent change in mortality, early 1980 to 2005 (men)

changes in DI participation within countries over time are driven by changes in health, at least as measured by mortality. This point is made more directly in figure I.5, also from Milligan and Wise (2012), which plots the change in mortality and the change in DI participation between 1980 and 2005 for the twelve participating countries and finds little evidence of a relationship between them. Trends in DI Participation versus Trends in Employment: While there is little evidence that changes in health are associated with changes in DI participation, we anticipate that changes in DI participation are associated with changes in employment at older ages. Here we explore the relationship over time by plotting the evolution of DI participation and employment rates at older ages within each country over time. A central goal of this phase of the project is to explore the relationship between DI programs and labor force participation through microeconomic analysis, as discussed below. The time-­series data here helps to provide motivation for the formal analysis to follow. The relationship between DI participation and employment in the participating countries is presented in figure I.6. In this figure the left axis measures employment and the right axis measures DI participation. As discussed above with respect to figure I.2, the DI participation rate for older men follows an inverted U-­shaped pattern in a number of countries, rising until the early-­to-­mid-­1990s and then falling, while several additional countries (for whom earlier data was not available) also have a decline in DI participation over the past several decades. The new insight from figure I.6 is that there is

Fig. I.6  Employment and DI rates for men, by country, for the age interval sixty to sixty-­four (except where noted)

12     Courtney Coile, Kevin Milligan, and David A. Wise

Fig. I.6  (cont.)

an inverse relationship between the DI participation and employment rates in virtually all of these countries. Specifically, in Canada, Denmark, Italy, the Netherlands, Sweden, and the United Kingdom, the relationship is quite clear; as DI participation increases the employment rate falls and as DI participation declines employment increases. The relationship is especially striking in Sweden, Canada, the United Kingdom, and Italy where the peak in DI participation (with a sharp increase and a sharp fall after the peak) is mirrored by a reverse relationship for employment. A similar relationship is also shown for Germany, but with greater fluctuation in the employment and DI trends over time. In the United States, the story is more complex. For men age sixty to sixty-­ four, the inverse relationship is evident in the 1970s, but over the past two decades both employment and DI participation have been rising. However, for US men age fifty to fifty-­four—­the ages at which a large number of men first receive DI benefits—­the inverse relationship is clear. A similar relationship (not shown) holds for the fifty-­five to fifty-­nine age groups in the United States. In three additional countries—­Belgium, Germany, and Spain—­the data are too noisy or the time series too brief to draw strong conclusions, although the data suggest a negative relationship at the beginning and at the end of the time period for which data are available in Belgium, at the end of the period in Germany, and perhaps at the end of the period in Spain. Nonetheless, the fact that we observe that employment moves in the opposite

Introduction    13

Fig. I.7  Share of men age fifty-­five to sixty-­four on DI in 2010, by health quintile Note: The data are from various years, 2008–2011, depending on the availability for each country. Data for Belgium, Denmark, Italy, the Netherlands, and Sweden are for ages fifty to sixty-­four. Data for Germany are for ages fifty to fifty-­nine.

direction of DI participation in most countries, in periods of both rising and falling DI participation and with the peak in DI participation lining up with the trough in employment in several cases, suggests a noticeable relationship between the two series. Health and DI Participation: Having explored how DI participation varies across countries over time, and with changes in health and employment over time, we next consider how DI participation varies by health quintile. The description of how the health quintiles are constructed is deferred to the second section of this chapter. The results are shown in figure I.7 for men age fifty-­five to sixty-­four. In all countries, there is a substantial DI gradient with respect to health, with those in the lowest health quintile dramatically more likely to be on DI than those in the middle or highest health quintile. This finding is of course consistent with the intended purpose of DI programs to provide income support to individuals with reduced work capacity. The figure also shows, however, that for people with similar levels of health (for example, those in the lowest health quintile in their own country), there are large differences across countries in the probability of being on DI. In the United Kingdom, nearly half of older men in the lowest health quintile are on DI, versus about one-­quarter of Danish men and one-­tenth of Japanese men in the lowest quintile. Among countries with similar rates of DI in the lowest health quintile—­such as the United States, Spain, and Sweden—­the share

14     Courtney Coile, Kevin Milligan, and David A. Wise

Fig. I.8  Share of men age fifty-­five to sixty-­four on DI in 2010, by education Note: Data are from various years, 2008–2011, depending on availability for each country. Data for Belgium, Denmark, Italy, the Netherlands, and Sweden are for ages fifty to sixty-­four. Data for Germany are for ages fifty to fifty-­nine. Low and high education groups are defined differently across countries.

of men in the middle health quintile who are on DI ranges from 3 percent in the United States to 13 percent in Sweden. Education and DI Participation: One feature of DI that may not be widely understood is the strong relationship between DI participation and education. Figure I.8 shows the share of men at ages fifty-­five to sixty-­four who are on DI by level of education across countries; the values for the highest and lowest education groups are shown on the graph, although definition of high and low varies across countries. In Denmark, Italy, the United States, and the United Kingdom, those individuals in the lowest education group are at least five times as likely to be receiving DI benefits as those in the highest education group. In other countries, the ratio of probabilities is somewhat lower, but still greater than two in every country. Differences in rates of DI participation by education group may reflect the fact that less educated individuals on average are in poorer health than those with more education—­a possibility that we explore in more detail below—­but likely also reflect economic circumstances such as weaker job prospects or higher replacement DI rates for workers with low lifetime earnings in systems with progressive benefit formulas. DI Participation by Education and Health: We return to the question of whether differences in DI participation by education are primarily due to health differences by calculating DI participation by health and education for those countries with large enough sample sizes to do so. Figure I.9 shows

Fig. I.9  Share of men age fifty-­five to sixty-­four on DI by health and education (by country) Note: Data for each country are generally pooled across multiple years to increase sample size and precision. Data for Belgium, Denmark, Italy, Netherlands, Spain, and Sweden are for ages fifty to sixty-­four. Education groups are defined differently in different countries.

16     Courtney Coile, Kevin Milligan, and David A. Wise

Fig. I.9  (cont.)

the participation percent by education for each health quintile in ten of the participant countries. In the lowest health quintile in the United States, 50 percent of persons with less than a high school degree are DI participants versus only 34 percent of those with a college degree. For those in the third health quintile, participation rates among college graduates and high school dropouts are 6 percent and 2 percent, respectively. In the United Kingdom, there are even larger differences by education in DI use by men in the same health quintile. In the lowest quintile, those in the low education group are over twice as likely to be on DI as those in the high education group (53 versus 22 percent); this is also true in the second quintile (23 versus 4 percent) and third quintile (6 versus 2.5 percent). A similar pattern is evident in the other countries, with Denmark and Sweden having particularly steep gradients, like the United Kingdom, and other countries reflecting gradients more similar to those in the United States. From these figures, we conclude that differences in DI use by education group are not due exclusively to differences in health. Rather, it appears that there are other factors such as differential labor market prospects or earnings potential that may explain the large differences in DI participation by education, conditional on health. Employment by Health and by Education: Finally, we explore the relationship between employment and health and employment and education, which are likely to vary across countries depending on the provisions of each country’s DI program. Employment rates by health quintiles are plotted for Denmark and Germany only—­for other countries the data necessary to compute an equivalent time series are not available. Figure I.10 shows that there are very significant differences across health quintiles in the probability that older men are employed. Although employment rates are higher at every level of health in Germany, the difference between the employment rates of those in the lowest and highest health quintiles is roughly the same in both countries, 20 to 25 percentage points. Figure I.11 presents employment rates at ages fifty-­five to sixty-­four by level of education, country, and year. This figure shows that there are very

Introduction    17

Fig. I.10  Employment by health quintile, for men age fifty-­five to sixty-­four (by country and year) Note: Data for Denmark are for ages fifty to sixty-­four.

large differences in employment by education. In most countries, the difference in employment between the highest and lowest education groups (where the definition of these groups varies by country) is at least 20 percentage points. Notably, these differences are of a similar magnitude to those seen across health quintiles in figure I.10. Thus education is strongly related to both DI participation and to employment at older ages, consistent with a causal link between employment and application for DI. Measuring Health Health is a central component of the analysis. Here we explain briefly the measure that is used and a key property of the measure. To maintain as much comparability across countries as possible, we use a health index developed by Poterba, Venti, and Wise (PVW) that has previously been used in several contexts (see, e.g., Poterba, Venti, and Wise 2013). The index, as set out by PVW, is the first principal component of twenty-­seven health indicators reported in the US Health and Retirement Study (HRS). Much of the analysis reported in this volume makes use of a nexus of comparable studies—­the English Longitudinal Study of Aging (ELSA), the Japan Study of Aging and Retirement (JSTAR), and the Survey of Health, Ageing and Retirement in Europe (SHARE), which includes eight of our participant countries: Belgium, Denmark, France, Germany, Italy, the Netherlands, Spain, and the United Kingdom. The similarity of these studies allows us to apply the PVW methodology across countries. To be more specific, in the current project we use a slightly modified version of the PVW index based on twenty-­five indicators that are common to the HRS and to all of the SHARE countries. Japan and the United Kingdom lack data on several of the indicators, so they use the same methodology with the remaining indicators. There are four countries that do not employ the PVW method in constructing health measures for their analysis. One is

Fig. I.11  Employment by education level, men age fifty-­five to sixty-­four (by country and year) Note: Data for Belgium and for Spain are for the age interval sixty to sixty-­four.

Introduction    19

Fig. I.11  (cont.)

Canada, which lacks detailed data on health in any survey that would meet the other requirements of this project and thus uses a simplified health measure (see country chapter for details). The others are Sweden, Denmark, and Germany, who have chosen to use non-­SHARE data to obtain a larger sample size for their analyses. For these four countries, therefore, the comparable health measure cannot be used. Nonetheless the comparable health measures for all SHARE countries are included in the discussion below. The health measures in non-­health-­index-­countries are not comparable to the index health measure. Also, in some countries, the precise index used in a country may differ slightly from the index used in this discussion of the properties of the index. The health measures and the weights (loadings) given to each measure in the index for each country (except Canada) are shown in table I.2A. Comparison of the weights across countries reveals striking consistency among the countries. That is, the ranking of the weights is very similar from one country to the next. This is especially apparent for the United States, the eight SHARE countries, and for the United Kingdom (based on ELSA data). Table I.2B shows the correlation of the weights for each pair of countries. All but two of the thirty-­two pairwise correlations for the United States and the SHARE countries are 0.95 or greater, and many are 0.97 or greater. Correlations between the rankings for the United Kingdom and each of the other countries and the ranking for Japan and each of the other countries are shown on the right-­hand side of the table. These correlations are based on the weights for the health indicators that are common to each country. For example, the correlations for Japan are based on the twenty-­two indicators that are common to the United States, the SHARE countries, and Japan. The correlations for the United Kingdom are based on the twenty variables that are common to the United Kingdom, the United States, and the SHARE countries. The pairwise correlations between the United Kingdom

20     Courtney Coile, Kevin Milligan, and David A. Wise

and the other countries for this smaller set of questions are 0.95 or greater for all countries except Japan, with a correlation of 0.92. In general, the correlations between Japan and the other countries are between 0.88 and 0.93 with one exception. When the “exact same” questions are used in each of the countries, the pairwise correlations are close to 1—­between .98 and .99—­for all of the countries except the pairwise correlations with Japan. The high correlations between the country loadings indicate that the relationships among the many health indicator responses are very similar across countries. For ease of analysis the index measures for each country are converted to percentile scores, with 1 the lowest and 100 the highest. For many comparisons the percentile scores are grouped into five quintiles. Many figures based on these quintiles are shown in the background section above. An important feature of the index is the strong correspondence to survival. For example, based on ELSA data in the United Kingdom, given the health index quintile in 2002 the survival rate in 2011 for persons in the lowest quintile is 59.7 percent, it is 72.6 percent in the second quintile, 81.9 percent in the third, 88.9 percent in the fourth, and 93.9 percent in the highest quintile. Based on HRS data in the United States, given the health index decile in 1992, the survival rate in 2008 ranges from 42.8 percent for those in the bottom decile to 71.4 percent for those in the fifth decile to 89.6 percent for those in the top decile. In the United States, the index in 1992 is also strongly related to future health events such as diabetes, lung cancer, health disease, stroke, hospital stay in 2008, and poor health in 2008 (Poterba, Venti, and Wise 2013). The following example points to the value of a health measure that can be constructed in a comparable way across countries, and provides some added support to the idea that the resulting health index values are reasonable. In figure I.12, we report the PVW health index by age and country, as measured relative to the US value. At ages fifty to fifty-­four, the health of women in the United States is worse than the health of women in most other countries. This finding continues at least through ages sixty to sixty-­four, but by the mid-­1970s, health in the United States is better than in all countries (with the exception of the United Kingdom). This finding is consistent with the conclusion of many analysts that health in the United States improves after Medicare eligibility at age sixty-­five and that expenditure on health care for the oldest old is relatively higher in the United States than in other countries. For men, shown in figure I.13, the general trend is similar but not as dramatic. Estimation A central goal of the analysis in this phase of the project is to estimate the relationship between the provisions of each country’s retirement programs and the labor supply (or retirement) behavior of older workers in

5,424

155,595

N

5,615

0.271 0.300 0.296 0.279 0.318 0.292 0.287 0.259 0.215 0.169 0.204 0.173 0.176 0.153 0.154 0.177 0.168 0.064 0.109 0.094 0.088 0.083 0.062 0.073 0.060

Sweden

5,431

0.270 0.297 0.299 0.260 0.297 0.309 0.275 0.284 0.202 0.176 0.210 0.124 0.195 0.106 0.144 0.221 0.190 0.059 0.105 0.095 0.090 0.082 0.080 0.085 0.061

Netherlands

4,198

0.264 0.298 0.292 0.258 0.303 0.301 0.291 0.265 0.236 0.196 0.213 0.169 0.196 0.122 0.108 0.160 0.184 0.114 0.098 0.108 0.097 0.094 0.079 0.020 0.035

Spain

5,416

0.288 0.292 0.283 0.273 0.284 0.288 0.273 0.259 0.241 0.199 0.200 0.193 0.182 0.142 0.132 0.134 0.203 0.080 0.127 0.121 0.119 0.110 0.065 –0.002 0.043

Italy

5,844

0.281 0.284 0.289 0.272 0.296 0.304 0.265 0.279 0.227 0.185 0.178 0.152 0.161 0.162 0.126 0.211 0.200 0.067 0.124 0.110 0.105 0.091 0.092 0.024 0.038

France

4,132

0.265 0.302 0.281 0.275 0.313 0.294 0.274 0.299 0.184 0.189 0.211 0.157 0.186 0.123 0.135 0.193 0.183 0.079 0.120 0.084 0.132 0.067 0.059 0.057 0.050

Denmark

6,739

0.280 0.294 0.297 0.279 0.288 0.289 0.264 0.276 0.192 0.201 0.186 0.137 0.177 0.159 0.132 0.204 0.236 0.087 0.114 0.087 0.097 0.089 0.071 0.024 0.061

Belgium

42,352

0.321 0.312 0.309 0.302 0.298 0.290 0.282 0.258 0.223 0.216 0.228 0.174 n/a 0.137 n/a 0.199 n/a 0.062 0.108 0.147 0.109 0.085 n/a n/a 0.044

United Kingdom

0.035

0.082 0.017 0.126 0.075 0.040 0.071 0.026

0.094 0.109

0.311 0.337 0.340 0.242 0.315 0.309 0.304 0.211 0.269 0.122 0.277 0.248

Japan

Note: The HRS values are based on data for all HRS cohorts for waves 1992 to 2008. The SHARE values are based on data for 2004 and 2006. The Japan index is based on pooled data from the first and second waves of JSTAR. The United Kingdom values are based on pooled data from 2002, 2004, 2006, and 2008. The precise index used in each country may differ slightly from the indices used here, which are based on the same health measures in each country, with the exception of the United Kingdom and Japan.

0.276 0.271 0.277 0.275 0.289 0.293 0.285 0.285 0.244 0.153 0.218 0.169 0.180 0.129 0.152 0.143 0.208 0.090 0.125 0.121 0.085 0.114 0.077 0.042 0.076

0.307 0.293 0.288 0.281 0.276 0.275 0.266 0.262 0.224 0.197 0.194 0.164 0.162 0.156 0.154 0.152 0.146 0.137 0.132 0.129 0.123 0.114 0.072 0.070 0.060

Difficulty walking sev. blocks Difficulty lift/carry Difficulty push/pull Difficulty with an ADL Difficulty climbing stairs Difficulty stoop/kneel/crouch Difficulty getting up from chair Self-­reported health fair or poor Difficulty reach/extend arms up Ever experience arthritis Difficulty sitting two hours Difficulty picking up a dime Back problems Ever experience heart problems Hospital stay Home care Doctor visit Ever experience psychological Ever experience stroke Ever experience high blood pressure Ever experience lung disease Ever experience diabetes BMI at beginning of period Nursing home stay Ever experience cancer

Germany

HRS

The PVW first principal component index for the United States (HRS) and SHARE countries

Question

Table I.2A

22     Courtney Coile, Kevin Milligan, and David A. Wise

Table I.2B

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

Correlations or principal component loadings for each pair of countries HRS

Germany

Sweden

Netherlands

Spain

Italy

France

Denmark

Belgium

1

0.951 1

0.961 0.968 1

0.925 0.949 0.973 1

0.961 0.972 0.966 0.952 1

0.962 0.974 0.955 0.919 0.978 1

0.949 0.959 0.969 0.964 0.966 0.965 1

0.949 0.952 0.977 0.978 0.966 0.949 0.968 1

0.939 0.953 0.961 0.964 0.961 0.956 0.984 0.971 1

United Kingdom 0.970 0.950 0.970 0.960 0.960 0.970 0.970 0.960 0.970 1

Fig. I.12  The PVW health index by country and age relative to the United States, for women

that country. The analysis in this phase of the project is closely related to the analysis in the second phase.1 Here, however, we give particular attention to the provisions of DI programs, as well as other pathways to retirement. More specifically, we want to understand how changing the provisions of a country’s DI program (and perhaps other programs) would affect retirement. To explore this, we first need to construct a retirement incentive measure that reflects how the provisions of a country’s social security, DI, and other relevant programs provide a greater or lesser return to continued work at a given age for each worker. Next, we assess whether this incentive mea1. See, in particular, the discussion on pages 10–15 of Gruber and Wise (2004).

Japan 0.900 0.910 0.930 0.900 0.910 0.920 0.880 0.880 0.850 0.920 1

Introduction    23

Fig. I.13  The PVW health index by country and age relative to the United States, for men

sure is empirically related to retirement behavior. Finally, we use the results of this estimation to simulate how a change to a country’s DI program (and the resulting change in the retirement incentive measure) would be expected to affect retirement. The key idea that underlies our analysis is the potential gain from postponing retirement from today’s age until some future age. This is the incentive to delay retirement. We first explain this incentive measure, assuming that there is only one pathway to retirement. We then explain the issues that arise when there are multiple pathways to retirement (e.g., social security and DI). We then discuss the other covariates included in the country retirement specifications. As the discussion below and in the country chapters makes clear, workers may face very different incentives for continued work depending on the provisions of retirement programs in their country as well as on individual characteristics such as potential earnings, earnings history, family structure, and other attributes. Retirement Incentive and the Option Value: To begin, assume that there is only one retirement program, social security. When a person retires he (or she) will receive a stream of benefits until death. If the person retires at age t, the present discounted value of benefits, or social security wealth, is given by SSWt . If the person retires one year later, the present discounted value of future benefits will be SSWt+1. The social security accrual from one year to the next is given by SSWt–1 – SSWt .

24     Courtney Coile, Kevin Milligan, and David A. Wise

That is, this measure describes the change in promised future social security benefits from working one additional year. Social security wealth will go up if an extra year of work is translated into a higher flow of benefits in the future, either because of the relationship between social security and lifetime earnings or because of actuarial adjustments that reward later retirement. Social security wealth may go down, though, if the extra benefits that accrue from the extra work are not large enough to compensate for the loss of any retirement benefits in that extra year of work. The net of the future extra benefit entitlement and the loss of benefits in that extra year of work is the one-­year accrual. One shortcoming of the accrual as a measure of retirement incentives is that there could be greater increases in social security wealth from delaying retirement by two years, three years, or more rather than by a single year; beyond some ages benefits may decline—­depending on the benefit formula in a given country. The gains associated with work beyond the current year will not be captured by this simple measure. Thus to fully appreciate the incentives inherent in the social security program, we must consider the path of benefits many years into the future. The benchmark approach we use for considering the entire future path of accruals is the “option value” (OV) 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 of retiring at the optimal date (which may be today, but more likely is 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. A simplified version of the option value measure at age t can be described by: Simplified OVt (r*) =

discounted discounted benefits if − benefits if retire at t retire at r*

+

discounted future wages . through age r*

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 age. 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 last bracket and the gain in pension benefits by the difference between the terms in the first 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. More detail on the option value specification is shown in the appendix on the option value model. 2. For a more detailed discussion, see Stock and Wise (1990).

Introduction    25

Multiple Pathways to Retirement: The discussion above assumes that there is only one pathway to retirement, but in all countries there are multiple pathways. In the United States there are two pathways—­Social Security and disability insurance (DI)—­but in other countries there are three or more pathways—­the social security “normal” retirement, DI, special unemployment insurance programs, or a special early retirement program. To estimate the OV incentive on retirement with multiple programs, we follow an instrumental variables-­like approach. For each program, we first estimate the OV measure for that program, essentially assuming that the worker will retire through that program and the only decision is at what age to retire. Next, we estimate the probability that the person assigns to each program as a possible pathway to retirement. Finally, we calculate the “inclusive OV,” which is the weighted average of the OVs for each of the possible programs. The probabilities to be assigned to each program are determined by the relationship between individual attributes and the likelihood that a particular program was chosen by similar workers in the past. For example, in the United States, the probability weight for the DI plan is determined by the probability that a person in each of four education levels was on DI anytime at ages sixty and sixty-­four in the relevant year (estimated using HRS data for the years 1992 to 2010). The exact method used for each country is described in the country chapters. This approach is an “instrumental variable” estimate of the expected OV faced by a given person. Figure I.14 shows the OVs by age for each country. The OV calculations are based on the detail in the appendix. For illustration, consider the programs in the United States and in Belgium. The United States has only two programs, DI and Social Security (SS). Belgium has four programs—­ Social Security, DI, unemployment insurance (UI), and early retirement (CER). Notice that in the United States, the OV of delaying retirement is much larger under the SS program than under the DI program. That is, the gain from delaying retirement is much greater under the SS program. Thus persons who consider the DI program as a route to retirement have a much greater incentive to retire at a young age than persons who consider SS as the only pathway to retirement. The inclusive OV is the weighted average of the SS and DI OVs. In the United States, the average DI weight is small so the inclusive OV is close to the SS OV. The OVs in Belgium are quite different. First note that the program OVs in general are much lower in Belgium than in the United States. Second, note that the inclusive OV is much lower in Belgium than in the United States. At age fifty, for example, the inclusive OV in the United States is about 33,200 but is only about 12,500 in Belgium. Thus it would appear that the average gain to delaying retirement is much less in Belgium than in the United States. It is important to understand that the estimated effect of the inclusive OV on retirement—­thought of as an instrumental variable estimate of the OV effect on retirement—­is taken as the effect of the OV on retirement

26     Courtney Coile, Kevin Milligan, and David A. Wise

Fig. I.14  Option values and inclusive OV by age, by country

and used in all estimates of the effects of program provisions and changes in program provisions on retirement. For example, this estimate is used to predict (simulate) the effect on retirement of having access only to the DI versus access only to the social security program. Estimation and Additional Covariates: Although the inclusive OV incentive measure is the key variable in the estimation, other individual attributes

Introduction    27

Fig. I.14  (cont.)

are also included. First recall that the OV depends on estimated individual earnings as described above. In addition, the specification for each country includes health—­typically controlling for health quintile based on the PVW index described above. One might expect health to be particularly important when contemplating retirement under the DI pathway. The specification also includes education level, gender, whether the person is married, whether the spouse works, total non-­social security assets, and occupation indicator variables. There is some variation depending on data availability in each country. Finally, each specification controls for age. Two versions are included. One includes indicator variables for each age, and the other includes age as a single continuous variable. The inclusion of age is particularly important when evaluating the effect of the OV on retirement. Quoting from the introduction to phase 2 of the project: 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. (Gruber and Wise 2004, 12) We return to this issue when discussing simulations below. Parameter Estimates: For each country, estimates are reported for several

28     Courtney Coile, Kevin Milligan, and David A. Wise

alternative specifications. For example, in some specifications separate indicator variables are included for each age; in others, a single linear age effect is included. In some specifications health quintile indicators are included, and in other specifications a single variable for health percentile is included. For some countries the sample sizes are large enough to obtain separate estimates by health quintile and by education level; in others, including most of the SHARE countries (if the SHARE data are used), the sample sizes are not large enough to estimate separate parameters by health quintile or by education. The most important coefficient is the estimated effect of the inclusive OV on the probability of retirement. The country estimates of this retirement incentive effect are sensitive to the differential variance in the OVs across countries. To account for this, the estimated effect of a one standard deviation change in the OV is reported in square brackets as well as the effect of a unit (10,000 “utility” units) change in the OV. In addition, in some specifications the percent gain in the OV from delaying retirement is estimated instead of the OV itself—­the percent gain from delaying retirement at age a is measured by the OV of delaying retirement at age a divided by the utility associated with retirement at age a. Like the standard deviation of the OVs, this measure may be more comparable across countries than OV units and thus help to make the results more comparable across countries. Estimates for each of the countries are reported in table I.3. Estimates are reported for two specifications. The first is the fourth specification in the first table of estimates presented in each of the country chapters. The second is the effect of the percent gain in the OV from delaying retirement. Several features of the estimates stand out. First, the estimated option value incentive measure is highly statistically significant in each of the countries, with the exception of Spain and Germany (using SHARE data). In these countries the sample sizes are apparently too small to obtain statistically significant results. The German estimates based on the much larger Socio-­ Economic Panel Survey (SOEP) data file are highly significant. Second, there is considerable variation across countries in the estimated effects. Even excluding the statistically insignificant estimates for two countries and the smallest estimates for the United Kingdom and Sweden, the estimated effects for the remaining countries vary by a factor of seven. In two countries the estimated effect of a unit (10,000) increase in the OV is to reduce the retirement rate by about 11 percent or more. In five countries the effect on retirement is between 3 and 5 percent. In the United Kingdom and Sweden the estimated effect is less than 1 percent. The estimated effect of a standard deviation change in the incentive measure also varies across countries, but less than the unit increase estimate. In eight countries these estimates are between 4 and 9 percent. In the remaining three countries with statistically significant estimates the values are between 1 and 3 percent. Third, in most countries there is very little difference in the estimated effect of the incentive

–0.006 (–0.001) [–0.028] –0.148 (0.022)

–0.0166 (0.0021) [–0.041] –0.0451 (0.0108)

–0.0315 (–0.0005)

–0.049 (0.020) [–0.023]

Italy

United Kingdom

Canada

Specification

(1): Specification (4), with age dummies Estimate Standard error Effect of OV std. change (2): Percent gain: Specification (4), with age dummies Estimate Standard error

–0.038 (–0.012)

–0.313 (–0.082)

–0.060 (–0.072)

–0.046 (–0.006) [–0.042]

France

–0.106 (–0.033) [–0.079]

Belgium

–0.119 (–0.049) [–0.091]

Netherlands

— —

–0.0015 (0.000) [–0.0126]

Sweden

–0.0186 (–0.0016)

–0.0423 (–0.0031) [–0.0525]

Germany

The effect of the retirement program incentive effect, inclusive OV, on retirement, by specification

(1): Specification (4), with age dummies Estimate Standard error Effect of OV std. change (2): Percent gain: Specification (4), with age dummies Estimate Standard error

Specification

Table I.3

–0.036 (–0.046)

–0.005 (–0.017) [–0.004]

Spain

–0.0806 (0.0015)

–0.0433 (0.0005) [–0.0438]*

Denmark

— —

–0.020 (–0.015) [—]

Germany (SHARE)

–0.0593 (.0124)

–0.0331 (.0011) [–0.056]

United States

–0.0384 (0.0122)

–0.0217 (0.006) [–0.045]

Japan

30     Courtney Coile, Kevin Milligan, and David A. Wise Table I.4

Estimated incentive measure effects by health quintile for selected countries

OV: Worst health quintile Standard error Effect of OV std. change OV: Second quintile Standard error Effect of OV std. change OV: Third quintile Standard error Effect of OV std. change OV: Fourth quintile Standard error Effect of OV std. change OV: Best health quintile Standard error Effect of OV std. change

United States

United Kingdom

Germany (SOEP)

Denmark

Sweden

–0.0594 (0.0038) [–0.073] –0.0353 (.0026) [–0.052] –0.0336 (0.0023) [–0.056] –0.0234 (.0018) [–0.044] –0.0197 (.0017) [–0.037]

–0.008 (0.002) [–0.062] –0.006 (–0.002) [–0.040] –0.003 –0.002 [–0.030] –0.005 (–0.002) [–0.050] –0.007 (–0.002) [–0.081]

0.0902 (0.0105) [–0.0707] –0.0453 0.0067 [–0.0576] –0.0285 –0.0043 [–0.0628] –0.0195 (–0.005) [–0.0628] –0.0219 (–0.005) [–0.0320]

–0.0639 (0.0015) [0.065] –0.0490 (0.0014) [0.0285] –0.0342 (0.0011) [0.0256] –0.0282 (0.0009) [0.0186] –0.0372 (0.0010) [0.0283]

–0.0022 (0.0001) [–0.0145] –0.0018 (0.0000) [–0.0142] –0.0013 (0.0000) [–0.0118] –0.001 (0.0000) [–0.0098] –0.0009 (0.0000) [–0.0097]

Notes: Germany (SOEP), Denmark, and Sweden do not use the PVW health index so that health comparability across all of the countries is not assured, although in each country the available measures can be used to rank persons by health.

measure in the specification with age indicators compared to the otherwise identical specification but with a single linear age measure—­these estimates can be seen in the country chapters. Finally, the estimated effects of other covariates vary substantially from one country to the other and many of the estimated effects are not statistically different from zero. The many estimates based on several additional specifications are shown in the country chapters. Although it is clear that persons in poor health are more likely to retire early through the DI pathway, whether the effect of the incentive measure on retirement should vary in one direction or another with health is not clear a priori. Some evidence, however, is provided in the country data. Table I.4 shows the estimated incentive measure effect by health quintile for several countries with sample sizes large enough to distinguish estimates by health. In four of the five countries the estimated effect of the incentive measure declines with health. In the United States the effect declines continuously from –0.0594 for those in the worst health to –0.0197 for those in the best health, in Germany from –0.0902 to –0.0219, in Denmark from –0.639 to –0.0373, and in Sweden from –0.0022 to –0.0009. In each of these countries the result is also shown clearly by comparing the effect of a standard deviation change in the OV for those in the best versus those in the worst health, shown by the estimates in the square brackets. The United Kingdom is an exception, showing essentially no relationship between the incentive measure and health. Recall that the health measures used in Germany, Denmark, and Sweden are based on the few selected health measures in the data files

Introduction    31

used in those countries and are not comparable to the PVW index measure used the United States and the United Kingdom. Nonetheless, the health measures used in the other three countries can be used to rank persons by health quintile. Note that the relationship between the incentive measure and health should not necessarily be expected to be the same in all countries. For example health is the central criteria for eligibility for DI in the United States, while the relationship may be less strict in other countries that may give more weight to labor market conditions, for example, to determine DI eligibility. The descriptive data above show a strong correspondence between health quintile and DI participation in each country, although the strength of the relationship varies from country to country, as shown in figure I.9. Simulations Each of the country chapters includes a series of simulations. Some simulations show the fit of the estimated specifications. For all countries these simulations show that the models predict well the proportion of persons that has retired by age. Other simulations are descriptive—­for example, showing employment by education or health by age. The most important simulations are used to predict the effect of the retirement program incentive effects on retirement. It is helpful to recall first the simulations that were done in the second phase of the project. The most important simulations in the second phase were used to predict the effect of increasing retirement program eligibility ages. We describe here two simulations—­S1 and S3—­that were reported in the introduction to the second phase (Gruber and Wise 2004). Both simulations show the effect of increasing the eligibility ages, but the estimation specification and the simulation methods differ. Simulation S1 is based on estimation that controlled for a linear measure of age in the specification and only the OV incentive measure (and the associated variables that determine the OV incentive) is used in the simulation.3 Simulation S3 uses age indicator variables in the estimation and, in addition, uses adjusted age indicators to simulate retirement under the program changes.4 The percent reduction in the proportion of men 3. The estimation in this earlier volume was also based on OV, though as noted above, the current analysis features a more careful modeling of DI and other pathways to retirement (thus, the OV measure used in phase 2 is not exactly the same as the OV inclusive measure used in the new simulations described below). 4. The estimated age indicator effects, as well as the 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 three 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.)

32     Courtney Coile, Kevin Milligan, and David A. Wise

Fig. I.15  Three-­year delay OV-­S1 and OV-­S3

out of the labor force (OLF) is shown in figure I.15. This figure reproduces the data in figure 16 of Gruber and Wise (2004, 29), and the details of the construction of the figure are discussed there. For the S1 simulation, the incentive measure for a country (the OV) is recalculated based on the OV that incorporates the implications of the delayed eligibility age. The shaded bars show the effect of only the change in the incentive implications of the three-­year delay. The average reduction in the proportion of men out of the labor force (OLF) is large—­28 percent. Underlying the average, however, are large differences across countries. For four countries the reduction was greater than 32 percent, for two countries the reduction was less than 4 percent, and was between 16 and 28 percent for the remaining countries. The simulated reduction in the proportion of men OLF is much larger if age indicators are used in estimation and the age effects for each age are moved up three years to correspond with the three-­ year increase in all program eligibility ages. It is not surprising that the effects of increases in the eligibility ages are large. For example, this simulation implies that the early retirement age in the United States increased from sixty-­two to sixty-­five and under S3 this reduced the OLF proportion by 36 percent. In most countries (although not in the United States because DI was not included in the analysis) increasing the eligibility age for retirement would also change the eligibility age for DI by three years as well. Now in this phase, with emphasis on DI, increasing the eligibility age for

Introduction    33

DI seems implausible in many if not most countries. Here we do not change the eligibility age, but instead ask how employment is affected differentially by the provisions of the DI pathway compared to the provisions of the regular social security pathway, and we consider the effect of changes in the provisions of DI programs, especially changes in eligibility stringency. The simulations are all based on the country estimates in table I.3, specification (4). For each simulation the first stage is to calculate OVs corresponding to the programs or program changes that are being compared. Then the estimated effect of the OV incentive effect from table I.3, specification (4) (together with the estimates for other variables in the specification) are used to simulate retirement at each age under each program or program change for each person in the sample. Then the implications for years of employment between ages fifty and sixty-­nine are calculated. Each country has reported the results of three simulations. The first simulation is intended to evaluate the effect of the differential incentive effects inherent in the provisions of each pathway on retirement—­if all persons faced only one of the pathway options. For the United States there are only two pathways—­Social Security or DI. For other countries there are three or more pathways. Each country has used the table I.3, specification (4) coefficients to predict each individual’s probability of retirement for each pathway—­using the DI OVs and then using the SS OVs for the United States. These estimates can also be found in the individual country chapters. For the Netherlands, for example, there are three pathways—­disability, unemployment, and retirement. The retirement probabilities (hazard rates) by age and the cumulative proportion of persons still working (survival rates) by age are shown in figure I.16. Separate lines are shown for each pathway in each country. The distance between the lines for the different pathways varies across countries, depending on the differences in the strength of the retirement incentives across the pathways. For illustration, consider the retirement rates and the survival rates for the Netherlands compared to the United States. The retirement rates are much greater in the Netherlands than in the United States—­at age sixty the retirement rates are 0.1 or lower for each pathway; in the Netherlands the retirement rates are close to three times as great, all greater than 0.27. Corresponding to the higher retirement rates at each age, the survival rate at each age is much higher in the United States than in the Netherlands. For example, at age sixty in the United States employment is much higher than in the Netherlands—­between 0.47 and 0.59 in the United States and between 0.21 and 0.38 in the Netherlands, depending on the pathway to retirement. The survival rates are only comparable across countries if the process begins at age fifty and are only shown for these countries. The hazard rates are provided for all countries for which the data are available. For each program the countries have calculated the mean predicted retirement by age and have used these data to calculate the expected years of work

34     Courtney Coile, Kevin Milligan, and David A. Wise

Fig. I.16  Retirement hazard rates and cumulative survival rates by age and by country

between ages fifty and fifty-­nine. For the United States, for example, the average years of work over the fifty to sixty-­nine age interval is simulated to be 10.18 years if everyone faced the DI OVs and 11.93 years if everyone faced the SS OVs. That is, on average, people work 17.3 percent more years when faced with the incentives inherent in the SS option rather than the incentives inherent in the DI option. In the Netherlands the simulated years of work

Introduction    35

Fig. I.16  (cont.)

in the fifty to sixty-­five age interval is 7.40 under the DI pathway, 9.02 under the unemployment pathway, and 7.47 under the retirement pathway. These simulated years worked between ages fifty and sixty-­nine for other countries are shown by pathway in table I.5. It is important to understand that these differences indicate the marginal effect of the DI incentive compared to the regular retirement incentive, hold-

DI N = 23

Netherlands 50–65

Germany 50–67

United Kingdom 50–69

Canada

United States 50–69

Country

Table I.5

(1) DI = base (2) SS Percent change vs. base (1) DI yrs. of work (2) SS yrs. of work Percent change (2)/(1) (1) DI (2) SS Percent change (2)/(1) (1) DI = base (2) UB (3) OA Percent change (3)/(1) Percent change (2)/(1) Percent change (3)/(2) Percent change vs. base (1) DI = base (2) UE (3) Retire Percent change (3)/(1) Percent change (2)/(1) Percent change (3)/(2) Percent change vs. base

Retirement programs compared

7.40 9.02 9.47 28.0 21.9 5.0

10.18 11.93 17.3 11.31 11.91 5.3 10.7 11.3 5.6 9.49 10.32 13.98 47.2 8.7 35.5

Yrs. of work if all persons faced the same retirement pathway option

8.94

7.02

9.13 9.79 7.2 8.94 9.96 13.62

8.33 9.64 15.7 10.22 10.83

Yrs. of work if all DI participants had faced the same retirement pathway

Simulations

23.8 8.52

13.0 7.56

21.37

4.2 11.06

2.8 10.19

7.69

9.51

3.8

1.9 9.39

10.61

10.1

5.1 10.41

9.17

Yrs. of work if one‑third to DI and two‑thirds to SS pathway

8.75

Yrs. of work if two‑thirds to DI and one‑third to SS pathway

Simulations: Effect of incentive measures alone on years of work between ages specified for each country (three simulations)

(1) CER (2) DI = Base (3) UI (4) OAP Percent change (4)/(1) Percent change (4)/(2) Percent change (4)/(3) Percent change vs. base (1) UE (2) DI = base (3) Normal retirement Percent change (3)/(1) Percent change (2)/(1) Percent change (3)/(2) Percent change vs. base (1) DI = base (2) SS Percent change vs. base (1) DI = base (2) Old age (3) Early retirement Percent change (3)/(1) Percent change (2)/(1) Percent change (3)/(2) Percent change vs. base (1) DI = base (2) Old age Percent change vs. base 13.17 13.93 5.8

4.50 4.80 6.7 12.00 12.67 12.31 5.6 2.6 2.9

4.09 4.96 5.50 34.4 21.3 10.8

5.36 5.65 5.71 7.54 40.67 33.45 32.05

11.35 11.89 4.8

3.6 3.8 5.6 6.9 7.25 7.26

4.293 4.188 4.766

4.53 4.66 4.65 5.51

3.3

1.8

10.06

5.06

11.73

2.8%

2.0%

11.55

7.09

3.6

1.9 7.04

3.73

8.74

4.68 3.67

4.554

17.60

0.43 4.384

5.48

4.68

Notes: For Germany the SOEF does not report DI application, thus estimates for persons in the worst health quintile are used in the right three columns of the table. For Japan there are too few DI applicants to simulate reliable estimates. For Spain none of the incentive estimates is significant, thus the simulations are not reported.

Sweden 50–69

Italy

Denmark 57–69

France 55–64

Belgium 50–64

38     Courtney Coile, Kevin Milligan, and David A. Wise

ing constant all other individual attributes included in the specification. In particular, it holds constant the estimated age dummies. For countries with multiple pathways the process would be repeated for each of the pathways. To be specific, we estimate the incentive effect of a retirement program—­the effect of OVinclusive—­with an equation like this: R = k + a OVinclusive + bAge + c Health + d Education + ˆ cˆ , dˆ and so forth. The estimate aˆ (the estimates reported in We estimate a, ˆ b, table I.3) is an IV estimate of the effect of the OV on retirement. For simulation we take aˆ as the estimate of the effect of OV on retirement and use it for all of the simulations. With a one-­year increase in age the effect on retirement ˆ is given by dR/dAge = a(dOV/dAge) + bAge, where the first term is negative ˆ (aˆ is negative) and the second term positive—­that is, the first term reduces the incentive to delay retirement and the second term increases the preference for retirement with advancing age. The likelihood of retiring advances with age because a reduction in the OV of continuing work is reinforced by the concomitant increase in age. If age is excluded from the specification, then to fit the retirement data the coefficient on OV will have to increase, and if the OVinclusive is eliminated from the specification the coefficient on age will have to increase to fit the retirement data. This is the identification issue mentioned above. In order to identify the correct effect of the incentive measure we must have an age specification that captures the true increase in preference to retire with age. One feature of the estimates that increases our confidence in the incentive estimates is that they are virtually the same whether the single linear age or indicators for each age are used to estimate the effect of age on the preference to retire. Instead of making calculations for all persons in the sample, the second and third simulations consider only persons who were observed to have chosen the DI option. The second simulation asks how much years of work would have changed for this group had the group faced the OVs of the regular retirement option instead of the OVs of the DI program. For the United States, among all those who applied for DI, years worked under the SS option is 15.7 percent greater than under the DI option (9.64 years versus 8.33 years; these values are lower than those for the full sample likely because DI applicants are less healthy than the population at large). For all those who received DI, work under the SS option would have been 16.2 percent greater under the SS option then under the DI option (9.87 years versus 8.49 years). Recall again that in phase 2 of the project we simulated the effect of delaying all program eligibility ages by three years, including the eligibility ages for DI and unemployment programs. In one of these simulations we used estimates with age dummies and in another we used estimates based on continuous age. These simulations suggested very large reductions in retirement, especially the simulations using age dummies in the estimation.

Introduction    39

The simulations proposed here do not consider raising the DI eligibility age, but rather direct attention to the incentive effects—­the OVs—­of the program provisions, and stringency provisions, conditional on the estimated age “preference” effects. It should not be surprising that the employment effect of changing the OV incentive effects is typically much smaller than changing the program eligibility ages. Increasing the eligibility age for DI for three years, for example, means that no one can claim DI benefits for these three years and thus cannot be on the DI program. This would cause great hardship to those who are truly disabled and undermine the insurance role of DI. That is why we do not consider changing the age of eligibility for the DI programs as we did in phase 2. The aim of the third simulation is to get an idea of the effect on retirement of greater stringency in DI acceptance. As in the second simulation, we focus on DI recipients (or applicants, if available). From that simulation, we have an estimate of expected working life if everyone follows the DI path and if everyone follows the SS path. We now make similar calculations to show the effect of making it harder for this group of people who are interested in using DI to access the program—­in effect changing the eligibility stringency. To do this, we first randomly assign two-­thirds of the group to the DI path and one-­ third to the SS path, calculate everyone’s expected probability of retirement, sum by age, and use that to generate an expected work life from ages fifty to sixty-­nine, as described above. We then repeat the process but randomly assign one-­third to the DI path and two-­thirds to the SS path. (If there are more than two paths the simulations are done for different combinations of programs, making different assumptions about which program persons use, if not to DI.) In the United States, the expected work life is 8.328 years if everyone takes the DI path (from the second simulation described above), 8.749 years with two-­thirds on the DI path, 9.166 years with one-­third on the DI path, and 9.635 years with all on the SS path (again from the second simulation). Not surprisingly, shutting down the DI path for one-­third of this sample has about one-­third the effect of shutting it down for the full sample of DI applicants/recipients. Again, the idea of this simulation is to simulate the work effect of making DI harder to access for a share of the population. The results of the simulations for most of the countries are reported in table I.5. The retirement programs that are compared for each country are shown in the first column of the table. The countries in the table are ordered by the average number of years worked—­between the ages shown—­for persons who retire under the “standard” retirement program—­ranging from 11.93 years in the United States and 11.3 years in the United Kingdom to 4.8 years in Denmark. The second column shows the years of work if all persons faced the same pathway option, using all the pathways available in a given country. For the

40     Courtney Coile, Kevin Milligan, and David A. Wise

United States, the years of work after age fifty would be 10.18 if everyone faced the DI incentives and 11.93 if everyone faced the social security incentives, a difference of 17.3 percent. The results differ across countries—­for example, the change in years of work for Canada is only 6.7 percent, which is one-­third the magnitude of the change in the United States. This in part reflects the size of the DI plan in Canada relative to the United States. The next column repeats the exercise, but uses the sample of disabled individuals only. The base number of years worked for this sample is smaller in all countries, and the percent impact of varying the incentives of this sample is smaller than for the entire sample in column (2). The last two columns show the results of the simulation that randomly assigns the incentives, to simulate the effect of making it more difficult for some DI applicants to access the program. The patterns in the results are expected from the calculations—­when two-­thirds of the sample is assigned to the DI incentives, the results look closer to the column (2) results than when only one-­third of the sample is assigned DI incentives. Overall, the simulations suggest that DI programs have a noticeable impact on retirement across countries. Conclusions This volume is the sixth phase of the ongoing project on retirement programs around the world. The focus is on the importance of disability ­programs (DI) and, in particular, the retirement incentive effects of DI programs compared to other retirement programs. This is the second of two phases on DI programs. The first DI phase (the fifth phase of the continuing project) presented analysis of historical trends in our group of countries intended to set the stage for the more formal analysis in the current volume. In the first DI phase, the countries summarized DI program reforms and considered how DI reforms were related to changes in health, in particular, measured by change in mortality. We also considered DI reforms as natural experiments that showed that exogenous reforms can have a very large effect on the labor force participation of older workers. The current phase is also closely related to the second phase of the project, also based on microeconomic analysis of the relationship between a person’s decision to retire and the program incentives faced by that person. In particular, in the second phase the countries considered the employment implications of increasing retirement program eligibility ages, including the eligibility ages for DI programs. The analysis showed that increasing eligibility ages would have very large effects on employment at older ages. In contrast, the current phase focuses on the retirement incentive effects of program provisions without considering changes in program eligibility ages. We give attention to the provisions of DI programs as well as the provisions of other pathways to retirement. The goal is to understand how changing the

Introduction    41

provisions of country DI programs in particular would change retirement. Each country estimated the relationship between program provisions and retirement incentives in their country using an extension of the option value model used in the second phase of the program. Several noticeable findings are based on background summary data. First, the proportion of men ages sixty to sixty-­four collecting disability benefits ranges widely across countries, ranging from 17 percent in Belgium to 16 percent in the United Kingdom, 14 percent in the United States, 6 percent in Italy and France, and 2 percent in Japan—­including Belgium and Italy that use a DI proportion different from the other countries. Second, the data show that in all countries, with the exception of the United States, there was large variation over time in DI participation rates with substantial decline in participation beginning in the early to mid-­1990s in many countries. For example, in Canada participation in the sixty to sixty-­four age group declined 49.6 percent between 1995 and 2009. In the United Kingdom, DI participation declined 49.6 percent between 1996 and 2012. In the United States, on the other hand, DI participation between 1990 and 2012 increased by over 30 percent. Third, variation in DI participation over time was unrelated to trends in health, which improved consistently over time based on declines in mortality. Fourth, and perhaps most striking, DI participation in all countries is very strongly related to education level, even controlling for health. Fifth, descriptive data show a noticeable inverse relationship between DI participation and employment over time. The measurement of health is a central component of the analysis. To maintain as much comparability across countries as possible we use the health index developed by Poterba, Venti, and Wise (PVW). The index as set out by PVW is the first principal component of twenty-­seven health indicators reported in the United States Health and Retirement Study (HRS). The index can be duplicated (approximately) through the nexus of comparable studies—­the English Longitudinal Study of Aging (ELSA), the Survey of Health, Ageing and Retirement in Europe (SHARE), and the Japan Study of Aging and Retirement (JSTAR). These surveys include each of the twelve participating countries except Canada. For reasons of sample size, however, alternative data sources have been used in Sweden, Denmark, and Germany and these data do not provide sufficient health data to construct the PVW index. Estimation is based on the regression counterpart to the Stock-­Wise option value analysis in which retirement is based on the gain (the option value) of delaying retirement. A unique feature of the estimation in this phase is the “inclusive option value” that allows estimation based on the provisions of all pathways to retirement in each country. Two features of the estimates stand out. First, the estimated option value incentive measure is highly statistically significant in each of the countries with the exception of two countries—­Spain and Germany (SHARE)—­where the SHARE

42     Courtney Coile, Kevin Milligan, and David A. Wise

country data files were not large enough to support precise estimation. Second, the estimated effect of the OV incentive measure is substantial in most countries. For example, a one standard deviation increase in the option value (used as a standard measure across countries) reduces the estimated retirement rate by between 4 and 6 percent in six countries, by between 8 and 9 percent in two countries, and between 1 and 3 percent in three countries. The most important results are in the form of simulations. First, simulations show that the model estimates fit the data very well—­which is to be expected in specifications in which age indicators are estimated. Second, simulations of retirement rates by age and survival in the labor force show very large variation across countries. Third, perhaps the most important simulations show the importance on retirement of differences in the provisions of each pathway to retirement in each country. These differences are estimated first by simulating the number of years worked between ages fifty and sixty-­nine if all persons faced only one of the pathways to retirement. For example, in the United States, years worked would be 10.18 if all persons faced the DI pathway provisions. If all persons faced the Social Security pathway, the average would be 11.93 years, an increase of 17.3 percent. In Belgium there are four pathways with estimated hours of work between ages fifty and sixty-­nine of 5.36, 5.65, 5.71, and 7.54 for the CER, DI, UI, and old-­age pension (OAP) pathways, respectively. Hours of work on the OAP pathway exceed hours on the CER, DI, and UI pathways by 40.67 percent, 33.45 percent, and 32.05 percent, respectively. Fourth, simulations show the effect on retirement of increasing the stringency of admission to the DI program. This simulation is especially relevant given the large reduction in DI participation in many countries since the late 1980s and the mid-­1990s. For example, if one-­third of the persons now on DI in the United States were instead eligible only for the Social Security program, the hours of work of current DI participants would be increased by 5.1 percent; if two-­thirds were eligible for the Social Security program only hours of work of current DI recipients would be increased by 10.1 percent. A comparable increase in the stringency of access to the DI program in the Netherlands would increase the years of work of current DI recipients by 7.69 percent and 21.37 percent, respectively. With large increases in life expectancy in all participating countries there is considerable interest in prolonging working lives. Indeed, there has been a large increase in the employment of men in most of the participating countries since the late 1980s and the mid-­1990s—­the same period over which DI participation has been declining in most countries. Future increases in working lives will depend on the capacity to work, which may depend on individual attributes such as education. The capacity to work will be the topic of the next phase of the International Social Security project.

Introduction    43

Appendix Appendix on the OV Incentive Measure Under the option value formulation, the value at age t of retirement at age r is given by Vt (r) =

r −1

S

s =t

s=r

∑ s −tEt(Ys) + ∑ s −tEt(kBs(r)),

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. 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 age. If the person retires at some later age he will gain from future wage earnings and from any gain in future pension benefits. If r* is the retirement year that gives the maximum expected gain, the option value is given by OVt(r*) =

r −1

S

S

s =t

s = r*

s =t

∑ s −t Et(Ys) + ∑ s −t Et(kBs(r*)) − ∑ s −t Et(kBs(t)) discounted utility

=

of future wage discounted utility +

of benefits if retire atr*

discounted utility −

of benefits if

.

retiree att

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 γ, β, and k. Instead, we assume these values: γ = 0.75, β = 0.03, and k = 1.5, which are somewhat different from estimates obtained by Stock and Wise (1990), especially the assumed value of β, which is much smaller than their estimate.

References Gruber, Jonathan, and David A. Wise. 1999. Social Security Programs and Retirement around the World. Chicago: University of Chicago Press.

44     Courtney Coile, Kevin Milligan, and David A. Wise —­—­—­. 2004. Social Security Programs and Retirement around the World: Micro-­ Estimation. Chicago: University of Chicago Press. —­—­—­. 2007. Social Security Programs and Retirement around the World: Fiscal Implications. Chicago: University of Chicago Press. —­—­—­. 2010. Social Security Programs and Retirement around the World: The Relationship to Youth Employment. Chicago: University of Chicago Press. Milligan, Kevin, and David A. Wise. 2012. “Introduction and Summary.” In Social Security Programs and Retirement around the World: Historical Trends in Health, Employment, and Disability Insurance and Reforms, edited by David A. Wise. Chicago: University of Chicago Press. Poterba, James, Steven Venti, and David A. Wise. 2013. “Health, Education, and the Postretirement Evolution of Household Assets.” Journal of Human Capital 7 (4): 297–339. PMCID: PMC4043284. Stock, James, and David Wise. 1990. “Pensions, the Option Value of Work, and Retirement.” Econometrica 58:1151–80. Wise, David A. 2012. Social Security Programs and Retirement around the World: Historical Trends in Health, Employment, and Disability Insurance and Reforms. Chicago: University of Chicago Press.

1

Disability Insurance Incentives and the Retirement Decision Evidence from the United States Courtney Coile

1.1

Introduction

The rolls of the US Disability Insurance (DI) program have risen dramatically since the program’s inception in 1956. Over the past two decades, the share of the population age twenty-five to sixty-four receiving DI benefits more than doubled, from 2.3 percent in 1989 to 5.1 percent in 2012 (Figure 1.1). The growth of the program is likely to continue, stabilizing at 7 percent of the nonelderly population, according to one projection (Autor and Duggan 2006a). The rising number of DI beneficiaries has jeopardized the program’s ability to pay benefits, with annual benefit expenditures reaching $140 billion in 2012 and the DI trust fund projected to be depleted by 2016. As the trustees of the program recently warned, “lawmakers need to act soon to avoid reduced payments to DI beneficiaries three years from now” (OASDI Trustees 2013). Concerns about the DI program have been amplified by the observation that the program’s growth does not appear to be driven by worsening population health. Over the period that DI participation doubled, the fraction of people reporting themselves to be in poor health or suffering from a worklimiting health problem was unchanged, if not declining (Milligan 2012; Duggan and Imberman 2008). These trends have led to renewed interest Courtney Coile is professor of economics at Wellesley College and a research associate of the National Bureau of Economic Research. This chapter was prepared for the NBER International Social Security project. I thank the organizers and other country teams for their suggestions. I also thank Peter Diamond, Jonathan Gruber, and Kevin Milligan, authors of US analyses in earlier volumes of the Social Security and Retirement around the World series that helped to inform the current chapter. For acknowledgments, sources of research support, and disclosure of the author’s material financial relationships, if any, please see http://www.nber.org/chapters/c13351.ack.

45

46

Courtney Coile

Fig. 1.1 DI beneficiaries as a share of population, age twenty-five to sixty-four (1957–2012) Source: Authors’ calculations based on table 5.D3 from the Social Security Annual Statistical Supplement and population data from the US Census Bureau (www.census.gov).

in understanding the causes of the rise in the DI rolls, as well as its consequences. The effect of DI on labor supply has been a subject of interest since Bound (1989, 1991) and Parsons (1991) reached different conclusions from comparisons of the earnings of accepted and rejected DI applicants. More recent work by Maestas, Mullen, and Strand (2013), French and Song (2012), and Chen and van der Klaauw (2008) has made use of plausibly exogenous variation in DI receipt coming from random assignment of DI applicants to medical examiners or similar sources. This study takes a different approach to exploring the effect of the DI program on labor supply, specifically labor force withdrawal or retirement. The methodology employed here builds on Coile and Gruber (2004, 2007), who construct several measures of the financial incentives for additional work arising from the structure of the Social Security (SS) program. One measure is the “option value” (OV), which captures the gain in utility resulting from retiring at the optimal future date, over and above the utility available by retiring today. Those studies find that having a larger financial incentive for continued work is associated with a reduced probability of retirement. However, these studies ignore the DI program, treating Social Security (and private pensions) as the only possible pathway to retirement. In the current study, I construct an “inclusive” option value measure that incorporates the financial incentives arising from both SS and DI, and esti-

Disability Insurance Incentives and the Retirement Decision

47

mate models that relate this new measure to the retirement transitions of workers age fifty to sixty-nine, using data from the Health and Retirement Study (HRS). To explore the effect of incentives on retirement conditional on health, I control for health using an index developed in Poterba, Venti, and Wise (2013). I explore whether the effect of incentives on retirement varies by health and education, both of which are strongly related to the probability of DI receipt. Finally, to put the magnitude of the findings into context and gauge the relevance of DI to retirement decisions, I use the regression estimates to simulate the effect of reducing access to DI. I have several key findings. First, the probability of DI receipt is strongly linked to education, even conditional on health. Second, the inclusive OV measure has a negative and significant effect on the probability of retirement; the effect is robust to choice of specification and varies by education and health. Finally, the simulations suggest that reducing access to DI would have large effects on the labor force participation of DI applicants. The remainder of the chapter is structured as follows. In the next section, I provide background on the US DI program and the past literature on DI and labor supply. Next, I describe the empirical strategy, notably how the inclusive OV measure is constructed, as well as the data used. I present descriptive statistics on the probability of DI receipt, and then present the main regression results. I conclude with a simulation of the effect of reducing access to DI and a discussion of the implications of the findings. 1.2 1.2.1

Background Institutional Features of Social Security and Disability Insurance

Disability insurance in the United States is part of the Social Security program. Eligibility for DI and the calculation of DI benefits is similar to that for SS, with a few key differences. Workers become eligible for Social Security retired worker benefits after ten years (forty quarters) of covered employment, which now encompasses most sectors of the economy. Benefits are determined by first calculating the Average Indexed Monthly Earnings (AIME), an average of the individual’s highest thirty-five years of earnings, indexed by a national wage index. Next, a progressive linear formula is applied to the AIME to get the primary insurance amount (PIA), where ninety cents of the first dollar of earnings is converted to benefits but only fifteen cents of the last dollar. Finally, the PIA is multiplied by an adjustment factor for claiming before or after the normal retirement age ([NRA]; currently sixty-six, but rising slowly to sixty-seven for those born in 1960 or later) to obtain the monthly benefit amount. Benefits are first available at age sixty-two but may be claimed as late as age seventy, and the adjustment factor for early or delayed claiming is

48

Courtney Coile

considered to be roughly actuarially fair.1 Before the NRA, workers face an earnings test if their earnings exceed a threshold amount, $15,480, in 2014. Benefits are available for spouses and survivors of retired workers, though a spouse who is also qualified for retired worker benefits receives only the larger of the benefits to which she (or he) is entitled. For the median earner, the Social Security replacement rate is 47 percent of average lifetime earnings (Biggs and Springstead 2008). While receipt of retired worker benefits upon claiming is automatic for an insured worker, the DI application process is more complex. First, in order to be disability insured, a worker must meet both “recent work” and “duration of work” tests, working in at least five of the last ten quarters (less if disabled by age thirty) and for up to forty quarters over the worker’s lifetime (depending on age at disability). An insured worker applying for DI must be determined to have a disability, defined as the “inability to engage in substantial gainful activity (SGA) by reason of any medically determinable physical or mental impairment(s) which can be expected to result in death or which has lasted or can be expected to last for a continuous period of not less than twelve months.”2 The review of a DI application can be a lengthy, multistep process— the initial decision is made by an examiner at a state disability determination (DDS) office, but denied applicants have up to four levels of appeal available to them. One recent study found that although only one-third of applicants were allowed in the initial determination, nearly twothirds were ultimately awarded benefits (Maestas, Mullen, and Strand 2013). Successful DI applicants begin receiving benefits five months after disability onset, and are eligible for Medicare after two years. Beneficiaries who earn more than the SGA threshold, $1,070 per month in 2014, lose DI eligibility. The disability screening process has been subject to changes over time. In the late 1970s, DDS offices tightened medical eligibility criteria in response to growing DI enrollments, resulting in a sharp increase in initial denial rates (Gruber and Kubik 1997). A 1980 law increased the number of “continuing disability reviews” (CDRs), leading to the termination of benefits for 380,000 individuals over the next three years (Rupp and Scott 1998). These actions generated a public backlash that led Congress to enact new legislation in 1984. While the new law did not change the statutory definition of disability, it shifted the focus of screening from medical to functional criteria, instructing examiners “to place significant weight on applicants’ reported pain and discomfort, to relax its strict screening of mental illness and to consider multiple nonsevere ailments (impairments) as constituting 1. Shoven and Slavov (2013) estimate that returns to delayed claiming have increased over time, particularly since 2000, while Munnell and Sass (2012) argue that the actuarial fairness of the Social Security adjustment factor has changed little over time. Coile et al. (2002) show there is a financial and utility gain from claiming delay for many individuals. 2. Social Security Act, Title II, Section 216 (http://www.ssa.gov/OP_Home/ssact/title02/0216 .htm#act-216-i), accessed May 11, 2015.

Disability Insurance Incentives and the Retirement Decision

49

a disability during the initial determination decision, even if none of these impairments was by itself disabling” (Autor and Duggan 2006b, 8). The 1984 law also put more weight on medical evidence provided by applicants’ own health care provider and less on that from the Social Security Administration’s medical examination. Several differences between SS retired workers and DI benefits are relevant for the discussion of financial incentives below. First and foremost, DI benefits are available (to a successful applicant) from the age of disability onset, while retired worker benefits are available only starting at age sixtytwo. Second, DI benefits are not subject to reduction for early claiming; thus, a worker claiming retired worker benefits at age sixty-two would receive 75 percent of their PIA (based on current rules), while a worker who was awarded DI benefits at age sixty-two (or any other age) would receive 100 percent of their PIA.3 Finally, there are some small technical differences in the calculation of the two benefits, such as a lower number of years of earnings and different indexing year (both due to the shorter career) used in the calculation of the AIME and PIA for DI benefits. 1.2.2

Relevant Past Literature

This chapter, like nearly any study of the US DI program, is motivated at least in part by the growth over time in DI enrollments, and thus the literature exploring the reasons for this trend is of interest. Changes in the stringency of medical screening are clearly one important factor. As figure 1.1 illustrates, fluctuations in DI enrollment over time match up with the dates of screening changes, with the DI participation rate falling by 20 percent between 1977 and 1984 (from 2.8 percent of the nonelderly population to 2.2 percent) following the initial tightening of eligibility criteria and increase in CDRs and rising again sharply following the 1984 law. The composition of the DI population has also shifted dramatically in the past two decades, with the number of beneficiaries with musculoskeletal and mental disorders growing by over 300 percent while the number with cancer and heart disease grew by only 30 percent; the explosive growth in the former group is consistent with the 1984 law’s relaxed screening of mental illness and greater emphasis on pain and workplace function (Autor and Duggan 2006a). Economic and demographic factors have also been put forward as possible explanations for the time-series trend. Autor and Duggan (2003) point out that the value of DI relative to potential labor market earnings has risen since the late 1970s because of the interaction between the DI benefit formula and rising income inequality, whereby DI benefits become relatively 3. The rise in the NRA makes it more attractive for early retirees to apply for DI when they retire, since the actuarial reduction for claiming retired worker benefits at age sixty-two is rising over time from 20 percent (for those born before 1938) to 30 percent (for those born starting in 1960). Li and Maestas (2008) find that the increase in the NRA has led to an increase in DI applications, particularly among those in poor health.

50

Courtney Coile

more generous if an individual’s earnings growth lags behind the average growth of earnings in the economy. Over the past two decades the increase in DI enrollment has been largest for those without a high school degree, consistent with their weakening position in the economy (Katz and Autor 1999). Another potential explanation is rising women’s labor force participation, which has made more women eligible for DI. As illustrated below, women’s DI participation rates rose more rapidly over this period than did men’s, lending some credence to this theory; however, Autor and Duggan (2006a) estimate that increased attachment to the labor force explains only one-sixth of the increase in women’s DI participation over time, suggesting that other factors may matter more. Finally, as mentioned above, changes in health do not appear to be a major driver of the growth in DI enrollment, since mortality rates have fallen over time while other health measures have generally been either flat or improving. A second strand of the literature that is highly relevant for the present analysis concerns the effect of the DI program on labor supply. The longterm decline in the labor force participation of older men that began after the end of World War II (before stabilizing and ultimately reversing starting in the early 1990s) coincided with the rapid growth of the DI program in its first two decades of existence, prompting analysts to explore the effect of DI on men’s labor force participation as far back as Parsons (1980). Estimating the effect of the DI program on labor supply is difficult because the counterfactual— how much DI recipients would have worked in the absence of the DI program— is unobservable. Comparing the labor force participation of DI recipients with that of the population at large is fraught because DI recipients are in worse health and may differ in other unobservable ways, introducing bias in the estimation. Bound (1989) offers a novel solution, using the postdecision earnings of rejected DI applicants as an upper bound estimate of the work capacity of successful applicants, the former group presumably being in better health than the latter. Finding that rejected DI applicants had labor force participation rates of less than 50 percent, Bound concludes that the work capacity of successful applicants is low. Subsequent papers (Parsons 1991; Bound 1991) have raised and debated potential problems with this approach. Rejected applicants may need to remain out of the labor force for years to avoid jeopardizing their appeals and may also suffer depreciation of human capital due to the interruption in their work career (which would not occur in the absence of a DI program). Lahiri and Wixon (2008) found that rejected DI applicants also tend to have intermittent work histories, further calling into question their use as a comparison group. More recent contributions to this literature have surmounted the usual endogeneity problem by identifying plausibly exogenous sources of variation in DI receipt. Maestas, Mullen, and Strand (2013) exploit variation in

Disability Insurance Incentives and the Retirement Decision

51

the allowance rates of DI examiners at the initial stage in the DI determination process. They find that among the roughly one-quarter of applicants on the margin of program entry, employment would have been nearly 30 percentage points higher in the absence of DI benefits. These effects are heterogeneous, ranging from no effect for the most impaired to a 50 percentage point effect for the least impaired. French and Song (2012) employ a similar methodology, using variation that arises from random assignment of DI cases to administrative law judges, a later stage in the DI determination process. Chen and van der Klaauw (2008) employ a regression discontinuity approach based on discrete changes in eligibility standards at various ages (e.g., age fifty-five) that are codified in the Medical-Vocational Guidelines and used for applicants when a disability determination cannot be made on medical grounds alone. The latter two papers obtain estimates roughly similar to those of Maestas, Mullen, and Strand (2013). Gruber (2000) differs slightly from the other papers in this group in that he focuses on the generosity of DI benefits. Making use of a differential increase in benefits in Quebec versus the rest of Canada in the 1980s to estimate a differencesin-differences model, he finds an elasticity of labor force nonparticipation with respect to DI benefits in the range of 0.3. The approach employed in this chapter takes a different tack, building on the analysis in Coile and Gruber (2001, 2004, 2007). As explained in more detail below, this approach involves calculating the financial incentive to continued work through the SS and DI programs (option value) and estimating its effect on retirement decisions. Rather than comparing labor supply outcomes of DI recipients and nonrecipients, as most of the above-referenced papers do, the approach taken here compares the labor supply outcomes of those with more and less to gain from continued work. As explained at greater length in the Coile and Gruber papers, there is substantial heterogeneity in the option value measure.4 While some of this heterogeneity arises from differences in characteristics such as age, marital status, and earnings (which may influence retirement decisions but can be included as control variables), much of it also arises from factors such as nonlinearities in the Social Security benefit formula and how they interact with the particulars of an individual’s earnings history. As we argue in those earlier papers, this is a fruitful source of variation for estimating the effect of Social Security on retirement. The innovation in this chapter, relative to those earlier studies, is to incorporate DI incentives in to the option value measure through the construction of the “inclusive option value” measure. 4. This is also true of the purely financial-based incentive measures that play a bigger role in these earlier studies, namely the “accrual,” or increase in lifetime present discounted value (PDV) of Social Security benefits arising from an additional year of work, and “peak value,” or change in PDV associated with working from the present age to the age at which PDV is maximized.

52

1.3 1.3.1

Courtney Coile

Empirical Approach Data

The data for the analysis comes from the Health and Retirement Study (HRS). The HRS began in 1992 as a survey of individuals then age fifty-one to sixty-one (born in 1931–1941) and their spouses, with reinterviews of these individuals every two years. Over time, new cohorts have been added to the survey to maintain a national panel of individuals over age fifty and their spouses.5 To date, 11 waves of data (1992–2012) have been collected; as the 2012 data has only recently been made available, this chapter uses the 1992–2010 data. The chapter uses the RAND HRS data file, a cleaned data set that links information over time and across family members and defines variables consistently over time. A key feature of the HRS is that it includes Social Security earnings histories for most respondents.6 This allows for the calculation of SS and DI benefit entitlements, which depend on the entire history of earnings. The HRS also contains richly detailed health information that is used in constructing the health index, as detailed below. The size of the HRS— over 30,000 individuals have appeared in one or more survey wave over the years— as well as the fact that it is a panel allows for the construction of a large sample of person-year observations. Specifically, the estimation sample includes observations for all men and women in any year from 1992 to 2009 in which they met three criteria: (a) they were age fifty to sixty-nine during the year; (b) they were in the labor force at the beginning of the year; and (c) they were observed in the subsequent survey wave, in order to be able to determine whether or not they retired that year. Thus an individual who was, for example, age fifty when first observed in the HRS in 1998 and retired in 2008 at age sixty would contribute eleven personyear observations to the sample, so long as he remained in the survey until 2010 (to determine whether he retired in 2008). The final sample includes 70,675 observations from 10,570 individuals. The labor supply outcome of interest in the chapter is retirement. Retirement is defined based on the labor force status reported at each wave, an individual being classified as retired when he or she has transitioned from working or unemployed at the previous wave to out of the labor force in the current wave, with the year of retirement assigned based on the date the individual reports at the current wave. Retirement is treated as an absorbing 5. The Asset and Health Dynamics among the Oldest Old (AHEAD) cohort (born before 1924) was added to the survey in 1998, when the previously separate AHEAD survey was merged with the HRS. The War Babies (1942–1947) and Children of the Depression (1924– 1930) cohorts were also added in 1998. The Early Baby Boomer cohort (1948–1953) joined the survey in 2004 and the Mid-Baby Boomer cohort (1954–1959) in 2010. 6. These data are restricted and available by application only.

Disability Insurance Incentives and the Retirement Decision

53

state, so that once an individual reports himself as out of the labor force after age fifty, any subsequent employment spells are not used in the analysis. 1.3.2

Pathways to Retirement

While in some other developed countries early retirement or unemployment insurance benefits offer a viable means of income support from the time a worker leaves his or her job until he or she becomes eligible for social security benefits, in the United States there are only two relevant pathways from employment to retirement: the traditional Social Security ([SS]; meaning retired worker) path and the disability insurance (DI) path.7 As noted above, SS benefits are available starting at age sixty-two. In the construction of the incentive measures, described in more detail below, SS benefits are treated as being claimed at the later of age sixty-two or when the individual retires. Although claiming is a separate decision from retirement and an individual could theoretically claim benefits either before retirement (once he or she has reached age sixty-two) or after, this assumption seems reasonable given that the SS earnings test, which is still in place for workers until they reach the NRA, depresses preretirement benefit claiming (Gruber and Orszag 2003) and that it is relatively rare for individuals to delay SS benefit receipt after retirement (Coile et al. 2002). The DI benefits are treated as being claimed at the time of labor force withdrawal, since there is no advantage to (or even mechanism for) delayed claiming.8 While this may be a reasonable assumption, it is clearly not realistic to assume that everyone can be a successful DI applicant. There is a medical screening process, and though it may be imperfect (as evidenced by the large number of denied applicants who are successful upon appeal, for example), some individuals— those in worse health, also potentially those who are older or in certain occupations due to the use of vocational guidelines in some cases— would be expected to have a higher probability of

7. Unemployment insurance (UI) benefits are typically available for only six months and only to insured workers who are laid off, limiting their value as a source of early retirement income. Coile and Levine (2007) suggests that UI benefits are not empirically important for the retirement decision, finding that workers who reach age sixty-two in a period of high unemployment are more likely to retire, but that the generosity of UI benefits has no effect on retirement transitions. They conclude that SS may be more relevant than UI in protecting older workers from the impact of a late-career employment shock. In addition, in theory, private pensions should be incorporated in the analysis as well, not as a distinct path to retirement but as an income source available to those individuals in the sample who are eligible for defined benefit (DB) pensions, whether they retire along the SS or DI path. Coile and Gruber (2007) calculate incentive measures using SS income only and using both SS and pension income and obtain very similar regression estimates from the two sets of measures, providing some justification for their omission here. 8. Successful applicants are eligible for benefits after a five-month waiting period from the onset of disability, as discussed earlier, but this detail is ignored in the analysis. The DI applicants often spend more than five months waiting for their final disability determination, but benefits are paid retroactively.

54

Courtney Coile

a success. A discussion of how the uncertainty in access to DI benefits is incorporated into the empirical analysis is deferred to the following section. 1.3.3

Option Value Calculations

To review, the goal of the analysis is to develop a retirement incentive measure that will reflect the financial incentives for continued work arising from both the SS and DI programs and to estimate its effect on retirement. To explain the chapter’s approach, in this section I first describe the standard SS-only option value measure used in prior analyses (Coile and Gruber 2004, 2007). I then explain how this will be expanded to an “inclusive OV” measure that incorporates DI benefits, including how the uncertainty about an individual’s ability to access DI is addressed. Finally, I explain other details relevant to the calculation of the inclusive OV measure. The option value (OV) approach was pioneered by Stock and Wise (1990) in order to model retirement incentives for workers with defined benefit (DB) pensions. Because DB pensions can have nonmonotonic accrual patterns, for example, very large returns to work in the year that pension vesting occurs or that the individual reaches the pension plan’s normal retirement age, the one-year change in the present discounted value (PDV) of pension wealth resulting from an additional year of work (the “accrual”) fails to capture the fact that by working this year, the employee is effectively purchasing an option to work in a future year with a larger accrual. Although nonmonotonicities in the accrual of SS benefits do not tend to be as large or frequent as those found for DB pensions, Coile and Gruber (2001) nonetheless show that they exist for SS as well. Option value is a forward-looking measure of the utility gain arising from working to the optimal future retirement date, in excess of the utility available by retiring today. Traditionally, OV has included only SS (and sometimes pension) benefits, but since the present analysis analyzes DI incentives as well, I use the notation OVSS to indicate the traditional measure that only includes SS. The OVSS calculation begins as follows: OVSS(R)ii = (1)

T 1 1 R   + probalive (wage ) probaliveit(k ∗ SSben(R)i ) ∑ ∑ it it t =0 (1 + )t t t = R (1 + ) 

− OVSS(R 0 ), where R refers to a future retirement date, R 0 refers to today, and T is the final period in which the individual could be alive. Also, OVSS(R) is essentially the sum of earnings until time R and of SS benefits (which are a function of R) from time R to time T, discounted for time preference and survival probability, where δ reflects the discount rate, γ reflects the curvature of the utility function, and k reflects the greater utility individuals receive from retirement income due to the utility of leisure. Unlike Stock and Wise (1990),

Disability Insurance Incentives and the Retirement Decision

55

who obtain values for the utility parameters by a structural estimation of their model, we assume that these three parameters take on the values of 0.03, 0.75, and 1.5, respectively.9 Equation (1) reflects the utility gain associated with retiring at some future date R, so the individual must repeat this calculation for all possible values of R and estimate: (2)

OVSS i = max{OVSS(R1 ) i ,OVSS(R 2 ) i , . . . ,OVSS(R max ) i}, R

where OVSS is the gain in utility arising from delaying retirement and receipt of SS benefits from the present time until the optimal date, the date at which utility is maximized. In our analysis, age sixty-nine is treated as the last possible retirement age considered by the worker. Having made this calculation for OVSS, it is straightforward to calculate OVDI in the same manner, temporarily ignoring the possibility that the DI path may be difficult to access for many individuals. In essence the OVDI calculation tells us, if one is going to retire via the DI program, what the optimal date (age) at which to do so is and how large the utility gain is from waiting until that optimal date. Having calculated OVSS and OVDI brings us to two related questions. First, how can we construct a single incentive measure that incorporates both?10 Second, what is the appropriate way to account for the fact that not everyone who might want to will be able to choose to retire down the DI path? It turns out that both questions have the same answer, which is to construct an inclusive OV measure that is a weighted average of the two individual measures, as follows: (3)

OVInclusivei = (DIprobabilityi ∗ OVDIi ) + ((1 − DIprobabilityi ) ∗ OVSSi ),

where OVInclusive is the key regressor in our retirement regressions. The obvious question that arises in its calculation is what value to use for DIprobability. In theory, this measure should reflect the probability that the DI path is a realistic option for a given individual. Our approach is to calculate the probability that people age fifty-five to sixty-four are receiving DI by year, sex, and education cell, and use these cell probabilities. This approach has the practical advantage that it requires relatively little data, making it feasible to apply in contexts where rich data such as the HRS is 9. An informal grid search over a range of possible values for the three parameters suggests that the likelihood function is relatively flat with respect to parameter choice. 10. One very relevant reason for preferring a single measure in the current context is that the results presented here will be combined with those from the other countries participating in the NBER International Social Security project, and the number of pathways may differ across countries. One of the important benefits of having analysts in a large number of countries undertake the same analysis (as nearly as possible) is the insights that can be derived when results are combined.

56

Courtney Coile

not available. While it would be possible, using the HRS, to go beyond this approach to estimate a predicted probability that any given individual would go on DI, incorporating health information that is surely relevant to DI application and receipt, an advantage of using cell averages is that it avoids the use of these potentially endogenous covariates. Additionally, since some regression specifications interact our incentive measure with health, it is awkward to also have health embedded in the construction of the incentive measure. In essence, one can think of this as similar to an instrumental variables approach, where we limit ourselves to the variation that is more plausibly exogenous to retirement to obtain a cleaner, if less precise, estimate of DIprobability. The actual values used for DIprobability are reported below. Finally, I briefly discuss a few salient technical details relevant to the calculation of OVInclusive; more information about these calculations can be found in the appendix to Coile and Gruber (2001). The worker’s potential future earnings must be projected to age sixty-nine in order to calculate OVSS and OVDI, as earnings enter directly in the OV measures. Following Coile and Gruber (2004), I grow real earnings by 1 percent per year from the last observed value. I estimate PIAs for all possible future retirement dates using a program that incorporates the Social Security benefits rules and has been cross-checked against the Social Security Administration’s ANYPIA model. The appropriate actuarial adjustment factor is applied in the calculation of SSBen(R). For married workers, OVSS and OVDI incorporate dependent spouse and survivor benefits, allowing for the probability that at any given age, either or both spouses may be surviving. The inclusion of spousal benefits is complicated by the fact that a spouse who is qualified for retired worker benefits is entitled to the greater of this or her dependent benefit, which will depend on her retirement date. A full modeling of joint retirement decisions is beyond the scope of this chapter, so I assume that any working wives (or husbands) retire at age sixty-two for the purpose of incorporating dependent benefits on the spouse’s record, a seemingly reasonable assumption, given that the median retirement age is sixty-two for married women who were working at age fifty. 1.3.4

Health Quintiles

An important goal of the larger project of which this chapter forms a part is to ask: Given health status, to what extent are differences in labor force participation within and across countries determined by the provisions of DI programs? To be able to answer this, it is necessary to control for health in the analysis, preferably in a way that incorporates as much information as possible and can be replicated across countries. The approach adopted here, which builds on Poterba, Venti, and Wise (2013) and is described at length elsewhere in this volume, is to construct a health index based on twenty-seven questions, including self-reported health

Disability Insurance Incentives and the Retirement Decision

57

diagnoses, functional limitations, medical care usage, and other health indicators. To do so, one first obtains the first principal component of these indicators, which is the “weighted average of indicators where weights are chosen to maximize the proportion of the variance of the individual health indicators that can be explained by this weighted average.” The estimated coefficients from the analysis are then used to predict a percentile score for each respondent, referred to as the health index. An individual’s health index value typically will vary by HRS survey wave, as updated health information points are incorporated. As Poterba, Venti, and Wise (2013) demonstrate, the health index is strongly related to mortality and to future health events such as stroke and diabetes onset, though not to new cancer diagnosis. In the analysis below, respondents are divided into health quintiles based on their health index scores. 1.4 1.4.1

Results Descriptive Analysis: DI Participation Rates

Before turning to the regression results, I present some figures on DI participation. Figures 1.2A and 1.2B show participation rates for men and women ages fifty to sixty-four since 1982, using data on DI beneficiaries from the Social Security Administration and population data from the US Census Bureau. Trends over time for older workers mirror those seen in figure 1.1 for the population at large. By 2012, one in seven men ages sixty to sixty-four (14.2 percent) is on DI, as is one in ten men at ages fifty-five to fifty-nine (10.6 percent), and one in fourteen at ages fifty to fifty-four (7.1 percent). The DI participation rates for older women have risen even more dramatically than for older men in the last three decades, doubling for the age sixty to sixty-four group, from 5.6 percent in 1982 to 11.4 percent in 2012, and tripling for women age fifty to fifty-four, from 2.0 percent in 1982 to 6.4 percent in 2012. Figures 1.3A through 1.3D show rates of DI receipt by education and health for men and women ages fifty-five to sixty-four. These and subsequent figures use data from the HRS;11 representative years from 1992 through 2008 are shown on the graph, though calculations are made for all years. The first thing to note is that the values shown on figures 1.3A and 1.3B are the DIprobability values used in the construction of OVInclusive, as they are year-sex-education cell average participation rates. Figure 1.3A shows a substantial DI participation gradient by education, with the lowest education group, high school dropouts, being five to six times more likely to be on DI than the highest education group, college 11. The data in these figures reflect all HRS respondents in the relevant age group, and are not limited to workers.

Fig. 1.2A

DI participation rates for men age fifty to sixty-four, 1982–2012

Source: Authors’ calculations based on table 5.D3 from the Social Security Annual Statistical Supplement and population data from the US Census Bureau (www.census.gov).

Fig. 1.2B

DI participation rates for women age fifty to sixty-four, 1982–2012

Source: Authors’ calculations based on table 5.D3 from the Social Security Annual Statistical Supplement and population data from the US Census Bureau (www.census.gov).

Fig. 1.3A Probability men age fifty-five to sixty-four in HRS have received DI, by education and year Source: Authors’ calculations from the HRS.

Fig. 1.3B Probability women age fifty-five to sixty-four in HRS have received DI, by education and year Source: Authors’ calculation from the HRS.

Fig. 1.3C Probability men age fifty-five to sixty-four in HRS have received DI, by health quintile and year Source: Authors’ calculation from the HRS.

Fig. 1.3D Probability women age fifty-five to sixty-four in HRS have received DI, by health quintile and year Source: Authors’ calculation from the HRS.

Disability Insurance Incentives and the Retirement Decision

61

graduates; in 2008, the rates were 22 percent for the former group and 4 percent for the latter. The rise in DI rates over time that was evident in earlier figures is present here as well for all education groups; the DI participation rate for high school graduates, for example, rises by 41 percent from 1992 to 2008, from 7.1 percent to 10.0 percent. Figure 1.3B shows that the DI participation gradient by education is, if anything, steeper for women; the rise in DI over time is also more pronounced, consistent with earlier figures. Figures 1.3C and 1.3D repeat the exercise, stratifying by health quintile (as defined above) rather than by education group. The DI participation gradient with respect to health is much steeper than that for education. This is not terribly surprising, in that there is a medical screening process for DI, so those in worse health (measured using data from the current survey wave) should be more likely to be on DI. Among men ages fifty-five to sixty-four in 2008, 46 percent of those in the lowest health quintile were on DI versus 9 percent for the second quintile, 3 percent for the third, and essentially no one in the top two quintiles. The strong relationship between DI receipt and the health index would seem to provide some reassurance that both the health index we construct is a useful summary statistic for health status and that the DI medical screening process is at least somewhat successful in identifying the least healthy. The graph for women is very similar, though the probability of being on DI for those in the lowest health quintile is somewhat lower, only 37 percent in 2008. One question raised by these figures is whether the correlation between education and DI receipt seen in figures 1.3A and 1.3B primarily reflects the effect of health, since low socioeconomic status is known to be correlated with poor health (Smith 1999), or whether there is a relationship between education and DI receipt even conditional on health. This question is answered in figures 1.3E and 1.3F, which show the probability of DI receipt by education and health, averaged across all years. The education gradient is substantially smaller, but remains nontrivial, with male high school dropouts in the lowest health quintile being 46 percent more likely to be on DI than college graduates in the same health quintile (50 percent vs. 34 percent), while female high school dropouts are 66 percent more likely to be on DI (38 percent vs. 23 percent). The education gradient is equally strong, if not stronger, in higher health quintiles, though the absolute rates of DI participation are quite small in the top two quintiles. Thus, we can conclude that education has a robust relationship with DI receipt. This is consistent with rising income inequality being one of the explanations for the rise in the DI rolls, as mentioned above. It is also consistent with finding that DI applications and awards tend to rise with the unemployment rate (Autor and Duggan 2003), since less educated workers experience higher rates of unemployment.

Fig. 1.3E DI participation by health quintile and education, men fifty-five to sixty-four, 1992–2009 Source: Authors’ calculations from the HRS.

Fig. 1.3F DI participation by health quintile and education, women fifty-five to sixty-four, 1992–2009 Source: Authors’ calculations from the HRS.

Disability Insurance Incentives and the Retirement Decision

1.4.2

63

Descriptive Analysis: Incentive Measures

Before examining the regression results, it is useful to take a closer look at the incentive measures that are the key regressors in those models. Figures 1.4A and 1.4B show the mean values of the OV measures by age for men and women. These figures are constructed by taking a sample of workers at age fifty and computing their incentive measures at all future ages through age sixty-nine; there is no concern of sample selection (e.g., higher income workers being less likely to retire) as the sample ages, as mean OV is calculated using data for all workers, regardless of their ultimate retirement decision. Starting with figure 1.4A, the first thing to note is that the mean for all of the OV measures (OVSS, OVDI, and OVInclusive) is positive, indicating that on average there is some utility gain associated with remaining in the labor force until the optimal future retirement date, whether the individual is contemplating retirement along the SS or DI path. For all measures, the mean value is declining with age, reflecting the fact that the closer one gets to the optimal retirement date, the smaller the utility gain associated with waiting until that date to retire.12 As far as the magnitudes, the OV measures are in utility units rather than in currency units, so the values do not have an easy interpretation, though higher values reflect a larger gain from retirement delay. The values of OVDI are lower than those for OVSS, for reasons I explain below, but have the same pattern of declining with age. The values for OVInclusive are much closer to those of OVSS than OVDI; this is expected, given that OVInclusive is a weighted average of the two and the average DIprobability in the sample is approximately 10 percent, putting more emphasis on OVSS in the calculation. The values for women, shown in figure 1.4D, are lower than for men, as women’s lower average earnings mean that they have less to gain from retirement delays (recall that the OV measures incorporate the value of earnings through retirement as well as the value of SS or DI benefits after retirement). However, the decline with age and relative magnitudes of the different measures display the same patterns observed for men. Some additional insight into these measures, and particularly into the relationship between OVSS and OVDI, can be gleaned from figures 1.4C and 1.4D. These report a simpler measure, the PDV of lifetime SS or DI benefits associated with each possible retirement date. The PDV measures reflect the financial (not utility) gain from additional work if one retires along either the SS or DI path, and include only changes in the value of benefits and not the additional wages that may result from additional work. As figure 1.4C indicates, PDVSS rises moderately with additional work through age sixty-two, the age of SS eligibility, as additional years of earn12. By construction, OV cannot be negative, but it will be zero once the individual has passed his or her optimal retirement date.

64

Courtney Coile

Fig. 1.4A

Mean OV by age for men

Fig. 1.4B

Mean OV by age for women

ings may replace zeroes or low-earnings years in the SS benefit calculation. After age sixty-two, the PDVSS grows more slowly, as an additional year of work is accompanied by a delay in the SS benefit claim that results in the loss of one year of SS benefits (lowering the PDV) but also in a higher actuarial adjustment and permanently higher SS benefits once receipt commences

Disability Insurance Incentives and the Retirement Decision

Fig. 1.4C

Mean PDV-SS and PDV-DI by age, men (2011 euros)

Fig. 1.4D

Mean PDV-SS and PDV-DI by age, women (2011 euros)

65

(raising the PDV); at the mean, the net of these two effects is positive, but modestly so.13 With a 3 percent discount rate, the series essentially peaks at or 13. These results will be sensitive to the choice of the discount rate, since the cost of remaining in the labor force for an additional year is borne now and the benefit is received in the future.

66

Courtney Coile

near the NRA. Here, the values (reported in 2011 euros, for consistency with other studies in this volume) do have a concrete meaning— working from age fifty to sixty-two raises the PDV of SS benefits by about 27,000 euros. The evolution of PDVDI with the age of retirement is much different— PDVDI starts at a much higher value than PDVSS, but declines much more sharply with age thereafter. The reasons for this relate to the differences between SS and DI benefits highlighted above. While additional years in the workforce can raise DI benefits by replacing a zero or low-earnings year with a higher-earnings year, as for SS, this effect is relatively less important for DI because DI uses a shorter averaging period.14 More importantly, DI benefits are available immediately upon DI award (after a five-month waiting period) and are not subject to actuarial adjustment. Therefore delaying onset of DI benefit receipt means a loss of benefits today, with no compensating increase in future benefits. For men, mean PDVDI falls from 270,000 euros if retirement occurs at age fifty to 151,000 euros if it occurs at age sixty-six. As expected, PDVSS and PDVDI for women have lower values but display the same patterns as for men. Returning to figures 1.4A and 1.4B, OVDI can be positive (and declining with age) even when PDVDI peaks at a retirement age of fifty because the OV measures include earnings as well as SS or DI benefits. The replacement rates from SS and DI are fairly low, both in absolute terms and by international standards, and so even though the OV calculation puts a greater value on a dollar of retirement income than a dollar of earnings because of the utility of leisure, it may still be optimal to delay retirement along the DI path even if DI benefits are immediately available in order to accumulate additional years of earnings. Nonetheless, the key point is that the sharply different profiles of PDVSS and PDVDI explain the much lower values of OVDI relative to OVSS in figures 1.4A and 1.4B— there is simply much less to be gained by remaining in the labor force for those retiring along the DI path, relative to the gains available from delaying retirement for those retiring along the SS path. 1.4.3

Regression Results

Finally, we turn our attention to the regression models and results. These models generally take the form: (4)

Rit =  0 +  1 OVit +  2 AGE it +  3 Healthit + 4X it + ε it ,

where retirement (R it ) is a dummy variable equal to 1 if the individual retires during the year (reports being out of the labor force at the following survey 14. To elaborate on this, a fifty-year-old considering retiring now through the SS path would likely have zeroes in the calculation of his PIA for SS benefits, as he is unlikely to have thirty-five years of covered earnings by this point. By contrast, for a fifty-year-old considering retiring now through the DI path, the PIA would be calculated based on only the highest twenty-three years of earnings, so it is less likely that this calculation would include zeroes.

Disability Insurance Incentives and the Retirement Decision

67

year and specifies this year as the year of retirement); OVit is the inclusive option value described above. We also use a “percent change” version of this variable by dividing the option value by the level of utility available by retiring today. The variable AGE represents either a set of age dummies or a linear variable for the individual’s age. The variable Health represents either a set of quintile dummies or the continuous health index. Finally, we include as a set of other controls (Xit) the individual’s marital status, citizenship status, education, occupation, industry, and the spouse’s employment status. The main regression results are presented in table 1.1A. The first key finding is that OVInclusive has a negative and statistically significant effect on earnings. An increase of 10,000 units (which is somewhat smaller than the mean value of OV, which is 14,526) would reduce the probability of retirement by 3.3 percentage points, or about 40 percent relative to the baseline retirement rate of 7.9 percent. The estimates also suggest that a one standard deviation change in the OV (a 14,770-unit change) would lower the probability by 5.6 percentage points. This result is quite consistent across specifications— using age dummies versus linear age or health quintiles versus the continuous health index has little effect on the results. The other coefficients on table 1.1A are much as expected. Health is an important determinant of retirement. In the models using health quintiles, relative to the poorest health group (omitted), those in higher health quintiles are 2.8 to 3.9 percentage points less likely to retire in any given year. The pattern of the four health quintile dummies suggests that the healthiest group has the lowest probability of retirement, though the difference between the lowest quintile and all others is more important than the differences between any of the other quintiles. The linear health index similarly suggests that better health (which is indicated with a larger index value) makes one less likely to retire, though the implied retirement gradient with respect to health is flatter using this continuous measure than that found using the quintiles. The probability of retirement rises with age, and the age dummies (not shown) exhibit the expected spikes at ages sixty-two and sixty-five. In table 1.1B, the standard OVInclusive measure is replaced with the percent change version of this measure. The results suggest that a 100 percent increase in OVInclusive would reduce the probability of retirement by 5.9 percentage points. A 100 percent increase in OVInclusive, evaluated at the mean, would represent something like a 14,000-unit increase. Thus, it seems about right that this effect (5.9 percentage points) is roughly similar to the one standard deviation change effect (5.6 percentage points), since that simulates a similar change in OVInclusive. The next set of tables explore whether the effects seen in tables 1.1A and 1.1B vary by health. In theory, it is not clear whether the impact of a given change in OVInclusive should have a bigger or smaller effect for someone in poor health. On the one hand, poor health may make individuals less likely

Age dummies Female

Age

Health quint 5 (highest) Health index

Health quint 4

Health quint 2 (second lowest) Health quint 3

OV_inclusive

 

Table 1.1A

0.0017 (.0002)

–0.0333 (.0011) [–0.056] –0.0282 (.0022) –0.0302 (0.0022) –0.0353 (.0022) –0.0388 (.0022)

(1)

Included

–0.0325 (.0011) [–0.055] –0.0281 (.0021) –0.0302 (0.0022) –0.0349 (.0021) –0.0385 (.0021)

(2)

Effect of inclusive OV on retirement

–0.0031 (.0022)

0.0019 (.0002)

–0.0338 (.0011) [–0.057] –0.0260 (.0022) –0.0283 (0.0022) –0.0326 (.0022) –0.0362 (.0022)

(3)

Included –0.0031 (.0022)

–0.0331 (.0011) [–0.056] –0.0259 (.0021) –0.0283 (0.0022) –0.0323 (.0022) –0.0360 (.0022)

(4)

–0.0007 (.00004) 0.0015 (.0002)

–0.0332 (.0011) [–0.056]

(5)

Specification

Included

–0.0007 (.00004)

–0.0325 (.0011) [–0.055]

(6)

–0.0037 (.0022)

–0.0006 (.00004) 0.0016 (.0002)

–0.0338 (.0011) [–0.057]

(7)

Included –0.0037 (.0022)

–0.0006 (.00004)

–0.0331 (.0011) [–0.056]

(8)

67,228 0.079 14,526 14,770

67,228 0.079 14,526 14,770

67,228 0.079 14,526 14,770

0.0044 (.0025) –0.0151 (.0022) 0.0002 (.0013) Included 0.0170 (.0040) 0.0100 (.0031) 0.0023 (.0032) 67,228 0.079 14,526 14,770

0.0040 (.0025) –0.0146 (.0021) 0.0000 (.0013) Included 0.0157 (.0039) 0.0091 (.0031) 0.0016 (.0031) 67,228 0.079 14,526 14,770

67,228 0.079 14,526 14,770

67,228 0.079 14,526 14,770

0.0040 (.0025) –0.0149 (.0022) 0.0002 (.0013) Included 0.0170 (.0040) 0.0100 (.0031) 0.0021 (.0032)

67,228 0.079 14,526 14,770

0.0037 (.0025) –0.0144 (.0021) 0.0000 (.0013) Included 0.0159 (.0039) 0.0092 (.0031) 0.0015 (.0031)

Notes: Coefficients are marginal effects of a 10,000-unit change in OV from probit models. Standard errors are shown in parentheses. The effect of a one standard deviation change in OV is shown in brackets (this is estimated as the effect of increasing inclusive OV from the current value –0.5 std. dev. to the current value +0.5 std. dev.).

No. of observations Mean ret. rate Mean of OV Std. dev. of OV

Educ.: Some college

Educ.: High school

Total assets (in millions of euros) Occup. dummies Educ.: t max(i)  Eit ⋅ i ⋅ (gYt −Tt − Ct ) if where Eit indicates whether individual i was employed (= 1), unemployed (= 0.6) or out of the labor force (= 0) in year t, and Eit represents future employment status; αi is the individual specific effect obtained from the fixed effects earnings regressions; g is a gender adjustment factor equal to 0.9 for women and 1.1 for men reflecting gender differentials in average earnings due to discrimination or productivity differences. Further, Yt are historical average gross earnings in year t, Tt is the historical average tax rate in year t, and Ct is the historical average contribution rate to social security, health, and unemployment insurance. Likewise,Yt are predicted average gross earnings in year t, Tt is the predicted average tax rate in year t, and Ct is the predicted average contribution rate to social security, health, and unemployment insurance. We use the predicted social security contribution rates published by the German statutory pension system, but as predicted tax rates we simply use the 2010 values; tmin(i) and tmax(i) are the first and last year an individual is observed in the SOEP. Earnings forecasts are made until 2035 (when the youngest cohort in our data reaches age seventy). In SHARELIFE the respondents are asked to recall any job spell, which lasted six months or longer, using the so-called Life History Calendar (see Schröder 2011). In addition, the first wage at the beginning of each job, characteristics specific to each spell (e.g., occupation, industry, part-time work, etc.), the current wage if the respondent is still working (at the time of the interview), and the last wage at the end of the main job (if the respondent is a retiree in 2008–2009) are obtained during the course of the interview. We use the methodology applied by Weiss (2012) to obtain the complete earnings history for each respondent for the German subsample of SHARELIFE: 1. In a first step, we take the imputed current wage from wave 1 and wave 2 to replace missing values for the current wage in SHARELIFE (for employment spells that started before wave 1 or wave 2). The remaining missing values for first wage, current wage, and last wage of the main job were

Health, Financial Incentives, and Early Retirement for Germany

307

imputed using predictive mean matching (first introduced by Little [1988]). Thus, we end up with complete data on the first job of each job spell, the current job, and the last wage of the main job, which are necessary to project the annual wage during each employment spell.3 2. In a second step, we regress the logarithm of the current wage (last wage of the main job, respectively) on potential labor market experience and its quadratic form, years of schooling and characteristics that are specific to the respective employment spell, that is, dummy variables for type of industry and for white-collar jobs. All explanatory variables are interacted with potential labor market experience and its quadratic form.4 3. The obtained estimated regression coefficients and the first wage of each job spell are then used to predict the wage at the end of each employment spell. Having the wage at the beginning and at the end of an employment spell and the length of each spell at hand we are able to calculate the annual growth of wages during an employment spell. 4. As a last step, we use the growth rate to compute the wage for each year during an employment spell. Weiss (2012) checked the validity of the wage prediction procedure using data from the SOEP. The provided evidence shows that this method for predicting wage works rather well. Net wages are translated into gross wages by using historical tax and social security contribution rates. Earnings points are calculated by comparing gross wages to historic average wages. Finally, we use the same method for SOEP and SHARE to make benefit forecasts for each individual at each planning age from fifty to sixty-nine until 2060, when the youngest cohort reaches age 100. We use the above earnings histories to compute disability pensions, unemployment benefits (UB1 and UB2), and old-age pensions at each retirement based on the relevant legislation. The Option Value of Postponing Retirement To compute the option value of continuing work in utils, we first convert income from work Y into utility using the instantaneous utility function u(Y) = Y γ with parameter γ = 0.75. Then we convert retirement income and/ or unemployment benefits B into utility using the instantaneous utility function v(B) = (kB)γ with parameter values γ = 0.75 and k = 1.5, where k reflects the relative value of leisure. These utility parameters are the same as in all other country chapters of this volume to ensure comparability. The choice 3. Observations were dropped for individuals whose wages at the end of the main job are coded in a foreign currency or are not codeable. Amounts given in German marks (DM) and East German marks are converted into euros. 4. Potential labor market experience is defined as the age in year t minus years of education and age at school entry.

308

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

of 1.5 is based on earlier US estimates (Gruber and Wise 2004) and appears to be very low for European countries. See, for example, Börsch-Supan et al. (2004), who use a grid search algorithm and obtain maximum likelihood estimates between 2.5 and 3.5 for Germany. This difference is exacerbated by a lower choice of γ than in Börsch-Supan et al. (2004). Next we compute the expected future lifetime utility at some planning age a from retiring at any future age R as (4)

EUi (B)Ra =

R −1

T 1 1 s u ( Y ) + s v (Bit ), ∑ it it t − a it t−a t = R (1 + r) t = a (1 + r)



where sit is the conditional probability of being alive in year t (obtained from life tables of the Federal Statistics Bureau) and r is a discount rate arbitrarily set to 3 percent. Here, Yit and Bit are earnings and benefits, respectively, in year t at age a + t. The first part of equation (4) reflects expected and discounted utility during one’s working life and the second part reflects expected and discounted utility during one’s retirement years. In our computations, the maximum attainable age T is arbitrarily set to 100. In t = 0 (at some planning age), the gain in expected lifetime utility from continuing working until age R (compared to receiving benefit B from the next period onward) is equal to (5)

R R a +1 OV( i B )a = EUi (B )a − EUi (B )a .

Thus, the option value of not retiring on benefit B at some planning age a is equal to the maximum gain (across all R) from delaying retirement: (6)

R OV( i B )a = max[OV( i B )a ]. R

Clearly, the larger the OV at any given planning age, the stronger the incentive to postpone retirement. 7.4.3

Weighting the Pathways— The Inclusive Option Value

The option value described in the preceding subsection is computed for a specific type of benefit, that is, for some specific pathway to retirement. However, since individuals in Germany have several options of retiring early, we want to estimate a weighted average of all option values to capture the overall incentive effect to retire. The weights should reflect the probability that either of these paths is a realistic pathway to retirement. Using the employment biographies in the German SOEP data, we estimate these weights— by year, sex, and education level— as the shares of the already retired population age fifty to sixty-nine who have retired through either of the following pathways: 1. Entering retirement on disability benefits (DI path). These are men and women who have changed directly from employment to retirement before the age of sixty and men who have changed from employment to retirement

Health, Financial Incentives, and Early Retirement for Germany

309

at age sixty to sixty-three. Formally, the latter group retires on old-age pensions for the disabled, but we classify these as retired through DI. 2. Entering retirement via unemployment (UB path). These are women who were unemployed before retiring at age sixty or older and men who were unemployed before retiring at age sixty-two and older. 3. Retiring on old-age pensions (OA path). These are women who change directly from employment to retirement at age sixty and older and men who change from employment to retirement at age sixty-three and older. We compute the inclusive option value of not retiring at age a as the weighted average of the three relevant option values: (7)

OV( i a) =  (DI) × OV(DI) i a +  (UB) × OV(UB) i a +  (OA) × OV(OA) i a.

Here, π(DI), π(UB), and π(OA) are the weights of the three pathways and OV(DI), OV(UE), and OV(OA) are the respective option values. For the SHARE data, we use the same approach with two deviations:

• We also include the basic pension intended to provide a minimum stan-

dard of living (Grundsicherung), which was introduced into the pension system in 2003. After calculating the expected earnings-related pension, we check if this pension is above the minimum level. If not, pension benefits are topped up to this level. • For all pathways we check if the respective pathway via DI or unemployment gives higher utility than the regular old-age pension. The assumption for this procedure is that this pathway can always be chosen. Only if DI or unemployment would deliver higher utility is the weighting applied. Based on the SOEP data, figure 7.5, panel (a) shows the option values (in 10,000 utils) of not retiring from each of the three pathways at each planning age between fifty and sixty-nine, separately, for men and women. Several points are worth being highlighted. First, average option values are always larger than zero, that is, given our parameterization of the utility function (standard across all country chapters in this volume) and continuing to work always increases expected lifetime utility, even at ages sixty-five and older. Thus, it seems surprising that anyone actually retires. This pattern is identical when looking at OVs calculated from the SHARE data (not shown here). The relative attractiveness of each pathway is reflected in the value of the OV conditional on age: the smaller the OV, the more attractive a pathway. For men up to age sixty, retiring on DI is the most attractive option on average, followed by unemployment and old-age pension. Considering the stylized benefit streams shown in figure 7.3, this is not surprising. The increase in the DI option value at age sixty seems puzzling at first sight. The reason for

310

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

Fig. 7.5 Average option values of postponing retirement and discounted income streams at planning ages from fifty to sixty-nine Notes: Panel (a): Average option values of postponing retirement via disability (DI), unemployment (UB), or old-age pension (OA) path at planning ages from fifty to sixty-nine. Panel (b): Average present discounted values of earnings/benefit streams when retiring via disability (DI), unemployment (UB), or old-age pension (OA) path at planning ages from fifty to sixtynine. Panels (c) and (d): Average inclusive option value by education and health quintile.

this jump is that from age sixty onward individuals actually do not receive DI proper, but old-age pensions for the disabled (OA-DI). In contrast to those planning to retire on DI (at ages younger than sixty), individuals planning to retire at any future age on OA-DI will not be credited further earnings points for the years they were unable to work. A summary of the present discounted value of the earnings and benefit stream from either pathway is shown in panel (b) of figure 7.5. These figures clearly illustrate the generosity of both DI and UE. The present discounted value of the earnings/benefit stream even falls with age because of the increase in pension benefits. The present discounted value of the oldage pension options rises continuously until the arbitrarily chosen end of the decision period. Figure 7.5 also shows the inclusive option value by age and education (panel [c]) and by age and health (panel [d]). There are substantial differences in option values across education groups, particularly among men. Basic track graduates have lower inclusive option values (thus higher incen-

Health, Financial Incentives, and Early Retirement for Germany

Fig. 7.5

311

(cont.)

tives to retire) than intermediate or academic track graduates at every age. Of course, this difference is due mainly to disparities in earnings streams across education groups. The variation across health quintiles is not as large, but it is clearly present. Individuals in the lowest health quintile (worst health status) have lower option values or higher early retirement incentives. This reflects differences in earnings across health groups, that is, an income gradient in health. Our analysis does not say anything about whether low earnings are due to bad health or bad health due to low earnings, as this is beyond the scope of this chapter. 7.5

Results

Our analytical SOEP sample consists of a panel of 4,109 working individuals with a total of 21,000 observations. The SHARE sample is much smaller with 813 individuals’ contribution to 2,118 observations in four waves. As mentioned before, each individual becomes part of the analytical sample at age fifty and remains in the sample until he or she self-reports as retired, becomes sixty-nine years old, dies, or is lost to follow-up for other reasons.5 5. In SHARE, a small number of spouses of respondents are between ages forty-five and fifty.

312

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

7.5.1

The Effect of Retirement Incentives

Before discussing the results of our retirement regressions, we describe our dependent variable. We consider as a retiree anyone who has self-reported being retired during the observation period. However, we consider as the year of retirement the year in which the individual stopped working. To give an example: if an individual claims to be retired for the first time at age sixty, but was employed for the last time at age fifty-seven, the individual is said to have retired at the age of fifty-eight, no matter if the individual was unemployed or out of the labor force or both during the two years between ending employment and becoming a retiree. In SHARE we are using self-reported retirement from the life history calendar. Respondents are asked for which reason they left their job. If they said “I retired” and this is the last job spell we observe, they are considered retired. This definition is identical to the other SHARE-based studies in this volume. Table 7.5 shows the probit regression results (marginal effects) of five specifications of the explanatory variables for the SOEP and the SHARE data, respectively. Model (1) includes a linear age effect and no covariates, (2) includes age dummies and no covariates, (3) includes a linear age effect and covariates, (4) includes age dummies and covariates, and (5) includes sex-specific age dummies and covariates. Model (4) is the common specification in this volume, of which we discuss the quantitative results. In general, results are clearer when using age dummies, but the qualitative findings are robust across the different specifications. The OV coefficient reflects the percentage point effect of increasing the inclusive option value by 10,000 utils. Note that these are marginal effects evaluated at the means of the explanatory variables. The effect is highly significant in all specifications based on the SOEP data. This also holds for the estimates based on the SHARE data when age is specified as a set of dummy variables (columns [2] and [4]). A 10,000-util increase in the option value reduces the probability of retirement by approximately 4.2 percentage points in the SOEP data and by 2 percentage points in the SHARE data. Put in relative terms, this means that a 73 percent increase in the option value relative to the overall mean of 13,770 reduces retirement hazards by about 80 percent compared to the overall average of 5.3 percent in the SOEP data. The SHARE-based estimate is substantially smaller: a similar 76 percent increase in the option value relative to the overall mean of 13,065 reduces retirement hazards by only about 36 percent compared to the overall average of 5.5 percent. The smaller effects in SHARE compared to SOEP may have various reasons. The SHARE data include the topping-up of low pensions via social assistance (Grundsicherung im Alter), which makes the option value flat for individuals with low pensions. Moreover, the samples are different: SHARE

Table 7.5

Probit regressions explaining the decision to retire (marginal effects), option value enters in absolute value Specification (1)

Option value (10,000 utils)

Health quintile 2 Health quintile 3 Health quintile 4 Health quintile 5 Linear age Age dummies Age dummies × sex Covariates

(2)

(3)

Panel (a). SOEP –0.0317 –0.0349 –0.0379 (0.0029) (0.0027) (0.0034) [–0.0323] [–0.0407] [–0.0404] –0.0140 –0.0127 –0.0136 (0.0025) (0.0022) (0.0024) –0.0179 –0.0153 –0.0173 (0.0025) (0.0022) (0.0024) –0.0181 –0.0158 –0.0170 (0.0025) (0.0022) (0.0024) –0.0162 –0.0140 –0.0153 (0.0026) (0.0023) (0.0025) X

(4)

(5)

–0.0423 (0.0031) [–0.0521] –0.0122 (0.0021) –0.0150 (0.0021) –0.0150 (0.0021) –0.0133 (0.0022)

–0.0421 (0.0033) [–0.0525] –0.0118 (0.0020) –0.0148 (0.0021) –0.0147 (0.0021) –0.0131 (0.0021)

X

X X

X

No. obs. Mean retirement rate Mean OV Std. dev. OV

21,027 0.052 13,704 7,549

21,027 0.052 13,704 7,549

20,915 0.053 13,771 7,513

20,915 0.053 13,771 7,513

OV (10,000 utils)

Panel (b). SHARE –0.010 –0.022 –0.007 (0.005) (0.006) (0.005)

–0.020 (0.007)

–0.018 (0.007)

–0.022 (0.011) –0.017 (0.011) –0.020 (0.011) –0.020 (0.011)

–0.032 (0.015) –0.021 (0.016) –0.026 (0.016) –0.028 (0.015)

–0.034 (0.015) –0.026 (0.016) –0.028 (0.016) –0.030 (0.015)

X

Health quintile 2 Health quintile 3 Health quintile 4 Health quintile 5 Age linear Age dummies Age dummies × sex Controls N Mean retirement rate Mean OV Std. dev. OV

20,915 0.053 13,771 7,513

X

X X X

–0.031 (0.015) –0.022 (0.015) –0.026 (0.016) –0.026 (0.015)

X

–0.022 (0.011) –0.017 (0.011) –0.019 (0.012) –0.021 (0.011) X

X

2,118 0.083 13,065.62 8,612.53

2,118 0.083 13,065.62 8,612.53

X

X

X X X

2,118 0.083 13,065.62 8,612.53

2,118 0.083 13,065.62 8,612.53

2,118 0.083 13,065.62 8,612.53

314

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

Table 7.6

Probit regressions explaining the decision to retire (marginal effects), option value enters in relative terms Specification

OV utility gain Health quintile 2 Health quintile 3 Health quintile 4 Health quintile 5 Linear age Age dummies Age dummies × sex Covariates No. obs. Mean ret. rate Mean OV gain Std. dev. OV gain

(1)

(2)

(3)

(4)

(5)

–0.0132 (0.0073) –0.0166 (0.0027) –0.0220 (0.0026) –0.0221 (0.0027) –0.0204 (0.0027)

–0.0137 (0.0074) –0.0164 (0.0025) –0.0210 (0.0025) –0.0214 (0.0025) –0.0197 (0.0026)

–0.0130 (0.0074) –0.0162 (0.0027) –0.0209 (0.0026) –0.0208 (0.0026) –0.0191 (0.0027)

–0.0139 (0.0077) –0.0161 (0.0025) –0.0203 (0.0025) –0.0205 (0.0025) –0.0187 (0.0025)

–0.0131 (0.0073) –0.0157 (0.0025) –0.0201 (0.0024) –0.0201 (0.0024) –0.0187 (0.0025)

X

X

X X

21,027 0.0522 0.6020 0.5404

20,915 0.0525 0.6050 0.5403

X

X

X X X

21,027 0.0522 0.6020 0.5404

20,915 0.0525 0.6050 0.5403

20,915 0.0525 0.6050 0.5403

includes respondents from East Germany and has a substantially later sample. Finally, the precision of measurement is different; SHARE has a more encompassing health measure, while SOEP has more precise earnings histories (see below). The numbers in square brackets in panel (a) show the average effect of a one standard deviation change in the inclusive option value— computed as the average predicted retirement probability when the option value is 0.5 standard deviations larger than its actual value minus the average predicted retirement probability when the option value is 0.5 standard deviations smaller than its actual value, averaged across all observations. The estimates indicate that a one standard deviation increase in OV reduces retirement rates by about 5.9 percentage points or nearly 100 percent of the average. The difference between the effect sizes found for the two alternative specifications can be partly explained by a strongly nonlinear (concave) relationship between OV and retirement hazards. The larger the option value at baseline, the smaller the effect of one util. Yet another estimated specification of the relation between the inclusive option value and the probability to retire can be found in table 7.6, shown for the SOEP data only. Here we show the effect of the utility gain if labor force participation is continued until the optimal age relative to the utility when retiring at the planning age. In other words, this is the percentage gain

Health, Financial Incentives, and Early Retirement for Germany

315

in future lifetime utility from choosing the optimal retirement age instead of retiring now. This measure has a few advantages over the absolute number of utils. For instance, absolute option values very much reflect the individual’s income level. A 10,000-utils option value of postponing retirement may thus mean a lot more (have a stronger incentive effect) to a low-income than to a high-income worker. Related to this, the relative utility gain might be more useful in international comparisons when income levels differ across countries. The reported coefficients reflect the effect size of a 100 percent lifetime utility gain from postponing retirement. Again, results show a decline in retirement rates when the lifetime utility gain from postponing retirement rises. Our estimates show that doubling this gain is linked with a 1.3 percentage point or 20 percent decrease in retirement hazard rates. Note, however, that standard errors are much larger (relative to the point estimates) than before, so that the coefficients are significant at the 10 percent level only. 7.5.2

The Effect of Health

We now turn to the effect of health on retirement rates. Table 7.5 shows the coefficients of health quintile dummies with those in the worst health quintile being the reference group. Respondents in higher quintiles are in better health. Our results clearly show that health has a significant relationship with retirement rates. Healthier individuals have lower retirement rates. The results also show that the relationship is highly nonlinear. Respondents in the baseline quintile have about 1.2 to 1.5 percentage points higher retirement hazard rates than those in the second to fifth quintiles (among which there is no big difference) in the SOEP sample. The SHARE sample exhibits a substantially larger effect of health on retirement, indicating that respondents in the baseline quintile have about 2.1 to 3.2 percentage points higher retirement hazard rates than those in the second to fifth quintiles. It is noteworthy that the nonlinear shape of the health effects is very similar for both the SOEP and the SHARE sample. Compared to the overall average retirement rate of 5.3 (8.3) percent, this suggests having a severe health shock and moving from the fifth to the first health quintile would increase retirement probabilities by more than 20 percent (35 percent) in SOEP (SHARE, respectively). “Mild” health shocks, for example moving from the fourth to the second health quintile, do not seem to have any effect on retirement rates. The larger marginal effects in the SHARE data compared to the SOEP, together with the substantially smaller effects of the option value, may reflect the more encompassing health measure in the SHARE data. The SOEP-based estimates of the financial incentives may therefore be an upper bound. Since the earnings histories in the SOEP are probably more precisely estimated than those based on the SHARELIFE data, the SHARE-based estimates of the financial incentives may rather be a lower bound.

316

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

7.5.3. The Effect of Financial Incentives by Health Status Having established independent effects of health and financial incentives on retirement rates, it is natural to ask whether the two interact. Specifically, the question is whether healthier individuals are more or less responsive to financial incentives. Casual reasoning suggests that the labor supply of sick individuals should be less elastic than the labor supply of healthy individuals simply because sick individuals’ choices are constrained by bad health. On the other hand, one might argue that individuals who are in bad health and consider early retirement might be more responsive to financial incentives than healthy individuals who have no plans to retire early. Both effects are not mutually exclusive and may even cancel each other out. Table 7.7, based on the SOEP data, shows the effect of the inclusive option value on retirement rates separately for each health quintile. This is estimated from separate retirement regressions for each of the health quintiles. We have estimated the effect of a 10,000-util increase, of one within-quintile Table 7.7

Marginal effects of inclusive option value on the decision to retire (obtained from probit models) by health quintile Specification

HQ1

Option value Std. dev. effect Option value gain

HQ2

Option value Std. dev. effect Option value gain

HQ3

Option value Std. dev. effect Option value gain

HQ4

Option value Std. dev. effect Option value gain

HQ5

Option value Std. dev. effect Option value gain

(1)

(2)

(3)

(4)

(5)

–0.0690 (0.0088) [–0.0498] –0.0207 (0.0163) –0.0351 (0.0076) [–0.0331] –0.0043 (0.0099) –0.0296 (0.0040) [–0.0369] –0.0232 (0.0132) –0.0155 (0.0050) [–0.0369] –0.0199 (0.0091) –0.0151 (0.0053) [–0.0189] –0.0193 (0.0088)

–0.0761 (0.0086) [–0.0582] –0.0221 (0.0163) –0.0364 (0.0068) [–0.0413] –0.0046 (0.0093) –0.0272 (0.0039) [–0.0508] –0.0230 (0.0119) –0.0173 (0.0044) [–0.0508] –0.0196 (0.0083) –0.0158 (0.0049) [–0.0268] –0.0179 (0.0080)

–0.0792 (0.0110) [–0.0587] –0.0146 (0.0151) –0.0439 (0.0079) [–0.0458] –0.0049 (0.0091) –0.0314 (0.0045) [–0.0444] –0.0206 (0.0133) –0.0173 (0.0058) [–0.0444] –0.0203 (0.0097) –0.0219 (0.0057) [–0.0218] –0.0272 (0.0097)

–0.0902 (0.0105) [–0.0707] –0.0164 (0.0152) –0.0453 (0.0067) [–0.0576] –0.0055 (0.0088) –0.0285 (0.0043) [–0.0628] –0.0214 (0.0122) –0.0195 (0.0050) [–0.0628] –0.0203 (0.0088) –0.0219 (0.0050) [–0.0320] –0.0248 (0.0086)

–0.0894 (0.0107) [–0.0722] –0.0147 (0.0142) –0.0469 (0.0074) [–0.0579] –0.0062 (0.0085) –0.0123 (0.0029) [–0.0722] –0.0111 (0.0056) –0.0235 (0.0066) [–0.0722] –0.0234 (0.0104) –0.0123 (0.0031) [–0.0341] –0.0139 (0.0051)

Health, Financial Incentives, and Early Retirement for Germany

317

standard deviation increase and of a 100 percent increase in future lifetime utility. The qualitative results are robust across all specifications, but not across different measures of the early retirement incentives. Let us begin by looking at the marginal effect of a 10,000-util increase in the option value to postpone retirement, that is, when the incentives are specified in absolute terms. Here, we find the biggest effect of the inclusive option value in the first health quintile. The estimated effects decrease up to the fourth quintile and then increase again (slightly). These findings appear to be consistent with the notion that sick individuals for whom early retirement is a more salient option are also more influenced by the financial incentives. At this point one should also keep in mind that average option values are positive everywhere and that those in good health have particularly high option values. Thus, making early retirement more or less attractive will probably not make a large difference to them. In contrast, those in worse health have lower option values anyway, so that changes in option values matter more. The opposite mechanism— sick individuals being so constrained in their labor supply decisions that financial incentives hardly matter— seems less prevalent. The estimated effects of a one within-quintile standard deviation change in the OV are qualitatively similar to those discussed in the preceding paragraph. However, the decline in the OV effect across quintiles is weaker. This may be partly due to the fact that the OV standard deviation is larger among the healthy (and richer) than among the unhealthy. In contrast to the findings above, where the sickest quintile stood out, it seems that it is now the healthiest quintile that is much less responsive to early retirement incentives than the rest. Finally, when looking at another relative incentive measure, the percent gain in lifetime utility, we find no clear pattern. If anything, the incentive effect seems to be slightly increasing with health status. It is, of course, unfortunate that different specifications yield somewhat different conclusions. The choice of specifications is rather ad hoc, and we have no theory that would tell us which specification to prefer. We believe a fair summary of the results shown in table 7.6 is that there is hardly any evidence that the labor supply of the sickest quintile is the least elastic at the extensive margin. Rather, depending on specification of the incentive variable, their reaction to financial incentives is at least as strong as the reaction of the healthier quintiles— if not stronger. 7.5.4

The Effect of Financial Incentives by Education Level

Our final set of results, again based on SOEP data, relates to differences in retirement behavior by education groups. As noted above, a good way to distinguish education levels in Germany is by type of school-leaving certificate or type of secondary school track attended: basic track, intermediate, and academic track. We have already seen that employment rates are lower for the less educated. Correspondingly, retirement rates are generally larger

318

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

Table 7.8

Marginal effects of inclusive option value on the decision to retire (obtained from probit models) by education Specification

Basic

Option value Std. dev. effect Option value gain

Intermediate

Option value Std. dev. effect Option value gain

Academic

Option value Std. dev. effect Option value gain

(1)

(2)

(3)

(4)

(5)

–0.0344 (0.0062) –0.0240 –0.0063 (0.0067) –0.0415 (0.0050) –0.0477 –0.0205 (0.0125) –0.0270 (0.0044) –0.0406 –0.0505 (0.0220)

–0.0392 (0.0063) –0.0291 –0.0068 (0.0070) –0.0409 (0.0047) –0.0518 –0.0212 (0.0122) –0.0230 (0.0033) –0.0620 –0.0387 (0.0172)

–0.0471 (0.0068) –0.0340 –0.0069 (0.0071) –0.0467 (0.0051) –0.0592 –0.0238 (0.0135) –0.0296 (0.0046) –0.0463 –0.0512 (0.0209)

–0.0511 (0.0067) –0.0391 –0.0069 (0.0073) –0.0457 (0.0049) –0.0627 –0.0254 (0.0134) –0.0238 (0.0034) –0.0686 –0.0398 (0.0159)

–0.0504 (0.0072) –0.0394 –0.0064 (0.0065) –0.0423 (0.0050) –0.0615 –0.0228 (0.0125) –0.0171 (0.0027) –0.0734 –0.0298 (0.0116)

among the less educated. Of workers who completed the basic track, about 5.5 percent retire annually, whereas of workers who completed the academic track, only 4.7 percent retire annually. As we have seen above, differences in average inclusive option values across education groups are substantial (see figure 7.5, panel [d]), reflecting well-known differences in lifetime income by education group. In terms of the strength of the option value effect on retirement rates, we find sizable differences across education groups and across specifications (see table 7.8). When estimating the percentage point effect of a 10,000-util increase using our standard specification (4), the size of the coefficient drops from 5.1 percentage points among basic track graduates to 2.4 percentage points among academic track graduates. However, when we standardize the effect size, for instance, by looking at the effect of one within-education group standard deviation increase in future lifetime utility, we tend to find the opposite result: academic and intermediate track graduates actually become more responsive than basic track graduates. This can be partly explained by the fact that the standard deviation in the utility gain is nearly twice as large in the highest than in the lowest education group. Also, when we look at the effect of a 100 percent lifetime utility gain, it is stronger among the better educated than among the less educated. 7.6

Understanding the Results and Their Implications

After having established that the health and financial incentive effects are qualitatively similar in the two data sets, we now exploit the longer time hori-

Health, Financial Incentives, and Early Retirement for Germany

319

zon and the richer policy variation in the SOEP data to better understand the implications of our results for pension policy. 7.6.1

The Model Fit

Our first step in understanding the above results is to show how well our estimated models fit the actual retirement behavior. Figure 7.6 compares, separately for men and women, the actual and predicted retirement hazard rates and the actual and predicted labor force survival rates (starting at age fifty). Note that the y axis for hazard rates is on a logarithmic scale. Both actual and predicted log retirement rates are practically the same. This was to be expected since our preferred regression specification contains age dummies interacted with sex that should be able to pick up much of the variation in retirement across the age distribution. Hazard rates increase almost linearly until age sixty-five. In other words, retirement hazards increase exponentially with age. The spikes at younger ages are due to sampling variation, but the spikes at older ages (sixty, sixty-three, or sixty-five) can be explained by the provisions of the pension system. At age sixty many women could retire on old-age pensions for women and men could retire on either old-age pensions for the unemployed or for the disabled. Age sixty-five is the regular retirement age. Figure 7.6, panels (e) to (h), show predicted retirement hazard rates for men and women by health quintile and by education. Again, this is shown on a log scale. The figures show that the main difference in retirement rates is between the first (least healthy) health index quintile and the other health quintiles. This holds across the entire age range but seems to be particularly pronounced below the age of sixty, that is, where the most salient early retirement option is DI. Differences in predicted retirement hazards by age also clearly show an education gradient. Academic track graduates have the lowest retirement hazards across most of the age distribution. The difference is very pronounced among men. In contrast, there is hardly any difference among women older than fifty-six. 7.6.2

Relationship between OVs and Retirement

We have seen above that the average inclusive option values are always positive. Thus, on average, there are at every age financial incentives to continue working. Another way to study the relationship between option values and retirement behavior is to find— for each individual— the age at which the incentive to retire reaches a maximum (hence the inclusive option value reaches a minimum) and to compute the cumulative proportion of respondents who have reached that age. Then we compare this proportion with the cumulative proportion of individuals who have retired. If reaching the minimum of one’s option value to postpone retirement affected actual retirement behavior strongly, we should see something like a one-to-one relationship between the two series. Figure 7.7 shows that the two series are positively related but at the end of the observation period (at age sixty-nine) only 20

Fig. 7.6 Model fit. Actual versus predicted average retirement rates and survival rates by sex, education, and health

Health, Financial Incentives, and Early Retirement for Germany

321

Fig. 7.7 Percent of individuals who have reached the minimum option value (maximum retirement incentive) and actual retirement rate (by age)

percent of individuals have reached their OV minimum, whereas more than 90 percent have retired. Thus, the predictive value of the OV for aggregate retirement rates seems limited. 7.6.3

Simulation of Pension Reform

In this section we show the result of a number of simulated changes to the pension system and compare them to the status quo. We will simulate four different scenarios by varying the option value to postpone retirement for each individual. First, we assume that everyone is able to retire on DI if they want to (maximum lenience), by giving a weight of one to the DI path and zero weight to the alternatives (100 percent DI probability). Second, we simulate the other extreme by taking away the possibility to retire on DI completely and giving zero weight to the DI path in the inclusive option value (0 percent DI probability, maximum stringency). Since in contrast to, for example, the United States, we have more than two retirement paths, and it is not a priori clear how to distribute the DI probability that prevails in the status quo to the other two paths. One possibility is to change the probabilities of the unemployment and the old-age pension path so that they remain proportional to the status quo. A behavioral interpretation of this possibility is that workers who are denied the DI path will partly choose the UB path and partly chose the OA path. We believe this is unrealistic. Rather, we think that workers will try to get on the “next best” retirement path, which is UB. Thus, in the 0 percent DI probability scenario, we keep the original OA weight constant but increase the UB weight.

322

Fig. 7.8

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

Simulation of pathway probabilities

Further, we model two intermediate scenarios by assuming a one-third and a two-thirds probability of receiving DI upon application. Again, this begs the question of how to choose the relative weights of the other two paths. We illustrate our modeling decision in figure 7.8. The x axis shows the assumed variation in DI probabilities and the y axis the implied DI, OA, and UB pathway probabilities. The vertical line shows the overall status quo probabilities. Starting from the status quo and increasing the stringency of the award process means going left, decreasing stringency means going right. When we increase the stringency, we assume that all workers who are denied DI consider taking the UB path instead. When we reduce the stringency, we assume that first workers who would otherwise have taken the “next best” UB path will consider DI. Only if there are no workers left from the UB path do workers who have taken the OA path consider taking the DI path. Retirement hazards and labor force survival rates are predicted using our baseline estimation equation (specification [4] in table 7.4). The results of our simulations can be seen in figure 7.9. Retirement hazard rates are largest under a 100 percent DI probability regime and smallest under a 0 percent DI probability regime. The other regimes are in between. To summarize the data shown in figure 7.9, we summed up the survival rates from age fifty to age sixty-nine to get at the expected number of work-

Health, Financial Incentives, and Early Retirement for Germany

323

Fig. 7.9 Simulated log hazard and survival rates under different scenarios (all individuals)

ing years at age fifty associated with each scenario. This is shown in table 7.9, part (a). The average number of working years over the fifty to sixty-nine age interval in our sample is 11.5 years among men and 11 years among women, which implies an average retirement age of 61.5 and 61 years, respectively. Note that this is very close to the average in administrative data as shown in figure 7A.1 in the appendix. Closing the DI path altogether will increase the number of working years by about 0.6 years among men but only 0.2 years among women. Opening up the DI path for everyone, for instance by reducing the stringency of the award process, so that everyone who would apply would be accepted, would decrease the average number of working years by 1.6 years among men and 2.4 years among women. Compared to the overall length of retirement of about seventeen years (Deutsche Rentenversicherung Bund 2012), these changes appear to be moderate. Next, we restrict the sample to the group of people who should have the most reason to apply for disability benefits, namely workers in bad health.

324

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

Table 7.9

Expected years of work (at age fifty) under various scenarios Status quo

0% DI prob.

One-third DI prob.

Two-thirds DI prob.

100% DI prob.

Part (a). Full sample Men Expected years of work Abs. diff. w.r.t. status quo Rel. change w.r.t. status quo Women Expected years of work Abs. diff. w.r.t. status quo Rel. change w.r.t. status quo

11.49

12.13 0.64 1.06

11.68 0.19 1.02

11.29 –0.20 0.98

9.94 –1.55 0.87

11.41

11.58 0.17 1.01

11.33 –0.09 0.99

10.15 –1.26 0.89

9.05 –2.36 0.79

10.80 –0.21 0.98

9.46 –1.55 0.86

9.61 –1.38 0.87

8.46 –2.53 0.77

Part (b). Ever in first health quintile Men Expected years of work 11.01 11.81 11.32 Abs. diff. w.r.t. status quo 0.80 0.31 Rel. change w.r.t. status quo 1.07 1.03 Women Expected years of work 10.99 11.15 10.88 Abs. diff. w.r.t. status quo 0.16 –0.11 Rel. change w.r.t. status quo 1.01 0.99

We do not show hazard rates but report only the average number of working years in association with our scenarios (table 7.9, part [b]). For this analysis, we include only individuals who were in the first health quintile at least once during the entire observation period. As it turns out, however, the estimated effect of changing the stringency of the disability benefits is only slightly smaller than the effect estimated for the full sample. This somewhat unexpected finding may be due to the fact that the common specification used for our simulations does not allow for differences in the size of incentive effects by health status. 7.7

Conclusion

In the light of continuing demographic change and ailing labor markets, pension reform remains high on the political agenda in many countries. Owing to a number of recent reforms, the German pension system appears to be on a more financially sustainable path today than it used to be ten or fifteen years ago. However, these reforms have reduced the generosity of the pension system and thus rescinded what many Germans used to view as considerable welfare state achievements. Reducing generosity has two effects on the social security budget (Börsch-Supan, Kohnz, and Schnabel 2007): a direct (mechanical) effect, by changing contributions and benefits for a given work history, and an indirect effect through behavioral responses to the reform, that is, more contributions and less benefits due to longer work-

Health, Financial Incentives, and Early Retirement for Germany

325

ing lives. These two effects can also be found on the level of the individual worker, and their relative strength depends on the elasticity of labor supply, which in turn may depend to a considerable extent on individual health. The aim of this chapter was to expand and complement earlier microsimulation studies on the German pension system by a more systematic treatment of health and disability. More than 20 percent of the workforce eligible for public pensions enters retirement first on disability pensions, at an average age of only slightly more than fifty years. While disability uptake rates have been fairly constant in the last three decades, important changes can be found with respect to the type of health problems that trigger early retirement. Retirement on disability pensions due to cardiovascular health problems has declined from nearly 40 percent to less than 10 percent and has been largely replaced by retirement due to mental health problems, which now are the primary diagnosis for disability in more than 40 percent of all early retirees (Börsch-Supan and Jürges 2012). We address two new questions in this chapter: First, to what extent do financial incentives to retire— measured by the option value to postpone retirement by one year— affect the retirement decisions of sick or disabled individuals? Put differently, does bad health reduce an individual’s labor supply elasticity at the extensive margin? The answer to this question has important policy implications. If the sick and disabled are not responsive to financial incentives because their labor supply has become inelastic, policies that aim at reducing the generosity of disability benefits and providing less incentives to retire early are only partially successful. They may primarily hurt those for whom disability pensions are an important part of the welfare system. In contrast, if disability pension recipients respond to reductions in generosity by postponing retirement, these reductions are less harmful than in the first case. Unfortunately, our empirical results with respect to that question are somewhat ambiguous, but the least common denominator is that among the least healthy financial incentives do not matter less than among healthier segments of the workforce. Given this result our next question is, how do changes to the stringency of the disability award process affect labor supply and average retirement age? Simulating the effect of those changes in stringency, such as stricter medical requirements, can be quite challenging. Therefore, we took a very simple approach and varied the likelihood of receiving disability pensions from 0 to 100 percent, that is, from a totally restrictive to a totally lenient reward process, and compared the average retirement age or the average length of the remaining working life to the status quo. We found that increasing the stringency of the award process so that the DI path is closed altogether will increase the expected retirement age by only 0.6 years among men and 0.2 years among women, based on SOEP data. In contrast, reducing the stringency to the extent that every applicant would be accepted would decrease the average number of working years quite substantially, namely, by 1.6

326

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

years among men and 2.4 years among women. These simulated changes seem moderate, in particular considering the fact that the SHARE-based elasticities were even smaller. However, what needs to be taken into account is that our simulations make an important assumption: individuals who are denied the DI path consider taking the next best path to early retirement (which is unemployment). Thus, making the disability benefit award process more stringent without closing this other fairly generous early retirement route would not greatly increase labor force participation in old age. There are several methodological caveats requesting further research. First, the SHARE data exhibit a substantially smaller marginal effect of the option value, and at the same time larger effects of the health variables, compared to the estimates based on the SOEP data. This may reflect the more encompassing health measure in the SHARE data. On the other hand, earnings histories are probably better captured by the SOEP data than by SHARELIFE. The SOEP-based estimates of the financial incentives may therefore be an upper bound, while the SHARE-based estimates may rather represent a lower bound. Second, the utility parameters embedded in the option value have not been estimated but are fixed at values based on earlier US estimates. While this ensures comparability to all other country chapters in this volume, the choice of the parameter describing the value of leisure appears to be very low for European countries. In future work, these parameters should be estimated by maximum likelihood methods. Third, and finally, the chosen functional form of the estimation equation is a combination of a probit model (and its underlying random utility logic) with an option value that has its own utility function embedded; in particular, a very specific functional form of the trade-off between labor and leisure. While this pragmatic approach has worked well in many circumstances, it is not an internally consistent model of labor supply that has failed in other circumstances (see Börsch-Supan 2012). Further research is necessary to shed light on the environments in which our approach is a reliable approximation of an internally consistent model.

Health, Financial Incentives, and Early Retirement for Germany

327

Appendix Table 7A.1 Age 58

Eligibility for UB1: Maximum number of months by age and year

until 1984

1985

1986–1987

Years 1988–1997

12

16

22

1998–1/2006 2/2006–2007 12 18

12 22

20

12

since 2008

12

26 15

18

26 24

18

32

18 32

Table 7A.2

24

Retirement rates by health, education, and sex (age fifty-five to sixty-four) Health quintile (%)

Education Men Basic track Intermediate track Academic track All Women Basic track Intermediate track Academic track All

1

2

3

4

5

All

46.93 35.90 30.98 42.85

41.38 30.37 20.78 35.61

33.40 28.79 18.04 28.75

27.95 22.50 16.46 23.64

25.75 21.97 13.87 21.62

37.64 28.30 19.92 32.09

36.68 37.95 33.10 36.57

29.42 27.23 22.56 28.18

24.63 24.05 17.51 23.52

23.77 19.97 15.38 21.59

23.05 21.72 12.47 20.94

29.37 27.21 20.91 27.86

Source: Calculation combines data from 1984–2009 SOEP.

Fig. 7A.1

Average retirement ages, West Germany (1960–2012)

Fig. 7A.2

Retirement rates over time by sex, age groups, education, and health

Source: Authors’ own computations using SOEP data.

Health, Financial Incentives, and Early Retirement for Germany

Fig. 7A.2

329

(cont.)

References Börsch-Supan,  A. 2005. “Work Disability and Health.” In Health, Ageing, and Retirement in Europe—First Results from the Survey of Health, Ageing and Retirement in Europe, edited by A. Börsch-Supan et al., 253–58. Mannheim, Germany: Mannheim Research Institute for the Economics of Aging. ———. 2011. „Gesamtwirtschaftliche Wirkungen der Rentenreformen.“ In Handbuch der gesetzlichen Rentenversicherung SGB VI, edited by  E. Eichenhofer, H. Rische, and W. Schmähl, 251–305. Munich: Luchterhand Verlag Gmbh. ———. 2012. “Note on the Stock-Wise Utility Function Used in their Option-Value Analysis.” Discussion Paper, Munich Center for the Economics of Aging. Börsch-Supan, A., and H. Jürges. 2012. “Disability, Pension Reform, and Early Retirement in Germany.” In Social Security Programs and Retirement around the World: Historical Trends in Mortality and Health, Employment, and Disability Insurance Participation and Reforms, edited by D. Wise. Chicago: University of Chicago Press. Börsch-Supan, A.,  S. Kohnz, and  R. Schnabel. 2007. “The Budget Impact of Reduced Early Retirement Incentives on the German Public Pension System.” In Social Security Programs and Retirement around the World: Fiscal Implications of Reform, edited by J. Gruber and D. Wise, 201–52. Chicago: University of Chicago Press. Börsch-Supan, A., and R. Schnabel. 1999. “Social Security and Retirement in Germany.” In Social Security and Retirement around the World, edited by J. Gruber and D. Wise. Chicago: University of Chicago Press. Börsch-Supan, A., R. Schnabel, S. Kohnz, and G. Mastrobuoni. 2004. “MicroModeling of Retirement Decisions in Germany.” In Social Security Programs and Retirement around the World: Micro-Estimation, edited by J. Gruber and D. Wise. Chicago: University of Chicago Press. Börsch-Supan, A., and C. B. Wilke. 2007. “The German Public Pension System: How It Will Become an NDC System Look-Alike.” In Pension Reform—Issues and Prospects for Non-Financial Defined Contribution (NDC) Schemes, edited by Robert Holzmann and Edward Palmer, 573–610. Washington, DC: The World Bank.

330

Jürges, Thiel, Bucher-Koenen, Rausch, Schuth, and Börsch-Supan

Commission for Sustainability in Financing the Social Security Systems (Kommission für die Nachhaltigkeit in der Finanzierung der Sozialen Sicherungssysteme). 2003. Final Report (Abschlußbericht). Berlin: Bundesministerium für Gesundheit und Soziale Sicherheit. Deutsche Rentenversicherung Bund. 2012. “Rentenversicherung in Zeitreihen. Ausgabe 2012.” DRV-Schriften Band 22. Dustmann, C. 2004. “Parental Background, Secondary School Track Choice, and Wages.” Oxford Economic Papers 56:209–30. Federal Ministry of Finance. 2013. Durchschnittsbelastung nach Tarifen 1958 bis 2014. Retrieved from www.bmf-steuerrechner.de/uebersicht_ekst/. Gruber, J., and D. Wise, eds. 2004. Social Security Programs and Retirement around the World: Micro-Estimation. Chicago: University of Chicago Press. Jürges, Hendrik. 2009. “Healthy Minds in Healthy Bodies. An International Comparison of Education-Related Inequality in Physical Health among Older Adults.” Scottish Journal of Political Economy 56 (3): 296–320. Jürges, Hendrik, and Kerstin Schneider. 2011. “Why Young Boys Stumble: Early Tracking, Age and Gender Bias in the German School System.” German Economic Review 12 (4): 371–94. Kemptner, Daniel, Hendrik Jürges, and Steffen Reinhold. 2011. “Changes in Compulsory Schooling and the Causal Effect of Education on Health: Evidence from Germany.” Journal of Health Economics 30:340–54. Little, R. J. A. 1988. “Missing-Data Adjustments in Large Surveys.” Journal of Business and Economic Statistics 6:287–96. Poterba, J., S. Venti, and D. Wise. 2010. “The Asset Cost of Poor Health.” NBER Working Paper no. 16389, Cambridge, MA. Riphahn, Regina T. 1999. “Income and Employment Effects of Health Shocks. A Test Case for the German Welfare State.” Journal of Population Economics 12 (3): 363–89. Romeu Gordo, L. 2006. “Effects of Short- and Long-term Unemployment on Health Satisfaction: Evidence from German Data.” Applied Economics 38 (20): 2335–50. Schröder, M. 2011. “Retrospective Data Collection in the Survey of Health, Ageing and Retirement in Europe.” SHARELIFE Methodology. Mannheim, Germany: Mannheim Research Institute for the Economics of Aging. Stock, J. H., and D. A. Wise. 1990. “The Pension Inducement to Retire: An Option Value Analysis.” In Issues in the Economics of Aging, edited by D. A. Wise, 205–30. Chicago: University of Chicago Press. Wagner, Gert G., Joachim  R. Frick, and Jürgen Schupp. 2007. “The German Socio-Economic Panel Study (SOEP)— Scope, Evolution, and Enhancements.” Schmollers Jahrbuch 127 (1): 139–69. Weiss, C. T. 2012. “Two Measures of Lifetime Resources for Europe using SHARELIFE.” SHARE Working Paper Series no. 06–2012.

8

Health, Disability Insurance, and Retirement in Denmark Paul Bingley, Nabanita Datta Gupta, Michael Jørgensen, and Peder J. Pedersen

8.1

Introduction

Labor force participation of older persons varies greatly both between countries and within countries over time. Individual health status, labor market conditions, and social security program provisions all play a role in this. Disability insurance (DI) programs are at the interface between social security provisions, labor market conditions, and health and may play an important role for many persons as they move from employment to retirement from the labor market. In principle, it may be the case that changes in DI participation rates reflect changing health and changing labor market conditions. However, trends in DI participation appear to be unrelated to changes in mortality and health. Differences in health between countries would need to be much larger than those revealed in comparable survey data in order to account for differences in DI participation (Milligan and Wise 2012). In many countries, DI effectively provides early retirement benefits before eligibility for other social security programs begin. This begs the main question: Given health status, to what extent are the differences in labor force participation for seniors across countries determined by the provisions of Paul Bingley is a research professor at the Danish National Centre for Social Research. Nabanita Datta Gupta is professor of economics at Aarhus University. Michael Jørgensen is a senior pension analyst at ATP in Denmark. Peder J. Pedersen is professor of economics at Aarhus University. This chapter is part of the International Social Security project at the NBER. The authors are grateful to the other participants of that project for useful comments and advice. We are also grateful to the Danish Strategic Research Council (dsf-09-070295) for financial support. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see http:// www.nber.org/chapters/c13326.ack.

331

332

Bingley, Datta Gupta, Jørgensen, and Pedersen

DI programs? Answering this question is a challenge because measuring health is notoriously difficult and DI programs interact with social security provisions in different ways across countries. Social security programs in general have been shown to provide strong incentives for older workers to exit from the labor market at certain ages (Gruber and Wise 2004). In the 2004 volume, incentives were characterized by an option value (OV) model that allows the expected future consequences of current work decisions to be accounted for (Stock and Wise 1990). This was implicitly an inclusive option value, in the sense that different pathways to retirement were included in a single summary measure of expected future consequences. Several countries with extensive DI programs, such as Sweden (Palme and Svensson 2004) and Denmark (Bingley, Datta Gupta, and Pedersen 2004), included a DI retirement pathway probabilistically as part of the inclusive option value. In the current volume, because DI programs are of primary interest, for the sake of greater comparability, DI pathways contribute to inclusive option values in a similar way across all countries. In order to control for health one needs to follow individuals over time either with repeated survey questions about self-assessed health or administrative data about health care usage. Different countries have different health data sources. Even the European countries participating in SHARE, which follow a survey protocol to maximize comparability across countries might have different modes of response between populations, which makes comparison response-by-response difficult. Most other countries in the volume use self-assessed health from surveys, whereas Sweden and Denmark use administrative records of health care usage for the sake of much greater sample sizes. Each of the studies calculates a single health index on the basis of the first principal component of their own sets of health measures. Most of the analyses are conducted on the basis of quintiles of these indices. Identification of incentive effects requires variation in pension program provision between individuals, and ideally within individuals over time by way of pension program changes or reform. We choose an observation period 1996–2008. That is from the first year that we can observe health care usage spanning the population based on administrative records, through the announcement of a major pension program reform in 1998, and beyond full enactment of the new law in 2006. From our descriptive analyses we can see clear gradients in DI participation rates by health quintile and by level of completed schooling. Those in worse health and with less schooling are more likely to receive DI at some point from age fifty. The gradient of DI participation across health quintiles is almost twice as steep as across levels of schooling. We find that pension program incentives in general are important determinants of retirement age. Individuals in poor health are significantly more responsive to economic incentives than those in better health, and those with low schooling are significantly more responsive to economic incentives than those with long

Health, Disability Insurance, and Retirement in Denmark

333

schooling. Hence low schooling and poor health are associated with greater DI participation, and those with low schooling and poor health are also most responsive to economic incentives. The remainder of the chapter is organized as follows. Section 8.2 shows background trends in labor force and DI participation over time by schooling and health. Section 8.3 presents the empirical approach, describing pathways to retirement, how they are weighted, describing the health index and the option value calculations. Section 8.4 presents results from estimating option value models of retirement controlling for health in various ways. Section 8.5 shows goodness-of-fit measures and conducts counterfactual simulations to illustrate some implications of the results. Section 8.6 summarizes and concludes. 8.2

Background

Previous studies have shown how trends in labor force participation for seniors have only a weak relationship with changes in mortality and other measures of health over time and across countries (Milligan and Wise 2012). Neither did there appear to be any relationship between the development of DI programs and changes in mortality and measures of health. These findings were on the basis of a broad view of disaggregated data covering a dozen countries and spanning several decades. In the current chapter we want to analyze how individual retirement behavior in Denmark is related to DI provisions, when controlling for individual variations in health and other characteristics. As background for this microanalysis, in this section we describe trends over time in DI participation, labor force participation and employment by age, and correlate these with individual characteristics: gender, health status, and educational attainment. In the population eighteen to sixty-four years old, the share receiving DI has been fairly constant at around 7 percent since 1990 (Organisation for Economic Co-operation and Development [OECD] 2008). This is quite low and stable relative to the situation in neighboring Nordic countries (OECD 2009). However, the relatively low DI participation rate in Denmark needs to be viewed in the context of competing transfer programs. Between 1992 and 1996, an early pension benefit (overgangsydelse) was available for the long-term unemployed age fifty to fifty-nine. This program removed many from the labor market who might otherwise have applied for DI. In 1998, an existing wage subsidy program for the disabled was expanded and relaunched ( flexjob). The disabled with some remaining work capacity were thus encouraged to stay in the labor market rather than exit on DI. Another relevant aspect in the development of DI in Denmark over recent years is the rather stable overall participation rate, with a growing proportion of new young claimants entering the program with psychiatric diagnoses (OECD 2013). A final aspect of DI in Denmark is that only very few reenter

334

Bingley, Datta Gupta, Jørgensen, and Pedersen

the labor market having once received DI (Høgelund and Holm 2006). This is surprising in light of Jonassen, Larsen, and Høgelund (2009), who find that of those with functional disabilities in 1995, 50 percent had improved functional ability in 2008. This was especially the case for the young and those starting out with a psychiatric functional disability. Time series of DI participation rates are shown in figure 8.1A for age groups fifty to fifty-four, fifty-five to fifty-nine, and sixty to sixty-four for men, and women are shown in figure 8.1B. Women have higher mean DI participation rates than men, and older groups have higher DI participation rates than younger groups. The youngest group has stable DI participation throughout the period for both genders, at 8 percent for men and 12 percent for women. Disability insurance participation has declined markedly for those age sixty to sixty-four, falling from 21 to 13 percent for men and a dramatic 36 to 17 percent for women. In the post-2008 years, not shown in figures 8.1A and 8.1B, DI shares are stable for the fifty and older group until 2013. The DI participation rates of figures 8.1A and 8.1B are now set alongside employment rates in figures 8.2A, 8.2B, 8.2C, 8.2D, 8.2E, and 8.2F, which show time series for age groups fifty to fifty-four, fifty-five to fifty-nine, and sixty to sixty-four, separately for men and women. A high degree of symmetry is evident, especially in the older group, whereby falls in DI participation are about two-thirds of the size of employment increases. Indeed, since 1999 employment is more common than DI participation for women age sixty to sixty-four. Overall, the share in retirement in this age group is higher than the share in employment, however, as the share of women in a SS program for early retirement is 40 percent of the age group by the end of the period we analyze. Associations with health status and schooling levels are shown in the next three figures. Figures 8.3A and 8.3B show DI participation rates for age group fifty-five to sixty-four by schooling for selected years, separately for men and women. There is a clear gradient in schooling in that those with lower education have higher rates of DI participation. Graduating high school approximately halves the DI rate, falling from 24 to 13 percent for men and from 35 to 17 percent for women in 1996. Subsequent educational attainment is associated with an approximately 3 percent reduction in DI rates for those with some college and another 3 percent reduction for graduating college. There is no discernible change in the educational gradient over time. In the following illustrations and for most of the econometric analysis, health status is characterized by quintiles of a health index. Calculation of the index is described in section 8.3.3. Figure 8.4A shows DI participation rates for age group fifty-five to sixty-four by health quintile for selected years for men, and women are shown in figure 8.4B. There is a clear gradient in health in that those with worse health have higher rates of DI participation.

Health, Disability Insurance, and Retirement in Denmark

Fig. 8.1A

DI participation by age group, men

Fig. 8.1B

DI participation by age group, women

335

Our quintile grouping resolves into three different DI rates, the worst quintile followed by quintiles 2 and 3 together with a lower DI rate, followed by better health quintiles with almost no DI recipients. The fall in DI rate from best health quintile to 2 and 3 is more marked than for schooling, falling from 48 to 25 percent for men and from 61 to 37 percent for women in 1996. There is no discernible change in the health gradient over time.

Fig. 8.2A

DI and employment for men, age fifty to fifty-four

Fig. 8.2B

DI and employment for women, age fifty to fifty-four

Fig. 8.2C

DI and employment for men, age fifty-five to fifty-nine

Fig. 8.2D

DI and employment for women, age fifty-five to fifty-nine

Fig. 8.2E

DI and employment for men, age sixty to sixty-four

Fig. 8.2F

DI and employment for women, age sixty to sixty-four

Fig. 8.3A and year

Men age fifty-five to sixty-four who have received DI by education

Fig. 8.3B and year

Women age fifty-five to sixty-four who have received DI by education

340

Bingley, Datta Gupta, Jørgensen, and Pedersen

Fig. 8.4A

Men age fifty-five to sixty-four who have received DI by health and year

Fig. 8.4B and year

Women age fifty-five to sixty-four who have received DI by health

The joint distribution of DI participation rates by schooling and health quintile together is shown in table 8.1 for age group fifty-five to sixty-four (men in the upper pane and women in the lower pane). Subpopulations with worst health and lowest schooling have highest DI participation rates, at 46 percent for men and 55 percent for women. At the opposite corner of the

Health, Disability Insurance, and Retirement in Denmark Table 8.1

341

Percent DI receipt age fifty-five to sixty-four by heath quintile and education health quintile

Men 1 Less than HS 2 HS grad 3 Some college 4 College All Women 1 Less than HS 2 HS grad 3 Some college 4 College All

1 (low)

2

3

45.58 30.16 22.75 17.22 34.11

25.63 12.97 8.56 6.98 16.06

22.84 9.25 6.08 4.11 12.72

55.11 32.92 25.78 16.98 42.12

35.59 15.63 11.43 7.28 23.55

33.46 11.96 8.30 5.05 20.72

4 4.03 1.81 1.81 1.23 2.41   7.32 2.44 1.83 1.33 4.25

5 2.14 0.94 1.10 0.73 1.30   2.50 0.86 0.69 0.49 1.48

All 20.63 10.70 7.01 4.71 12.83   28.91 12.66 9.51 5.66 19.02

table, men and women in the best health and with a college degree both have a DI rate of less than 1 percent. Within each health quintile there is still a marked schooling gradient in DI participation rates. Similarly, within each educational level there is still a marked health gradient. Health is the most important marginal distribution, with 17 percent of men and women receiving DI of those in worst health with a college degree, whereas only 2 percent of men and women in best health and less than high school participate in DI. Information from table 8.1 is further split by selected years in figure 8.5A, which presents DI participation rates for the age group fifty-five to sixtyfour jointly by schooling and health for men, and women are shown in figure 8.5B. The joint gradient in DI participation rates by health and schooling is maintained proportionally throughout, with worst health and lowest schooling men and women in 1996 at 57 percent, falling to 37 percent by 2008. The fall of one-third for this group over twelve years is similar in magnitude to the DI participation rate difference for those in worse health between some high school and some college. In the final two sets of background figures, employment rates are associated with schooling and health. Figure 8.6A shows employment rates for age group sixty to sixty-four by schooling over time for men, and women are shown in figure 8.6B. Men have higher employment rates than women. Indeed, men with some college have similar employment rates to women with a college degree, and men with a high school degree have similar employment rates to women with some college. There are similar upward trends in employment rates for the three education groups without a college degree. In 2008, for example, the range of mean employment rates across levels of schooling is narrower for men, ranging from 48 to 80 percent, than for women, ranging from 26 to 70 percent.

Fig. 8.5A DI recipients by education and health quintile, age fifty-five to sixtyfour (men)

Fig. 8.5B DI recipients by education and health quintile, age fifty-five to sixtyfour (women)

Health, Disability Insurance, and Retirement in Denmark

Fig. 8.6A

Employment by education, age sixty to sixty-four (men)

Fig. 8.6B

Employment by education, age sixty to sixty-four (women)

343

Figure 8.7A shows employment rates for age group sixty to sixty-four by health quintile over time for men, and women are shown in figure 8.7B. There is a clear health gradient in employment rates, with those in worst health having lowest employment rates, and those in the best two health quintiles having highest employment rates. Employment rates across all health quin-

344

Bingley, Datta Gupta, Jørgensen, and Pedersen

Fig. 8.7A

Employment by health quintile, age sixty to sixty-four (men)

Fig. 8.7B

Employment by health quintile, age sixty to sixty-four (women)

tiles for men and women increase uniformly over the sample period. The increase in employment rates from 1995 to 2008 by about 20 percent points is similar to the difference in moving from the two worst health quintiles to the second best. In summary, our years of observation (1995–2008) covers a period of

Health, Disability Insurance, and Retirement in Denmark

345

falling DI participation, increasing labor force participation, and increasing employment for seniors, especially those age sixty to sixty-four. There are steep gradients in health, with those in worse health more likely to participate in DI and less likely to be in employment. There are similar and almost as steep gradients across the schooling distribution, with those without a high school diploma more likely to participate in DI and less likely to be in employment. 8.3

Empirical Approach

Our goal is to estimate the relationship between DI provisions and retirement age, given health status. In order to do this we need to consider all transfer programs relevant for the transition from work to retirement for seniors. These different pathways to retirement need to be combined in a weighted average measure that summarizes their relative potential importance. An inclusive option value framework will be introduced to characterize incentives implicit in the programs to retire at different ages. Finally, we need to condition on health in a way that is comparable across data sets and countries. The following four subsections present these elements of our empirical approach. 8.3.1

Pathways to Retirement

There are three main pension programs supporting income in retirement that are relevant for our analysis. First is a disability insurance program ( førtidspension, hereafter DI) available for those age eighteen to sixty-six and later eighteen to sixty-four who have permanent social and/or health impairments that reduce work capacity. Second is a contribution-based but largely tax financed postemployment wage program (efterløn, hereafter SS), which is essentially unemployment insurance benefit without a job search requirement available for ages sixty to sixty-six and after 2006 from sixty to sixty-four. Third is old-age pension ( folkepension, hereafter OAP), which is a demogrant available from age sixty-seven and after 2006 from age sixty-five, based on years of residence. Our period of analysis (1995–2008) is chosen to span reforms in DI stringency and SS/OAP incentives in order to provide variation by which to identify the effects of program provisions on the retirement age for older workers. The SS program was introduced in 1979 for ages sixty to sixty-six and existed largely unchanged until reforms in 1992 and 1999. The 1992 rules are relevant for the first part of our sample period. Eligibility from 1992 to 1995 required membership of an unemployment insurance fund for at least twenty of the last twenty-five years. An individual was allowed to work for a maximum of 200 hours. If the 200 hours was exceeded, it resulted in a permanent disqualification from the program. The political motivation for the 200 hours restriction was the idea that youth unemployment would be

346

Bingley, Datta Gupta, Jørgensen, and Pedersen

reduced by cutting the labor supply. This, however, turned out not to be the case as shown in Bingley, Datta Gupta, and Pedersen (2010). For individuals claiming SS at ages sixty to sixty-two, the benefits for the first two years were at the level of unemployment insurance and reduced to 80 percent for the last four years. Delaying SS until age sixty-three or older gave benefits at 100 percent of the maximum unemployment insurance benefit level until age sixty-six. This policy obviously incentivized retiring at age sixty-three rather than at younger or older ages. In 1995 unemployment insurance fund membership history requirements were increased to twenty-five out of the previous thirty years. Until 1999, only payouts from life annuities in occupational pensions were means tested. An SS reform was announced in March 1999 and enacted in July 1999. Means testing of payouts or returns from all contributory pensions— whether they were actually paid out or not— was introduced for those claiming SS at ages sixty and sixty-one. Those eligible for SS and not retiring now accumulate a quarterly USD 2,200 bonus beginning at age sixty-two. This reform shifted the retirement age incentive spike from age sixty-three down to age sixty-two. The previous limitation of working at most 200 hours per year was removed and replaced by a high effective marginal tax rate. The UI fund membership history requirements were further increased to thirty out of the last thirty-five years. Contributions were unbundled from UI and became separately elective. An important element of the 1999 reform was the reduction in OAP age from sixty-seven to sixty-five. Those age sixty and older at enactment (born before July 1939) were unaffected and could first claim OAP at age sixtyseven, whereas those born later could claim OAP from age sixty-five. The change in OAP age was implemented from July 2004 through June 2006 and the maximum age for claiming SS benefits changed accordingly. This policy change was obviously running against the trend of pension reforms typically increasing the age of eligibility. The interpretation is fiscal considerations, in that the great majority of the sixty-five- and sixty-six-year-olds were in the DI or the SS program with benefits significantly higher than in OAP. The DI program has existed in essentially the same institutional form in the period 1984–2002, but with some stringency tightening in the 1990s. It was available to those with permanent social or physical work impairments depending on three levels of severity/generosity. During this period, benefit levels were closely linked to the overall level of wages, but several stringency measures were introduced at different times. Three stringency reforms can be distinguished. First, during 1995–2002 a series of selective municipal award audits were undertaken, whereby each year two out of Denmark’s fifteen counties were chosen and a random sample of new benefit awards was drawn for reassessment of eligibility. Second, in 1997 central government refunds to municipalities were reduced for expenditure on DI to individuals age sixty and older, bringing refunds into line with those for younger age

Health, Disability Insurance, and Retirement in Denmark

347

groups. Third, in 1998 municipalities were required to first consider whether other locally administered programs, such as work rehabilitation or a program with disability wage subsidies, might be relevant before processing an application for DI. In 2003, the government simplified DI for new awards by reducing the number of levels from three to one, but also introduced an array of condition- and needs-specific financial additions. These additions make net changes to incentives due to the reform difficult to characterize for systematic analysis. Other relevant related programs for those in short-term poor health, with short-term or permanent work impairments but some remaining work capacity, are sickness benefits (sygedagpenge), rehabilitation benefits (revalidering), and disability wage subsidies ( fleksjobs), respectively. We do not consider these programs as pensions financing retirement because they involve some degree of attachment to the labor market. Nevertheless, they are worth mentioning because of their relevance at the interface between health, work, and retirement. Work sickness absence benefits and rehabilitation are awarded temporarily. Disability wage subsidies are a payment at the level of the minimum wage for permanently reduced work capacity. Individuals in this program are classified as employed, or unemployed and seeking work, and therefore not retired for modeling purposes. 8.3.2

Calculating the Probabilities of Different Pathways

An option value incentive measure needs to combine provisions across different potential pathways to retirement. In order to integrate DI we need to impute to each person a probability that DI is a realistic option. These probabilities can then be used as weights to combine pathways into a single inclusive option value measure. We use a stock measure of calculating DI probabilities from the proportion participating in DI by different cells combining individual characteristics. Cells are calculated for those age fifty-five to sixty-four by level of schooling, gender, and year. Selected years of these DI weights are presented in figures 8.3A (men) and 8.3B (women). 8.3.3

Health Index and Health Quintiles

A continuous health index needs to be created and divided into quintiles so as to be comparable with other countries. Poterba, Venti, and Wise (2011) propose such an index be calculated from the first principal component of twenty-seven health indicators from the Health and Retirement Study (HRS). In the Danish administrative registers, we use the first principal component from hospital discharge records and prescription medicine purchases. The principal components analysis is conducted on the population age fifty to eighty during the years 1994–2007. From hospital records we consider all encounters, for both day patients and overnight stays. Each encounter has a primary diagnosis code (ICD-10)

348

Bingley, Datta Gupta, Jørgensen, and Pedersen

and duration. We aggregate diagnoses to the three-digit level, giving 160 distinct diagnoses after twelve diagnoses with fewer than 100 cases are dropped. Durations of hospital stays are summed over a two-year period within each diagnosis for each person. In other words, hospitalization is characterized for each person as length of hospital stay over the previous two years with each of 160 primary diagnoses. From prescription medicine records we consider all purchases from high street pharmacies. Each purchase has a drug code (ATC-5) and dosage. We aggregate drug codes to the three-digit level, giving 170 distinct drug types after eight drug types with fewer than 100 persons purchasing are dropped. Dosages are normalized according to World Health Organization (WHO)defined daily dosages and summed over a two-year period within each drug type for each person. In other words, drug consumption is characterized for each person as number of standard daily doses over the previous two years for each of 170 drug types. Principal components are calculated over hospitalizations and prescriptions together in two-year periods. For example, when modeling retirement behavior in 1996, principal components would be calculated for 1994–1995; for behavior in 1997, principal components would be based on 1995–1996, and so forth. The first principal component forms our health index. Figure 8.8A shows mean centile of the health index over age by gender, and schooling level is shown in figure 8.8B. By convention, a higher centile is taken to indicate better health. Men have a higher mean health centile than women. Note that it is conventional to observe that men have better selfreported health, less health care usage, but higher mortality than women of a similar age. Health declines with age and the gender health gap narrows from 5 centiles at age fifty to 1 centile at age seventy. The gender health gap at age sixty corresponds to the mean health decline over four to five years. According to our health index, based on health care usage, a woman age sixty is as healthy as a man age sixty-four. This is in spite of her having higher expected longevity. Figure 8.8B shows a health centile gradient in schooling with those with lowest schooling having worst health. The schooling health gradient narrows from 10 centiles at age fifty to 6 centiles by age seventy, however the 3 health centile difference of moving from high school graduation to some college persists. 8.3.4

Option Value Calculation

The goal of our analysis is to estimate the relationship between pension program provisions and age of retirement. Incentives implicit in pension program provisions can be characterized by the potential gain from postponing retirement until future ages. In order to do this we follow the option value approach of Stock and Wise (1990) and extend this to explicitly allow for different potential pathways to retirement in the form of an inclusive option value measure.

Health, Disability Insurance, and Retirement in Denmark

Fig. 8.8A

Health index mean by gender

Fig. 8.8B

Health index mean by education

349

From the vantage point of each age a while in work, there are several possible pathways ( pa = 1, . . ., PA) to retirement, each with an associated utility stream V dependent upon age of retirement time r. A pathway constitutes a number of years of continued work, denoted in the first summation of equation (1), followed by the number of years receiving pension benefits

350

Bingley, Datta Gupta, Jørgensen, and Pedersen

specific to that pathway until death at age A, denoted by the second summation of equation (1). Expected utility at each future age s from the vantage point of each age Ea is weighted by the probability of survival to that age ps|a and discounted βs–t back to the present. While working, wage income ω(s) is received at each age, while retired benefit income Brk(s) is received at each age dependent upon pathway and age of retirement. The utility function includes a parameter for leisure κ, which scales retirement benefits relative to earnings. Both incomes in work and retirement are raised to the power γ representing risk aversion: (1)

Ea {Vka(r)} =

r −1

A

s=a

s=r

∑ ps| a s − a( (s)) + ∑ ps| a s − a(Brk(s)).

For each retirement pathway pa, the future age of retirement at which the expected discounted utility stream is maximized is denoted r*. The comparison is between expected utility streams associated with all retirement ages until maximum age of retirement R. The option value of staying in work at the present age a compared to following eventual retirement pathway pa is defined as the difference between the maximum of expected utilities from future retirement ages along that pathway compared to retiring now: (2)

OVka ≡ max [Ea{Vka (r *)}] − Ea{Vka (a)}. a < r* ≤ R

Having defined the OV of staying in work from the vantage point of each age a for each retirement pathway pa, it remains to weight each pathway with the probability Pk so that it represents a set of relevant alternatives for each individual. An inclusive OV measure combines routes weighted by the probabilities that they are relevant as follows: OVa =

(3)

K

∑ Pk OVka .

k =1

This inclusive option value measure makes explicit the extension to the Stock and Wise (1990) option value approach that allows us to incorporate several different routes to retirement. This can be cast in a regression framework further allowing for differences in health status. Consider retirement status R for person i of age a in health quantile j. This is assumed to be a function θj of exogenous individual characteristics Xia and a function δj of inclusive option value OVia ; Hj is a measure of health and εiaj is an error term: (4)

Riaj =

J

∑ [j Xia + j OVia H j ] + ε iaj . j =1

Equation (4) is estimated as a probit model for year-to-year retirement. Retirement behavior is characterized as an optimal stopping problem in that an individual remains out of the labor force once retired. Benefit collection and retirement are assumed to be synonymous. Pathways from the labor force to OAP could be direct or via DI, SS, or a private pension drawdown.

Health, Disability Insurance, and Retirement in Denmark

351

Individuals are selected at ages fifty-seven to sixty-six and must be working in the first year of observation. We assume a maximum age of retirement R at sixty-seven and force those who are still working at age sixty-six to retire at sixty-seven on OAP. We use population life tables for survival probability s from age a published in 2009 by age and gender for ages fifty-eight to ninetynine and impose zero survival probability at age 100. After retirement, an individual leaves the data set. Exits from the data set due to death, migration, or change of marital status are treated as missing at random. Observations for individuals who leave the data set are used in estimation until the year before the exit and the last observation is classified as working. Potential earnings profiles are assumed to be flat from age fifty-seven, with 1 percent real growth. Option value calculations assume knowledge of the pension and tax system as in place at the vantage point of observation. Individuals form expectations on the basis of that system and any future changes that had already been announced at that time. For the sake of comparison with other countries, preference parameters are fixed at the levels found in US data as discount rate β = 0.97, utility of leisure κ = 1.5, and risk aversion γ = 0.95. It is informative to present examples of these option values in order to fix ideas. Figure 8.9A shows mean option value for the 1941 cohort by age for each retirement path as well as for inclusive option value combining all pathways for men, and women are shown in figure 8.9B. Option value falls with age. The fall is from a higher base for men compared to women, but the proportional fall over age is similar. The DI option value declines smoothly, whereas SS option value slows its decline just for age sixty-one and resumes a decline thereafter more slowly. This reflects an absence of age-related conditions for DI, but a postreform penalty for SS at age sixty-one due to means testing of private pension wealth, followed by bonus payments for delaying SS retirement for each quarter beyond age sixty-two. The ranking between OV profiles differs between women and men. For women the DI OV is lower than SS, while the opposite is found for men. The 1941 cohort of men typically have higher occupational pension wealth than women. As a consequence, benefits from SS are means tested to a higher degree for men than for women. Women, however, have a higher prospective rate of compensation from SS than men due to lower wages on average in a setting where benefits do not depend on preentry wages. 8.4

Results

In this section we present estimates of the models constructed in the previous section. Option value is the main explanatory variable of interest, and it is informative to first see how this evolves over age alongside retirement age to understand how it is driving the retirement decision modeling. Figure 8.10A shows the percent of men or women having reached maximum

Fig. 8.9A

Mean OV by age for 1941 cohort (men)

Fig. 8.9B

Mean OV by age for 1941 cohort (women)

utility, or minimum option value, by age for the prereform 1938 cohort, and the postreform 1941 cohort is shown in figure 8.10B. Also shown is the percent retired by age for men and women. The percent having reached minimum option value is higher for men than women and rises faster over age for women. The pattern is similar pre- and postreform, but with a bigger share reaching minimum option value by each age prereform.

Health, Disability Insurance, and Retirement in Denmark

353

Fig. 8.10A

Percent having reached minimum OV and retired by age (1938 cohort)

Fig. 8.10B

Percent having reached minimum OV and retired by age (1941 cohort)

The remainder of the section presents estimates of option value coefficients and controls for different specifications and samples. Table 8.2A shows estimates from retirement probit regressions with option value as the key explanatory variable and health measures as controls. Each column is for a separate regression to check sensitivity of results to the inclusion of linear age versus age dummies, inclusion of additional covariates, and to different

OV_inclusive   Health quint 2 (second lowest) Health quint 3   Health quint 4   Health quint 5 (highest) Health index   Age Age dummies Female   Married  

 

Table 8.2A

       

–0.0907*** (0.0006) –0.0033*** (0.0008) –0.0101*** (0.0008) –0.0098*** (0.0008) –0.0020* (0.0008)     X

(1)

       

X

–0.0591*** (0.0005) –0.0190*** (0.0006) –0.0256*** (0.0006) –0.0228*** (0.0007) –0.0150*** (0.0007)    

(2)

Effect of inclusive OV on retirement

0.0016* (0.0006) –0.0198*** (0.0010)

–0.0775*** (0.0007) –0.0007 (0.0008) –0.0089*** (0.0008) –0.0116*** (0.0008) –0.0047*** (0.0008)     X

(3)

X 0.0051*** (0.0005) –0.0184*** (0.0008)

–0.0433*** (0.0005) –0.0162*** (0.0006) –0.0249*** (0.0006) –0.0250*** (0.0007) –0.0180*** (0.0007)    

(4)

(5)

       

–0.0903*** (0.0006)                 –0.0060*** (0.0005) X

Specification

       

X

–0.0589*** (0.0005)                 –0.0063*** (0.0005)

(6)

–0.0001 (0.0006) –0.0198*** (0.0010)

–0.0773*** (0.0007)                 –0.0061*** (0.0005) X

(7)

X 0.0024*** (0.0005) –0.0188*** (0.0008)

–0.0434*** (0.0005)                 –0.0064*** (0.0005)

(8)

1,296,332 0.207 0.120 9,898 10,110

         

         

1,296,332 0.081 0.120 9,898 10,110

           

           

1,296,332 0.103 0.120 9,898 10,110

(0.0006) –0.0194*** (0.0008) –0.0509*** (0.0011)

0.0542*** (0.0006) –0.0000*** (0.0000) X –0.0024***

1,296,332 0.229 0.120 9,898 10,110

(0.0005) –0.0154*** (0.0006) –0.0362*** (0.0008)

0.0428*** (0.0005) –0.0000*** (0.0000) X –0.0045***  

1,296,332 0.081 0.120 9,898 10,110

         

1,296,332 0.205 0.120 9,898 10,110

       

1,296,332 0.103 0.120 9,898 10,110

(0.0006) –0.0195*** (0.0008) –0.0510*** (0.0011)

0.0543*** (0.0006) –0.0000*** (0.0000) X –0.0025***

Note: Coefficients are marginal effects of a 10,000-unit change in OV from probit models. Standard errors are shown in parentheses. ***Significant at the 1 percent level. **Significant at the 5 percent level. *Significant at the 10 percent level.

No. of observations Pseudo R2 Mean ret. rate Mean of OV Std. dev. of OV

Spouse retired   Total assets   Occup. dummies Educ: