Factors Affecting Worker Well-Being : The Impact of Change in the Labor Market 9781784411497, 9781784411503

Citation preview

FACTORS AFFECTING WORKER WELL-BEING: THE IMPACT OF CHANGE IN THE LABOR MARKET

RESEARCH IN LABOR ECONOMICS Series Editor: Solomon W. Polachek IZA Co-Editor: Konstantinos Tatsiramos Volume 29:

Ethnicity and Labor Market Outcomes Edited by Amelie F. Constant, Konstantinos Tatsiramos and Klaus F. Zimmermann

Volume 30:

Jobs, Training, and Worker Well-Being Edited by Solomon W. Polachek and Konstantinos Tatsiramos

Volume 31:

Child Labor and the Transition between School and Work Edited by Randall K. Q. Akee, Eric V. Edmonds and Konstantinos Tatsiramos

Volume 32:

Who Loses in the Downturn? Economic Crisis, Employment and Income Distribution Edited by Herwig Immervoll, Andreas Peichl and Konstantinos Tatsiramos

Volume 33:

Research in Labor Economics Edited by Solomon W. Polachek and Konstantinos Tatsiramos

Volume 34:

Informal Employment in Emerging and Transition Economies Edited by Hartmut Lehmann and Konstantinos Tatsiramos

Volume 35:

35th Anniversary Retrospective Edited by Solomon W. Polachek and Konstantinos Tatsiramos

Volume 36:

Research in Labor Economics Edited by Solomon W. Polachek and Konstantinos Tatsiramos

Volume 37:

Labor Market Issues in China Edited by Corrado Giulietti, Konstantinos Tatsiramos and Klaus F. Zimmermann

Volume 38:

New Analyses of Worker Well-Being Edited by Solomon W. Polachek and Konstantinos Tatsiramos

Volume 39:

Safety Nets and Benefit Dependence Edited by Ste´phane Carcillo, Herwig Immervoll, Stephen P. Jenkins, Sebastian Ko¨nigs and Konstantinos Tatsiramos

RESEARCH IN LABOR ECONOMICS VOLUME 40

FACTORS AFFECTING WORKER WELL-BEING: THE IMPACT OF CHANGE IN THE LABOR MARKET EDITED BY

SOLOMON W. POLACHEK State University of New York at Binghamton and IZA

KONSTANTINOS TATSIRAMOS University of Nottingham and IZA

United Kingdom
North America
Japan India
Malaysia
China

Emerald Group Publishing Limited Howard House, Wagon Lane, Bingley BD16 1WA, UK First edition 2014 Copyright r 2014 Emerald Group Publishing Limited Reprints and permission service Contact: [email protected] No part of this book may be reproduced, stored in a retrieval system, transmitted in any form or by any means electronic, mechanical, photocopying, recording or otherwise without either the prior written permission of the publisher or a licence permitting restricted copying issued in the UK by The Copyright Licensing Agency and in the USA by The Copyright Clearance Center. Any opinions expressed in the chapters are those of the authors. Whilst Emerald makes every effort to ensure the quality and accuracy of its content, Emerald makes no representation implied or otherwise, as to the chapters’ suitability and application and disclaims any warranties, express or implied, to their use. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library ISBN: 978-1-78441-150-3 ISSN: 0147-9121 (Series)

ISOQAR certified Management System, awarded to Emerald for adherence to Environmental standard ISO 14001:2004. Certificate Number 1985 ISO 14001

CONTENTS LIST OF CONTRIBUTORS

vii

EDITORIAL ADVISORY BOARD

ix

PREFACE Solomon W. Polachek and Konstantinos Tatsiramos

xi

EXPLAINING THE REVOLUTION IN U.S. FERTILITY, SCHOOLING, AND WOMEN’S WORK AMONG HOUSEHOLDS FORMED IN 1875, 1900, AND 1925 Matthias Cinyabuguma, William Lord and Christelle Viauroux INTEGRATING RETIREMENT MODELS: UNDERSTANDING HOUSEHOLD RETIREMENT DECISIONS Alan L. Gustman and Thomas L. Steinmeier THE ROLE OF DEGREE ATTAINMENT IN THE DIFFERENTIAL IMPACT OF JOB CORPS ON ADOLESCENTS AND YOUNG ADULTS Maria Bampasidou, Carlos A. Flores, Alfonso Flores-Lagunes and Daniel J. Parisian INSECURE, SICK AND UNHAPPY? WELL-BEING CONSEQUENCES OF TEMPORARY EMPLOYMENT CONTRACTS Vincenzo Carrieri, Cinzia Di Novi, Rowena Jacobs and Silvana Robone

v

1

79

113

157

vi

CONTENTS

THE EFFECT OF LAND TITLE ON CHILD LABOR SUPPLY: EMPIRICAL EVIDENCE FROM BRAZIL Mauricio Moura and Rodrigo Bueno

195

THE CHANGING TIME USE OF U.S. WELFARE RECIPIENTS BETWEEN 1992 AND 2005 Marie Connolly

223

DOES HIGHER EDUCATION QUALITY MATTER IN THE UK? Arnaud Chevalier

257

BUSINESS VISITS AND THE QUEST FOR EXTERNAL KNOWLEDGE Massimiliano Tani

293

LIST OF CONTRIBUTORS Maria Bampasidou

University of Florida, USA

Rodrigo Bueno

University of Sao Paulo, Brazil

Vincenzo Carrieri

Department of Economics and Statistics and CELPE, University of Salerno, Italy; Health Econometrics and Data Group, University of York, UK

Arnaud Chevalier

Royal Holloway University of London, London, UK

Matthias Cinyabuguma Department of Economics, University of Maryland, Baltimore County, USA Marie Connolly

De´partement des sciences e´conomiques, University of Quebec in Montreal, Cirpe´e, Cirano

Cinzia Di Novi

Department of Economics, Ca’ Foscari University of Venice, Italy

Carlos A. Flores

Department of Economics, California Polytechnic State University at San Luis Obispo, USA

Alfonso Flores-Lagunes Department of Economics, State University of New York at Binghamton, USA and IZA Alan L. Gustman

Department of Economics, Dartmouth College, USA

Rowena Jacobs

Centre for Health Economics, University of York, UK

William Lord

Department of Economics, University of Maryland, Baltimore County, USA vii

viii

LIST OF CONTRIBUTORS

Mauricio Moura

George Washington University and Harvard University, USA

Daniel J. Parisian

Department of Economics, State University of New York at Binghamton, USA

Silvana Robone

Dipartimento di Scienze Economiche, Universita` di Bologna, Italy and Health, Econometrics and Data Group, University of York, UK

Thomas L. Steinmeier

Texas Tech University, USA

Massimiliano Tani

School of Business, UNSW Canberra, Australia and IZA

Christelle Viauroux

Department of Economics, University of Maryland, Baltimore County, USA

EDITORIAL ADVISORY BOARD Orley C. Ashenfelter Princeton University

Daniel S. Hamermesh University of Texas

Francine D. Blau Cornell University

James J. Heckman University of Chicago

Richard Blundell University College London

Alan B. Krueger Princeton University

David Card University of California

Edward P. Lazear Stanford University

Ronald G. Ehrenberg Cornell University

Christopher A. Pissarides London School of Economics

Richard B. Freeman Harvard University

Klaus F. Zimmermann IZA and University of Bonn

ix

PREFACE Big changes have been going on in labor markets over the last century. First, for most countries fertility rates declined. Second, women’s labor force participation has risen, but men’s has not. Third, college attendance increased, again more so for women than men. Finally, not well spelled out, transportation has become cheaper, and as a result, business travel increased. This volume contains eight articles pertaining to the causes and consequences of changing labor markets. Of the eight articles, the first two deal with changing demographics; the next four with new or changing government programs and policies; the final two articles deal with education and knowledge acquisition. In the United States, women’s labor force participation increased dramatically from 17% in 1890 to 58% in 2012. College graduation rates increased six-fold from 5% in 1940 (for white males) to over 30% in 2008. Births per thousand population declined from 30.1 in 1910 to 13.8 in 2009. At the same time, men’s labor force participation fell from 83% in 1890 to 70% in 2012. The gender wage gap narrowed from 41% in 1960 to about 23% today. Are these trends related, and is there a unified model to explain them? In the first article, Matthias Cinyabuguma, William Lord, and Christelle Viauroux devise a model that synthesizes various strands of the burgeoning household labor supply, fertility, human capital, and macroeconomic growth literature consistent with these trends. They then calibrate the model based on historical data. They find that changes in the net cost of children, concomitant with rising relative women’s to men’s wages, explains the decline in fertility and the rise in schooling. When one retires constitutes another change. With more women working for pay, a new increasingly observed pattern is spouses often retire together during the same year, this despite the younger age of most wives. Such a pattern implies at least some husband
wife coordination, for example, motivated by the leisure time of both spouses being complements in the maximization of utility. As a result, modeling individual retirement as an individual decision might be inadequate, given that a family model appears more applicable. Early family models made a number of simplifications, in particular, that retirement decisions constitute a binary choice, xi

xii

PREFACE

namely to retire, or not. In the second article, Alan L. Gustman and Thomas L. Steinmeier relax this and a number of other restrictions. As a result, they obtain richer more general results, including spikes in retirement at 62 and 65 along with the spike in couples’ retirement in the same year. In addition, given the robustness of their model, they are able to address a number of other phenomena unexplained by conventional models. One example constitutes differences in wealth accumulation by families with similar earnings opportunities. The role of government in the US economy also changed dramatically. Federal spending was about 3% of GNP in 1900 but rose to slightly less than 20% in 2010. Combined federal, state, and local US government spending was about 8% of GNP in 1900 and about 37% in 2010. The same increase is true for Sweden and many other countries. Concomitant with these increases were a host of government programs. Early programs were designed to educate the population, to promote agriculture, to protect the poor, and to safeguard workers. Some government programs work like expected, others have unintended consequences. In either case, government programs need to be scientifically evaluated to assess their worth. The next four articles deal with assessing aspects of government programs. The Job Corps is the largest training program for disadvantaged youth 16
24 years old in the Unites States. It provides job placement, residential services, social skills, vocational instruction, and the opportunity to earn a GED or high school diploma. A number of studies already found a positive impact of the Job Corps in raising earnings and increasing the probability of employment. However, no study to date has broken down the impact, either to which particular subsamples of the disadvantaged gain more, or to examine which aspects of the Job Corps’s instruction and socialization techniques provide the most benefits. In the next article, Maria Bampasidou, Carlos A. Flores, Alfonso Flores-Lagunes, and Daniel J. Parisian utilize a new statistical technique to obtain surprising, yet important, results bounding the effects of particular Job Corps interventions. Often government programs have secondary effects. Economic agents, such as the firm, respond to legislation, and these responses also have consequences. For example, many firms across Europe now require long periods of temporary employment. This trend is partly in response to redundancy legislation, which forces firms to show cause to terminate an employee. In Europe, from 1990 to 2011, temporary employment increased from 11% to 15%. One country exhibiting a particularly quick rise in temporary employment is Italy. In 1990 a mere 5% of Italy’s workers were temporary; by 2011 this figure was 13%
a 150% gain. One question is

Preface

xiii

the effect on workers of having to take temporary employment. In the next article, Vincenzo Carrieri, Cinzia Di Novi, Rowena Jacobs, and Silvana Robone examine this question. They concentrate on psychological health and happiness mostly for the young and focus on gender differences. Using a propensity score matching estimator with data from the Italian “Health Conditions and Use Health Service” they find an interesting asymmetry between men and women. Temporary contracts damage psychological wellbeing of young men and individuals without family economic support, but not the mental health of women. This result has implications regarding women’s labor force participation and occupational choice as well as for government policies regarding temporary employment and welfare support for young employees. For developing countries, providing property rights, particularly titles to land, is often viewed as an important requisite to instigate economic development, growth, and poverty reduction. One purported mechanism through which property rights work is to provide credit access because property can be used as collateral, but the evidence is mixed. Another mechanism recently gaining more attention is through the labor market. The privilege of not having to guard one’s living quarters enables adult household members to shift away from work at home towards supplying more time to the labor market. Interestingly, recent evidence at least for Peru indicates more adult labor supply leads to less child labor, which can result in more children in school, more human capital accumulation, and higher long-term growth. But the evidence on this process is still relatively scant. In the next article, Mauricio Moura and Rodrigo Bueno utilize a unique data set for two Brazilian cities comparing similar neighboring communities in metropolitan San Paulo, first in March 2007, and then in August 2008, before and after inhabitants in one of the cities received land titles. The other city did not receive land titles until 2012. Using a difference-in-difference estimation technique, Moura and Bueno are able to find the effect of property titling, particularly on child labor. Changes in government policies can have other effects, as well. In the early 1990s many believed the United States was in the throws of a welfare crisis. An increasingly large number of women joined and remained on the welfare rolls. Further, the likelihood of exiting seemed to diminish with time on AFDC. By the mid-1990s, Bill Clinton, working with Congress, introduced welfare reform (the Personal Responsibility and Work Opportunity Reconciliation Act passed in 1996) designed to force welfare recipients to get off welfare rolls and get back to work. Welfare beneficiaries dropped dramatically, in part because of this reform, and in part

xiv

PREFACE

because of improved economic conditions. One concern was the impact on children. If mothers were enticed back to work, what would happen to mothers’ time spent with their kids? Clearly changes in time devoted to childcare could influence child well-being. Using simple difference-indifference techniques along with ATUS data, past studies claimed time spent with children surprisingly increased. In the next article Marie Connolly employs propensity score techniques along with five new time-use studies and is able to question the strength of these past results. As in past studies, she finds time spent at work increased, but time spent with children did not. As academics, we all recognize knowledge is important. Given the large body of research on education, we are aware how individuals acquire knowledge through formal schooling and on-the-job post school training, but there are still a number of unanswered questions about the acquisition of knowledge. The next two articles deal with such issues. A number of studies point to widely varying rates of return to education. In part, this variation stems from differing student abilities, fields of study, and economic conditions at graduation, but some of this heterogeneity in school outcome arises for other reasons. One, in particular, is school quality. Using 2009 earnings data for a set of 7000 United Kingdom students who graduated in 2003, Arnaud Chevalier employs a generalized propensity score selection-on-observables technique to isolate the effect of school quality. He finds, on average, a one standard deviation increase in school quality is associated with one to three percent higher earnings. This premium is nonlinear with graduates from the most prestigious institutions benefiting 5% more than graduates from third quality quartile institutions. Less well known is how firms, rather than individuals, acquire knowledge. Surely, businesses can purchase technology outright, surely they can invest in R&D, but certainly there must be other methods of knowledge propagation. In the final article, Massimiliano Tani examines an aspect of corporate knowledge propagation, not yet analyzed in the literature. As such, it is path breaking. He seeks to make inroads on how business travel increases a firm’s knowledge base. He does so via a unique survey of business travelers taken at airports. From this data he estimates a two-equation empirical model to establish that business travel enhances corporate knowledge. He then determines the type business travel that enhances knowledge the most. As with past volumes, we aim to focus on important issues and to maintain the highest levels of scholarship. We encourage readers who have prepared manuscripts that meet these stringent standards to submit them to

Preface

xv

Research in Labor Economics (RLE) via the IZA website (http://rle.iza.org) for possible inclusion in future volumes. For insightful editorial advice, we thank Aslan Akay, Thomas Andre´n, Christian Brinch, Lorenzo Cappellari, Ste´phane Carcillo, Julia Darby, Nele de Cuyper, Marloes de Graaf-Zijl, Ross Finnie, Patrick Emerson, Bjo¨rn Gustafsson, Jo¨rgen Hansen, Stephen Jenkins, Katarina Katz, Marike Knoef, Sebastian Ko¨nigs, Deborah Levinson, Stephan Lindner, Peter R. Mueser, Pimrak Pakdeethai, Andreas Peichl, Rosario Sa´nchez-Pe´rez, Jessamyn Schaller, John Robst, Arne Uhlendorff, Robert Valletta, Ulrika Vikman, April Wu, Christoph Wunder, and Zhong Zhao. Solomon W. Polachek Konstantinos Tatsiramos Editors

EXPLAINING THE REVOLUTION IN U.S. FERTILITY, SCHOOLING, AND WOMEN’S WORK AMONG HOUSEHOLDS FORMED IN 1875, 1900, AND 1925 Matthias Cinyabuguma, William Lord and Christelle Viauroux ABSTRACT This paper addresses revolutionary changes in the education, fertility and market work of U.S. families formed in the 1870s
1920s: Fertility fell from 5.3 to 2.6; the graduation rate of their children increased from 7% to 50%; and the fraction of adulthood wives devoted to market-oriented work increased from 7% to 23% (by one measure). These trends are addressed within a unified framework to examine the ability of several proposed mechanisms to quantitatively replicate these changes. Based on careful calibration, the choices of successive generations of representative husband-and-wife households over the quantity

Factors Affecting Worker Well-Being: The Impact of Change in the Labor Market Research in Labor Economics, Volume 40, 1
78 Copyright r 2014 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-912120140000040001

1

2

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

and quality of their children, household production, and the extent of mother’s involvement in market-oriented production are simulated. Rising wages, declining mortality, a declining gender wage gap, and increased efficiency and public provision of schooling cannot, individually or in combination, reduce fertility or increase stocks of human capital to levels seen in the data. The best fit of the model to the data also involves: (1) a decreased tendency among parents to view potential earnings of children as the property of parents and (2) rising consumption shares per dependent child. Greater attention should be given the determinants of parental control of the work and earnings of children for this period. One contribution is the gathering of information and strategies necessary to establish an initial baseline, and the time paths for parameters and targets for this period beset with data limitations. A second contribution is identifying the contributions of various mechanisms toward reaching those calibration targets. Keywords: American family; quantity-quality trade-off; convergence; high school movement; married female labor force participation rate JEL classifications: I21; J13; J22; N31; N32

INTRODUCTION The American family has never been a static institution. That said, the latter decades of the 19th century and first decades of the 20th century were ones of exceptional change. Over this period households came to choose much smaller families, to forego child labor, to view extensive formal education among dependent children as essential preparation for adulthood, and to allocate a significant portion of the adulthood of wives to market labor. Indeed, comparing households formed in the 1870s and those formed in the 1920s, fertility fell by half (from about 5.3 to 2.6 children); the proportion of their children who graduated high school increased from about 7% to 50%; employment rates among male youth ages 10
15 fell from about 30% to 6%; and, the fraction of adulthood devoted to marketoriented work by wives increased several fold (from 7 to 23 percent by one measure).

Explaining the Revolution in U.S. Households

3

Based on a calibration of family decisions across three generations, we conclude that changes in the net cost of children offer the best explanation of the decline in fertility and rise in schooling over this period. Further, an important element of this increase in net costs seems to derive from evolving norms which reduce the perceived appropriateness of sending children to work. The narrowing of the gender wage gap is fully capable of accounting for the rise in married women’s market work (indeed, marriage bars during the Great Depression make our calibrated increase greater than that which occurred). The remainder of this Introduction further describes our approach and contribution. Labor economists have developed sophisticated models of household labor supply, fertility, and human capital investments, including applications to the period under consideration. Macro economists have also examined these issues, as well as longer-term determinants of growth. Fewer of these macro formulations have narrowly focused on this period. Traditional economic historians have emphasized data collection and the importance of institutions and norms as well as markets. A first contribution of this paper is to synthesize these literatures, creating a useful bridge to each subfield. From labor economics, a model of household optimization which includes a broad range of household decisions is developed. Three successive generations of parents and their children are linked via careful calibration, emphasizing the dynamics and tools favored by macro economists. The necessary microeconomic and macroeconomic historical data is assimilated to calibrate the framework. The results are interpreted in light of all three literatures, with the institutions and norms studied by economic historians assuming a prominent role. A second contribution is the identification of those mechanisms best able to explain the family dynamics within the calibration, as well as those seen to have little explanatory power. At the household level, parents maximize utility over the quantity and quality of children, the work of wives and children, and the production of household public goods. Explicit (and commonly employed) functional forms are assumed for household utility, household production and human capital production. An important consideration was a choice of functional forms which yield closed-form solutions for all choice variables. This allows clear, and hopefully intuitive, pricetheoretic interpretations of all results. Successive generations are linked as the human capital bequest by one generation of parents proves important to the decisions made by their children once they become adults. A wide range of data is assimilated to restrict the feasible parameterization of the model for households formed in 1875, 1900, and 1925. (This requires some

4

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

discussion of adjacent cohorts: The initial baseline for 1875-parents requires knowledge of the human capital investments in them, as children, by the parents of 1850, while the earnings of households forming in 1950 are greatly influenced by their human capital bequeathed by the parents of 1925.) Nested versions of the model are simulated to assess the contributions of various mechanisms, individually and in combination, toward fitting the time paths of the target variables. Finally, the results and their significance are viewed in light of the broader literature and other plausible mechanisms. The calibrations indicate that rising wages, declining mortality, a declining gender wage gap, and increased efficiency and public provision of schooling cannot, individually or in combination, reduce fertility or increase stocks of human capital to levels seen in the data. Calibrations hit the targets only when additional mechanisms are addressed. In particular, the best fit of the model to the data also involves: (1) a decreased tendency among parents to view potential earnings of children as the property of parents and, (2) rising consumption shares per dependent child. These results suggest that for standard “quantity-quality” models to provide a good quantitative fit, they need to incorporate parental utility from child consumption and welfare during dependency. Several potential explanations are discussed as to why child consumption shares increased and the willingness to work children decreased. Additional research is required to assess their relative contributions. A final contribution is establishing the consistency of a calibration based on microeconomic determinants of the gender wage gap with results from the growth accounting literature. This increases confidence in the separate results and provides an important example where methods and data from one subfield are usefully employed to address an important issue in another. Of course, not every potentially important mechanism can be included in a given model. Further, the quantitative responses to parametric change may prove sensitive to how the various mechanisms which are considered are modeled. Thus, the paper also informally discusses mechanisms not addressed within the model included; implications of differing degrees of responsiveness to those mechanism which are addressed are also considered. The paper is organized as follows. The next section presents a brief description of the historical period, the stylized facts to be explained, those factors deemed to explain them, and a selective review of literature. The section “Modeling the Household” presents the model. The section “Assignment of Parameter and Target Values, and Calibration of the Initial Baseline” explains the calibration strategy. This is followed by a presentation

Explaining the Revolution in U.S. Households

5

of calibration results in the sections “Missing the Targets: Income and Schooling Costs; Mortality Decline” and “Reaching the Targets: Increasing Disregard for the Potential Earnings of Dependent Children, and Greater Child Consumption.” The final section summarizes and concludes.

THE HISTORICAL SETTING AND LITERATURE REVIEW The literature on the high school movement, fertility decline, gender wage gap, mortality decline, and rise in market work and earnings among married females for the United States is immense. The following survey is, of necessity, highly selective. Stagnant, then Growing Demand for Educated Workers The “second” industrial revolution of the Post-Civil War decades involved the innovation and spread of large-scale unskilled-labor-saving capital equipment powered by, first steam, then electricity. New machines substituting for human muscles raised the relative productivity of women in manufacturing (see Galor & Weil, 1996). Simultaneously, the demand rose for skilled bluecollar males to design new machines and service existing ones. Through 1900 high-skilled blue-collar workers were trained primarily “on-the-job.”1 For this reason, the returns to experience were high in the late nineteenth and early 20th century (see Goldin & Polachek, 1987). Conversely, most jobs through 1900 required only the basic literacy and numeracy acquired in common or grammar school. Nineteenth century returns to a year of education were low, in parts depressed by the shortness of the school year, a curriculum which was only loosely linked to market skills, meager expenditures on books, buildings, and teachers, and an occupational structure with relatively few positions requiring “advanced” education. In this light, Connolly (2004) emphasizes the quantitative significance of unskilled human capital within the late 19th century economy, especially in the South. Those few jobs requiring additional schooling (such as bookkeeper, clergy or school teacher) were often filled by the children of upscale households who might attend the odd public high school, a private boarding school, or college. Nevertheless, at century’s end (and perhaps much earlier) there was a significant premium to skill (Goldin & Katz, 2008). New large-batch and continuous processing technologies meant larger firms with layers of

6

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

bureaucracy to record sales and expenses, issue payrolls, market goods, manage financial assets, provide legal services, etc. This greatly increased demand for clerical and other white-collar workers with business skills, such as basic bookkeeping and typing (Goldin & Katz, 2008). With no comparative advantage in the provision of this basic knowledge, firms began to require appropriate schooling of those seeking skilled positions. With growing demand for educated workers, more parents desired public high schools and this increase in potential class sizes lowered the prospective cost of public provision. Both provision and enrollment soared. As draftsmen and machinists came to acquire more human capital in “shop” courses at school before commencing work, on-the-job training and the return to experience on shop floors fell. Similarly, females learned business skills in the classroom instead of at work. Thus, over the first few decades of the 20th century, work experience became a less important determinant of pay while education became more important (Goldin, 1990 and Goldin & Polachek, 1987). For this reason, from 1890 to WWII the increase in total human capital, obtained either on-the-job or at work, would have risen less rapidly than trends in the level of and returns to schooling (controlling for experience) would suggest. Goldin and Katz (2008) view the wage premium to skill and education as the outcome of a “race” between skill-biased technological change (SBTC), which increases the demand for skilled labor, and increases in the supply of skilled workers which, for this period, meant high school graduates. They conclude SBTC accelerated in the later 19th century and then proceeded at a fairly steady rate throughout the 20th century. During the early acceleration, high school enrollments remained low and the skill premium was bid up. Into the 1910s that premium remained high as increases in the demand for, and supply of, skills were roughly offsetting. However, high school graduation rates soared after 1910 and Goldin and Katz (2008, p. 316) find that “from 1910 to 1930 the skill premium fell by 1.28% per year on average.” Doepke and Zilibotti (2005) and Doepke and Tertilt (2009) argue SBTC affected the political process, resulting in child labor laws, reduced fertility, increased public spending on health (and therefore lower child mortality) in the late 19th and early 20th centuries.2

The Growth in Education In the latter decades of the 19th century, a family’s secondary earners were more often older children rather than wives. Especially in poor families,

Explaining the Revolution in U.S. Households

7

parents would not forego child earnings for “advanced” education even if tuition was heavily subsidized. Consequently, the demand for public high schools remained weak. High schools were, if public, urban, or if private, attended by children of the affluent. After about 1910, though, as incomes rose, population densities increased, the roles of children were redefined, and the wage premium to high school completion remained high, public secondary education spread widely among boys and girls. The high school graduation rate rose from 9% to 50% during the “high school movement” between 1910 and 1940 (Goldin & Katz, 2008, p. 195). The quality of education was also rising as real expenditures per youth aged 5
19 more than tripled between 1910 and 1950 (see Table B1). Since parents incur most costs of educating dependent children, such as tuition, foregone earnings and housework of children, while their children in adulthood reap the higher salaries education provides, most formulations of schooling through the teenage years assume parents are altruistic toward children (Becker, 1981). That high school attendance was low in the late 19th century, despite a sizeable wage premium to graduates, suggests some form of credit market imperfection (Becker & Tomes, 1986; Galor & Zeira, 1993). Becker and Tomes note that since parents cannot legally assign debt to children, parental finance of children’s education reduces parental consumption. Consequently, altruistic parents of limited means
relative to the cost of the fully efficient level of child education
are forced to make difficult trade-offs between own consumption and investments in child quality. When this results in investments in children that are inefficiently low, parents are said to be “transfer-constrained” (Lord, 2002, chapter 6). Among constrained households, all altruistically motivated intergenerational transfers are human capital bequests (no financial bequests). Rangazas (2000) simulates a model of U.S. economic growth based on altruistically motivated human and physical capital transfers. He finds that calibrations reflecting transfer constraints better match U.S. growth characteristics than do calibrations in which parents always make the fully efficient investments in their children. Consequently, in the framework developed below, transfers are in human capital form, only.3 Lord and Rangazas (2006) model the demographic transition and rise of schooling investments since 1800. Their framework assumes altruistic parents and emphasizes the role of declining wealth from family enterprise in reducing fertility. They note that as education has risen, so has the earnings gap between mature adults and their teenage children. For this reason, the potential earnings forgone by schooling children have fallen relative to adult potential earnings. With diminishing marginal utility of consumption,

8

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

the opportunity cost of schooling children
the utility loss from lower parental consumption
declines. The framework developed below includes a similar mechanism. An important aim of growth accounting is to attribute observed growth in output per worker or per hour worked to underlying determinants. The role of improvements in the quality of the labor force via increased education has been estimated in a variety of ways. Abramovitz and David (2000) find only a small role for human capital in late 19th century U.S. growth. Goldin and Katz (2008) find that education contributes more than 14% toward the growth in output per labor hour between 1915 and 1940. Turner, Tamura, and Mulholland (2011, 2013) have recently estimated a much larger role for human capital. In the section on calibration of the initial baseline we use novel methods and data and find an estimate close to that of Goldin and Katz.

The Gender Wage Gap Polachek (1975) presents an early analysis of the division of labor by gender within the family. He shows that an initial comparative advantage which induces gender specialization leads to reinforcing follow-up investments. These increase productivity differentials over time and cement early time allocation patterns. Suppose, as assumed in the model below, physical strength confers a greater bump to productivity in market work than child care. Then males would have greater labor market experience and women would assume all child care responsibilities. Goldin and Polachek (1987) and Goldin (1990) examine the earnings of women relative to men and why they changed over time. They report that from 1890 to the 1930s there was a significant narrowing of the gender wage gap from .46 to .56, and to .60 by 1970. In Goldin and Polachek (1987) this narrowing is quantitatively partitioned into roles for changing amounts and returns to experience, increases in the level of and returns to education, and changing rewards to other gender characteristics (male strength in particular). Increases in education are found to be most important, perhaps accounting for more than 40%. Women born in the 20th century devoted a larger portion of their lives to market work, reducing the differential in the quantity of experience. Goldin (1986, 1990) and Goldin and Polachek (1987) argue that relatively more experience for females coupled with relatively lower returns to experience for males, accounts for about 30% of the decline in the gender gap. The spread of steam, and then later electric, power replaced human muscles

Explaining the Revolution in U.S. Households

9

on the farm and in the factory. The shift in total labor demand toward office workers further reduced the proportion of jobs for which large muscles were a big advantage. The falling premium to strength lowered the relative productivity and pay of males by a smaller amount than changes in experience (Goldin & Polachek, 1987). The findings of Goldin and Polachek (1987) and of Goldin (1986) play a key role in the calibration of the model in the section on calibration of the initial baseline.

The Increased Market Work of Wives Goldin (1990) stresses the rise of the clerical sector for increasing female labor force participation (marriage bars during the Great Depression slowed the increase). Female clerical workers were more likely than female manufacturing workers to return to work as their children matured. Consequently, the life cycle labor force participation of married females increased as the proportion of females working in clerical occupations soared. Indeed, the married female labor force participation rate (MFLFPR) rose from 4.6 in 1890 to 21.6 in 1950 (Goldin, 1990, p. 17, Table B1).4 Albanesi and Olivetti (2007) argue that technological improvements related to the bearing and nursing of children were instrumental to the rise in the labor force participation of mothers. Conversely, Mokyr (2000) argues that new understandings of the role of hygiene in preventing sickness and death led turn-of-the-century mothers to devote more time to housework, delaying the acceleration of mother’s market work. Galor and Weil (1996) suppose that the capital deepening accompanying the second industrial revolution decreased the return to strength, narrowing the gender wage gap, reducing fertility and increasing MFLFPRs. According to Greenwood, Seshadri, and Yorukoglu (2005) the rise of labor-saving capital goods in the household (clothes washers, dryers, vacuum cleaners, dishwashers, etc.), in combination with diminishing marginal utility of non-tradable goods produced in the household, reduced the value of mother’s time in the household. These appliances reduce the reservation wages of females and increased MFLFPRs in the middle of the 20th century.5,6

Mortality High baseline mortality before the last decades of the 19th century was spiked by periodic epidemics of cholera, typhoid, yellow fever, influenza

10

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

and other infectious diseases. However, beginning around 1880 mortality commenced a rapid descent. The white infant mortality rate, which was a staggering 214.8 in 1880, had declined to 120.1 in 1900 and to 26.8 by 1950 (Haines, 2000, Table 4.3); rates among black children were even higher. Children who survived infancy were at lower, but still significant, risk of death. Of 100 children born in 1880, an additional 12 died between the ages of 1 and 15 (Murphy, Simon, & Tamura, 2008, Tables 14 and 15). This mortality transition was facilitated by massive public investments in clean drinking water and hygienic waste removal as well as advances in scientific understanding (Preston & Haines, 1991).7

Fertility Fertility was declining in the United States from at least the early 19th century. Jones and Tertilt (2008) and Murphy et al. (2008) analyze self reports of retrospective fertility of ever-married women coded in various Census years. Women born in the 1850s who eventually married attained adulthood circa 1880 and bore about 5.3 children. Females born only a half century later, with fertility centered about 1930, ended up bearing only about 2.6 children
just half that of their grandmothers. Children surviving into adulthood fell by a smaller percentage as deaths during infancy and childhood declined from horrific levels. Jones and Tertilt (2008) find a strong negative relationship between the occupational income of fathers and household fertility. A similarly strong negative relationship is found between the education of the husband and/or wife and fertility. If there is positive assortative spousal mating on education, all of these findings are consistent with Becker’s (1981) observation that children require significant time, and that as the value of time (especially mother’s wages) increases, children become more expensive. Then, so long as each child is treated the same, and parents care about both the quantity and quality (i.e., earnings in adulthood) of children, higher wages would reduce fertility and simultaneously induce a substitution toward child quality. In Soares and Falcao (2008) parents receive utility from surviving children and, as child mortality declines, so does fertility.8 Doepke (2005) examines reductions in child mortality for fertility using several variants of the Barro-Becker (1989) model of intergenerational altruism. In each variant fertility does fall, but the number of surviving children increases.

Explaining the Revolution in U.S. Households

11

Our calibrations support the findings of Doepke; other factors prove responsible for the decline in net reproduction rates. Lord and Rangazas (2006) conduct a quantitative assessment of a theory of long-run growth in the United States
from 1800 to 2000
which is able to reproduce central features of the quantity-quality trade-off. Their macroeconomic transition model includes fertility, student time in human capital production, and multigenerational family business. Family business generates a wealth effect which increases fertility. The decline in family business accounts for about 40% of the decline in the nineteenth century. Our results are contrasted with their findings.

Female Empowerment and Increased Concern for Children This period was one of rising female power. The right of women to vote was formalized by the 19th amendment to the United States Constitution in 1920. Most states in the preceding decades had passed laws enabling women to keep monies earned while married and to enter into contracts
the principle of coverture was in rapid retreat. More women were working for pay, earning higher wages relative to men than had their mothers. Further, mothers spent less time debilitated in pregnancy and were taught how to reduce the rate at which their children succumbed to illness. Miller (2008) and Doepke and Tertilt (2011) review convincing evidence that women place relatively more weight than men on child expenditures and welfare. Shifts in income from husbands to wives tend to reduce expenditures on alcohol, tobacco, and men’s clothing while increasing expenditures on children’s food and clothing. Therefore, the increase in the relative power of females circa 1900 may have increased investments in the quality and well-being of children. Greater investments per child make children more expensive, lowering fertility. This mechanism is reinforced if the share of parental resources devoted to each child’s consumption during dependency is increased.

Declining Economic Role of Children Massachusetts introduced the first state compulsory schooling law in 1852 but by 1910 41 states had such laws, while 40 had passed child labor legislation (Goldin & Katz, 2008, p. 191). Puerta (2009) finds compulsory schooling laws increased school enrollment in affected areas 7% (and

12

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

reduced fertility by 15% among women of reproductive age). Other researchers have found somewhat smaller effects (Goldin & Katz, 2008). Zelizer (1985) argues that as the economic contribution of children declined, there was a shift in the perception of parents toward children; they became “emotionally priceless.” In her view, child labor as a source of household income became reprehensible.9 Moehling (2005) finds that even by the first decades of the 20th century children’s private consumption was increasing in their earnings. As children worked less in and out of the home, organized leisure increased. Boy and cub scouts, girl scouts and brownies, boys and girls clubs of America, Demolay, Pop Warner football, and American Legion baseball are among the youth organizations which had their origins in the first decades of the 20th century. Our research builds on the foregoing literature in a variety of ways. Below we present a model of two-parent households in which fertility, child labor and human capital, household production, and MFLFP are all endogenous. This framework is calibrated to the households formed circa 1875, 1900, and 1925 including information on changes in the gender wage gap, returns to schooling, levels of human capital inputs, and infant and child mortality. The model’s flexibility and careful calibration allow us to assess many of the explanations for family change considered above.

MODELING THE HOUSEHOLD Determinants of Human Capital For each generation of adults there are four determinants of adult human capital; schooling, labor market experience, unskilled labor, and gender. The human capital of an adult male in period t is ½h0m þ h^t Emt while that of a female is ½h0ft þ h^t Eft . Here, h^t is units of schooling human capital bequeathed by the parents of t − 1 to their children. h^t is assumed to be equal across males and females. h0ft (respectively h0m) is the stock of “unimproved” human capital available for use while a dependent youth in t − 1 and as an adult in t. It is associated with nature’s endowment, learning-bydoing or observation prior to market work for females and males respectively. As explained in the section on calibration of the initial baseline, h0m is assumed constant through time, while h0ft increases as the premium to strength declines (yet h0m > h0ft for all t). Such unskilled human capital includes any minimum legal or cultural requirements of parents to provide

Explaining the Revolution in U.S. Households

13

food and attention to their children. Thus, it is the “no-schooling” stock of human capital. Eft (respectively Emt) indicates how schooling and unskilled human capital are augmented by work experience for females and males respectively. The potential earnings of males and females are determined by the market valuation of their stocks of human capital. The market places the same value wt on units of unskilled and skilled human capital, whether provided by males or females. Thus the potential earnings of a male beginning adulthood in t are wt h0m Emt þ wt h^t Emt = wt ½h0m þ h^t Emt = wt hmt

ð1Þ

where hmt = ½h0m þ h^t Emt is the male’s stock of human capital in adulthood. Similarly, the potential adult female earnings are wt ½h0ft þ h^t Eft = wt hft

ð2Þ

Combining Eqs. (1) and (2) yields the potential household earnings wt ðhft þ hmt Þ = wt ht The ratio of the earnings of an adult female working full time with average experience to those of an adult male working full-time is γt =

wt hft hft = wt hmt hmt

ð3Þ

Hence, the gender wage gap is10 1 − γt

ð4Þ

Production of Schooling Human Capital Schooling human capital is acquired during dependency and deployed during adulthood. Parents forming households at t choose the quality and quantity of their children’s education. The quality, xt, is determined by goods inputs such as teachers and books, while the quantity, st, is the fraction of a child’s youth devoted to schooling. Parents also choose the time spent on human-capital enhancing activities by mother et, the effectiveness of which depends on her human capital hft. Thus, the schooling human

14

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

capital produced in accordance with parental choices in t, and which children can deploy as adults in t + 1 is given by h^t þ 1 = bt sθt s xθt x ðhft et Þθhe

ð5Þ

where bt is an efficiency scalar, and θs, θx, θhe ∈ (0,1) are production function parameters (elasticities). Multiplicative human capital production functions have been a workhorse in labor economics for decades (Ben-Porath, 1967; Heckman, 1976; Lord, 1989). This specification implies own-price elasticities of 1, and cross-price elasticities of 0. Other implications are discussed in the sections on “calibration of the initial baseline” and “missing the targets.”

Preferences Parents care about the number of children surviving to adulthood nt (half of whom are boys), the earnings in adulthood of those children, and the consumption of household produced goods. These sentiments are embodied in the utility function Ut for parents beginning adulthood in t nt 2 nt þ σlnwt þ 1 ðh0ft þ 1 Eft þ 1 þ h0m Emt þ 1 Þ 2

Ut = lnGt þ ψlnwt þ 1 h^t þ 1 ðEmt þ 1 þ Eft þ 1 Þ

ð6Þ

where Gt is the consumption of household production goods. Potential earnings of children in adulthood derive from two sources. The third term is the earnings across all surviving children derived from unschooled (i.e., unskilled) human capital. Parent’s relative taste for these “unimproved” earnings is σ; such earnings may be increased by choosing to have a larger number of surviving children nt. The aggregate earnings of adult children associated with schooling human capital is given by the second term. The intensity of preference for such earnings is captured by ψ, and these earnings may be increased by having more surviving children nt and/or by investing more in their education (which increases h^t þ 1 ). Andreoni (1990) argues that altruists get a “warm glow” from their own contributions to a recipient, and therefore place a higher valuation on the marginal dollar of own contributions to a recipient’s well-being than to those from other sources. His findings suggest that there is no reason to

Explaining the Revolution in U.S. Households

15

expect σ and ψ, to be equal. Both ψ and σ embody a taste for quantity of children. However, only ψ reflects a taste for improving the quality (i.e., schooling) of individual children. For this reason an increase in ψ/σ is viewed as an increased relative preference for quality of children over quantity of children. If the preferences of parents were purely altruistic toward their children, parental utility would depend upon the utility of children (cf. Becker, 1981) rather than the potential earnings in adulthood of children However, since the utility of children is likewise increasing in the their potential earnings (and thus consumption) these preferences and those of children are closely aligned. Furthermore, the formulation above facilitates discussion of alternative mechanisms in the results section, while avoiding the nearly insuperable data demands of infinitely lived economic agents. With logarithmic preferences, the utility function is strictly quasi-concave and monotonically increasing in each argument. Parental choices are made over Gt, nt, and h^t þ 1 , and are constrained in various ways, which we now explain.

Constraints All adults marry for life upon reaching adulthood and immediately make all decisions for the new household’s remaining life. Fathers work full time. Mothers allocate time in their adulthood among household production, market work, and children. The market earnings of fathers, mothers, and older children are spent on family consumption and developmental inputs. By accounting for these uses of time and goods we develop below an overall budget constraint for the family. The Life Cycle and Time Use Period and Mortality Structure. Childhood is spent under the direction and care of parents. Childhood and adulthood each last one period, but adulthood is imagined to last twice as long as childhood.11 Not all live births result in a child who survives to adulthood. The number of live births to a mother at time t required to produce 1 child surviving to adulthood in t + 1 equals d1t. d1t exceeds one for two reasons. First, some children die within the first year of life (infant mortality). Indeed, a significant portion of all infant mortality is neonatal, occurring in the first weeks of life (some other conceptions are carried nearly to term and naturally aborted late, or perhaps stillborn). Second, some children who survive infancy die before reaching adulthood. d1t reflects both types of mortality, so that as either

16

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

declines so will d1t.·d2t is the number of children surviving infancy necessary to produce 1 child reaching adulthood; d2t reflects only youth mortality. Mother’s Time Allocation. Mothers devote time to household production, raising children and the labor market. Each live birth requires significant mother’s time to activities largely unrelated to the child’s quality, whether the child survives infancy or not. Even deaths occurring within the first year of life impose large time costs of ρ for mother in terms of lost productivity during pregnancy, recovery following delivery, time to nurse and tend while the infant survives, and time grieving an infant’s demise. Each child surviving infancy imposes additional time costs of ρ on mother during its dependency largely unrelated to child quality. These include “picking up” after children, laundry, dish washing, etc. Since most such chores require little skill, we assume that the time required is independent of the stock of mother’s human capital. Mothers devote et units of time to the development of human capital in each young child surviving infancy. This “quality” time includes activities such as reading and talking to, and educational play with, the young child. It also can reflect, as in Mokyr (2000), time spent learning about and preparing safe and nutritious foods, household cleaning directed at reducing the population of bacteria and viruses in the household, or monitoring activities designed to protect the child from accidents. We suppose that the productivity of mother’s time devoted to such human capital development increases linearly in her human capital. zt units of mother’s time are allotted to household production. zt is combined with market goods ct to produce household consumption goods Gt. These goods are consumed by parents throughout their adult lives; Gt also includes any household public goods which are enjoyed by children as well as parents.12 Mothers may also devote time to the labor market, mt (such time is not determined by where it is performed
home/factory/ office/store
but by its pecuniary motivation). In combination, these uses of time are constrained by the 1 unit of time at mother’s disposal. Thus, mother’s time use must satisfy nt ½d2t ðρ þ et Þ þ d1t ρ þ mt þ zt = 1

ð7Þ

Children’s Time Budget. Dependency lasts one period. Each of the d2tnt children surviving infancy has Tt < 1 units of productive time, since very young children cannot work at all and older children lack the stamina and strength and concentration to work full time (Lord & Rangazas, 2006).

Explaining the Revolution in U.S. Households

17

Parents decide how much time, lt, older, potentially wage-earning, children should contribute to the household budget through market-oriented work (performed in the market or at home) and how much time st to spend in schooling. Hence, the time constraint faced by each child is given by st þ lt = Tt

ð8Þ

Sources and Uses of Money Income In addition to goods used in household production there are goods outlays on the quantity and quality of children. Parents spend d2tτtwtht for each surviving child on clothes, housing, and other child consumption items that tend to mechanically increase with a family’s standard of living, yet have little effect on child quality (such goods are the numeraire). Although we believe such expenditures to be common, they are little-treated in the literature; they also prove important to the calibration. Parents also spend money for children’s schooling or developmental inputs xt, with each unit costing Pt. Public financing of primary schooling made a given family’s cost independent of usage by the late 19th century, (unless one desired to supplement the short school year of the time). For older children attending high school, tuition was less subsidized. Overall, the cost across all goods inputs (including books, educational toys and broadening vacations) is less than one, and was falling over time with the further expansion of public schooling. Total goods expenditures across all children are given by nt d2t ðPt xt þ wt ht τt Þ

ð9Þ

Market earnings for a husband beginning adulthood in t are wthmt. The potential earnings of the wife (i.e., should she devote all time to market labor) are wthft. Older children can work in t, but if they do, offer only unskilled human capital to the market while dependents; for females, h0ft + 1. Due to less strength and concentration as compared to adults, children earn only μtwt per unit of unskilled human capital, with μt∈(0,1). These potential earnings in period t are therefore d2tμtwtnth0tTt, where h0t + 1 is average unskilled human capital across males and females (h0m + h0ft + 1)/2. Actual earnings of children are below potential earnings to the extent children spend time st in school. Altogether potential household money income is wt ht þ d2t μt wt nt h0t þ 1 Tt

ð10Þ

18

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Combining the results from Eqs. (7), (8), (9), and (10), the family’s overall budget constraint is expressed as13 wt ht þ d2t μt wt h0t þ 1 Tt nt = d2t μt wh0t þ 1 st nt þ nt ½d2t ðρ þ et Þ þ d1t ρhft wt þ nt d2t ðPt xt þ wt ht τt Þ þ zt wt hft þ gt

ð11Þ

The household’s potential labor income is given on the left-hand side. The right-hand side gives the total spending on, respectively, the implicit costs of schooling older children, the implicit cost of mother’s time devoted to quality and quantity of children, the money outlays for kids’ education and consumption, the implicit costs of mother’s time devoted to household production, the goods used in household production. Household Production We assume that household production is governed by the equation Gt = gνt ðhft zt Þ1 − ν

ð12Þ

As specified, the productivity of the wife’s time in household production is increasing in her human capital. This is certainly plausible, but below we see that mother’s optimal choice of household production time zt is independent of hft. We have noted that fathers work full time in marketoriented labor and that older children work when not in school. Of course, especially in the 19th century, fathers and children were also engaged in household production. To the extent they work “at home,” their labor efforts are implicitly priced at their market wage with the cost included in gt. Consequently, the model does not require us to distinguish where the work of children and fathers is performed or whether work performed at home is for family consumption or sale to the market. Similarly, domestic servants are hired inputs and likewise included in gt. As men and children work more outside the home, and as domestic servants are released, intermediate market goods (e.g., store-bought flour and clothes, and washing machines) become more important. The multiplicative production function implies own-price elasticities of 1 and cross-price elasticities of 0. Optimization Parents of generation t choose the quality of children h^t þ 1 based on input choices xt, st, and et quantity of surviving children nt and public goods

19

Explaining the Revolution in U.S. Households

consumption Gt based on input choices zt and gt, so as to maximize their utility function given by Eq. (6), subject to constraints Eqs. (7) and (11). The Lagrangian Lt is written: Lt = ln gνt ðhft zt Þ1 − ν þ ψ ln wt þ 1 bt st θs xθt x ðhft et Þθhe ðEmt þ 1 þ Eft þ 1 Þ þ σ ln wt þ 1 h0t þ 1 ðEmt þ 1 þ Eft þ 1 Þ 
þλ

nt 2

nt 2

d2t μt wt nt h0t þ 1 ðTt − st Þ − zt wt hft − nt ½d2t ðρ þ et Þ þ d1t ρhft wt − gt



þ wt ht − nt d2t ðPt xt þ wt ht τt Þ ð13Þ

The first order conditions (FOCs) for the optimal choices of gt, zt, xt, et, st, and nt are: v=gt = λ

ð14Þ

ð1 − vÞ=zt = λwt hft

ð15Þ

θx ψ=xt = λnt d2t Pt

ð16Þ

θhe ψ=et = λnt d2t wt hft

ð17Þ

θs ψ=st = λnt d2t μt wt h0t þ 1

ð18Þ


 ðψ þ σ Þ=nt = λ ½d2t ðρ þ et Þ þ d1t ρhft wt − d2t μt wt h0t þ 1 ðTt − st Þ þ λd2t ðPt xt þ wt ht τt Þ

ð19Þ

These FOCs reveal standard intuitions. Eqs. (16
18) govern the demand for human capital inputs. They all balance the left-hand-side marginal utility of accumulating human capital (and therefore child earnings in adulthood) against the utility cost from foregone parental consumption of doing so. Notice that in each equation this cost is increasing in fertility ntd2t, so that as stressed by Becker (1981) the price of child quality is increasing in the quantity of children. Further, in Eqs. (16) and (17) which govern the developmental inputs for perishable children, this price of quality per surviving child is increasing in d2t since the higher is child mortality, the more children must be born in order to produce a surviving one. The cost of mother’s and older children’s time inputs are increasing in their respective wages. Similarly the goods input prices enter into their FOCs for goods. Eq. (19) governs the choice of number of surviving children. Notice that all

20

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

human capital inputs enter into the price side of this expression. So, in Becker’s symmetry, the price of child quantity is increasing in child quality. Additionally, this price of quantity also increases in the various fixed costs associated with each surviving child (both goods and time, for both young and older children). Solving the system of optimality conditions above yields the explicit demand functions discussed below. The Quality and Quantity of Children Proposition 1. Parental investments in child quality are given by youth schooling inputs xt and st and mother’s time devoted to children’s human capital production et. The quantity of surviving children is nt (so that fertility is d1tnt). These investments are given by:

nt =

  θx ψwt d2t ht τt þ hft ðd2t ρ þ d1t ρÞ − d2t h0t þ 1 μt Tt 
  P xt = i = s; x; he d2t Pt ψ 1 − i θi þ σ

ð20Þ

  θs ψ d2t ht τt þ hft ðd2t ρ þ d1t ρÞ − d2t h0t þ 1 μt Tt 
 P  i = s; x; he st = d2t μt h0t þ 1 ψ 1 − i θi þ σ

ð21Þ

  θhe ψ d2t ht τt þ hft ðd2t ρ þ d1t ρÞ − d2t h0t þ 1 μt Tt 
 P  et = i = s; x; he d2t hft ψ 1 − i θi þ σ

ð22Þ


 P  ht ψ 1 − i θ i þ σ  i = s; x; he ð1 þ ψ þ σ Þ d2t ht τt þ hft ðd2t ρ þ d1t ρÞ − d2t h0t þ 1 μt Tt

ð23Þ



Proof: See Appendix A

’

Discussion. Notice that all of these optimal solutions share a similar structure. Consider first the human capital input, or “quality,” variables xt, st, and et. The numerators differ only in that each contains the exponent in human capital production for that input; each denominator includes the “own” input price, but is otherwise the same. The common term inside the braces in the numerator for each expression is the cost, net of potential benefits, of an additional child surviving to adulthood independent of quality (fixed costs of child consumption and mother’s time inputs for quantity minus potential child earnings). The common term inside the rounded brackets in the denominator reflects the cost of increasing quality (which is

Explaining the Revolution in U.S. Households

21

lower the higher are the returns to scale in human capital production). Thus, an increase in the numerator relative to the denominator is associated with a higher relative price per surviving child, and leads to an increase in the child quality variables. Notice, also, that these common terms are “flipped” in the expression for surviving children, so that a higher relative price per surviving child leads to a reduction in nt. These considerations are central to the “quantity-quality” trade-off in both the family-level and economic growth literatures (Becker, 1981 and Galor, 2005). Selected Comparative Statics. Suppose that schooling human capital rises across generations but that unskilled human capital remains constant. Then the human capital of adult males hmt and females hft increase (relative to t − 1) while unskilled human capital of children is unchanged. Higher parental human capital increases the ‘fixed’ costs of child consumption and of mother’s time inputs related to quantity; potential child earnings
which depend on the quantity of unskilled labor
are constant. Thus, the net costs of child quantity increase, inducing a substitution away from quantity toward quality (xt, st, et all increase and nt falls). Viewed differently, nt falls as the net cost of children increases by a larger percentage than potential parental earnings. Ceteris paribus, then, hmt + 1 and hft + 1 will also increase, increasing xt + 1, st + 1, and et + 1 and so on. This effect becomes weaker through time, though, as parental human capital rises relative to that of their dependent children. Thus, as Lord and Rangazas (2006) note, there is a supply-side element associated with any initial rise in human capital which carries forward into future generations. In their framework, as parental earnings rise relative to potential child earnings, there is a declining opportunity cost of schooling children (the utility loss from forgone parental consumption is lower at higher levels of consumption). Jones and Tertilt (2008) show that, empirically, fertility and income have varied inversely since at least the middle of the 19th century in the United States. Since human capital has risen over this time, and human capital increases income, this paper’s model is consistent with that pattern. As in Lord and Rangazas (2006), Becker and Barro (1988), and Barro and Becker (1989), as examples, increases in wt have no effect on fertility, due to offsetting income and substitution effects. Infant Mortality. Parent’s choices over quantity and quality are affected by their expectations of infant and child mortality. A reduction in infant mortality reduces d1t with no effect on d2t. Since lower infant mortality means fewer times mother must expend ρ units of time in order to produce a

22

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

surviving child, the cost of a surviving child falls with d1t. Since schooling and other human capital investments do not occur during infancy, those costs are unaffected by changes in infant mortality. Overall, though, the relative price of human capital inputs rise, so that quantity of children is substituted for quality: xt, st, and et fall while nt rises. This result is also obtained in Doepke (2005) and Barro and Becker (1989). The number of children ever born to a cohort, or just fertility, is d1tnt. Inspection of Eq. (23) reveals that there is an ambiguous effect of reduced infant mortality on fertility. That is, even though the number of surviving children demanded has risen, the fact that fewer births are required to produce a surviving child makes the effect on births unclear. Youth Mortality. Suppose now that youth mortality declines, while infant mortality is unchanged. This reduction in d2t also reduces the number of children born required to produce a surviving child d1t; indeed, d2t and d1t would fall by the same percentage.14 This would not affect any of the quality variables as the numerator and denominator would each fall the same percentage. The relative prices per child surviving to adulthood of quantity and quality are unchanged. Notice that there is also no change in the number of births d1tnt or the number surviving to childhood d2tnt. However, the number of surviving children nt clearly falls (in percentage terms the decline is 1 divided by the percentage decline in d2t and d1t). In the current framework, child quality variables are unlikely to rise as mortality falls. This differs from Soares and Falcao (2008) who put less emphasis on the fact that falling mortality reduces the costs of quantity (as well as quality). Notice that goods inputs xt increase with the wage per unit of human capital wt, whereas the time inputs st and et do not. All components of the “mechanical” costs of quantity include human capital variables and so increase with wt. However, the price of the time inputs of mothers in et and of children in st is proportional to wt. With the numerator and denominator increasing by the same percentage, wt drops out of their solutions. However, the cost per unit of xt is Pt. Consequently, wt remains and xt increases with wt/Pt. Finally, notice that the solutions for the quality variables are increasing in θx, θs, and θhe. Costs of Children. A child is more expensive the greater is the share of potential parental earnings spent on each child’s consumption, τt. An increase in that share also reduces the relative price of child quality. Thus, the quality variables xt, st, and et are all increasing in τt, while the number of surviving children nt declines. The potential economic benefit of children

Explaining the Revolution in U.S. Households

23

is increasing both in the total time children are available for work, Tt, and in their wage per unit of human capital as compared to parents, μt. When these are lower, the net costs of children of given quality (fixed costs minus potential benefits) are higher. Thus, lower values for these parameters induce substitution toward the quality variables and away from fertility. Goods Inputs and Mother’s Time in Household Production Proposition 2. Goods inputs in household production are gt =

vwt ðhmt þ hft Þ ð1 þ ψ þ σ Þ

ð24Þ

Optimal mothers’ time in household production is given by zt = Proof: See Appendix A

ð1 − vÞðhft þ hmt Þ ð1 þ ψ þ σ Þhft

ð25Þ ’

Eq. (24) reveals that an increase in the household’s potential wage earnings, arising from any combination of higher wages, or higher human capital for males or females, serves to increase the use of market goods in household production. Notice that if human capital of females hft increases by a larger percentage than that of males, the time mothers spend in household production zt falls. That is, a reduction in the gender wage gap induces mothers to reduce time in household production, and increase time devoted to market work. Intuitively, the more expensive is mother’s time input, the less of it is used in household production (as shown in the denominator). This is only partially offset by a wealth effect (present in the numerator). Note, though, that if hmt were to increase with no change in female human capital, the derived demand for zt would increase (as in De Vries, 2008). That is, households with high-earning husbands demand lots of household public goods and so demand lots of mother’s time input. Note that zt is independent of infant and youth mortality. Taking the ratio of Eq. (25) to Eq. (24) shows that an increase in the wife’s human capital reduces the ratio of her time input to goods inputs, so that household production becomes more goods intensive over time. The goods’ intensiveness of household production also increases with increases in the wage per unit of human capital, even if hft is constant. Recall that the time inputs of children and domestics are valued at their wages and

24

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

included in gt. Consequently, we can infer that the increased expenditures on store-bought goods inputs characterizing the second industrial revolution exceeded in magnitude the reduced expenditures on child and domestic inputs. The mother’s time constraint (Eq. (7)) shows that her labor market time increases with endogenous reductions in household production, child investment time, and the number of surviving children; it also increases if the exogenous time costs of child quantity (ρ and ρ) fall over time. Calibration exercises reveal the relative importance of these different sources of change in market orientation.

ASSIGNMENT OF PARAMETER AND TARGET VALUES, AND CALIBRATION OF THE INITIAL BASELINE The model above was calibrated to examine the quantitative significance of several proposed determinants of the revolution in households formed between 1875 and 1925. A meaningful calibration requires that the model’s parameters be “pinned down” to capture a fact or otherwise be reasonably restricted. Otherwise, if numerous parameters are “left free,” sensitivity analysis may produce such a large range of possible impacts that few meaningful conclusions are possible. This section also produces estimates of the changes in schooling human capital-endogenous targets which are not directly observed. Table 1 helps link the household decisions and notation within the model to the life cycles of the various cohorts. Consider, for example, the birth cohort of 1850, which forms households in 1875. Column (2) reviews the composition of the human capital used in adulthood by that cohort. Assuming the adult work life is 40 years, males work full time in the labor market between ages 25 and 65, or the years 1875
1915. Column (3) notes that over those same ages and years, households choose the household production inputs (the wife’s time and market goods) and the wife’s time in the labor market. Column (4) suggests the interpretation that children are born to that household when parents are between the ages of 26 and 30
and thus over the years 1876
1880. Finally, in Column (5) children commence school at age 6, with the first starting when parents are age 31 in 1881. The potential schooling period is ages 6
19, so the last child would be finished

Married Male Working Full-Time Utilize HCa (Ages 25
65)

Choose Household Production Inputs and Mother’s Labor Supply (Ages 25
65)

Have Children (Ages 26
30)

Choose HC Inputs Schooling HC (Ages 31
45)

1875
1915 hm1875 = ðh0m þ h^1875 ÞEm1875 hf1875 = ðh0f1875 þ h^1875 ÞEf1875

1875
1915 z1875, g1875

1876
1880 d11875n1875

1881
1900 x1875, s1875 e1875, h^1900

1900
1940 hm1900 = ðh0m þ h^1900 ÞEm1900 hf1900 = ðh0f1900 þ h^1900 ÞEf1900

1900
1940 z1900, g1900

1901
1905 d11900n1900

1906
1925 x1900, s1900 e1900, h^1925

1925
1955 hm1925 = ðh0m þ h^1925 ÞEm1925 hf1925 = ðh0f1925 þ h^1925 ÞEf1925

1925
1965 z1925, g1925

1926
1930 d11925n1925

1931
1950 x1925, s1925 e1925, h^1950

Cohort Born

1850

1875

1900

a

Choice Variables and HC Entering Adulthood.

m1875

m1900

m1925

Explaining the Revolution in U.S. Households

Table 1.

HC: Human Capital.

25

26

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

by 1900.15 Parents implement their human capital investment plans over those same years. Some values for parameters and targets are easier to assign than others. For example, published infant and youth mortality rates in different years
the basis for d1t and d2t
can reasonably be assumed to be fairly accurate. Some, such as the gender wage gap in different years (which determines γt) have been carefully estimated and are used with confidence. Similarly, targets for the goods and time inputs in human capital production xt and st can be inferred from government publications and the prior literature. The taste parameters are difficult to estimate directly, but are chosen so as to achieve specific values for the quality and quantity of children in the initial baseline. Among the more difficult items to determine are the time paths for the efficiency parameter in human capital production bt; unskilled human capital among females; schooling human capital h^t ; and the wage per unit of human capital wt. Fortunately, careful estimates by Goldin and Polachek (1987) of the gender gap through time and of the determinants of the narrowing of that gap allow inference of bt, h0ft, and h^t . Then, with h^t in hand, estimates of the growth in income across cohorts are developed and then used to determine wt. The method of calculating bt, h0ft, h^t , and wt are fully described in the text below. For many parameters and targets, only the main results are presented in the body of the text, with additional details found in Appendix B or footnotes. For easy reference, Table 2 lists all parameters and targets, including a few words on how they are determined. Gender Wage Ratio γt and Experience Impacts Emt and Eft Goldin and Polachek (1987) and Goldin (1990) calculate the time-path of γt and also estimate the roles played by changes in unskilled human capital, schooling human capital, and experience to the narrowing of that gap over time. Their study thus provides an excellent, consistent framework for the determination of various parameters and targets that would otherwise be difficult to ascertain. Goldin and Polachek (1987) find that the female to male ratio of earnings among full-time employees across six occupations closed from .463 in 1890 to .556 in 1930, further narrowing to .603 by 1970. Most of the narrowing occurred by the 1930s and Goldin (1990, p. 62) notes this ratio in the economy as a whole “was virtually stable from 1950 to around 1980.”16

27

Explaining the Revolution in U.S. Households

Table 2. Variables h^t xt st lt d1t*nt nt mt zt γt Variables g t

et Parameters θs θx θhe bt Emt Eft Pt wt μt hom hoft Tt τt vt

Parameters and Variables Description. Targets

Units of schooling HCa bequeathed by the parents of t
1 to their children HC goods input per child; inferred from published estimates Student HC time input per child; inferred from published estimates Child’s time working during dependency; Tt − st Number of live births to a mother at time t; published estimates Number of children surviving to adulthood; published estimates Fraction of time mothers devote to the labor market; inferred from published estimates Mother’s time input in household production; restricted by values of mt, ρt, ρ t Hourly wage of adult female relative to that of adult male; inferred from published estimates Other Goods input in household production; optimal solution, Eq. (24) Mother’s HC time input per child; optimal solution, Eq. (22) Human capital Elasticity of HC with respect to st, commonly employed value from literature Elasticity of HC with respect to xt, commonly employed value from literature Elasticity of HC with respect to mother’s time input (hftet) Efficiency parameter in HC production, inferred from ht + 1 target Impact of work experience on life earnings of adult males; published estimates Impact of work experience on potential life earnings of adult females; published estimates Cost per unit of xt, inferred (see “Other Parameters,” “Falling µt, Markets and Norms” sections) Wage per unit of HC; inferred subject to constraint Hourly wage of child relative to that of adult male; inferred from published estimates stock of unskilled HC for males; assumed, determines scale stock of unskilled HC for females; inferred under constraint Potential work time of children in t; inferred under constraint Per child private consumption share of potential adults earnings; inferred from contemporary values goods inputs in household production; set to meet baseline mt

28

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Table 2. Parameters d1t d2t Parameter ψ σ Parameters ρ ρ a

(Continued ) Mortality

Number of births to a mother required to produce 1 surviving child; derived from published estimates Number of children surviving infancy to produce 1 surviving child; derived from published estimates Taste Determines relative taste for skilled earnings of children; “pins down” ht + 1 Determines relative taste for earnings derived from unskilled HC; “pins down” nt Other Mother’s non-HC time cost per child surviving infancy; inferred from time use surveys Mother’s non-HC time cost per birth; inferred from time use surveys

HC: Human Capital.

For purposes of calibration, the gender wage ratio confronting adults forming households in t, γt, is based on the economy-wide figure for t + 15
when that birth cohort is age 40, and roughly in mid-career. Then γ1857 = .463, the economy-wide gender wage ratio in 1890 as measured by Goldin and Polachek. Averaging their estimates for 1890 and 1930, for parents of 1900 (birth cohort of 1875) γ1900 is set equal to .52. The cohort born in 1900, becoming parents in 1925, might confront γ1925 = .57 (given the slow increase between 1930 and 1970). For the cohort born in 1925, γ1950 = .603, the economy-wide value for 1970; the decisions of this cohort are not targeted, but the gender wage gap it confronts reflects the human capital investments of the parents of 1925. The average impact of experience on human capital among households formed in 1875 (more generally in period t) is likewise based on that in the entire labor market in 1890 (in year t + 15). As calculated by Goldin and Polachek (1987) this is Em1875 = 2.53 for males and Ef1875 = 1.62 for females.17 For the successive cohorts of parents their results suggest that for males Em1900 = 2.63, Em1925 = 2.19, and Em1950 = 2.01. For females these impacts are Ef1900 = 1.55, Ef1925 = 1.48, and Ef1950 = 1.41. The slight reduction among females occurred even as the average experience among working women was increasing significantly. The downward trend among males and females is consistent with the discussion of the section “The Historical Setting and Literature Review,” which noted that over this period there was a substitution away from employer and industry specific on-the-job training toward the acquisition of general human capital in schools. Thus,

Explaining the Revolution in U.S. Households

29

human capital among males was rising over this period only because the increases in schooling human capital were (appreciably) greater than the decline in experience human capital. Among females, the decline in experience human capital was offset by increases in both schooling and in the productivity of their unskilled human capital. Determination of h^1875 and h0f1875 Information on the years of, and rate of return to, schooling are used to establish an estimate for h^1875 . The U.S. Census did not begin collecting information on educational attainment or income until 1940. The average years of schooling for the birth cohort of 1850 was perhaps six years. Smith (1986, Table 1) reports figures for pre-1865 birth cohorts by race. Upon weighting by population, this implies schooling of about 6.3 years. His data relies on replies from the 1940 census on education completed. Consequently, this figure is possibly inflated upwards both by survivor bias and (over)reporting bias. Bleakley, Costa, and Lleras-Muney (2013, Fig. 4) plot educational attainment among a sample of white Union Army soldiers from 1863 (the Gould Sample) with educational attainment a bit less than six years for those born in the mid-1840s.18 Making inferences primarily from microeconomic evidence in Goldin and Katz (2008) it is assumed that the rate of return to a year of schooling was 3.1% for the cohort born in 1850.19 The low returns to education at that time reflect some combination of the short school year and low expenditures per student (both discussed below). Labor demand considerations presumably played a role as well. With a return of 3.1%, a person born in 1850 with the average of six years of education earns 20% more than a fellow cohort member with no education, (1 + .031)6 = 1.2. If attainment was a bit lower than six years, the return would be a bit above 3.1%. Such a comparison in contemporary society would be nonsensical, as individuals with no education are in no way otherwise the same as those with average education. However, in the middle of the 19th century those without formal education need not have been “defective” (President Abraham Lincoln had about one year of formal education and his successor, Andrew Johnson, had no years of formal education). The unskilled human capital of a male parent h0m is set to 10, which simply establishes the scale for human capital. The 20% premium to an educated male worker thus implies that h^1875 = 2. As discussed in a subsequent section, results are little affected if we instead used higher (lower) values for

30

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

h^1875 , such as h^1875 = 2:8 (1.2), so long as other parameters are calibrated to be consistent with those values. There is now enough information to determine the unskilled human capital of females in 1875, h0f1875. To see this, recall the expression for the gender wage ratio (Eq. (3)) and apply it to parents of 1875. Then substitute in the values established above for γ1875, h^1875 , Em1875, Ef1875, and h0m to yield ^

γ 1875 =

½h0f1875 þ h1875 Ef1875 ^

½h0m1875 þ h1875 Em1877

  h0f1875 þ 2 1:62 = = :463 10 þ 2 2:53

which indicates that h0f1875 = 6.68. This value implies a premium to strength for males for the initial period of almost 50% which is both huge and plausible (cf. Goldin & Polachek, 1987, p. 147).20

Narrowing of Gender Wage Gap Due to Changes in Experience The reasons for the significant closing of the gender wage gap, and their relative importance, have been examined by Goldin and Polachek (1987) and Goldin (1986, 1990). Goldin and Polachek (1987) emphasize the changing roles for experience, unskilled human capital (i.e., strength) and schooling for this period.21 They estimate that the role of education (reflecting both changes in the rate of return to, and level of, education) in narrowing the gap is 50% more important than that of experience. They also conclude that a reduction in the premium to male strength was at least as important as that of experience. Supposing equal roles for experience and changes in the reward to strength, each explains about 29% of the total narrowing, with education then explaining about 43%. Recall Goldin and Polachek (1987) found Em1875 = 2.53 and Em1950 = 2.01, and for females Ef1875 = 1.62 and Ef1950 = 1.41. The ratio Eft/Emt rose from .64 to .70, or by 9.37% between parents of 1875 and 1950. Ceteris paribus, this increases the gender wage ratio to .463(1.0937) = .506, or by .043, or about 30% of the total increase in γt, .043/(.603 − .463). Schooling Target h^1950 , as Implied by Micro Evidence on the Gender Wage Gap As noted, increases in schooling human capital are targeted to contribute roughly 43% of the narrowing of the gender wage gap from .463 to .603

Explaining the Revolution in U.S. Households

31

Thus, increases in h^t alone must raise the wage ratio to .463 + (.43) (.603 − .463) = .523. The required value bequeathed by households formed in 1925 (and used by their adult children beginning in 1950), h^1950 , is then obtained from "

# 6:68 þ h^ 1950 1:62 = :523 10 þ h^ 1950 2:53

This produces a level of schooling capital for our last birth cohort, 1925, of h^1950 = 8:2. This makes the schooling human capital of the 1925 birth ^ cohort, compared to the 1850 birth cohort, equal to hh^ 1950 = 8:2 2 = 4:1; that is, 1875 schooling human capital must increase a little more than four-fold to generate the postulated narrowing of the gender gap associated with education.22

Unskilled Human Capital h0f1950 Finally, the premium to men’s strength was declining, enough to raise the wage ratio by 28% of the narrowing from .463 to .603, or to .463 + (.28) (.603 − .463)=.502. Assuming this was associated with rising valuations for unskilled female labor, we can solve for h0f1950 from   h0f1950 þ 2 1:62 = :502 12 2:53 Thus, by the birth cohort of 1925 the valuation of female unskilled human capital has risen to almost three-quarters that of men, h0f1950 = 7.41. Rendall (2010) confirms that a significant premium to strength existed into the 1980s. Notice that the restrictions on levels of male and female human capital from the gender wage ratio expression do not rule out increases in human capital from factors that change h0m, hft, and h^t þ 1 equiproportionally, as may occur with a general improvement in health (decline in morbidity). This point is pursued further in a later section.

Targeted Schooling Human Capital and Role of Education in Growth Accounting This section demonstrates this targeted increase in schooling human capital
from h^1875 = 2 to h^1950 = 8:2 (i.e., the schooling bequest of parents starting families in 1925)
is consistent with the contribution of education

32

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

found in the growth accounting literature. Comparison is made with several prominent recent findings. First, Turner et al. (2011, 2013) create original long-run data series for the United States, including output per worker and human capital. They use this data to engage in growth and development accounting exercises by states, regions and time periods. Over the course of the 20th century, they find that, nationally, education contributed 30.1% of total growth in output per worker (2013, Table 6), assuming a labor share of two-thirds. For the subperiods 1880 − 1920 and 1900 − 1960 the contributions are estimated to be 42.4% and 27.7% (2011, Table 6). That large contribution over the first sub-period is especially surprising in light of the extensive capital deepening and reorganization of production typically associated with the second industrial revolution. They abstract from unskilled human capital, making for a faster rate of growth in human capital from schooling than when-as in our approach-unskilled human capital is quantitatively important. They also assume a 10% rate of return to a year’s schooling, which is about double the rate of return to a year of common school (grades 1
8) found for 1915 by Goldin and Katz (2008, Table 2.5).23 In our paper, the contribution of education to growth is the growth in human capital due to schooling relative to income growth across cohorts born in 1850, 1875, 1900, and 1925 (and their educational inputs at age 15, or years 1865, 1890, 1915, and 1940). This produces a contribution of education to growth over the period of 20.9%.24 Although this is lower than the contribution as estimated by Turner et al., it is somewhat above other well-known estimates based on labor productivity growth, or output per man-hour. Since hours worked declined dramatically from the late 19th century into the middle of the 20th century, output per man-hour increased much more rapidly than did output per worker. For this reason, a given role for education is lower when expressed in light of labor productivity growth than in terms of output per worker. In order to compare our results with those based on output per man-hour, the rise in cohort output per worker is adjusted to reflect the decline in hours worked over this period. Sundstrom (2006, 47
48) reports that hours worked per week declined from about 62 in 1890 to 38 in 1960. This implies output per man-hour increased about 63% more than output per worker over this period. Expressed in terms of growth of output per man-hour, the schooling human capital target in our paper contributes about 15.7% of growth. This percentage is quite similar to Goldin and Katz (2008, Table 1.3) who find education explains about 14.3% of the rise in output per man-hour between 1915 and 1940. They use the previously discussed Iowa sample,

33

Explaining the Revolution in U.S. Households

weighting income differences at each level of education by the educational shares in the labor force. Their methodology is unaffected by the quantitative significance of unskilled labor. Abramovitz and David (2000, Table 1.5) conduct growth accounting over several intervals, including 1890
1927 and 1929
1966. Their figures imply that over the entire period 1890
1966, the contribution of education is about 10.3%. Unlike Turner et al., they find the contribution of education was especially low in the earlier period, which includes much of the second industrial revolution but only the first few years of the high school movement. Overall, despite significant differences in methodology, the targeted increase in schooling human capital from our parameterization fits easily within the range of recent prominent estimates. In the results section, growth accounting is revisited when results of this paper are compared with other quantitative assessments.

Human Capital Productivity Parameters, Inputs, and Role of Curriculum The production function for schooling human capital was given by h^t þ 1 = bt sθt s xθt x ðhft et Þθhe

ð26Þ

Values for the θs from the literature along with constructed estimates of the inputs enable us to infer the rise over time in the efficiency parameter bt. Returns to Scale in Human Capital Production Using schooling inputs is superior to using years of schooling, as the time spent in school and the educational inputs per year have increased dramatically over time. Are observed increases in schooling inputs consistent with the four-fold increase in h^t calculated above? To address this requires (i) measures of the schooling inputs for the 1850 and 1925 birth cohorts, and (ii) values for the exponents in the human capital production function (Eq. (26)). The exponent on an input in the human capital production function is its elasticity of human capital with respect to the input. All empirical evidence indicates that the time (or quantity of school) margin st is appreciably more productive than are schooling inputs such as teachers or books, the xt, which reflect school quality (Browning, Hansen, & Heckman, 1999 and Lord & Rangazas, 1993). A consensus estimate for goods is θx = .10; perhaps a little lower in recent times and possibly a little higher in earlier periods. A value of .10 has also been employed for mother’s time input hftet,

34

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

θhe = .10 (Rangazas, 2002). A broader range of values has been estimated for θs with most falling between .5 and .7. We employ an intermediate value of θs = .6 (see Browning et al., 1999 and Lord, 1989 for additional discussion). θs + θx + θhe = .8 are therefore the returns to scale in human capital production.25 Schooling Inputs: Expenditures and Time Table 3 shows the targets for goods and time inputs chosen by parents for their children’s human capital. xt/x1850 is the (constant dollars) ratio of the schooling expenditures for the children aged 5
19 of generation t relative to those of 1850.26 The input choices of 1850-parents are exogenous parameters which help determine the human capital of their children, the parents of 1875; the calibration chooses the 1875, 1900, and 1925 values for xt and st. Schooling expenditures and school attendance for those aged 5
19 are measured when the children are age 15. Thus, for parents forming households in 1875 their schooling input choices are measured when their first child would be age 15, in 1890, and given the notations x1875 and s1875. Table 3 shows both goods and time inputs increase dramatically. x1925/ x1850 is 14.9 (or 16.6 if we include college expenditures for those 18 and 19 years old in the last cohort). Rangazas (2002, Table 1, p. 935) reports that the share of GDP devoted to primary and secondary education rises from 1.0% in 1880 to 2.4% in 1940. In the initial baseline x1875 is chosen so that the ratio of schooling expenditures to father’s life earnings should is a bit above 1.0% (since women and children also contributed to earnings). Since our figures include college expenditures for the last cohort of parents, the targeted share for x1925 is 2.6%. Time in school triples between the 1870 and 1940 school years, with the fraction of days in a year attended in 1940 exceeding 30%; thus s1925 = .309. These figures are comparable to those Table 3. Households Formed in Year t

Goods and Time Schooling Inputs. 1850

1875

1900

1925

Year HC input observed (children age 15), t + 15

1865

1890

1915

1940

xt x1850 st

1

2.11

5.88b

14.9 (16.6)c

0.09

0.129

0.201

a

a

HC: Human Capital. Average of values for 1910 and 1920. c The value in parenthesis includes college expenditures among 18 and 19 years olds. b

0.309

Explaining the Revolution in U.S. Households

35

produced by Rangazas (2002, Table 2, p. 936). (Since youth is only half of adulthood, the targets for the calibration are the st values divided by 2.)27 Increase in h^t between 1850 and 1875 Birth Cohorts Recall, the schooling human capital bequeathed by parents of 1850 was (exogenously) determined to be h^1875 = 2:0. In the calibration, the parents of 1875 choose the schooling capital their children take into adulthood, h^1900 . As noted, the schooling inputs for a birth cohort are measured when cohort members are age 15. In Table 3, the schooling input per population member aged 5
19 increased 43% and the goods input 111% between 1865 and 1890. Given the elasticity of human capital with respect to time inputs is .6, and for goods inputs .1, all else the same schooling capital would rise by 37% between those cohorts. However, there was also some multifactor productivity growth in human capital production over this period as graded schools, curriculum reforms, and openness in the schooling of blacks began to occur. It is assumed that b increased 10% over this period, a bit less than .4% per year (.00388) (far below the average rate of multi-productivity growth in the entire economy for this period). Altogether, this implies an increase in schooling capital of a bit more than 50% between these cohorts, from h^1875 = 2:0 to h^1900 = 3:0. The efficiency parameter b1875 (based on 1890 schooling inputs) is then set to produce a value for h^1900 of about 3.0. In the model developed above, an increase in the schooling human capital of children from one generation to the next, increases the schooling inputs those children choose in adulthood for their children. Consequently, the stipulated increase in schooling human capital between the birth cohort of 1850 and that of 1875 impacts the subsequent time-path of the endogenous variables. The larger is that stipulated increase, the greater is the simulated increase in h^t for the next two generations. Thus, if one believes the increase in h^t between the 1850 and 1875 birth cohorts is appreciably above (below) 50%, other factors would have to contribute less (more) to an explanation of the rise in human capital across all generations studied (moderate-sized changes in the estimate of the 1865 inputs have only a small impact on the entire change between 1865 and 1890). This sensitivity is discussed in the results section. Empirical Increase in h^t with Implications for Increase in bt How much of the total projected increase in h^t , from 2 to 8.2, is due to increases in inputs, and how much to multifactor productivity change in the schooling sector? The portion due to increases in inputs is calculated using the time paths for goods and time inputs discussed above (see

36

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Tables B1 and B2) along with the production function elasticities (the θs). This increase derives from increases in schooling inputs and/or their quality. Using the inputs from the tables discussed above (including the 1865 adjustments to 1870 values), the assumed human capital production elasticities for inputs, and an assumed doubling in mom’s effective human capital input,28 we calculate schooling human capital for the birth cohort of 1925 which is 3.0 times that of the 1850 birth cohort. Recall, though, the increase in schooling human capital necessary to account for 43% of the narrowing of the gender wage ratio was by a factor of 4.1 (i.e., 8.2/2). The gap between the three-fold increase in schooling human capital based on observed inputs and that required to appropriately narrow the gender wage ratio in our framework then is associated with an increase in the efficiency parameter in human capital production bt. Consequently, the required increase in efficiency is 37% (i.e., 3.0(1.367) = 4.1). There is little quantitative guidance as to the rate of total factor productivity advance in the production of schooling human capital over this long interval. However, there are several reasons to believe it was non-trivial. First, this period witnessed the transformation from one-room school houses to graded schools across the country. Second, changes in curriculum dramatically increased the correspondence between the knowledge imparted in schools and the knowledge useful in the marketplace. For example, typing and business courses proliferated at the expense of Latin and dancing classes (Goldin & Katz, 2008). The baseline simulation increases bt by 10% across each generation; over the 75-year period, this corresponds to multifactor productivity growth in schooling of almost .4% per year. Although measures of economy-wide multifactor productivity growth for this period vary widely, this implied growth rate in bt is appreciably below such estimates. This is appropriate as education is typically
certainly later in the 20th century
envisioned to be a relatively low productivity growth sector.29

Cohort Income Change and the Wage Per Unit of Human Capital The growth accounting exercise discussed above relied on estimates of income across cohorts. The calibration also requires values for the wage per unit of human capital confronting a household formed at time t, wt. Values for wt may be determined endogenously, even though the framework is partial equilibrium. The key to tractability is that the human capital of husbands and wives is known when the household is formed (since

37

Explaining the Revolution in U.S. Households

schooling is already completed and the returns to experience are exogenous). The few steps required to determine wt are explained below. First, in Table 4, column (2), an estimate of the average income per worker over the working years
or permanent income
is shown for workers forming households in t, ypt. This permanent income per worker is a weighted average of the earnings of males and females of t,
 
  ypt = wt ½Lm;t hm;t þ 1 − Lm;t hft = wt Lm;t hm;t þ 1 − Lm;t γ t hmt where Lm,t is the proportion of the labor force that is male, and where the second equality uses the fact that hf ;t = γ hm;t .30 Table 4 indicates the values for Lm,t and shows the estimates of the male earnings for each cohort; values for γt were given in “Gender Wage Ratio γt and Experience Impacts Emt and Eft” section. Notice that wt is not uniquely determined for cohorts beyond the 1875households, but rather depends upon how the simulations affect human capital. However, once wt is known, the choice variables of t may be solved for. Knowledge of the choice variables xt and st allows calculation of h^t þ 1 , and the procedure is repeated for the next cohort. Table 4. 1 Age 25 in t 1875 1900 1925 1950 a

Calculation of Life Cycle Permanent Income Per Full-time Male, wthm,t. 2 ypt

3 a

7.56d 9.66e 13.79f 33.8g

Lm,t

b

.832h .794i .754 .627

4

5

γt

½wt hm;t E

.463 .52 .58 .603

8.30 10.71 15.37 39.67

6 c wt hm;t w1875 hm;1875 E

 c

1.00 1.29 1.84 4.78

Real output per worker in thousands of year 2,000 dollars is from Murphy, Simon, and Tamura (2008), Table 1. All figures are adjusted by labor’s share of GDP, assumed constant at 70% throughout the period. b The male share of the labor force is from Historical Statistics of the United States: Table Ba417
424
Labor force participation, by sex and race: 1850
1990. c ½wt hm;t E is the empirical estimate for male life cycle earnings in t; i.e., a parameter. d Average of 1880, 1890, and 1900. e Average of 1900, 1910, 1920, and 1930. f Average of 1930, 1940, and 1950. g Average of 1950, 1960, 1970, and 1980. h Average of 1880 and 1900. i Average of 1910 and 1920.

38

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Market-Oriented Work of Wives White wives seldom worked outside the home in the late 19th century. However, by the 1920s a significant proportion of new brides would
with interruptions
devote many years of adulthood to market labor. Table 5, column (3), reports life cycle participation rates (LCPR) for white married females of different birth cohorts. Figures are derived from Roberts’ (2007, Fig. 1.9); for each birth cohort he sums Census participation rates for several age brackets during its member’s adult years, then divides by the number of brackets (see also Goldin, 1990, chapter 2). LCPRs increased dramatically, from 2.5% among white females born 1855
1864 (attaining adulthood about 1880) to 21% among for those attaining adulthood about 1930. However, as stressed by Goldin (1990), Census figures prior to 1940 understate the market-oriented labor of wives working from home in family businesses.31 She estimates the undercount for 1890 and Sobek (1997) performs similar corrections for 1880, 1900, 1910 and 1920. Their cross-section participation adjustments are shown in column (4); these fall steadily after 1890. Column (5) presents an adjusted LCPR, termed ALCPR. It consists of the sum of the LCPR and (.75) times the adjustment in column (4). Scaling by .75 is done to account for Goldin’s finding that those working from home typically only worked part time; in contrast, before around 1940 she finds that most of those working outside the home did so full time. These adjusted participation rates more than double between the birth Table 5. Life Cycle Labor Force Participation of White Married Females. 1 Birth Cohort

2 Attain Adulthood

3a LCPR

4b Adjustment

5c ALCPR

6 AVG

1855 − 1864 1865 − 1874 1875 − 1884 1885 − 1894 1895 − 1904 1905 − 1914

1880 1890 1900 1910 1920 1930

2.5 3.5 4.4 6.8 11.8 21.0

9.8 9.9 8.7 6.4 6.4 3.2d

9.9 10.9 10.9 11.6 16.6 23.4

6.2 7.2 7.6 9.2 14.2 22.2

Figures from Roberts (2007, Fig. 1.9). Averages are computed across age groups of 20
29, …, 60
69. b Goldin (1990, chapter 2), Sobek (1997, Table 2.5). c Column (5) is sum of column (3) plus 0.75 times column (4) (see text). d There is no adjustment measure for 1930, the last census year before modern concept. Figures from Roberts (2007, Fig. 1.9). a

Explaining the Revolution in U.S. Households

39

cohorts of 1855
1864 and 1905
1914, rising from 9.9% to 23.4%. Those agnostic about such adjustments may prefer the simple average of the LCPR and ALCPR, or AVG. These are provided in column (6), and rise from 6.2 to 22.2% for those attaining adulthood in 1880 and 1930, respectively.32 Institutional responses to the Great Depression are not accounted for within the model. In particular, Goldin emphasizes that during the Great Depression marriage bars were extended to numerous sectors of the economy, reducing the employment of married women (Goldin, 1990, chapter 6). And, since work interruptions reduce the value of prior work experience upon re-entering the workforce, participation may have remained lower even after marriage bars were eliminated in the 1940s.33 Absent those marriage bars, LCPR for the new brides of 1925 would be higher. Consequently, since our model does not factor in marriage bars, it would not be surprising if the simulations overshoot empirical participation. Conversely, Goldin (1990, pp. 154
157) examines survey data from young women born between 1944 and 1954 regarding their expected future participation rates and finds that when rates have increased rapidly young women have underestimated their future participation. Less clear is (1) whether such underestimation occurred for the earlier cohorts we consider, and (2) whether their parents
who in our model control the human capital investments in children
may have better anticipated their daughters’ life cycle work. Mortality In 1900 the infant mortality rate was 16.24%, while the mortality rate for those ages 1
19 was 3.23%. With about 19.5% of children ever born dying during dependency, about d1,1900 = 1.24 live births were required to produce a child surviving dependency, while d2,1900 = 1.03.34 By 1925 the infant mortality rate had fallen to 7.54%, while the mortality rate for those aged 1
19 was 1.03%. Consequently, d1,1925 = 1.098 and d2,1925 = 1.014. Murphy et al. (2008, Tables 13
15) report that infant mortality was 17.1% in 1880, with an additional 12% of births dying between ages 1 and 15. These figures are used to produce d1,1875 = 1.47 and d2,1875 = 1.14. Fertility Targets The U.S. Census irregularly collected data on fertility in the first half of the 20th century. Jones and Tertilt (2008) use these Census responses to

40

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

estimate the children ever born in earlier periods. For the cohort born between 1851 and 1855, attaining adulthood in the 1870s, children ever born is 5.3 (based on responses in the 1900 Census). For the birth cohorts of 1876
1880 (chosen to reflect children born to parents setting up households in 1900), children ever born was 3.25 (based on responses in the 1940 census). For the birth cohort of 1901
1905, estimated from responses in the 1950 Census, fertility had fallen to 2.59.

Mother’s Time Allocation to Child Quantity Recall that each infant requires a proportion ρ of mother’s time for activities unrelated to child quality, while each older child imposes a time cost of ρ. Consider first the time costs beyond infancy, ρ. Ramey (2009) exploits time use surveys conducted in the 1920s to estimate how housewives’ time spent in home production varied with the number and ages of children. A woman with no children and at least some high school spent 44 hours per week in home production. The presence of children increased mother’s time in housework, with older children requiring less time: If the youngest child was between one and five years, Ramey finds housewives spent almost seven extra hours per week and if the child was between 6 and 15 years of age, the housewife spent an extra two hours per week. Thus, in 5 of the 19 years (ages 1
5), seven hours per week are devoted per child; from ages 6 to 19, that is, in 14 of the 19 years of dependency beyond infancy, two hours per week are devoted to children. Assume all of this time is spent on activities related to the quantity of children
(rather 
7  than
14
their  quality). Given the 5 2 þ 70-hour work week, mothers spend 19 70 19 70 = :048, or 4.8% of her time during the each child’s dependency from age 1 through age 19. Since the length of dependency is only half that of adulthood, ρ = .024. What about the time required per infant ρ? Ramey found each child under age one added 17 hours to the housewives’ work week. Indeed, Albanesi and Olivetti (2007) estimate breast feeding alone required about 14
17 hours per week the first year. They also find that episodes of incapacitation of mother during pregnancy and/or following childbirth early in the 20th century were more prevalent than today. They find the average pregnancy was associated with 4.5 unproductive months. Assuming productivity is reduced by 60% during the incapacity, this increases the time cost by an average of about seven weeks per pregnancy. All of the prepregnancy time loss and some portion of the post-pregnancy time costs should be added to the Ramey figures. Seven weeks represent 490 hours,

Explaining the Revolution in U.S. Households

41

which divided by 52 weeks implies that incapacity adds a little over nine hours per 
1 week
26 to the 17 from Ramey, for a total of 26. Consequently, ρ = ð:5Þ 20 70 = :009. The average infant absorbs about 1% of mother’s adult time endowment. Relative Wage of Dependent Children Recall that μth0t + 1Tt are the potential earnings contribution of a dependent child toward the family budget. Potential earnings are reduced when children are sent to school, and explicitly realized when children are employed in wage labor. When children are engaged in the household production of Gt their time is valued at the market wage and this expense is reflected in the goods cost of the household production good. The direct monetary contributions of children were significant in the late 19th century, but had become insignificant by the middle of the 20th century. Their earnings contributions declined mainly because the high school movement increased the time older children devoted to human capital accumulation. The ratio of a child’s wage per unit of human capital to that of an adult male in 1875 is assumed to be μ1875 = .4. This estimate is consistent with the evidence of Parsons and Goldin (1989) upon dividing the earnings per child of different ages from 10 to 19 and gender by their probabilities of working, and then averaging. As motivated above, μt is assumed to decline from μ1875 = .4 to μ1925 = .3. Other Parameters Parental expenditures Ptxt on the goods inputs per child xt are independent of the price; from Eq. (20) if Pt is 10% lower, xt is 10% higher. Because of the expansion of public education, and in particular the high school movement for the period in question, the price to parents of schooling inputs fell through time. The precise rate at which it fell is uncertain. As an initial assignment P1875 = .5, as most high schools were still private at that time, outlays for some books and other home inputs were not subsidized, and transportation costs were not trivial. As the high school movement proceeds, Pt falls. It is assumed that by 1940 Pt = .2, a reduction of 60%.35 Unsurprisingly, it is the percentage reduction rather than the initial value, that is, important to results. P1925 = .2 yields simulated goods inputs close to those in the data for x1925. The choice of P1925 is re-considered in the discussion of the “best fit” calibration.

42

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

The 1875 value for the time potentially available for dependent children to work is set to T1875 = .2. Notice that since dependency is only one half the length of the one period of adulthood, T1875 could not exceed .5. Further, if children in the first six years of life cannot work and children age 19 have left the home, the available maximum falls below one third. Finally, children between ages 6 and 12, say, lack the stamina to work full time and/or attentively, suggesting a value of approximately .2; and, within the narrow range of feasible values, .2 produces the best fit for the initial baseline. Other parameters are more difficult to measure and are chosen to produce the desired initial baseline. The exponent 1 − v on mother’s time in household production in 1875, z1875, is set to .26 which yields the previously estimated figure for time mothers devoted to market work m1875 = 5.8%.36 There is little direct evidence on the portion of lifetime parental income devoted to the private consumption goods of dependent children. Modern estimates of the non-human capital outlays per child in 2006 for middleincome families are around 6% (United States Department of Agriculture [USDA], 2006). The shares of household income devoted to housing and transportation have increased dramatically from early in the 19th century, while the shares of income devoted to food at home and clothing have declined.37 Some portion of those increases in housing have gone to reducing the number of children per bedroom (child privacy may be a superior good). Similarly, a portion of the higher income share devoted to transportation involves help financing cars and car insurance for teenage children, or transporting younger children to “play dates.” Although information on expenditures on toys and child entertainment is scarce, it seems probable that expenditure shares have increased for them as well. Supposing that private consumption expenditures on children were a superior good in the 20th century, we set τ1875 = .04 per child for the parents of 1875. Since contemporary fertility is lower, τ1875 = .04 leaves the share of household income devoted to child consumption across all children roughly constant. The unobserved taste parameters ψ and σ are chosen so as to pin down initial schooling inputs (especially ψ) and fertility (especially σ).

MISSING THE TARGETS: INCOME AND SCHOOLING COSTS; MORTALITY DECLINE This paper examines quantitative implications of several proposed explanations of changes in the behavior of married households formed between

43

Explaining the Revolution in U.S. Households

1875 and 1925. The calibration proceeds as a series of nested exercises, often adding features not considered in the preceding one. This results section begins with a description of the initial baseline.

Experiment 1: Growth of Public Schooling, and Changing Returns to Experience The Initial Baseline For this and subsequent experiments the choices made by the parents of 1875 are as targeted for the baseline: Parameters are set such that h^1900 , the schooling human capital chosen for children by the parents of 1875 is, 3.04, about 51% above the h^1875 = 2 embodied by those parents. Those parents choose about 5.3 births, send children to school about 13% of student’s time endowment, while wives’s devote 7.4% of their adult time endowment to the labor market, and about e1875 = .026% of their time endowment raising the human capital of each child. Aggregate schooling goods inputs are close to 1% of GDP.38 The parameter values common to all experiments is given in Table 6.39 Table 6. Parameters Yt θs θx θhe bt Pt T Emt Eft ρ ρ h0m h0ft

Input Table for Parameters That are Common to all Runs. Year 1875

Year 1900

Year 1925

1000 .6 0.10 0.10 1.23 0.50 0.20 2.53 1.62 0.024 .009 10.0 6.67

1286 .6 0.10 0.10 1.40 0.35 0.20 2.36 1.55 0.024 .009 10.0 6.93

1840 .6 0.10 0.10 1.54 0.20a 0.20b 2.19 1.48 0.024 .009 10.0 (10.0)b 7.19 (7.45)c

An alternative value for P1925 is discussed in “Falling µt, Markets and Norms” section, and Table 9. b T is permitted to change across periods in “Child Labor Restrictions and Compulsory Schooling Laws Reduce Tt” section. c Values in adulthood for children of 1925-parents.

a

44

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Experiment 1 Results Experiment 1 examines implications of changes in parental resources and the direct costs and efficiency of educational investments. First, it incorporates the exogenous rise in life cycle parental earnings (or “permanent income”) across generations. A second feature is the inclusion of the empirical experience profiles of men and women Emt and Eft. Third, the per-unit price Pt of goods inputs in human capital production xt is allowed to fall over time reflecting the end of school rate bills and the rise of public sector provision of high school. Finally, the efficiency parameter bt in human capital production rises over time as previously described. This simulation misses most targets by wide margins as shown in Table 7. Fertility declines only from 5.3 to 5.1, as opposed to the target of 2.6. The time input of students st barely increases, rising to only 103% of the its 1875 value, whereas the increase is 233% in the data (from .129 to .309). Goods inputs xt perform somewhat better rising to 500% of the initial baseline value compared to the target of 700%. Table 7. Variables

Experiment 1; Income, Public Schooling, Experience.a Year 1875

Year 1900

Year 1925

3.04

3.71

(a) Parental characteristics variables 2.0 h^t wt 1.0 w1875 0.463 γt

1.27

1.86

0.502

0.537

(b) Parental choice variables 2.36 xt st .128 et .0026 5.29 nt 5.29 d1t*nt g 796 t .738 zt .074 mt 3.04 h^t þ 1 γt + 1 .502

4.58 .136 .0025 5.15 5.15 1052 .698 .119 3.71 0.537

11.8 (16.5)b .134 (.309)b .0024 5.12 5.12 (2.60)b 1540 .668 .151(.233)b 4.46 (8.2)b 0.58 (.603)b

a

This experiment allows for rising exogenous income, falling price of schooling inputs Pt, and changing profile of premium to experience Em,t and Eft as in Table 6. Also, d1t = d2t = 1 since mortality is not considered in this experiment. b The figures in parentheses are targets for the variable in 1925. For x1925 the target is seven times the 1875 value; for st see Table B2; fertility target in text; for mt see Table 5; targets for h^1950 and γ1950 are given in the text.

Explaining the Revolution in U.S. Households

45

xt increases for two main reasons. First, the posited decline in Pt from .5 confronting the parents of 1875 to .2 for 1925-parents by itself increases xt by 150%. Second is the large increase in the wage per unit of human capital: The exogenous increase in male earnings across generations arises from some combination of higher wt and more human capital. With only a small increase in total parental human capital, the rise in wt from that confronting the parents of 1875 to those of 1925 is 87%, roughly equal to the rise in earnings across the generations. Ceteris paribus, this increases xt by 87%. In combination those two price changes account for 92% of the increase in xt. Parents of 1875 bequeath h^1900 = 3:04 to their children, whereas parents of 1925 bequeath h^1950 = 4:46 units of schooling human capital; the target though is h^1950 = 8:2. To appreciate why the fit of the model is so poor, consider again the solutions for xt (20), st (21), nt (23), and zt (25). The moderate increase in h^t þ 1 (5) which does occur is primarily due to three factors. First, is the aforementioned increase in xt.40 Second, the postulated 21% rise in multifactor productivity in the schooling sector increases h^t þ 1 by 21%. Larger increases in s1925 would also go a long way toward increasing schooling human capital since the output elasticity for st (θs = .6) is so much greater than that for goods inputs (θx = .1). Neither Pt nor wt appear in the expression for the student’s time input st or fertility. (Actually, wt enters in both the numerator of st via parental income and the denominator through the opportunity cost of student’s time, and so cancels out). If there were a smaller degree of substitutability between xt and st, reductions in Pt would increase st. Cinyabuguma, Lord, and Viauroux (2009) consider the implications of a human capital production function in which xt and st are perfect complements. In that case, although st then increases when Pt falls, the increase in xt is smaller and the implications for h^t þ 1 are essentially the same. The effect of higher parental human capital (which increases the costs of children but not their potential benefits) is weak in Experiment 1. Despite the small simulated increase in schooling human capital, the stocks of human capital in adulthood hmt and hft fall because the (exogenous) returns to experience declined. When the schooling targets are met in Experiment 3, the increase in schooling human capital is more than twice as large as the negative impact of experience among males, and about three times as large among females. Finally, mother’s labor market participation rises from 7.4% to 15.1% of the life cycle endowment, compared to a target of 23.3%. The rise in mother’s market time is driven by the narrowing of the gender wage gap which reduces time devoted to household production, zt.

46

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Since increasing schooling human capital narrows the gender gap, the fit for mother’s target will improve in experiments more closely approaching the h^1950 target. Further discussion of mother’s market work is postponed until Experiment 3. In summary, Experiment 1 reveals that additional factors must be examined if the simulated model is to fit the targets. As in the current study, Galor (2005) finds limited evidence that increases in per capita income to some threshold level contributes much to the demographic transition. “The simultaneity of the demographic transition across Western European countries that differed significantly in their income per capita suggests that the high level of income that was reached by those countries … had a limited role in the demographic transition (Galor, 2005, p. 228).” Similarly, Bleakley and Ferrie (2013) find that 19th-century random wealth shocks to families in Georgia had little effect on the education of those families’ sons or grandchildren. Also, Doepke (2004) points out that subsidies to education (here a reduction in Pt) need not cause a decline in fertility. “Since a subsidy lowers the cost of children, an education subsidy increases fertility among parents who would have sent their children to school even without the subsidy” (Doepke, p. 361). The findings of this paper thus echo similar findings for different times and places. Conversely, Lord and Rangazas (2006) conduct a quantitative assessment of a theory of long-run growth in the United States
from 1800 to 2000
which does reproduce central features of the quantity-quality tradeoff.41 The discrepancy between their rapid increase in schooling from higher income and education subsidies in the first decades of the 20th century and our much smaller response is largely a consequence of the specification of the human capital production function. Their framework contains two periods of schooling for dependent children. Significantly, the schooling of younger children is exogenous. Consequently, increases in human capital result only from increases in the schooling of older children. On one hand, they assume a lower elasticity associated with schooling time than we do. On the other hand, the percentage increases in schooling are much larger in their framework, as they are defined relative to the endogenous schooling (i.e., that among older students) only. In particular, suppose younger children spend .085 of their time endowment in school and total schooling time increases in the late 19th century from .105 to .12. Then, in their human capital production function, the schooling input has risen from .105
.085 = .02 to .12
.085 = .035; an increase of 75%. In our framework the percentage increase would only be 26.3%, from .095 to .12.42 Also, whereas skilled human capital is only about 16% of the total human capital among 1875-parents in our calibration, it is about 70% in

Explaining the Revolution in U.S. Households

47

theirs. To quantify the impact of these two differences, we use the targeted time inputs over our period in their human capital production function. Then, assuming the same growth rate of per capita income as in our calibration, education’s share of total income growth is about 30% for that period under their formulation, or almost 50% higher than ours. This is problematic as, aside from the outlier estimates of Turner et al., 30% appears “too large.” Since that framework and the one in this paper vary in other ways, we simply note that it is unclear whether their results would persist if the role of schooling capital in growth was lower.43

Experiment 2: Reduced Child Mortality Experiment 2 augments the parameter changes in Experiment 1 by incorporating the significant reductions over time in infant and child mortality (and thus in d1t and d2t). As discussed in the literature review and model development sections, declining mortality may increase the number of surviving children nt, as the significant costs of pregnancy must be incurred less frequently to procure a child surviving into adulthood. Nevertheless, children ever born d1tnt are likely to fall given the decline in d1t. Inspection of the expressions for the child quality variables xt, st, and et suggests the impact effect of mortality decline depends upon the relative percentage changes in d1t and d2t Empirically, the percentage decline in d1t exceeds that in d2t. The number of children ever born required to produce a child surviving to adulthood d1t fell from d1,1875 = 1.47 to d1,1925 = 1.098, a decline of 25.8%; this exceeds the decline of 10.7% in the number of children surviving infancy required to produce a child who reaches adulthood d2t (d2,1875 = 1.14 while d2,1925 = 1.018). Thus, there is a greater percentage decline in the fixed costs of producing a surviving child d2t ht τt þ hft ðd2t ρ þ d1t ρÞ than in the potential benefits d2th0t + 1μtTt. This induces a substitution away from quality toward quantity. Thus, the prospects for improving the fit of the simulation through this channel appears limited. Indeed, incorporating declining mortality decline produces results broadly consistent with Experiment 1. Results are displayed in Table 8. As anticipated, the percentage increases in xt and st from 1875 to 1925 are a bit smaller than the small increases found in Experiment 1. The simulation reveals that compared to the trivial reductions in fertility in Experiment 1, the reductions in mortality induce a larger fertility d1tnt decline, from 5.3 to 4.4 for those parents of 1925. Fertility remains far above the target of 2.6 and the decline in fertility which does occur is entirely a consequence of

48

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Table 8. Variables

Year 1875

(a) Parental characteristics variables 2.0 h^t

Experiment 2, Mortality.a Year 1900

Year 1925

3.00

3.63

wt w1875 γt

1.0

1.27

1.87

0.463

0.502

0.536

(b) Parental choice variables xt st et nt d1t*nt gt zt mt h^t þ 1 γt + 1

2.34 .128 .0025 3.63 5.33 849 .768 .074 3.00 .502

4.48 .133 .0024 3.93 4.87 1121 .727 .122 3.63 0.536

11.4 (16.5)b .129 (.309)b .0023 4.00 4.40 1641 .700 .157(.233)b 4.32 (8.2)b 0.577 (.603)b

a

In addition to those parameter changes of Experiment 1, the mortality exercise incorporates the time profiles of infant and child mortality as reflected in d1,1875, d1,1900, and d1,1925; and d1,1875; and d2,1875, d2,1900, and d2,1925. b The figures in parentheses are targets for the variable in 1925. For x1925 the target is seven times the 1875 value; for st see Table B2; fertility target in text; for mt see Table 5; targets for h^1950 and γ1950 are given in the text.

the decline in infant mortality, as reflected in the lower d1t. Indeed, as anticipated, there is a modest increase in the number of surviving children nt (from 3.6 to 4). The finding that falling infant mortality cannot account for the decline in net fertility rates is not new. Doepke (2005) shows analytically in a series of models that infant mortality decline may be expected to increase the number of surviving children (i.e., net fertility) while typically reducing total fertility. Doepke (2004) assesses the quantitative significance of those models via a calibration of fertility in England as infant mortality declines from 1860. His calibrations support the hypothesized inability of infant mortality decline to explain the decline of net fertility. Doepke also presents an extensive review of empirical studies across a broad range of countries. The empirical results are more mixed than his theoretical findings, or the quantitative findings for England. Nevertheless, he concludes that overall the empirical evidence supports the inability of infant mortality decline to explain the demographic transition. Our calibration for the United States

Explaining the Revolution in U.S. Households

49

over this period may be seen as complementing both Doepke’s theoretical findings (2005) and his quantitative findings for England (2004). More generally, Galor (2005, pp. 225
227) in his influential review of unified growth models and empirical evidence concurs that infant mortality decline was not the underlying cause of the demographic transitions across Western Europe.

REACHING THE TARGETS: INCREASING DISREGARD FOR THE POTENTIAL EARNINGS OF DEPENDENT CHILDREN, AND GREATER CHILD CONSUMPTION Quantity and Quality of Children The inability of the experiments conducted so far to reproduce stylized facts concerning education and fertility raises the question of what mechanisms may be capable of explaining those facts. Viewing the optimal solutions for the quality variables (xt, st, and et), and for quantity nt, reveals several candidates for strengthening the simulated quantity-quality tradeoff. First, recall that lowering Pt increases xt. However, while reductions in Pt in Experiments 1 and 2 have a powerful effect on xt, impacts on h^t þ 1 and nt are modest. Second, increases in mother’s time per child unrelated to child quality ρ or ρ would lower nt and increase investments in child quality. However, Olivetti (2006) and Greenwood et al. (2005) present evidence that from near the end of the interval considered here ρ and ρ were trending downward. Third, the taste parameters ψ and σ could be changed to improve the fit of the model to the targets. Economists, though, have an appropriate aversion to invoking a change in tastes with no motivation beyond producing desired results. In that spirit our calibrations will not adjust preferences, while the discussion acknowledges mechanisms capable of altering preferences over this long period. Yet to be considered are several parameters affecting the net cost of children’s potential earnings: Potentially, the quantity-quality targets might be achieved by some appropriate combination of a lower wage per unit of human capital of children relative to adults μt, lower time children are available for work Tt, and/or an increase in each child’s private consumption share τt. Below it is shown that such a combination exists and, we argue, is quite plausible. Parents, recall, view each dollar of potential earnings of children d2tμtwtnth0t + 1Tt as a one dollar expansion of the household’s potential

50

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

income. This assumption was warranted for middle-class households in the decades following the Civil War when the second earners in families tended to be children rather than mothers (Parsons & Goldin, 1989). Conversely, in recent decades there is little presumption that middle-class parents have the right to compel children to work and then remit earnings to the parents (Zelizer, 1985). Consequently, over this transitional period norms apparently changed in such a way as to frown upon child labor undertaken for the benefit of the family. Indeed, Moehling (2005) finds that by the second decade of the 20th century even daughters who turned over their paychecks to parents nevertheless enjoyed higher private consumption. Further, Zelizer (1985) argues that by the middle of the 20th century the jobs and earnings of child workers were viewed by parents as appropriate only as human-capital-enhancing experiences (learning the value of money, working and saving toward a goal, fulfilling obligations, etc.) rather than as contributions to the family coffers. Below we consider how the rise of such norms might be reflected within the model. Child Labor Restrictions and Compulsory Schooling Laws Reduce Tt Evidence of changing norms is the proliferation of legislation affecting children’s schooling and work. Massachusetts introduced the first state compulsory schooling law in 1852 and, by 1910, 41 states had such laws, while 40 had legislated restrictions on child labor (Goldin & Katz, 2008, p. 191). Assuming some groups would need to change behavior to comply, and that compliance was enforced, the impact on work and schooling could be modeled as a reduction in Tt.44 Puerta (2009) examines implications in the United States between 1850 and 1920. He finds compulsory schooling laws increased school enrollment in affected areas 7% relative to otherwise similar areas which did not pass such legislation. These impacts would be concentrated among the working and lower classes, among whom the effects may have been far above 7%. Further, such legislation may reflect new, more restrictive middle-class norms regarding child labor. Such norms also effectively lower Tt and may have preceded and been quantitatively more significant than the actual legislation. Prominent quantitative assessments by Doepke (2004) and Lord and Rangazas (2006) conclude that legislation affecting child labor and schooling was central to the decline in fertility during England’s demographic transition after 1840. The Forster Act of 1870 had made primary education the duty of the public sector. That Act significantly increased the public subsidization of education. In 1874, the Factory Acts were amended, raising the minimum age at which children could work and

Explaining the Revolution in U.S. Households

51

extending that provision to all industries. Doepke notes child labor restrictions “unambiguously increase the cost of children, and therefore lead to lower fertility (362)” and that fertility decline in England accelerated shortly after these changes. He conducts simulations of policies and demographic transitions in South Korea, Brazil and England, contrasting the effects of policies that subsidize education and restrict child labor. Unlike child labor restrictions, he shows subsidization policies can actually increase fertility by reducing the cost of children who would have been attending school without the subsidies. His results suggest child labor restrictions are an important aspect of English fertility decline. Lord and Rangazas attempt to reproduce through calibration the relatively unique longer-term pattern of English fertility-increasing fertility during the late 18th century through the first decades of the 19th century, and then its long decline. They are largely successful in explaining the rise in fertility into the early 19th century. However, the “Doepke effect” mentioned above, whereby fertility is encouraged by increased public subsidization of children already in school, helps keep fertility relatively constant after 1840. Only by invoking compulsory schooling legislation is fertility pushed down, setting off the demographic transition. Rather than a change in norms, Doepke and Zilibotti (2005) provide an explicit economic self-interest rationale for child labor restrictions. They point out that 19th-century working-class families had conflicting economic interests regarding child labor restrictions. On one hand, the labor earnings of own children led parents to oppose such restrictions. This effect increases with family size. On the other hand, since child labor is a substitute for unskilled adult labor, restrictive child labor legislation would confer higher wages to working-class fathers. This trade-off, they argue, was altered by SBTC, which raised the return to, and demand for, children’s education. But higher quality per child increases the cost of child quantity, so fertility falls. With smaller families, the child earnings foregone by restricting child labor are reduced. Consequently, the effect on father’s earnings comes to dominate and the working-class votes in favor of child labor restrictions. Thus, in response to SBTC, child labor restrictions are voted in by families already planning smaller families. Most researchers, including Goldin and Katz (2008) and Puerta (2009), have found only a modest role for Compulsory Schooling and Child Labor legislation. In the analysis below, reductions in Tt reflect some unspecified combination of new norms among the middle-class and/or binding restrictions affecting the potential work of lower-class children.

52

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Falling μt, Markets and Norms The potential earnings of children would also be lowered by a reduction in their wage per unit of human capital relative to adult males, μt.45 Lower μt reduces the opportunity cost to parents of schooling children and powerfully increases st; it also lowers nt by reducing potential earnings of children. The simulations above have yielded values for s1925 well below, and for nt well above, the target. Lower μt could help reach those targets. μt falls if the relative wage within the market falls, or if parents allow working children to keep an increasing share of their earnings. It could be that each of these was important at some point. According to Zelizer (1985), technical change was reducing the demand for unskilled youth from the late 19th century. The affected employments she discussed
cash boys and girls in grocery and department stores, for example, mainly affected children below high school age. By this argument, μt was falling among younger children in the late 19th century. This coincides with a sharp decline in their labor force participation rates: participation among boys aged 10
15 fell from 32.5% in 1880 to 6.4% in 1930 (among girls the decline was from 12.2% to 2.9%) (Whaples, 2010). Further, during the Great Depression employment opportunities dried up, reducing μt even among older teens. This likely helps explain the further expansion of
especially male
high school enrollments during the 1930s. Thus, it is possible that market forces alone were driving down μt and that this could explain some portion of the rise in schooling and the decline in fertility. However, Zelizer’s evidence of decreasing demand for child labor is quite limited and it is, overall, unclear how technological change was affecting μt among young workers before the Depression. Another possibility
indeed, Zelizer’s principal thesis
is that parental norms were changing as children transformed from being “economically valuable to emotionally priceless.” “While in the 19th century, the market value of children was culturally acceptable, later the new normative ideal of the child as an exclusively emotional and affective asset precluded instrumental or fiscal considerations (Zelizer, p. 11).” The new norm prohibited child labor undertaken for the benefit of the family. This lower propensity to view potential child earnings as parental property constitutes a reduction the rate parents “tax” the earnings of children. In the model, this is a lower “effective” value for μt; over time, an increasing share of market earnings becomes property of the child. Increasing disregard of children’s earnings may have been an evolution of the separate spheres ideology. An aspect of separate spheres household organization in the 19th century
women at home, men working away

Explaining the Revolution in U.S. Households

53

from home
was the notion that the home was sacred and the market profane. Thus, according to Degler (1980, pp. 73
74), “(e)xalting the child went hand in hand with exalting the domestic role of women; each reinforced the other.” Exhortations for a family wage
a salary that could support a male wage earner and his dependents would extend the firewall between the gentle home and harsh market
were also rising over this period. In Zelizer’s words, “the expulsion of children from the “cash nexus” … although shaped by profound changes in the economic, occupational and family structures, was also part of a cultural process of “sacralization of children’s lives (Zelizer, 1985, p. 11).” These considerations are now combined to produce Experiment 3. First, suppose that μ1875 = .40, as in Experiments 1 and 2, falls to μ1900 = .35, and then to μ1925 = .30. (Thus, a son born to parents forming households in 1925 who earned $100 per month would be able to retain $25 for use on private consumption whereas a son of 1875-parents would remit everything to parents.) Second, suppose the time potentially available for children to work on parents account falls 75% from the 1875 level, from T1875 = .20 to T1900 = .10 and then T1925 = .05. The 1925 value is consistent with parents continuing to view two teenage years as suitable for labor on behalf of the family. Third, as discussed in the data section, we allow the share of parental income devoted to private consumption per child τt to rise over time (such expenditures include clothes, toys, spending money for entertainment, and separate bedrooms).46 Here, τ1875 = .40 as in prior experiments but then rises to τ1900 = .045 and τ1925 = .052. Thus the value for households forming in 1925 remains below the contemporary estimate of .06 motivated in the data description.47 Experiment 3 uses these time profiles for Tt, μt and τt, along with the other Experiment 2 parameters. Results are reported in Table 9. Fertility at 2.63 and the .324 share of childhood devoted to schooling are extremely close to the targets of 2.6 and .301; likewise h^1925 = 8:19 as compared to the target of 8.2. Given the calibration strategy, it is perhaps unsurpising that the gender wage ratio is exactly met at .603. Setting the undetermined P1925 = .23. produces x1925 = 16.7 (the target is 16.6). The parameter values necessary to achieve the targets are not precisely pinned down. In particular, the targets can also be met by offsetting higher values for τt with higher values for Tt. As one example, nearly the same results are achieved when τ1900 = .05 and τ1925 = .055, if then T1925 = T1900 = .12 (with T1875 = .20). Although not immediately obvious, an increase in τt also contributed to the decline in U.S. fertility in Lord and Rangazas (2006). They calibrated per child consumption as a share of adult’s life cycle earnings, as we do,

54

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Table 9. Variables

Experiment 3; Reaching the Targets.a

Year 1875

(a) Parental characteristics variables 2.0 h^t

Year 1900

Year 1925

2.99

5.01

wt w1875 γt

1.0

1.27

1.70

0.463

0.502

0.549

(b) Parental choice variables xt st et nt d1t*nt gt zt mt h^t þ 1 γt + 1

2.33 .128 .0025 3.63 5.33 849 .768 .074 3.00 .502

6.08 0.206 .0033 2.89 3.59 1121 .727 .159 5.01 0.549

16.7 (16.6)b .320 (.309)b .0038 2.40 2.63 (2.60)b 1761 .580 .327 (.233)b 8.19 (8.2)b 0.603 (.603)b

a

In addition to those parameter changes of the Mortality Experiment 2, this exercise incorporates a time profile for dependent’s wage per unit of human capital retained by parents, μ1875 = .40, μ1900 = .35, and μ1925 = .30; for the time available for dependent children to work as perceived by parents T1875 = .20, T1900 = .10 and T1925 = .05; and for the share of potential parental earnings devoted to private consumption per child τ1875 = .04, τ1900 = .045, and τ1925 = .052; finally, P1925 = .23. b The figures in parentheses are targets for the variable in 1925. For x1925 the target is seven times the 1875 value; for st see Table B2; fertility target in text; for mt see Table 5; targets for h^1950 and γ1950 are given in the text.

and they kept that share constant across generations. However, in their framework total household life wealth included unearned family business income. As the relative importance of family business declines in the development process, the share of child consumption in total family resources increases, as in the simulations above. Mother’s Time Devoted to Household Production In Experiment 3, the proportion of wives’ adult life cycle devoted to market work rises from m1875 = .074 to m1925 = .327. This latter figure exceeds the target of 23.4% (Table 5, columns (5)). This section accounts for the increase within the simulation and then addresses the overshooting of target participation. Within the model, mt changes for two types of

Explaining the Revolution in U.S. Households

55

reasons: (1) a changing mix of the quantity and quality of children; and (2) the rise in the relative wage of women. Recall that each conception requires ρ = .009 units of mother’s life cycle time endowment through infancy, and an additional ρ = .024 of time per child unrelated to child quality through dependency. The 2.7 fewer conceptions and 1.2 fewer children surviving infancy free up 5.3% of mother’s time for market. But, starting from m1875 = .074 in 1875 this only increases m1925 to .127 or 12.7%. Additionally, time devoted to child quality per older child et increases over 60%, from e1875 = .0025 to e1925 = .0038, but the low levels combined with the decline in net fertility results in little change in m1925 from this channel. Most important to the increase in m1925 is the decline in the gender wage gap. This reduces the time wives devote to household production (25) and accounts for the majority of the increase in m1925. As discussed in “Market-Oriented Work of Wives” section, it is unsurprising that the calibration would overshoot the empirical target for m1925. As noted there, many employers (very openly) implemented marriage bars during the Great Depression of the 1930s; many employed married women were “let go” and other married women potentially interested in working were not considered by employers. This effort to increase employment opportunities for male heads of households directly reduced the participation of married women. Given some degree of persistence in market work, it is probable that subsequent participation was reduced as well. Viewed in this light, the model’s overshoot of participation is unsurprising.48

Discussion Extensions Yielding Similar Implications Greater Power of Wives. Doepke and Tertilt (2011) review convincing empirical evidence indicating greater relative control of family resources by wives increases expenditures on children. These expenditures are for private consumption goods such as clothing and for other goods, such as food, that have human capital dimensions.49 These results suggest consumption of dependent children increases with the bargaining power of wives, rather than simply the pooled potential household wealth. In a household bargaining model, suppose mother’s preferred τt and the “taste for quality” parameter ψ are larger than father’s. Then the household’s bargaining solution in response to a declining gender wage gap may yield higher values for τt and ψ. However, Doepke and Tertilt make clear there are numerous

56

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

reasonable ways to model household bargaining and the implications of bargaining can prove sensitive to reasonable alternative assumptions. Nevertheless, such an approach may offer a complementary explanation of the “family revolution.” Similarly, government policies may more nearly reflect the preferences of females once they gain the franchise. Miller (2008) presents convincing evidence that the enfranchisement of women in the United States contributed to legislation and expenditures designed to improve the welfare and reduce the mortality of children. For example, programs which disseminated knowledge of the germ theory of disease empowered mothers to protect their children from often-fatal illnesses (Mokyr, 2000). Greater Love of Children? Another possible factor altering sentiments was the decline in infant and child mortality. The early modern English family historian Lawrence Stone suggests that when infant and child mortality rates are high, it is “folly to invest too much emotional capital in such ephemeral beings (Stone, 1977, p. 105).” As infant and child mortality rates plummeted after 1880 parents may have felt it safer to form early and strong attachments to their children. Becker and Barro (1988) assume utility per child is higher when the number of children is lower, corresponding to higher altruism per child when families are smaller. If parents also care about the consumption of dependent children, smaller families would be associated with larger private consumption shares per child. This makes each child more expensive and reduces the relative price of quality. In our framework, these mechanisms could be approximated by an increase in τt. Now combine several of the points from above: Greater political power for females may have contributed to the decline in infant mortality. This, in turn, might reduce the emotional distancing of parents from highly perishable infants in the high-mortality regime (as described by Stone, 1977). As a final link, more intense emotional bonding may result in greater altruism per child (as in Becker & Barro, 1988). Translated into the parameters of our framework, greater female power could produce some combination of changes in τt, μt, Tt and ψ capable of inducing the family transformation. Social historians agree that over time parents perceived that their efforts contributed more to children’s survival and success. Into the 19th century, many believed that God had pre-ordained who would be saved, leaving little role for parents in children’s salvation (Calvinism). However, the early

Explaining the Revolution in U.S. Households

57

19th century rise of Romanticism and Arminianism (salvation through faith) provided more scope for parents to mold the character and fate of their children. Further, Darwin’s writings reduced the average belief in an active God, while the Germ Theory of Disease made parents aware their children’s survival depended importantly upon parental efforts.50 If the death of a child was not solely God’s Will, then perhaps the conception of a child might also be in the domain of parental choice. Parents began to consciously control fertility and perceive the survival and development of a child was their province. Andreoni (1990) develops a model of “impure” altruism. In his framework, an altruist gains more utility when personally responsible for an increase in the income of someone about whom the altruist cares than if the higher income arises from another source. Similarly, as parents perceive more responsibility for the life and welfare of their children, they may value improvements in their life and welfare more than before. This view is also consistent with reductions in Tt and μt and increases in τt and ψ.

Other Possibilities Morbidity Morbidity, like mortality, also fell significantly from the late 19th century.51 Reductions in morbidity increase effective human capital, by raising the productivity of time, and/or the amount of time (as with an increase in the number of years in adulthood one is capable of working). Given the various uses of time in the model there are many ways morbidity change could influence parental choices. Some morbidity reductions principally affect youth, as the reductions in hookworm in the American South described by Bleakley (2007). Others, such as fewer musculoskeletal problems and increased remaining life beyond dependency principally affect adults. Elimination of the sorts of childhood afflictions that lead to lifetime scarring, as evidenced by rising final height in adulthood and IQs, may affect the productivity of time both when young and when old. There are reasons to suspect the effects of lower morbidity on the quantity and quality of children were modest over this period. To see this, suppose reduced morbidity increases the effective unskilled and skilled human capital of adults, and the unskilled human capital of dependent children all by same factor D > 1. Inspection of the optimal solutions for st, et and nt reveal that there is no impact effect from scaling all human capital terms by some constant D.

58

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Nevertheless, even this type of age-neutral morbidity decline has some behavioral implications. In particular, reductions in morbidity increase xt to Dxt . However, as seen in Experiment 1, with θx small, this has little effect on h^t þ 1 . Also, eradication of hookworm among the young, for example, would increase student’s effective time at school, as well as work. Similarly, mother’s effective time input in offspring’s human capital production would rise. Thus, h^t þ 1 would increase, with knock-on effects for choice variables in the subsequent period. Unreported calibration results show the full impact is small when morbidity decline increases Dt by a total of 18% over three generations. If morbidity decline affects different components of human capital asymmetrically, the impacts could be larger (moving closer toward or further away from the targets). For example, with limited empirical support, Cinyabuguma et al. (2009) report simulations in which morbidity decline increases the effective human capital of adults much more than that of dependent children. In that event, a significant increase in child quality and decrease in child quantity can arise. Robust progress relating to morbidity decline awaits improved empirical understanding of how various components of human capital have been affected over time. Skill Premiums and Skill-Biased Technological Change Another potential explanation of the rapid rise in schooling that began shortly before 1920 is that SBTC increased the wage premium to skill and thereby encouraged public high school provision and attendance. This paper’s model was loosely adapted to allow for changing skill premiums and calibrated to assess the role of those changes. In particular, the wage per unit of schooling capital was made to be increasing, but concave, in schooling human capital. The rate at which this wage increased with schooling was calibrated in accordance with the empirical skill premiums during the early adulthood of the children.52 The results suggest an acceleration of the skill premium played little role in increasing education or, via the quantity-quality trade-off, reducing fertility. The reason is straightforward: The skill premium was declining at a rapid rate nationally throughout the central years of the high school movement. Indeed, Goldin and Katz (2008, p. 316) report that the skill premium declined 1.28% per year between 1910 and 1930. Relatedly, Katz and Margo (2013) detail significant skill premiums and SBTC even from the antebellum period.53 If neither SBTC nor high skill premiums were a new phenomenon, it is unclear why they should have stimulated attendance just as the high school movement began. The issue does not disappear if the

Explaining the Revolution in U.S. Households

59

focus is on rates of return to schooling rather than skill premiums. Of course, the rate of return depends on costs as well as benefits. And, public provision of high school lowered tuition while child labor restrictions and new norms stigmatizing the work of children lowered the cost of schooling children. Thus, it is possible that the rate of return to schooling was rising even as the wage premium was falling. Even if that is true, these cost reductions were already incorporated in Experiment 3 via lower Pt, μt and Tt; that experiment reached the targets without invoking rising skill premiums. Despite the foregoing, we believe it likely that SBTC was important to the family revolution. For one thing, SBTC reduced the premium to muscles, which reduced the gender wage gap. As stressed in this paper, this reduction in the gender wage gap pulled married women into the labor force. More generally, Katz and Margo (2013, Tables 4 and 6) present time
series evidence on the economy’s occupational distribution from the 19th century. They demonstrate that there was an acceleration in the share of employments characterized as “white collar” after 1910, as the high school movement was unfolding. Our model is poorly suited to address this general skill upgrading within occupations associated with SBTC. As one example, Goldin (1990) emphasizes SBTC contributed to an expansion of clean, interesting, and respectable office jobs for females in the early decades of the 20th century. She notes that before such jobs were available, the work of a wife in a factory or domestic service was evidence that the husband had failed as provider. These “higher-status” female clerical jobs, however, were perhaps intrinsically meaningful to wives, and caring husbands would “allow” their wives to work if they desired. Clerical work, though, required additional education. As applied to our theory, increased human capital among married females increases the cost of time required to raise each child, contributing to the decline in fertility. As a second example, Goldin and Katz (1999) argue that the significant regional differences in the timing of the high school movement were a product of regional differences in the stocks of social capital. Many parents nationally desired advanced schooling for their children that they might obtain the opportunities afforded by SBTC. However, public provision required the imposition of property taxes, and only relatively homogenous communities were willing to take on that burden. In summary, an adaptation of our model to allow for skill premiums did little to explain the family revolution. However, compelling changes in the occupational structure lead us to believe that SBTC played an important

60

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

role in the increase in schooling.54 It is also possible SBTC played some role in fertility decline.

SUMMARY AND CONCLUSION This paper analyzed the great reduction in fertility, rise in schooling, and acceleration of the movement of married women into the paid labor force that occurred between the later portion of the 19th century into the first decades of the 20th century. First, a model of household decisions over fertility, schooling of children, household production and married female labor supply was developed. In that model, transfer-constrained parents make all transfers to children via human capital bequests and the quantitative significance of unskilled human capital in the 19th century is made explicit. An initial baseline for the model was calibrated based on historical data. Then a series of careful simulations assessed the quantitative importance to family change of several mechanisms proposed in the literature. Experiment 1 invoked falling prices for educational goods inputs (i.e., increased public provision of education), rising parental incomes, changing returns to work experience, and rising multifactor productivity in the production of schooling. In combination, these induced a significant increase in the goods input. However, there was little effect on schooling time, schooling human capital, or fertility. These results reinforce findings in the literature of weak effects on fertility of income (Galor, 2005) or subsidies to education (Doepke, 2004). Experiment 2 extends the first by incorporating the empirical decline in infant and child mortality. This does little to improve the fit of the model. Intuitively, falling infant and child mortality reduce the price of surviving children and this effectively offsets the decline in the price of quality (via schooling) for surviving children. Although fertility declines, net (of mortality) fertility is largely unchanged. This limited effect on net fertility of declining infant and child mortality finding is consistent with theoretical and empirical evidence presented in Doepke (2005) and Galor (2005). Experiment 3 examined factors which increase the net cost of children. Specifically, it incorporated lower values of the opportunity cost to parents of schooling children (lower μt), a smaller window of time parents were willing to work children (lower Tt), and a higher budget share for private consumption per child (τt). These mechanisms prove capable of achieving the model’s targets for fertility and human capital. The reduction in μt had

Explaining the Revolution in U.S. Households

61

a powerful effect on increasing st, while achieving the targeted reduction in fertility required higher τt in addition to lower Tt and μt. Our preferred explanation of the declines in μt and Tt and increase in τt includes a change in parental norms so as to make child labor less acceptable and/or increase parental altruism toward children. Additionally, there is evidence that child labor and compulsory schooling legislation had modest direct effects on schooling and indirect effects on fertility. Such legislation may itself be a response to changing norms. Alternatively, Doepke and Zilibotti (2005) suggest SBTC led narrowly self-interested adult males to favor restrictions on child labor. We acknowledge the that the quantity-quality trade-off quite possibly involved additional influences. However, limited current understanding of how morbidity decline affects the productivity of time at different ages and in different uses precludes a definitive assessment of its impact. This paper does clarify why an age-neutral reduction in morbidity may have only negligible effects on schooling and fertility. Also, we noted that SBTC as reflected in skill premiums would not have much explanatory power in our framework since premiums were falling during much of the high school movement. Although not captured by our framework, the growing proportion of occupations combining high skill and social status suggests that SBTC played some role in the high school movement. Finally, this paper assumes unitary household preferences. Empirically, it appears wives (as compared to their husbands) have a stronger relative preference for child welfare. Household bargaining models in the presence of a decline in the gender wage gap often predict an increase in spending on children, as proposed in this paper. An analysis employing several different bargaining assumptions would prove a valuable complement to the current contribution. Market work among wives increases significantly in each experiment. The principal mechanism is the decline in the gender wage gap which induces substitution away from household production toward market work. The time savings from reduced fertility also contributed (especially in Experiment 3). The calibrated increase in married women’s market work exceeds the empirical rise. This, though, is not surprising: During the Great Depression many employers imposed marriage bars, which directly reduced current employment among married women and quite possibly reduced their employment in later years once the bars were lifted. The results of this paper suggest fruitful directions for our future research. In the current framework, an increase in the exogenous private consumption share of dependent children was necessary for a calibration to

62

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

meet all targets. An analysis which makes that share endogenous would increase confidence in the quantitative significance of that mechanism. Second, alternative hypotheses generating the reduced interest among parents in appropriating child earnings will be modeled and evaluated. Third, the existing framework will be extended to analyze the regional convergence of southern to northern incomes, schooling, and fertility which characterized this period.

NOTES 1. To the extent such human capital was general, rather than firm- or industryspecific, and since risks of separation were significant, workers presumably paid for some portion of their training costs through lower wages (Becker, 1975). 2. Doepke and Zilibotti note that men find child labor harmful in the aggregate, as it substitutes for their own labor and depresses the wages of adult men. As SBTC proceeds, it increases the return to education and families respond by having fewer children, each of whom they educate more. With fewer children, the foregone child earnings in their own family from child labor restrictions are reduced, which makes parents more willing to support legislative restrictions. They provide informal evidence that patterns of fertility, education and child labor legislation are consistent with their theory. Doepke and Tertilt argue that SBTC encouraged men to surrender economic and political power to women. Intuitively, husbands prefer to maintain bargaining power over their own wives. Working in the opposite direction, men want their daughters and grandchildren to have good education and want their sonin-laws to be well-educated. Additionally, female voters put more weight on child quality than do men. As SBTC raises the return to education, these benefits to offspring and grandchildren come to outweigh the loss of reduced bargaining power in their own home and men vote to increase women’s rights. They present suggestive stylized facts supporting these hypotheses. 3. Many small estates are an “accidental” consequence of unannuitized life cycle savings; to incorporate them would require modeling transfers under conditions of uncertainty. 4. Antecol (2011) analyzes more recent trends in female employment. Light and Omori (2012) discuss the decline in life-long marriage in recent decades, with implications for female employment. 5. The mechanisms Greenwood et al. suggest only become operative at the end of the period addressed in this paper. Adshade (2012) and Rotella (1980, 1981) envision that an exogenous increase in high school attendance induced skill-biased technical change in office machines. Ferna´ndez, Fogli, and Olivetti (2004) propose a role for culture in the rise of MFLFPRs. They find that males whose own mothers had worked are more likely to prefer a spouse who works. 6. Soares and Falcao (2008) consider linkages among adult longevity and MFLFP. Increased longevity lengthens the period over which investments in human capital can be recouped. This increases human capital investments by females,

Explaining the Revolution in U.S. Households

63

inducing them to substitute away from fertility and increase market work. However, Hazan (2009) shows that increased adult longevity in the United States was associated with lower, rather than higher, life cycle market work among men. 7. Once the germ theory of disease gained acceptance, a concerted effort was undertaken by governmental and charitable institutions to inform mothers of the risks of germs and of those steps mothers could take to protect the health of their children. Practices such as washing hands before eating, quarantining those who are ill, boiling water, pasteurizing milk, and keeping living areas clean, boosted health and reduced mortality. Vaccines against cholera and typhoid arrived late in the nineteenth century, while vaccines against diphtheria, whooping cough, and tuberculosis became available early in the twentieth century. The discoveries of sulfa drugs in the 1930s, then mass production of penicillin in the 1940s, helped further reduce mortality and perhaps morbidity (Preston & Haines, 1991). Morbidity is considered further in “Morbidity” section. 8. They note that lower child mortality increases the returns to parental investments in quality, so that investments per child increase. In their framework the increase in parental investments per child more than offsets the decline in fertility, so that total time investments in children increase. Thus, lower child mortality reduced the MFLFP. 9. Children could still have small jobs and chores, but only insofar as these help develop character and good work habits. Any earnings would be retained by the children in order to develop the ability to manage money. 10. Changes in the gender wage gap over time are more important than the initial level of the gap. This paper assumes that changes in the gender wage gap arise from changes in schooling, experience, and the premium to strength (as in Goldin & Polachek,1987). 11. Some early versions of the paper allowed for multiple periods in childhood and adulthood. For the issues addressed the insights gained were dominated by the costs of increased complexity. Were the framework extended to a general equilibrium growth model, for example, it would be necessary to include a consumptionsavings margin. The fact adulthood is twice as long as dependency is addressed in the calibration section. 12. With logarithmic preferences mother’s time allocation proves independent of whether household productivity is increasing in her skill level; of course Gt and utility are higher when her skills matter. 13. The price per unit of the human capital goods input is less than 1 because much schooling is publicly provided. Often the taxes to pay for public education come from levies on real property. The budget constraint abstracts from those and other taxes since the burden of additional notation swamps any insight gained. The schooling taxes would range from about .3% of life wealth in the initial period to about 1.5% in the last period (see calibration discussion). One approach is to envision taxes are levied on wage earnings and to then simply view the wage per unit of human capital as being net of tax. 14. Suppose the youth mortality rate is my. Then d2t = 1/(1 − my). If the infant mortality rate is mi, then d1t = d2t/(1 − mi) = 1/[(1 − my)(1 − mi)] so that changes in youth mortality change d2t and d1t by the same percentage.

64

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

15. Less realistically, but in strict conformity with the model, could imagine all children are born in 1875, commence school at age six in 1881 and complete schooling in 1895. 16. However, if the ratio is instead based on hourly earnings among full-time workers there is a further increase to .662 by 1970 as full-time men come to work longer hours than full-time women, especially after 1940. 17. The experience numbers have been converted from the log form in which Goldin and Polachek presented them. 18. This estimate of 6 years for the birth cohort of 1850 appears consistent with the time series attainments reported in Goldin and Katz (2008, Fig. 1.4), which is a bit over seven years for the Native-Born birth cohort of 1876. 19. Goldin and Katz (2008, Table 2.5, pp. 78
79) present returns to schooling among males 18
65 based on a 1915 Iowa state census. The returns to those younger than 35 are reported separately, enabling us to infer the returns to those age 35 and older. Slightly more than half of the males in that sample are older than 35. The returns for this 35
65 group of males are 3.73% for each year of common school (very few attended high school at that time). The mid-range male in that subsample would be age 50, and so born in 1865. From at least 1865 through the first decades of the twentieth century the returns to primary school were increasing (see Card & Krueger, 1992 and Goldin & Katz, 2008). For example, in the Goldin and Katz (2008) Iowa sample, the returns for those 18
34 had risen to 4.83% for common school. This suggests a rate of return for males born in 1850 below the 3.73% figure. Also, the market rate of return for females would surely have been below that for males, as would the return to all schooled in the South (see Wright, 1986). Altogether, a return of about 3 percent for each year of common school seems appropriate for those 1850; a value of 3.1% fits the baseline calibration nicely. Bleakley, Costa, and Lleras-Muney (2013, p. 5) report a rate of return to formal schooling of only about 1% for the mid-19th century. This return would apply principally to education obtained by cohorts born in the 1830s and earlier. This is consistent with the Goldin and Katz evidence that the rate of return to schooling was low but rising. 20. Goldin and Polachek (1987) note that “data on piece-rate earnings in 1895 indicate that males earned on average 30 percent more than did females (i.e., the wage ratio was 77), when the piece rate was identical for both, and when both worked at the same job, in the same factory.” They point out that this constitutes a lower-bound on the reward to greater male strength since it was only in those occupations where physical differences were less important that men and women worked together. And, the premium to strength for teenagers in the 1860s was presumably greater than in 1895. Alternatively, Goldin (1990) argues that there may have been a modest amount of discrimination within manufacturing in the 19th century. Thus, if such discrimination for unskilled attributes was 15%, say, then the productivity of unskilled females in the typical job would have been well above 6.68. This type discrimination could be captured in the model by deflating a “true” female unskilled productivity by a discrimination factor. However, since the approach developed below is based on changes in the gender wage ratio unrelated to discrimination, abstracting from discrimination is not important to the issues examined. Interestingly, Goldin argues that in the first few decades of the

Explaining the Revolution in U.S. Households

65

twentieth century discrimination rose significantly, as firms began to rely on internal labor markets. In these clerical jobs, she argues, all women were passed over for training opportunities, since most women were expected to quit upon pregnancy, if not marriage. Thus, those married women who did continue to work received less on-the-job training and fewer promotions, and this was reflected in their earnings. 21. Goldin and Polachek find that narrowing of the gap was primarily due to rising wages for females within occupational groupings
especially clerical and professional
rather than changes in the occupational distribution (the latter explaining only about 2.6 of the 10 rise in the ratio between 1890 and 1930). Goldin (1990) argues that discrimination seems to emerge only after 1940, especially in the clerical sector. Polachek (1975) also addresses the role of discrimination. 22. The years of schooling together with the rate of return to schooling determine h^1875 . In general, the smaller is h^1875 the larger is the multiple of increase in h^1875 required to have schooling explain 43% of the narrowing of the gender wage gap, and the greater is the resulting role for educational efficiency bt. 23. They point to estimates that a one-year increase in the average education within a state increases output per worker by 10%. However, the factors producing an increase in a state’s average level of education
improved infrastructure, commercial growth, etc.
may overstate the effect on an individual of increasing own education by a year with given state infrastructure (Topel, 1999). 24. Among men, holding experience and unskilled human capital constant, formal schooling increases total human capital from [10 + 2](2.53) = .30.36 to [10 + 8.2] (2.53) = 46.04, or about 52%. Recalling that unskilled human capital among females begins at 6.68, the percentage increase in human capital from schooling alone is 71%. Using an estimated average of male and female labor for shares between 1890 and 1965 (about .73 for males), the increase in schooling human capital across these cohorts would be about 57%. Over a 75-year period, this corresponds to an average annual growth rate of human capital from educational deepening of about .6%. If labor’s share of income is .7, this implies an average annual contribution to output per worker of (.6)(.7) = .42%. Since growth in output per worker across these cohorts averaged 2.01% (discussed below), the relative contribution from educational deepening is .42/2.01 = .209, or 20.9%. 25. If the returns to scale were lower, the percentage increase in h^t from the historical increase in goods and time inputs would be smaller. Consequently, to achieve the targeted h^1950 the percentage rise in bt would need to be larger. Education has long been considered a low productivity growth sector. Lowering θs somewhat and increasing bt a bit had little effect on any interpretations in the experiments considered below. 26. The data underlying Table 3 is presented in Appendix B, Tables B1 and B2. 27. The following motivates our choice of values for that birth cohort. For 1860 and 1870 census data indicates the proportion age 5
14 enrolled were 69.4 and 75.1. Since it is hard to imagine enrollments were rising much during the Civil War, the 1860 figure is used for 1865. This provides a rise in enrollments of 75.1/69.4 or 8.2% between 1865 and 1870. Suppose there was also a modest increase of 5 days attended per enrolled student per year in the several years following the Civil War. Then, the 1865 school-year student (born in 1850), is assumed to have spent 9% of

66

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

his time endowment in school. It is similarly assumed that there is a 15 percent increase in expenditures on goods inputs between 1865 and 1870. Differences of a few percentage points in these schooling input estimates for the 1850 birth cohort would have little impact on any results. 28. Mother’s effective time input is her human capital times her time input. This effective time input approximately doubles between 1875 and 1925 in our preferred calibrations, with some probable increase between 1850 and 1875. Since the exponent θhe = .1, even a several percentage points difference in the growth of this effective time input from that postulated would have very small effects. 29. Recall that the increases in inputs from the birth cohort of 1850 to that of 1875 are imputed to calibrate the model. If one only focuses on the increase in h^t between those chosen by parents of 1875 and of 1925 the implied growth in bt is essentially unchanged. 30. This expression also assumes that labor force participants work full time. Goldin (1990) argues that female employees typically worked full time through 1940. 31. Before 1940, Census participation questions differed appreciably from the modern participation concept. Prior to 1940, the question was that of one’s ‘gainful occupation’ (though the wording varied a bit from Census to Census). Goldin (1990) notes that many women in the nineteenth century engaged in marketoriented work on their husband’s farm or kept boarders; under the modern conception of labor force participation they would be counted as in the labor force. However, many viewed themselves as principally housewives, and reported this ‘occupation’ to Census takers. 32. Goldin (1990) notes considerable heterogeneity among women’s labor market participation. Those employed in any year tended to have significant participation persistence. Similarly, those without children worked more than those with children. This heterogeneity has declined over time and the initial extent should not be overstated. For example, in a Women’s Bureau Survey from 1940, Goldin calculates that those between the ages of 40
49 in 1939 (and thus born between 1880 and 1889) had 15.5 years of work experience if currently working, and 7.6 if not currently employed (Table 2.5, p. 31). 33. Goldin (1990) stresses that there was considerable heterogeneity among women regarding their labor market behavior. While few (especially white) married women worked outside the home in the late nineteenth and early twentieth centuries, those currently employed tended to have significant persistence in the labor force, and those not working currently tended to have not accumulated much work experience since marriage. Similarly, those without children worked more than those with children. This heterogeneity has declined over time and the initial extent should not be overstated. For example, in a Women’s Bureau Survey from 1940, Goldin calculates that those between the ages of 40
49 in 1939 (and thus born between 1880 and 1889) had 15.5 years of work experience if currently working, and 7.6 if not currently employed (Table 2.5, p. 31). 34. Information on mortality by age since 1900 is available in the HSUS table Ab988
1047. 35. Given that taste parameters are also used to pin down the initial xt, see below, the results would be essentially unchanged if the 60% reduction had instead

Explaining the Revolution in U.S. Households

67

occurred from an initial price of .6 or .4. In each case goods inputs would increase 150% from their initial level, ignoring any indirect effects of rise in inputs in one period for those purchased in later periods. 36. To maintain the desired initial market time for mother, this value is adjusted trivially across the baselines of the different experiments. 37. http://www.bls.gov/opub/mlr/2001/05/art3full.pdf 38. Husbands in 1875 had normalized lifetime earnings of 1000. Adding the earnings of wives increases income somewhat, say to 1100. If young adults with dependent children and old adults each have the same income, goods inputs across all children divided by parental income provide the GDP share. Each child surviving infancy uses x1875 = 2.3, and about 4 children survive infancy. 39. The taste parameters are ψ = .55 and σ = .23 in Experiment 1. To maintain other features of the initial baseline once mortality decline is included in Experiment 2 and thereafter, ψ = .40 and σ = .174. 40. Card and Krueger (1996) examine the link between earnings in adulthood and schooling expenditures per student within a year in the early decades of the American South. They find increases in earnings in adulthood of somewhat greater than 1% for every 10% increase in expenditures per student on instruction. Their findings are at the upper-end of estimates for the U.S. over this period. Even so, part of the higher earnings effect occur because increases in school quality increased the years of schooling children choose to receive. Thus, a 10% increase in expenditures per student per year translates to a larger percentage increase over a student’s entire school tenure. Once this effect is taken into account their results are even consistent with the elasticity θx = .1 assumed in the calibration. Nevertheless, ‘the baseline’ experiment was repeated using a value of θx = .15 to examine the sensitivity of the results to changes in this parameter. This only very modestly improved the fit of the model, as st and nt were little affected. Given the modest effects of the higher value and the sentiment that for lifetime resources θx = .1 is the ‘best guess,’ subsequent experiments are based exclusively on that value. 41. As in this paper, there is exogenous income growth and decline in the price of schooling. Unlike this paper, their framework includes multigenerational family business and life cycle savings. Family business creates an additional source of wealth which made for high fertility in 1800. 42. They offset this overly large increase somewhat by adding a small constant to schooling time. 43. In their preferred calibration, declining family business income accounts for about 40% of the fertility decline in the nineteenth century, but has little effect in the twentieth century. These reductions in family business income have no effect on the schooling of children in their model. Fertility decline and schooling increases continue into the twentieth century. 44. Schooling and work legislation removes low-wage younger children from the labor force; the average μt among older children not covered by the legislation is (all else the same) increasing. Consequently, modest reductions in Tt produce a full effect on potential child earnings less than proportional to the reduction in Tt. 45. Recall that a reduction in the relative hourly wage differs from a reduction in μt. The relative hourly wage μth0t/ht can also fall if the human capital of adults is increasing (holding μt constant).

68

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

46. This is private consumption; recall, children also benefit from the household production of communal goods included in Gt. 47. Zelizer (1985, 153) reports on a 1930 study indicating that the cost of raising a child to age 18 was $7,425, while income was approximately $2500. The ratio is 2.97. Haveman and Wolfe (1995, Table 1, p. 1830) find annual costs of children to parents in 1991 are $7,579 whereas mean income of a household in the middle quintile is $30,148. Multiplying the annual cost by 18 produces a ratio of costs to income of 4.56. 48. Allowing for an exogenous Progressive Era increase in et as mothers perceive their productivity in the production of child health has risen
as claimed by Mokyr (2000)
would lower the calibrated m1925. This would reduce the gap between the calibrated and empirical m1925. The calibrated rise of m1925 would also be smaller if the elasticity of substitution between goods and time in household production was less than that implied by the functional form for household production, which is 1. 49. Evolutionary arguments also suggest that men are less concerned with the quality of children than women (Diamond, 1997). In Doepke and Tertilt (2009) men grant the franchise to women in response to rising rates of return to human capital. More power for wives leads to increases in the bargaining power of their daughters and in the education of their grandchildren, which grandparents like. Cvrcek (2007) argues that increased employment among single females in the last decades of the nineteenth century increased their bargaining power and share of marital output. Doepke and Tertilt (2011) illustrate a noncooperative bargaining model in which a narrowing gender wage gap alters the mix of household public goods produced via household production functions. In their framework, higher wages for wives make their time input more expensive and can reduce the household supply of time-intensive public goods such as children even in the absence of a change in preferences. In Chiappori’s (1992) cooperative marriage bargaining model, husbands and wives have different preferences for household public goods (such as quantity and quality of children). As a wife’s bargaining power increases, there is increased weight given to her preferences. 50. Dye and Smith (1986, p. 347) report that in 1915 [s]ocial reformer Florence Kelley articulated the new consciousness: “So long as did not know that children need not die … [W]e strove for resignation, not intelligence. A generation ago we could only vainly mourn. Today we know that every dying child accuses the community. For knowledge is available for keeping alive and well so nearly all that we might justly be said to sin in the light of the new day when we let any die.” 51. Bleakley (2009) concludes that the combined efforts of the eradication of hookworm and malaria in early 20th century southern states increased income by 25% compared to unaffected states. Costa (2009) finds that between 1910 and the 1990s functional disabilities declined by 0.6% per year among men age 60
74, while over a similar period “the average decline in chronic respiratory problems, valvular heart disease, arteriosclerosis, and joint and back problems was about … 0.7% per year (2009, p. 2).” Declines in malnutrition, sequelae from polio, tuberculosis, malaria, smallpox, cholera, typhoid, other diarrheal episodes, exposure to animal waste products, industrial and other work related accidents further contributed to morbidity decline. The decline in these physiological insults was reflected in the increasing heights of adult men entering Amherst College which rose from 169.9 cm

69

Explaining the Revolution in U.S. Households

in 1870 to 178.1 cm in 1935. (Historical Statistics of the United States, Vol. 2, p. 582, Series BD661.) This suggests that even upper middle-class households benefited from reduced physiological insults in the nineteenth century. At the intersection of mortality and morbidity, remaining years of life expected by a 20-year-old male in 1870 were 41, rising to 45 in 1919
1921, and to about 49 in 1949
1951, potentially increasing the life cycle supply of human capital (Hazan & Zoabi, 2006). 52. Units of unskilled human capital were paid wt while units of skilled or schooling human capital received w^ t . With a skill premium associated with SBTC, w^ t is increasing in the stock of schooling capital as captured by ɛ

w^ t = wt h^t

with ɛ > 0. Thus the potential earnings of a male become  ɛ h i 1þɛ Emt = wt hmt wt h0m Emt þ wt h^t h^t Emt = wt h0m þ h^t

ð27Þ

53. Their research reveals that while early manufacturing was (famously) de-skilling, other sectors paid significant skills premiums. Overall, they characterize the aggregate economy as experiencing SBTC even during the antebellum period. 54. At least one previous calibration exercise of the United States has invoked SBTC. Greenwood and Seshadri (2002) devise a highly stylized two-sector model to produce the demographic transition and structural transformation. A central building block is their assumption that skilled labor is not useful in the agricultural sector, but is productive elsewhere in the economy. Then SBTC increases education and income of those in the “manufacturing” sector, inducing structural transformation and fertility decline. However, for the period of interest to the current paper, the return to education was at least as high among farm owners and operators as among blue-collar or white-collar workers (Goldin & Katz, 2008, Table 2.5). Their assumption may be more appropriate for more recent times. 55. Current expenditures per public elementary and secondary school pupil in average daily attendance in 1941 were $675 in 1982
1984 dollars. Call this expenditures per enrolled student (HSUS Bc924, pp. 2
482). The comparable figure is $2,503 for 1941 in 1982
1984 dollars for enrolled undergraduate students
educational and general expenditures per student (HSUS table Bc966). Thus, the ratio of college to elementary and secondary students is 3.71. About 20% of the 1925 birth cohort ever enrolled in college, whereas about 8% graduated. We assume the enrollment rate averages about 15% for 18
19-year olds. Two years of college are added onto that from K-12. So, (2/15)(2503)(.16) = 53.4; this increases expenditures for 1940 to 463 + 53 = 516.

ACKNOWLEDGMENTS We are grateful to Marie Steele for wonderful
and cheerful
research assistance. We also thank Peter Rangazas, Marianne Wanamaker, and

70

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

seminar participants at the Western Economic Association meetings, Meetings of the Cliometric Society, and the Public Policy Seminar at UMBC for useful comments.

REFERENCES Abramovitz, M., & David, P. A. (2000). American macroeconomic growth in the era of knowledge-based progress: The long-run perspective. In S. L. Engerman & R. E. Galman (Eds.), The Cambridge economic history of the United States (Vol. 3, pp.1
92). Cambridge: Cambridge University Press. Adshade, M. (2012). Female labour force participation in an era of organizational & technological change. Canadian Journal of Economics, 45(3), 1188
1219. Albanesi, S., & Olivetti, C. (2007). Gender roles and technological progress. CEPR Discussion Paper No. DP6352. Andreoni, J. (1990). Impure altruism and donations to public goods: A theory of warm-glow giving. Economic Journal, 100(401), 464
477. Antecol, H. (2011). The opt-out revolution: Recent trends in female labor supply. In S. W. Polachek & K. Tatsiramos (Eds.), Research in labor economics (Vol. 33, pp. 45
83). Research in Labor Economics. Bingley, UK: Emerald Group Publishing Limited. Barro, R. J., & Becker, G. S. (1989). Fertility choice in a model of economic growth. Econometrica, 57(2), 481
501. Becker, G. S. (1975). Human capital: A theoretical and empirical analysis, with special reference to education. Chicago, IL: University of Chicago Press. Becker, G. S. (1981). A treatise on the family. Cambridge, MA: Harvard University Press. Becker, G. S., & Barro, R. J. (1988). A reformulation of the economic theory of fertility. Quarterly Journal of Economics, CIII(1), 1
25. Becker, G. S., & Tomes, N. (1986). Human capital and the rise and fall of families. Journal of Labor Economics, 4(3), S1
S39. Ben-Porath, Y. (1967). The production of human capital and the life cycle of earnings. Journal of Political Economy, 75, 352
365. Bleakley, H. (2007). Disease and development: Evidence from hookworm eradication in the American south. Quarterly Journal of Economics, 122(3), 73
117. Bleakley, H. (2009). Economic effects of childhood exposure to tropical disease. American Economic Review, 99(2), 218
223. Bleakley, H., Costa, D., & Lleras-Muney, A. (2013). Health, education, and income in the United States, 1820
2000. NBER Working Paper No. 19162. National Bureau of Economic Research, Cambridge, MA. Bleakley, H., & Ferrie, J. P. (2013). Shocking behavior: Random wealth in antebellum Georgia and human capital across generations. NBER Working Paper No. 19348. Browning, M., Hansen, L. P., & Heckman, J. J. (1999). Micro data and general equilibrium models. In J. B. Taylor & M. Woodford (Eds.), Handbook of macroeconomics (Vol. 1A, pp. 543
633). North Holland: Elsevier. Card, D., & Krueger, A. B. (1992). Does school quality matter? Returns to education and the characteristics of public schools in the United States. Journal of Political Economy, 100(1), 1
40.

Explaining the Revolution in U.S. Households

71

Card, D., & Krueger, A. B. (1996). School resources and student outcomes: An overview of the literature and new evidence from north and south Carolina. Journal of Economic Perspectives, 10(4), 31
50. Chiappori, P. (1992). Collective labor supply and welfare. Journal of Political Economy, 100(3), 437
467. Cinyabuguma, M., Lord, W., & Viauroux, C. (2009). Health and the revolution in household behavior 1880
1940: Fertility, education and married female labor supply. UMBC Working Paper No. 09108. Connolly, M. (2004). Human capital and growth in the post-bellum south: A separate but unequal story. Journal of Economic History, 64(2), 363
399. Costa, D. (2009). Why were older men in past in such poor health? In D. Cutler & D. Wise (Eds.), Health in older age: The causes and consequences of declining disability among the elderly. Chicago, IL: University of Chicago Press for NBER. Cvrcek, T. (2007). Mothers, wives and workers: The dynamics of white fertility, marriage and women’s labor supply in the United States 1870
1930. In T. Cvrcek (Ed.), Three Essays in Economic History, PhD dissertation. Degler, C. (1980). At odds: Women and the family in America from the revolution to the present. Oxford: Oxford University Press. De Vries, J. (2008). The industrious revolution: Consumer behavior and the household economy. Cambridge: Cambridge University Press. Diamond, J. (1997). Why is sex fun? The evolution of human sexuality. New York, NY: Basic Books. Doepke, M. (2004). Accounting for fertility decline during the transition to growth. Journal of Economic Growth, 9(3), 347
383. Doepke, M. (2005). Child mortality and fertility decline: Does the barro-becker fit the facts? Journal of Population Economics, 18(2), 337
366. Doepke, M., & Tertilt, M. (2009). Women’s liberation: What’s in it for men? Quarterly Journal of Economics, 124(4), 1541
1591. Doepke, M., & Tertilt, M. (2011). Does female empowerment promote economic development? NBER Working Paper No. 19888. Doepke, M., & Zilibotti, F. (2005). The Macroeconomics of child labor regulation. American Economic Review, 95(5), 1492
1524. Dye, N. S., & Smith, D. B. (1986). Mother love and infant death, 1750
1920. Journal of American History, 73(2), 329
353. Ferna´ndez, R., Fogli, A., & Olivetti, C. (2004). Mothers and sons: Preference formation and female labor force dynamics. Quarterly Journal of Economics, 119(4), 1249
1299. Galor, O. (2005). The transition from stagnation to growth: Unified growth theory (pp. 171
293). North Holland: Handbook of Economic Growth. Galor, O., & Weil, D. (1996). The gender gap, fertility, and growth. American Economic Review, 89(2), 374
387. Galor, O., & Zeira, J. (1993). Income distribution and Macroeconomics. Review of Economic Studies, 60(1), 35
52. Goldin, C. (1986). The earnings gap between male and female workers: An historical perspective. NBER Working Paper No. 1888. Goldin, C. (1990). Understanding the gender gap: An economic history of American women. New York, NY: Oxford University Press. Goldin, C. (1999). A brief history of education in the United States. NBER Historical Working Paper No. 119.

72

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Goldin, C., & Katz, L. (1999). Human capital and social capital: The rise of secondary schooling in America, 1910 to 1940. Journal of Interdisciplinary History, 29, 683
723. Goldin, C., & Katz, L. (2008). The race between education and technology. Cambridge, MA: Harvard University Press. Goldin, C., & Polachek, S. (1987). Residual differences by sex: Perspectives on the gender gap in earnings. American Economic Review, 77(2), 143
151. Greenwood, J., & Seshadri, A. (2002). The U.S. demographic transition. The American Economic Review, 92(2), 153
159. Greenwood, J., Seshadri, A., & Yorukoglu, M. (2005). Engines of liberation. Review of economic studies, 72, 109
133. Haines, M. R. (2000). The population of the United States, 1790
1920. In S. J. Engerman & R. E. Gallman (Eds.), The Cambridge economic history of the United States (Vol. 2). Cambridge: Cambridge University Press. Haveman, R., & Wolfe, B. (1995). The determinants of children’s attainments: A review of methods and findings. Journal of Economic Literature, 33(4), 1829
1878. Hazan, M. (2009). Longevity and Lifetime Labor Supply: Evidence and Implications. CEPR DP 5963. Hazan, M., & Zoabi, H. (2006). Does longevity cause growth? A theoretical critique. Journal of Economic Growth, 11, 363
376. Heckman, J. J. (1976). A life-cycle model of earnings, learning, and consumption. Journal of Political Economy, 84(4, part 2), S11
S44. Jones, L., & Tertilt, M. (2008). An economic history of fertility in the U.S.: 1826
1960. In P. Rupert (Ed.), Frontiers of family economics (Vol. 1, pp. 165
230). Frontiers of Family Economics. Bingley, UK: Emerald Group Publishing Limited. Katz, L. F. & Margo, R. A. (2013). Technical change and the relative demand for skilled labor: The United States in historical perspective. NBER Working Paper No. 18752. Light, A., & Omori, Y. (2012). Can long-term cohabiting and marital unions be incentivized? In S. W. Polachek & K. Tatsiramos (Eds.), Research in labor economics (Vol. 36, pp. 241
283). Research in Labor Economics. Bingley, UK: Emerald Group Publishing Limited. Lord, W. (1989). The transition from payroll to consumption receipts with endogenous human capital. Journal of Public Economics, 38, 53
73. Lord, W. (2002). Household dynamics: Economic growth and policy. New York, NY: Oxford University Press. Lord, W., & Rangazas, P. (1993). Altruism, deficit policies, and the wealth of future generations. Economic Inquiry, 31(4), 609
630. Lord, W., & Rangazas, P. (2006). Fertility and development: The roles of schooling and family production. Journal of Economic Growth, 11, 229
261. Miller, G. (2008). Women’s suffrage, political responsiveness, and child survival in American history. Quarterly Journal of Economics, 123(3), 1287
1326. Moehling, C. (2005). ‘She has suddenly become powerful’: Youth employment and household decision-making in the early twentieth century. Journal of Economic History, 65(2), 414
438. Mokyr, J. (2000). Why was more work for mother? Knowledge and household behavior 1870
1945. Journal of Economic History, 60(1), 1
41. Murphy, K., Simon, C., & Tamura, R. (2008). Fertility decline, baby boom, and economic growth. Journal of Human Capital, 2(3), 262
301. Olivetti, C. (2006). Changes in women’s hours of market work: The role of returns to experience. Review of Economic Dynamics, 9(4), 557
587.

Explaining the Revolution in U.S. Households

73

Parsons, D., & Goldin, C. (1989). Parental altruism and self-interest: Child labor among late-nineteenth century American families. Economic Inquiry, 27(4), 637
659. Polachek, S. W. (1975). Potential biases in measuring male-female discrimination. The Journal of Human Resources, 10(2), 205
229. Preston, S. H., & Haines, M. R. (1991). Fatal years: Child mortality in late nineteenth-century America. Princeton, NJ: Princeton University Press. Puerta, M. J. (2009). The fewer, the merrier
Compulsory schooling laws, human capital and fertility in the U.S. Princeton University Working Paper No. 102057. Ramey, V. (2009). Time spent in home production in the twentieth-century United States: New estimates from old data. Journal of Economic History, 69(1), 1
47. Rangazas, P. (2000). Schooling and economic growth: A king-rebelo experiment with human capital. Journal of Monetary Economics, 46(2), 397
416. Rangazas, P. (2002). The quantity and quality of schooling and U.S. Labor productivity growth (1870
2000). Review of Economic Dynamics, 5(4), 932
964. Rendall, M. (2010). Brain versus BRAWN: The Realization of Women’s Comparative Advantage. University of Zuric, Paper No. 491. Roberts, E. (2007). Her Real Sphere? Married Women’s Labor Force Participation in the United States 1860
1940. Dissertation, University of Minnesota. Rotella, E. (1980). Women’s labor force participation and the decline of the family economy in the United States. Explorations in Economic History, 17(2), 95
117. Rotella, E. (1981). From home to office: U.S. Women at work 1870
1930. Ann Arbor, MI: UMI Research Press. Smith, J. P. (1986). Race and human capital: Reply. American Economic Review, 76(5), 1225
1229. Soares, R., & Falcao, B. (2008). The demographic transition and the sexual division of labor. Journal of Political Economy, 116(6), 1058
1104. Sobek, M. (1997). A century of work: Gender, labor force participation, and occupational attainment in the United States, 1880
1990. Dissertation, University of Minnesota. Stone, L. (1977). The family, sex, and marriage in England 1500
1800. New York, NY: Harper and Row. Sundstrom, W. (2006). Hours and working conditions. In Historical statistics of the United States (Millenial ed., Vol. 2, pp. 47
48). Cambridge: Cambridge University Press. Topel, R. (1999). Labor markets and economic growth. In O. Ashenfelter & D. Card (Eds.), Handbook of labor economics (Vol. 3A, pp. 2943
2984). Amsterdam: Elsevier Science B.V. Turner, C., Tamura, R., & Mulholland, S. E. (2011). How important are human capital, physical capital and total factor productivity for determining state economic growth in the United States, 1840
2000. MPRA, 32856. Turner, C., Tamura, R., & Mulholland, S. E. (2013). How important are human capital, physical capital and total factor productivity for determining state economic growth in the United States. Journal of Economic Growth. Online First Article (unassigned), 1
48. United States Department of Agriculture. (2006). Costs of children. Retrieved from http:// www.cnpp.usda.gov/Publications/CRC/crc2006.pdf Whaples, R. (2010). Child labor in the U.S. Retrieved from http://eh.net/encyclopedia/article/ whaples.childlabor Wright, G. (1986). Old south, new south: Revolutions in the southern economy since the civil war. New York, NY: Basic Books. Zelizer, V. A. (1985). Pricing the priceless child. Princeton, NJ: Princeton University Press.

74

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

APPENDIX A: PROOFS OF PROPOSITIONS 1 AND 2 Condition (19) gives: 1 1 ½d2t ρ þ d1t ρhft wt − d2t μt wt h0t þ 1 Tt þ d2t wt ht τt Þ = λnt ðψ þ σÞ 1 þ ðhft wt d2t et þ d2t μt wt h0t þ 1 st þ d2t Pt xt Þ ðψ þ σÞ Conditions (16), (17) and (18) can be rewritten: 1 θx ψ = xt λnt d2t Pt

ðA:1Þ

1 θhe ψ = et λnt d2t wt hft

ðA:2Þ

1 θs ψ = st λnt d2t μt wt h0t þ 1

ðA:3Þ

When introducing the expressions (A.1)
(A.3) above and simplifying gives: ½d2t ρ þ d1t ρhft − d2t μt h0t þ 1 Tt þ d2t ht τt Þ 1 = wt ≡ wt A t ½σ þ ψ ð1 − ΣθÞ λnt

ðA:4Þ

where At ≡

½d2t ρ þ d1t ρhft − d2t μt h0t þ 1 Tt þ d2t ht τt Þ ½σ þ ψ ð1 − ΣθÞ

ðA:5Þ

The expressions (A.1)
(A.3) can be rewritten: xt =

θx ψwt At ; d2t Pt

et =

θhe ψ At ; d2t hft

st =

θs ψ At d2t μt h0t þ 1

ðA:5Þ

which, when substituting for At gives (20), (21) and (22). From (14) and (15), we get that ð 1 − vÞ v = =λ zt wt hft gt

ðA:6Þ

75

Explaining the Revolution in U.S. Households

or gt =

vwt hft zt ð 1 − vÞ

ðA:7Þ

Furthermore, using (A.6) in (A.7), we get: gt = w t At vnt nt =

gt hft = zt vwt At ð1 − vÞAt

ðA:8Þ

where the third equality uses the expression of (A.7). Now plugging the expressions of Eq. (A.5) in Eq. (11) and solving for ht gives ht =

  hft zt ðθs þ θhe þ θx ÞAt ψ þ d2t ðρhft þ ht τt − μt h0t þ 1 Tt Þ þ d1t ρt hft ð1 − vÞAt 0 1 νh ft A þ zt @hft þ ð1 − vÞ

where we use (A.8). Finally, 

zt = hft

ht ð1 − vÞ   d2t ðρhft þ ht τt − μt h0t þ 1 Tt Þ þ d1t ρt hft þ1 ðθs þ θhe þ θx Þψ þ At

which, using the expression of (A.5) and simplifying gives (25). Finally, using (A.7), (A.8) and substituting for At gives (23) and (24).

76

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

APPENDIX B: GOODS AND TIME INPUTS Table B1 depicts the time-path of schooling expenditures. Column (1) indicates the year of the expenditures. Column (2) provides the expenditures per pupil enrolled in public primary and secondary schools in 1982
1984 constant dollars. Column (3) is the school enrollment rate (public and private) of those aged 5
19, including postsecondary. The product of columns (2) and (3) yields column (4), expenditures per population member aged 5
19. This requires the reasonable assumption that expenditures per public and private student are roughly equal. For 1940, college expenditures among those 18 and 19 are added to the total to make comparable to student attendance data discussed below. This adds about 10% to the 1940 total.55 Column (5) is the ratio of expenditures per pupil enrolled in some year compared to that in 1870; column (6) is the ratio of expenditures per 5
19 population member in some year to those in 1865. Rangazas (2002, Table 1, p. 935) reports that the share of GDP devoted to primary and secondary education rises from 1.0% in 1880 to 2.4% in 1940. In the initial baseline it is envisioned that the ratio of schooling expenditures to father’s life earnings should be a bit above 1.0% (since women and children also contributed to earnings). Since our figures include college expenditures for 1940, the targeted share is 2.6% for that year.

Table B1. 1 Beginning school year 1865a 1870 1875 1880 1885 1890 1895 1900 1905 1910 1915 1920

2 $ per pupil enrolledb

74 84 78 106 121 155 165 200 264 284 323

Expenditures on Students.

3 4 % Expenditures enrolledc per pop. 5
19 0.484

31.1 35.82

0.578

45.08

0.543

65.70

0.505

83.33

0.592

156.29

0.643

207.68

5 1870$ per pupil enrolled

1.00 1.14 1.05 1.43 1.64 2.09 2.23 2.70 3.57 3.84 4.36

6 Product/1865 product 1.00 1.15 1.45 2.11 2.68 5.03 6.68

77

Explaining the Revolution in U.S. Households

Table B1. 1 Beginning school year 1925 1930 1935 1940

2 $ per pupil enrolledb 463 573 540 620

(Continued )

3 4 % Expenditures enrolledc per pop. 5
19 0.699

400.53

0.748

463.76 (516)d

5 1870$ per pupil enrolled 6.26 7.74 7.30 8.38

6 Product/1865 product

12.87 14.9 (16.6)

a

Estimated (see text). Table Bc909
925 Public elementary and secondary school expenditures from Historical Statistics of the United States, Millennial Edition (Vol. 2). New York, NY: Cambridge University Press, contributed by Claudia Goldin. c Numbers are displayed as proportions. Table CG.A.15 School enrollment of 5
19-year olds per 100 persons, by sex and race: 1850
1994; Goldin (1999) “A Brief History of Education in the United States.” Percentage includes attendance at public and private schools and also home schooling (so long as deemed comparable to regular schooling and led to a degree). d Educational and general expenditures per undergraduate in 1941 (HSUS table Bc966). b

Table B2 provides information about time spent in school by children aged 5
19 in various years. Column (1) indicates the school year. Column (2) is the average days attended per enrolled public school student. Column (3) is the percent of the 5
19-year-old population enrolled in school (public and private, includes college enrollment). The product of columns (2) and (3) yields column (4) the days attended per member of 5
19 population (assuming days attended are equal for private and public students). The fraction of days attended in a year are then provided in column (5). School time triples from the 1870 to the 1940 school year. Schooling time of a member of a given birth cohort is proxied by the column (5) fraction corresponding to when birth cohort members are aged 15. Then, for example, members of the birth cohort of 1925 (viewed in 1940) devote 30.9% of their time endowment to school. For the 1900 birth cohort we average the figures from 1910 to 1920 to get 20.2%; the figure for the 1875 birth cohort is then 12.9%. These figures are comparable to those produced by Rangazas (2002, Table 2, p. 936).

78

M. CINYABUGUMA, W. LORD AND C. VIAUROUX

Table B2. 1 School year beginning 1865a 1870 1875 1880 1885 1890 1895 1900 1905 1910 1915 1920 1925 1930 1935 1940

Children’s Time in School.

2 Product/365

3 Percentage enrolledb

4 Days attended

5 Product

79.4 79.4 80.0 84.1 86.6 94.8 98.0 106.0 111.8 120.9 125.9 135.9 144.0 146.3 150.7

0.484

38.43

0.09 0.105

0.578

46.24

0.12

0.543

47.02

0.129

0.505

49.49

0.136

0.592

66.19

0.181

0.643

80.95

0.222

0.699

100.62

0.276

0.748

112.69

0.309

Source: Goldin (1999), Table CG.A.6. a Estimated (see text). b Numbers in this column are displayed as proportions.

INTEGRATING RETIREMENT MODELS: UNDERSTANDING HOUSEHOLD RETIREMENT DECISIONS Alan L. Gustman and Thomas L. Steinmeier ABSTRACT This paper advances the specification and estimation of econometric models of retirement and saving in two earner families. The complications introduced by the interaction of retirement decisions by husbands and wives have led researchers to adopt a number of simplifications. Our analysis relaxes these restrictions. The model includes three labor market states, full-time work, partial retirement, and full retirement; reverse flows from states of lesser to greater work; an extended choice set created when spouses make independent retirement decisions; heterogeneity in time preference; varying taste parameters for full-time and part-time work; and the possibility of changes in preferences after retirement.

Factors Affecting Worker Well-Being: The Impact of Change in the Labor Market Research in Labor Economics, Volume 40, 79
112 Copyright r 2014 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-912120140000040003

79

80

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

Keywords: Retirement; retirement in two earner families; household retirement; family retirement; pensions; Social Security JEL classifications: C61; D31; D91; E21; H55; I3; J14; J16; J26; J32

INTRODUCTION Researchers continue to make progress toward a comprehensive econometric model of retirement and saving. Our aim in this paper is to make contributions on two levels. First, we contribute to the analysis of retirement and saving at the family level by introducing a level and complexity in the decision making of each spouse not found in previous work. Second, we bring together into a model of family retirement details regarding the choice set, preferences, and constraints previously found only in models of retirement in one earner households, and typically not found together in a single model, even of one earner households. Among the features in our family retirement model: (1) Outcomes include full-time work, partial retirement, and full retirement. (2) Saving and retirement are jointly determined. (3) The analysis is structural. Preferences and constraints are specified separately. (4) One cannot borrow on future income or Social Security. Thus liquidity constraints are incorporated into the analysis. (5) There is heterogeneity in time preference. This means that the response rates to future rewards from wages or from postponing Social Security or pension claiming differ among members of the population. It also means, consistent with the distribution of wealth, that some will be well prepared for retirement and others poorly prepared. (6) There are two utility functions, two decision makers, and one budget constraint. Each spouse makes his or her own decisions. However, the decision makers are strategic; their decisions are made with awareness of the other spouse’s decision making, preferences and opportunity set, implications for the income and wealth of the household, for joint consumption, and for the welfare of the spouse. (7) In this specification, the solution to the model is not as obvious as in the case with a single (unitary) utility function. Within periods, household solutions take account of the reactions of one’s spouse. Between periods, household decision making is solved through backward induction. We show that this set up sometimes provides

Integrating Retirement Models

81

different results from a solution based on a pure-strategy Nash equilibrium. For example, strategic decisions can allow the husband and wife to avoid prisoners’ dilemma-type outcomes. (8) The interactions of the decisions of each spouse are fully integrated into the estimation procedure. (9) The analysis is stochastic, including stochastic life expectancy and interest rates. (10) People may reduce their work effort overtime, then subsequently increase it. Some people may return to the labor force after retiring fully, or increase their hours of work after partially retiring, as circumstances change, with some events foreseen and others not. Others may change their decisions if retirement is less satisfying than anticipated, or as decisions of husbands and wives interact in ways that have not previously been analyzed. (11) Incentives from pensions and Social Security are fully modeled on an individual basis for each individual in the sample, including the very sharp, nonlinear incentives from defined benefit pensions that still account for two-thirds of the pension wealth of those on the cusp of retirement. To do this, the estimation makes use of linked data in calculating the opportunity set, including detailed pension plan descriptions obtained from employers. It also fully models the Social Security benefit formula using matched Social Security records. It does not use typical or representative pension or Social Security incentives. (12) There are minimum hours’ constraints at work. Many jobs, typically held in the prime working years and paying higher wages, require full-time work or none at all. Partial retirement is not allowed on the career job unless specified in the answers to the survey. (13) The wage offer on a career job depends on tenure and on hours of work. Typically when one leaves a career job, the wage offer drops substantially, as it often does for part-time work. To contribute to the analysis of decision making at the family level, we model alternative rankings of husbands and wives among many potential outcomes. This involves a richer array of outcomes than has been considered in earlier models of family decision making, including our own. Each spouse decides on full, partial, or nonretirement, the path to be taken in future years, implications for the behavior of their spouse, and responses to changes in that behavior. The model incorporates a number of relevant dimensions of behavior, importantly the major flows among the various retirement states, including reverse flows.1 In the process, the model reproduces such major features of the retirement hazard as the spikes in retirement at ages 62 and 65. The model also reproduces the joint distribution of retirement by husbands and wives, reflecting the coordination of couples’ retirement decisions.2

82

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

In addition, it reproduces the wealth accumulated for retirement by each household. Our econometric model is estimated with data from the Health and Retirement Study (HRS). It is sufficiently detailed to facilitate policy analysis, allowing us to change the specification of the equations representing the incentives created by Social Security, pensions, labor market opportunities, including the availability of partial retirement or spouse employment opportunities, and other constraints on individual behavior. Our model is sufficiently rich to allow us to estimate the likely effects of policy on the retirement and saving of those at different parts of the income and wealth distributions. The section “The Dynamic, Stochastic Model” presents the stochastic, dynamic model of retirement and saving decision making by two earner families. The decision process in the two earner families is further explored in the section “The Decision Process.” Data and estimation are discussed in the section “Data and Estimation.” The section “Simulations” presents simulations with the model, including a number of policy analyses. The final section concludes.

THE DYNAMIC, STOCHASTIC MODEL The utility functions of the two spouses are functions of consumption (a public good within the household) and labor supply over the lifetime. For the husband, the value function at time t is give recursively by:   h V h ðmt ; tÞ = α1½Cðmt ; tÞα þ eXt βh þ ɛh uhL Lh ðmt ; tÞ; Lw ðmt ; tÞ þ e−δ

3 X

pðmt ; mt þ 1 ÞV h ðmt þ 1 ; t þ 1Þ

mt þ 1 = 1

where mt is an indicator of whether both spouses are still living at time t (mt = 1), only the husband is living at time t (mt = 2), or only the wife is living (mt = 3). In this function, C(mt, t) is consumption at time t in survival state mt, and Lh and Lw are the leisure (i.e., retirement) amounts of the husband and wife, respectively. p(mt, mt + 1) is the probability that the household in state mt at time t will be in state mt + 1 at time t + 1, and uhL determines the value of leisure to the husband. Note that the value of leisure to the husband may depend on the amount of leisure of the wife in the same period. The exponential form preceding uhL is a multiplicative factor

83

Integrating Retirement Models

for the value of leisure. It consists of a standard linear form Xβ plus an individual effect ɛ which reflects the strength of the husband’s preferences for retirement over work. The elements of X contain a constant, age, and health status, with the corresponding parameters βho , βh1 , and βh2 . As age increases, work gradually becomes more onerous and retirement more desirable. When the utility of retirement exceeds the utility of consumption from the income earned from work, the individual retires. Leisure can take on three values associated with full-time work, partial retirement, and full retirement. The value of retirement for the husband can be modified if the wife is also retired. For the case where the wife is working full time, the basic value of leisure uhL for the husband is normalized to zero if he is working full time, unity if he is fully retired, and uhP if he is partially retired. We assume the utility function is concave in leisure. If partial retirement is equated with approximately half-time work, the value of uhP should fall between one-half and one. The closer the value is to one, the greater the value of partial retirement is relative to full retirement, and the more frequently and longer should be the spells of partial retirement. The model allows the value of retirement to be increased if the husband prefers to spend time with the wife, and the wife is also retired. In the case without partial retirement, this can be accomplished simply by adding an additional variable to the Xβ linear form reflecting whether the spouse is retired. With partial retirement, however, the picture is a little more complex. The general idea is that the wife’s retirement adds to the utility of the husband’s retirement only up to the point of the husband’s retirement. If the husband is partially retired, it doesn’t matter whether the wife is fully retired or partially retired, since the additional leisure of the wife, if she is fully retired, doesn’t add anything to the husband’s leisure during the time he is at work. If the husband is fully retired, however, it does make a difference whether the wife is partially retired or fully retired, since only the part of his leisure that he shares with the wife is augmented. If the augmentation factor is η, the values of the function uhL as a function of its two arguments can be written as follows:

Husband’s Retirement Status

Wife’s Retirement Status

Full-time work Partial retirement Full retirement

Full-Time Work

Partial Retirement

Full Retirement

0 0 0

uhP

1 ηuhP +(1 − uhP ) η

ηuhP ηuhP

84

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

If the husband is working full time, the value of his leisure is zero regardless of the retirement status of his wife. If the husband is partially retired, the value of leisure is uhP if his wife is working full time, and this gets multiplied by a factor of η if his wife is working no more than part time. If the husband is fully retired, the value of leisure is unity. This again is multiplied by a factor of η if his wife is also fully retired. If his wife is partially retired, we can divide the husband’s leisure into two parts. His partial retirement leisure has a value of uhP , which is multiplied by η because his wife is also partially retired. His remaining leisure has a value of 1 − uhP , but this leisure does not get multiplied because the wife is not there for this leisure. The total value of the husband’s leisure is the sum of these two parts. The value function of the wife is symmetric:   w V w ðmt ; tÞ = α1½Cðmt ; tÞα þ eXt βw þ ɛw uwL Lh ðmt ; tÞ; Lw ðmt ; tÞ þ e−δ

3 X

pðmt ; mt þ 1 ÞV w ðmt þ 1 ; t þ 1Þ

mt þ 1 = 1

where the superscripts and subscripts w refer to the wife’s utility and leisure. The budget constraint for the family is given by the asset evolution equation: At þ 1 = ð1 þ rt ÞAt þ ð1 − Lht ÞWth þ Bht þ ð1 − Lwt ÞWtw þ Bwt þ It − Cm;t All quantities in the equation are measured in real terms. Assets are constrained to be non-negative. They grow at the real interest rate rt. The second term on the right side is the husband’s earnings, and the fourth term is the wife’s earnings. The third and fifth terms are the husband’s and wife’s pension and Social Security benefits, respectively. Although not indicated by the notation, these benefits depend on the past work and retirement decisions. In the case of Social Security, these can even depend on the past work and retirement decisions of the spouse. The term It is any inheritances that the household may receive, and the last term is household consumption. Note that consumption is dependent on the survival state of the household, and that the budget constraint must hold regardless of the mortality experience. Social Security and pension benefits are included in the asset evolution equation on the date they are received, and not on the date they are accrued. Benefits are assumed to be collected on the earliest date possible under the rules of Social Security and the pension plan. For Social Security, benefits can be collected after the early entitlement age even if the

Integrating Retirement Models

85

individual is working, as long as the earnings test does not completely erase the benefits. A person who loses the entire benefit to the earnings test must wait until the full retirement age to collect benefits. Pensions can be collected no earlier that the date the individual leaves the job, and they may be delayed if the individual has not worked long enough to be eligible for immediate benefits. Individuals may leave one full-time job and immediately take another full-time job in order to become eligible for benefits from the first job, although the second full-time job will generally have lower wages associated with the loss of tenure. The model includes a stochastic real interest rate rt, which is assumed to come from the actual distribution of asset returns, as documented by Ibbotson Associates (2002). Time periods are annual and asset returns are uncorrelated overtime, which is approximately the case. Mortality outcomes reflected in the underlying transition probabilities p(mt, mt + 1) are also stochastic. The time preference parameter δ is heterogeneous over the population of households and is treated as a fixed effect whose value is estimated for every household in the population.3 The two epsilon terms in the linear forms multiplying leisure, ɛh and ɛw, come from normal distributions with mean zero and standard deviations of σh and σw, respectively, and with a correlation parameter ρɛ. These terms allow husbands and wives to have different preferences for leisure versus work. The correlation allows for the possibility that one reason spouses in households may retire at about the same time is that the two spouses share the same attitudes toward leisure versus work. For estimation purposes, the initial values of these two epsilon terms are treated as random effects. After the individual first leaves full-time work, the value of leisure may suffer an unanticipated change. For instance, it may fall if the individual finds that leisure is not as pleasurable as anticipated, leading the individual to return to work after retiring. To provide for this possibility, the model allows the value of the epsilons to change overtime after the initial retirement, with correlation parameters overtime of ρh and ρw for the husband and wife, respectively. The model assumes that the spouses do not anticipate that the values of leisure might change after retirement.4 uhP and uwP , the values of partial retirement, are stochastic from the estimator’s point of view, but not from the individual’s point of view. We h h assume they come from an exponential distribution f ðuhP Þ = ceγt uP defined over the interval 0.5
1, where the factor c is necessary to make the integral of the distribution over the allowable range equal to 1. The higher γ ht , the greater the probability the husband’s value of uhP is close to 1, making it more likely that he will go through a period of partial retirement.

86

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

Conditional on employment, the probability of partial retirement increases with age. Accordingly, we allow the value of γ ht to increase with age according to the equation γ ht = γ ho þ γ h1 aht , where aht is the husband’s age at time t. We assume the husband’s relative preference for partial retirement, and therefore the husband’s relative position within this distribution, does not change even as the entire distribution changes to higher values of uhP . The values of uwP for the wife are treated symmetrically. The sequence of events in the model is as follows. In every period, the couple starts with a level of assets. If the husband was still in his career job in the previous period and if that job had a defined contribution pension, he may also have a given level of defined contribution pension assets. If he was not in the career job in the previous period but had a defined benefit pension in that job, he may also have a pension benefit amount, which may or may not be currently collectable. The same situation with regard to pension amounts also applies to the wife. Further, the household may be eligible to collect a given level of Social Security benefits, either currently or sometime in the future. Given their current situation, the spouses make their decisions as to what their retirement status (working full time, partial retirement, or full retirement) will be during the current period, and how much they will save or spend down from their accumulated assets. If an individual was working in the career job in the previous period and decides not to work in the career job this period, any defined contribution pension assets are combined with regular assets and assumed to be available for consumption. At the end of the current period, a random draw is taken from the distribution of asset returns, and assets plus any defined contribution amounts are assumed to grow at that rate of asset returns. One of the spouses may die, causing a transition between the survival states. Depending on the retirement status decisions made in the current period, the amount of futuredefined benefit pensions and Social Security amounts may also be increased. The model assumes that between ages 50 and 70, spouses make their decisions about working full time, partially retiring, or fully retiring. Before age 50, each spouse works at whatever their earnings were. After age 70, they are retired. With two utility functions, two decision makers, and one budget constraint, the solution of the model is more complex than in the case with a single utility function. The mechanism by which spouses make their labor supply decisions is described in the next section. At all ages, consumption decisions are made given income in the current period and the assets owned when the couple entered the period. The state variables of the model are collected into the vector S = {A, ɛh, w ɛ , Jh, Jw, dch, dcw, dbh, dbw, dfh, dfw, ssh, ssw}. The J’s are binary variables

Integrating Retirement Models

87

indicating whether the individual is still in the career job. The dc’s are levels of defined contribution pension assets, and the db’s are annual defined benefit pension amounts. The df’s are binary variables indicating whether the defined benefit pension amounts are currently collectable or whether they are deferred until a later date as specified in the plan. The ss’s indicate the individual’s own Social Security benefit; the family Social Security benefit is a function (and not necessarily the sum) of these two variables. The real value of pension and Social Security variables are initially set when the individual leaves the career job. For the pension variable, this value is adjusted downward overtime, since few pensions are fully indexed for inflation.5 For the Social Security variable, the value for future periods may be increased if the individual returns to full-time work or engages in partial retirement work, an increase that is due either to a reduction in the early reduction amount or an increase in the delayed retirement amount. However, Social Security benefits in the current period may be reduced if the individual works either full time or part time and the earnings from that work exceed Social Security’s earnings test limit. The model is solved through the usual backwards induction method. Given values of the individual effects and the parameters of the model, the decision vector and its value are calculated for all possible combinations of the state variables in the final period. Next, the immediately preceding period is considered, and the decision vector and its value are again calculated for all possible combinations of the state variables, allowing for the stochastic effects which affect subsequent values. The continuous variables (assets, the epsilons, pension, and Social Security amounts) are broken down into a vector of discrete amounts, and, if required, any amount between these discrete amounts is interpolated. The integration related to the stochastic events (asset returns) in the expected values is done numerically. The process is repeated backwards until the first period.

THE DECISION PROCESS This section describes the process by which the husband and wife make their decisions. Within each period, each spouse maximizes his or her value function. Both are aware that their decision may influence the decision of their spouse. We call this decision process, in which each spouse is aware that their choice can influence their spouse’s choice, strategic decision making.

88

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

Strategic Decisions with a Work/Retirement Choice To begin this analysis, it will help to look first at a framework where each spouse is deciding either to work full time or retire. This leaves the analysis of partial retirement until later in the section. In such a framework, the joint decision result has four possibilities: both spouses work; both spouses retire; the husband works and the wife retires; and the husband retires and the wife works. These possibilities can be represented by the following grid: Husband’s Decision

Wife’s Decision

Work Retire

Work

Retire

A C

B D

Thus under decision B the husband retires and the wife works, and similarly for other outcomes. Both the husband and the wife have utility values associated with each alternative. Let VBh be the husband’s value if the wife works and the husband retires. This value will include the expected utility of decisions in future periods, given the income and consumption in the current period and any restrictions that the current decision to work places on future decisions. For instance, the decision of the husband to be retired in the current period may mean that in any future periods the wage rate for full-time work will be lower than if the husband had continued in his career job. Similarly, VCw would denote the wife’s value if she retires and the husband works full time. In the strategic framework, either spouse may choose to retire, but if they do so they are locked into that decision for at least that period. This reflects the difficulty in general of reversing the retirement decision over a short period of time without incurring undue transition costs. If either spouse does choose to retire, the other spouse is assumed to have the opportunity to make a similar choice in the period. As an example, suppose VAh < VBh and VCh < VDh , so that the husband prefers to retire whatever the wife’s decision. Also suppose VAw < VCw and VBw < VDw , so the same is true for the wife. In this situation, if the husband retires, the wife will certainly retire, and vice versa, and both spouses know as much. The Nash equilibrium is clearly choice D, with both spouses retiring, and this choice is a definite possibility for the strategic case. However,

Integrating Retirement Models

89

it is possible that choice A would be preferred to choice D by both spouses. The intuition here is that if one spouse retires, that spouse knows that s/he will lose not only own earnings but the earnings of the other spouse as well. In this case, neither of them would have the incentive to retire unless the other retired first, and as a result the decision will be to remain at choice A, which is preferred by both. If either spouse prefers choice D to choice A, however, that spouse will retire and the other spouse will follow. The key difference between the strategic framework and the Nash framework is that the strategic framework allows for the possibility that one spouse will react to the other spouse’s choice. In the Nash framework, the decisions are simultaneous, without the possibility that either spouse will react to the other spouse’s choice. In essence, the model treats retirement as an absorbing state within a period, so that the second spouse has the opportunity to react to the first spouse’s decision to retire, and the first spouse takes the second spouse’s reaction into account in making his or her decision to retire. An exhaustive (and possibly exhausting) discussion of all the possibilities of this framework is found in Gustman and Steinmeier (2009), but a couple of other examples will be illuminating. Suppose that VAh > VBh and VCh < VDh , so that the husband prefers to work if the wife works and retire if she retires, and that VAw < VCw and VBw > VDw , so that the wife prefers to retire if the husband works and work if the husband retires. Also, if the wife is working, having the husband working does not decrease the leisure term in her utility function (since Lw = 0 in this case) but does provide additional income from which the wife also benefits. This means that VAw > VBw , which in turn is sufficient to establish a strict preference ordering of the four choices for the wife: C > A > B > D. She knows if she retires first, the husband will follow suit, resulting in combination D, her least preferred outcome. Thus, she will not want to be the first to retire. This leaves the husband with the choice between combinations A and B. Given his preferences in this case, he will choose not to retire. Thus, the final result is combination A, with both spouses working. Note that there is no simple Nash equilibrium in this case, but even so, the strategic decision process yields a definite result. A second example occurs if VAh < VBh and VCh > VDh for the husband and w VA < VCw and VBw > VDw for the wife. Both partners prefer to do the opposite of the other spouse’s choice. The wife’s preferences are the same as in the last example, so again her preference ordering among the four combinations is C > A > B > D. Analogous reasoning for the husband establishes B > A > C > D for his preference ordering among the four combinations. The husband would prefer to retire first, since if he does so the wife would

90

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

continue working and the final result would be B, his most preferred combination. The wife would also like to be the first to retire, since in that case the final result would be C, her most preferred combination. But if both of them retire in an effort to be first, the result would be D, the least preferred combination for them both. This is the one case where the strategic reasoning by itself does not yield a definite result. We assume that in this case the spouses would cooperate enough to avoid the worst result and neither would retire, leading to combination A, which is the second choice for both spouses. However, note that if either spouse shows much of a preference for joint retirement, this case is unlikely to occur, since for that spouse a preference for retiring if the other spouse is working will almost surely not be reversed if the other spouse retires. The strategic framework may lead to interactions between husbands and wives not previously analyzed. For instance, consider a situation where the husband has been working and the wife retired in previous years, and now the husband is on the cusp of retirement. In the current year, both spouses would prefer to be retired, given the retirement choice of the other spouse, but both would prefer choice A rather than choice D. Hence, neither wants to retire lest the other retire too, leading to a less preferred choice for both. In essence, if either spouse chooses retirement he or she loses not just his or her earnings, but also the earnings of the other spouse as well. As a result, neither will want to choose retirement, and the wife will return to work in order to keep the husband working as well.

Decisions with Partial Retirement The addition of partial retirement introduces a third choice for both the husband and the wife, but the general method of analysis remains the same. To analyze the choices with partial retirement, we denote the nine possible outcomes as follows:

Husband’s Retirement State

Wife’s Retirement State

Working Partially retired Fully retired

Working

Partially Retired

Fully Retired

A D G

B E H

C F I

Integrating Retirement Models

91

Again, each of these states has a utility value associated with it for both the husband and the wife. Each spouse knows that the other spouse may react to their decision, and they make their own decisions with regard to the other spouse’s reactions. The general approach is similar to the case previously described with only the two choices of working or retirement. At the beginning of each period each spouse calculates the utility of either partially or fully retiring, given the other spouse’s reaction. This may be a somewhat more complicated process than occurs if there is only a work/retire decision. For instance, if the husband partially retires first, the wife may partially retire as a result. But that may not be the end of the story. In reaction to the wife’s partial retirement, the husband’s preference may be to fully retire, in which case the wife may or may not follow suit. The chain of reasoning to conclude how both spouses will decide is analogous, though a bit more involved than the situation in the previous subsection. Each spouse then ascertains whether it would be better to go ahead and either partially or fully retire, given the other spouse’s reaction, or continue working and allowing the other spouse to make the decision either to continue working or retire. If both spouses conclude that they are better off working, or if only one spouse decides to retire first, then those decisions determine the final outcome. There are cases where both spouses conclude that is in their interest to be the first to retire, but again these instances would not be expected if either spouse values joint retirement very much. As before, we assume that if such instances were to arise, the spouses would reach an accommodation wherein both spouses would settle for a combination which was, while not the most preferred by either, better than the combination which would be reached if both spouses tried simultaneously to achieve their most preferred combination. It may be helpful to illustrate a specific case in more detail. First, consider the husband’s decision to retire, either partially or fully. If he retires fully, the wife is faced with the choices C, F, and I. Suppose that VCw < VFw < VIw . Then if the husband retires fully, the wife will do so too, and the result will be combination I. Now consider what happens if the husband retires partially. In this case the wife’s choices are B, C, E, F, H, and I. Suppose that in addition to the previous preferences of the wife for C, F, and I, VBw < VEw < VHw , VIw < VEw , and VHh < VIh < VEh . If the husband partially retires, the wife would certainly want to at least partially retire since she prefers either E or H to B and either F or I to C. If she fully retires, however, she knows that the husband will also fully retire, since he prefers I to H. If she partially retires, she knows that he will stay partially retired since both prefer

92

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

E to H. So if he partially retires, he expects that she will partially retire also. Since he prefers the ultimate outcome E if he partially retires first to the ultimate outcome I if he fully retires first, he would choose partial retirement if he were the first to choose. Now consider the wife’s decision to retire, either partially or fully. Suppose that in addition to the preferences in the previous paragraph, VDh < VEh < VFh and VGh < VHh < VIh . By the same reasoning as in the previous paragraph, if the wife is the first to partially retire, the final outcome will be E, and if she is the first to fully retire the final outcome will be I. Since she prefers E to I, she will choose partial retirement if she were the first to choose retirement. The final piece of the decision is whether either of them will be the first to choose to retire. The husband knows that if he is the first to retire, he will choose partial retirement and the final outcome will be E. Likewise, the wife knows if she is the first retire, she will choose partial retirement and the final outcome will be E. If neither one of them retires, the final outcome will be A. Suppose that VEh < VAh and VAw < VEw . The husband would be fine to remain at A, but the wife finds E preferable. She partially retires, and the husband follows suit. If, on the other hand VEh < VAh and VEw < VAw , neither spouse would have an incentive to retire first, and the final outcome would be combination A. While the retirement decisions in the model are made strategically by the spouses, the savings and consumption decision is nonstrategic. The spouses have the same time preference and the same consumption parameter α, so the consumption component of the two utility functions is identical. Given the retirement decisions, consumption is chosen so as to equate the marginal utility of current consumption with the expected marginal utility of savings. Given that the primary focus of this paper is on retirement behavior, we leave the introduction of separate time preference rates and different consumption parameters to future work.

DATA AND ESTIMATION Estimation in this paper is based on data from the initial cohort of the HRS, a panel survey of roughly 7,600 households with at least one individual born from 1931 to 1941. The first wave of the HRS was in 1992, with reinterviews every two years thereafter. Both spouses in HRS households were interviewed separately, even if one spouse was in the eligible birth

Integrating Retirement Models

93

cohort but the other was not. The HRS also has two important supplements, which are available on a restricted basis. First, Social Security earnings records are attached for about 75% of the sample, allowing fairly precise estimates of Social Security earnings and benefits for this part of the sample. Secondly, for about two-thirds of those respondents who indicated they had a pension on their current job, the survey obtained and coded the pension’s summary plan description. This enables a much more precise determination of the retirement incentives of pensions than is normally obtainable from the respondents themselves. Our analysis uses data from the first six waves of the survey, so that in the last wave the eligible birth cohorts were 61
71 years of age. By age 61 half of the individuals were retired, with virtually all retired by age 71, so this period includes the overwhelming majority of retirements for the sample individuals. Observations are from households of married couples where both spouses appear to have career jobs from which retirement is a meaningful concept.6 Social Security benefits are calculated from two state variables (one for each spouse) for the basic benefits, which is essentially the primary insurance amounts (PIA) times any applicable early retirement penalty or delayed retirement credit. Benefits are reduced if the individual is working and subject to the earnings test, and the state variables are updated to reflect changes in the penalty or credit as a result of the earnings test. Spouse and survivor benefits are also calculated from these two state variables. Defined benefit pension amounts are treated similarly, as a state variable reflecting the amount of the pension benefit. This variable is effectively determined when the individual leaves the pension job. Defined contribution pension amounts are treated as a separate state variable reflecting the fact that these amounts are not available for consumption until the individual has separated from the job. Like the asset variable, the returns to defined contribution plans are treated as stochastic. In this model, neither the wage nor health status is treated as stochastic. The estimation method is the generalized method of simulated moments (MSM). In this procedure, a group of moments is gathered into a column vector m. These moments are generally the difference between some observed statistic, such as the percentage retired as of a specific age, and the percentage that is simulated for the sample using specified values of the parameter. In general, these moments come from an asymptotically normal distribution with a mean value of zero. The estimation procedure seeks the P parameter values which minimize q = m0 W−1m, where W = ni= 1 mi m0 i . The mi vectors are the moments of the individual observations, and the W

94

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

matrix is essentially the observed variance
covariance matrix of the moments. Variances of the estimates are calculated from var(Θ)= [G0 W−1G]−1, where Θ is the vector of parameters and G is the derivative of the moments with respect to the parameters. If the model is correctly specified, m is distributed around zero, and q should have a χ2 distribution with λ − k degrees of freedom, where λ is the number of moments and k is the number of parameters estimated. To construct the moments for a specific observation using a specific set of parameters, we follow a two-step procedure. The first step is to calculate the approximate value of the time preference parameter δ using the observed retirement dates and the value of wealth as of the beginning of the survey. If one or both spouses have not retired, we use the expected retirement date, and if the retirement from full-time work was into partial retirement, retirement from partial retirement work is calculated in the same way. This establishes the earnings stream of the household, and we solve the remaining consumption model with stochastic returns and mortality. This simplified model is solved by backwards induction to yield the saving decision at any age for any level of wealth. Wealth is then started at zero and accumulates using the observed returns and savings decision. The value of δ is adjusted up or down until the computed level of wealth at the beginning of the survey just matches the observed wealth. In general, it is not possible to establish a unique value for an individual effect such as δ in a model with unobserved stochastic elements, as is the case with the full model. However, extensive experimentation with the full model shows that the range of values of δ which yield particular retirement dates and a particular wealth level is fairly narrow.7 Further, the value of δ calculated from the simplified model is generally within that range. The intuition for this result is that given the observed opportunity set, the unobserved leisure preferences are fairly closely tied to the retirement dates, so the assumption of fixed retirement dates does not greatly affect the calculation of the time preference δ. The second step is to solve the full model using the value of δ calculated from the first step. This is done using the method of backwards induction discussed at the end of the section “The Dynamic, Stochastic Model.” It yields the decision rules applicable to any set of the state variables. Random draws are made from the distributions of ɛh, ɛw, uhP , and uwP , given the parameters σh, σw, ρɛ, γ ho , γ h1 , γ wo , and γ w1 of those distributions. These values are combined with the observed rates of returns and the decision rules to calculate the work-retirement patterns used in the moments. Then another draw is made for ɛh and ɛw, and the process is repeated 10,000

95

Integrating Retirement Models

times. Each time, another set of retirement states for both spouses is calculated, and the values associated with the various moments are updated. Once the model has been estimated, the calculation of the 10,000 simulations does not add appreciably to the time required, so we use this number in order to make the simulated moments close to the theoretical moments. In order to make the calculations feasible, new draws of uhP and uwP are not made for each of the 10,000 simulations per observation, but new draws of these values are made for each observation. The moments used in the simulation are chosen to provide identifying information on the parameters. The moments used are as follows:

Description of Moments The percentage retired from full-time work at ages 54
66 Husband Wife

Number of Moments 13 13

The percentage completely retired at ages 55, 58, 60, 62, and 65 Husband Wife

5 5

The percentage retired from full-time work at ages 55, 58, 60, 62, and 65 among families in the bottom third of potential earnings Husband Wife

5 5

The percentage retired from full-time work at ages 55, 58, 60, 62, and 65 among families in the upper third of potential earnings Husband Wife

5 5

The percentage retired from full-time work at ages 55, 58, 60, 62, and 65 among those in poor health Husband Wife

5 5

The percentage of reversals where the respondent was working full time after having been partially or fully retired in the previous interview Husband Wife

5 5

The percentage of couples in each interview where both spouses were working full time

6

The percentage of cases where one partner was working full time in one of the later interviews and the other was retired in one of the earlier interviews

6

Total moments

88

96

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

The first 26 moments help to establish the overall pattern of retirements governed by βho , βh1 , βwo , βw1 , σh, and σw. The next 10 moments reflect the amount of partial retirement by age, and hence help to establish the values of γ h0 , γ h1 , γ wo , and γ w1 , which determine the distribution of the value of partial retirement. The next 20 moments help to establish the value of α. The higher the value of α, the later will be the retirement of high earning individuals relative to the retirement of lower earning individuals. The next 10 moments, which involve individuals in poor health, help to determine βh2 and βw2 . The following 10 moments, which have to do with reversals, help to establish the values of ρh and ρw, the changes in leisure preferences after retirement which could induce a return to work. The next six moments,

Table 1. Parameter Symbol

Parameter Estimates.

Description

Husband’s parameters Parameters for the value of leisure Constant βho Age βh1 Own health βh2 Spouse retired augmentation ln ηh Standard deviation of ɛh σh Correlation of ɛh after retirement ρh Parameters for the value of partial retirement Constant γ ho Age γ h1 Wife’s parameters Parameters for the value of leisure Constant βwo Age βw1 Own health βw2 Spouse retired augmentation ln ηw Standard deviation of ɛw σw Correlation of ɛw after retirement ρw Parameters for the value of partial retirement Constant γ wo Age γ w1 ρɛ α

Correlation between ɛh and ɛw Consumption parameter Number of observations Q value

Estimated Value

Absolute t-Statistic

−10.37 0.12 6.00 2.63 4.75 0.86

106.00 5.43 5.84 2.20 7.44 14.01

−0.68 0.24

1.47 2.58

−10.35 0.19 4.25 1.18 5.04 0.87

49.73 4.16 3.77 0.75 8.42 18.12

−6.14 0.06

4.02 0.32

0.67 −0.47

1.70 4.80 851 65.09

97

Integrating Retirement Models

which relate to joint retirement, help to determine the values of ηh and ηw, which govern how much the leisure of one spouse is augmented by the presence of the other spouse. The final six moments, which measure how often one spouse retires early and the other spouse retires much later, help to establish the value of ρɛ, the correlation between ɛh and ɛw. With high values of ρɛ, cases where one spouse retires early and the other much later occur less often than when ρɛ is low. Parameter estimates are reported in Table 1. Most are significant at conventional levels.8 Of particular interest are ηh and ηw, which describe how the retirement of one spouse affects the desirability of retirement of the other spouse. Consistent with our earlier work (Gustman & Steinmeier, 2004), it appears that the husband places a somewhat higher value on having the wife retired than vice versa. The standard errors, however, suggest that this difference is probably not statistically significant. For 70 degrees of freedom, the 5% critical value of the χ2 distribution is 90.53 and the 1% critical value is 100.43. The estimated value of 65.09 is well below these critical values, which indicates that there is little evidence that the model does not fit the data well, at least in the dimensions measured by the moments previously listed. Another way of looking at this is that the probability value of this q value is 0.64, which suggests that by chance, the q value would be higher than the estimated amount more than half of the time.9

SIMULATIONS Having obtained estimates for the parameters of the model, we now turn to simulations with the model. First we present the “base” simulation, which uses the actual budget sets for the individuals in the sample. Two features of these simulations are noteworthy. They reproduce the spike in retirement of husbands and wives, confirming the tendency of couples with different ages to retire together. They also reproduce the spike in retirement at age 62, the most prominent feature of the retirement hazard, yet a feature not reproduced in most models of family retirement, nor in most models of retirement that ignore joint decisions of husbands and wives. Table 2 shows the retirement outcomes from the base simulation. Focus on the three columns indicating the percentages of husbands and wives at each age working full time, partially retired, or fully retired. As expected, the percentage working full-time falls with age while partial retirement

98

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

Table 2. Age

Retirement States in Base Simulation, By Age and Gender.

Percent Pseudo-Retiring From FT work

Husbands 55 3.5 56 2.5 57 4.3 58 4.2 59 5.7 60 7.2 61 6.4 62 14.7 63 6.9 64 6.8 65 7.9 66 5.3 67 3.9 Wives 55 56 57 58 59 60 61 62 63 64 65 66 67

4.6 4.3 5.0 5.6 6.1 6.6 6.0 10.5 6.3 5.1 4.8 3.6 3.1

Percent

Observed Pct Pseudo-Retiring

From all work

In FT work

Partially retired

Fully retired

From FT work

2.7 1.8 3.0 3.4 4.3 5.4 4.9 10.7 5.4 5.7 7.0 5.0 3.5

87.4 84.9 80.6 76.4 70.7 63.5 57.2 42.5 35.6 28.8 20.8 15.5 11.6

3.4 4.1 5.4 6.1 7.5 9.3 10.8 14.7 16.2 17.3 18.3 18.6 19.0

9.2 11.0 14.0 17.5 21.8 27.2 32.1 42.8 48.2 53.9 60.9 65.9 69.4

2.1 3.9 3.2 3.1 4.7 5.2 5.5 13.9 5.8 8.0 6.9 6.2 3.1

3.2 3.1 3.5 4.1 4.4 5.4 5.0 8.8 6.9 6.3 6.2 5.4 4.5

76.0 71.7 66.7 61.1 55.1 48.5 42.5 32.1 25.7 20.6 15.8 12.2 9.2

6.9 8.1 9.6 11.1 12.8 14.0 14.9 16.7 16.1 14.9 13.5 11.7 10.2

17.1 20.2 23.7 27.8 32.1 37.5 42.5 51.3 58.2 64.5 70.7 76.1 80.6

6.1 2.8 7.0 4.8 4.0 3.5 7.1 12.7 8.8 4.8 6.3 1.4 4.8

increases. By age 62, fewer than half the husbands are still working full time, while for the wives the percentage working full time falls below half at age 60. Both husbands and wives have substantial partial retirements. For the husbands, the percentage partially retired grows steadily until it reaches 19% at age 67. For wives, partial retirement grows at earlier ages, reaching a peak of around 16.7% at age 62, and declines thereafter. The first two columns of the table are “pseudo” retirements, the differences between the figures at adjacent ages. For instance, 32.1% of husbands

Integrating Retirement Models

99

were fully retired at age 61 and 42.8% at age 62. The difference of 10.7% is taken as the pseudo-retirements from work at age 62. It is really a net result of individuals completely retiring at that age less the individuals who had been retired at 61 but who returned to work at age 62. In any case, the prominent feature of these numbers is the spike of retirements at age 62, both for the husbands and to a lesser extent for the wives. For comparison, the last column of the table gives the observed “pseudo” retirements from fulltime work. The spike in the simulations is approximately the same size as in the raw data for the husbands, but falls a little short for the wives. The model is a little less successful in capturing the secondary spike around age 65 for the husbands, and not very successful at capturing the secondary spike for the wives. Indeed, the raw data in column 6 shows retirement from full-time work for husbands is 8.0% at age 64 and 6.9% at age 65, but these differences are not statistically significant.10 Table 3 decomposes full-time work and partial retirement. Full-time workers are divided into individuals who are still working in career jobs, having never previously fully or partially retired, and those who previously fully or partially retired and have returned to full-time work. For both husbands and wives, the percentage of the sample who have returned to full-time work rises gradually up until about age 60, peaks at around 9 percentage points in the early 60s, and tails off thereafter. At its peak at age 61, more than 16% (9.4/57.2) of full-time husbands are individuals who have returned to full-time work after a period of full or partial retirement. For wives, the figure is even higher at 21% (8.8/42.5). Although after age 61 the numbers of full-time workers decline both in the career jobs and in the return jobs, the percentage of full-time workers who have been previously retired continues to grow until it reaches 28% for husbands at age 67 and 30% for wives at the same age. For partial retirement, the decomposition is between those who partially retired from full-time work without going through a period of full retirement and those who were fully retired at some previous point. From the table, it is clear that the majority of part-time work is done by individuals who are on a traditional path of moving from full-time work to partial retirement and then to full retirement. But a significant portion of part-time work is by individuals who previously were fully retired. For the husbands, a little over a third of part-time workers in their mid to late 50s have previously retired. This figure increases during the early 60s until by age 67 around half of the part-time husbands have previously been fully retired. Although wives exhibit the same trend, the percentages of previously retired part-time workers is somewhat lower, particularly in the mid-1950s.

100

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

Table 3. Returns to Work in Base Simulation, by Age and Gender. Age

Percent in Main Job

Percent to FT Work after Retiring

Percent in PT Work after FT Work

Percent in PT Work after Retiring

Husbands 55 82.3 56 78.8 57 73.8 58 68.5 59 62.2 60 54.5 61 47.7 62 35.9 63 28.9 64 22.9 65 16.1 66 11.6 67 8.4

5.1 6.1 6.8 7.9 8.5 9.0 9.4 6.6 6.7 5.9 4.8 4.0 3.2

2.3 2.6 3.5 3.9 4.8 5.9 6.7 9.9 10.0 10.0 10.2 9.7 9.4

1.1 1.5 1.9 2.2 2.7 3.4 4.1 4.8 6.2 7.3 8.1 8.9 9.6

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

6.1 7.0 7.7 8.3 8.6 8.8 8.8 6.3 5.0 4.3 3.9 3.5 2.8

4.9 5.8 6.7 7.6 8.5 9.0 9.3 11.0 9.8 8.6 7.6 6.3 5.3

2.0 2.3 2.9 3.5 4.3 5.0 5.6 5.7 6.3 6.3 5.9 5.4 4.9

69.8 64.7 59.0 52.9 46.5 39.7 33.8 25.8 20.8 16.3 11.9 8.7 6.3

Table 4 indicates the frequency with which husbands and wives transition at least once from a state of greater retirement to a state of lesser retirement. Roughly a third of both husbands and wives returned to full-time work after having partially or fully retired. The percentage is slightly higher among wives than among husbands. Around a quarter of both husbands and wives return to part-time work after a period of full retirement. These are not mutually exclusive categories, so one cannot simply add them up. Some individuals will go through both transitions. Accordingly, the last row of the table suggests that a little over two-fifths of husbands transition from greater retirement to less retirement, as do almost half of the wives.

101

Integrating Retirement Models

Table 4.

Respondent Returning to Work in Base Simulation, by Gender. Husbands Wives

Percent returning to full-time work after full or partial retirement Percent returning to part-time work after full retirement Percent returning to full-time work after full or partial retirement or returning to part-time work after full retirement

Table 5.

34.4 25.7 48.5

Distribution of Differences in Retirement: Years in Base Simulation.

Husband retires first

Wife retires first

29.0 23.7 43.2

Difference in Retirement Dates (Years)

Percent of Households

10 + 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6 7 8 9 10 +

19.3 3.1 3.3 3.4 3.5 3.5 3.7 3.6 3.6 3.5 13.5 2.2 2.8 2.9 3.2 3.5 3.3 3.4 3.0 2.7 9.1

Note: The retirement date is the year the individual first retired from full-time work.

Table 5 addresses the frequency with which both spouses retire at the same time. In the top part of the table the husband retires before the wife, while in the bottom part, the wife retires before the husband. Although on average the husband is a couple of years older than the wife, the husband retires first in only about half the cases, while the wife retires first in around 36% of the cases. In the remaining 13.5% of the cases, both spouses retire in the same year. Note that this spike in joint retirement is around 10
11

102

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

percentage points higher than the surrounding figures, where the age of retirement differs by a year or two in either direction. This is almost surely a result of the spouse retirement variables in the model (the η’s), which increase the value of leisure of one spouse if the other spouse is also retired. In contrast, the correlation in leisure preferences should lead to more of a smooth hump in the joint retirement distribution as opposed to a spike. Table 6 reports on the distribution of time preferences. Not quite half of the households have a time preference rate of less than 5%, while around a third have time preference rates of 50% or higher. The latter group essentially has no financial assets other than forced savings, meaning that they are effectively consuming all of their available income.11 Only around a sixth of the households fall in the middle, with time preference rates between 5% and 50%, and most of those are in the bracket from 5% to 10%. Clearly a model that assumes all households have a uniform time preference rate will yield very misleading results for a third of the households, while models allowing two mass points for the time preference distribution, one at a low rate and the other at a high rate, may be a lot closer to the truth. Even here, there is a very substantial difference between the amount of wealth that would be accumulated by an individual with a 1% time preference rate relative to an individual with a 3% time preference rate, and a model with only two mass points of time preference is unlikely to reflect that difference. Now we turn to two policy simulations: first increasing the early entitlement age to 64; and second changing the availability of opportunities for partial retirement. Our model explains the spike in retirements at the Social Security early entitlement age as a result of liquidity constraints and the high-time preference rates found for many in the sample. Even though Social Security compensates for differences in retirement age on a roughly actuarially fair basis, those with high-time preference will consider these adjustments to be inadequate. This leads some to retire as soon as Social Security benefits first become available. A policy delaying the age of eligibility for early retirement benefits should move high discounters who accepted benefits as soon as available at age 62 to delay their retirement. This can be seen in the simulations in which the early entitlement age is increased from age 62 to age 64. Consider first Table 6.

Distribution of Time Preference Rates in Base Simulation.

Time preference rate Percent of households

0
5% 44.5

5
10% 12.6

10
25% 8.0

25
50% 1.9

>50% 33.0

103

Integrating Retirement Models

the husbands. As seen by comparing columns 1 and 2 in Table 7, most of the retirement spike formerly at age 62 moves to age 64. In the base simulation, husbands’ retirement at age 62 was about 8 percentage points higher than average retirement in the surrounding years. In the simulation with the increased early entitlement age, the spike at age 62 falls to less than 1 percentage point, while the spike at age 64 is about 6 percentage points higher than the average of the surrounding years. Retirement in other years is not greatly affected. The effect on wives is very similar, although perhaps not as pronounced because the age 62 retirement spike is not as pronounced for the wives. Overall, it appears that increasing the early entitlement age is relatively effective in moving most of the spike in retirement at age 62 to the new age of early entitlement. With regard to the sources of the increase in full-time work at 62 and 63 when the early entitlement age is raised to 64, some of the additional full-time workers come at the expense of part-time work, but most of them come from the full retirement category. A second policy simulation considers the effect of changing the availability of opportunities to partially retire on overall work effort. One argument made in favor of expanding the opportunities for part-time work is to allow older Americans to work into years when health or other issues might make it impossible or difficult to work full time. To the extent that Table 7.

Simulated Effect of Increasing Early Entitlement Age on Percent Pseudo Retiring from Full-Time Work.

Age

Percent Pseudo Retiring from Full-Time Work In Base Simulation ER Age 62

ER Age Increased to 64

In Base Simulation ER Age 62

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

3.5 2.5 4.3 4.2 5.7 7.2 6.4 14.7 6.9 6.8 7.9 5.3 3.9

ER Age Increased to 64

Wives 3.4 2.5 4.2 4.2 5.8 7.4 6.1 7.6 7.3 13.7 7.6 5.3 4.1

4.6 4.3 5.0 5.6 6.1 6.6 6.0 10.5 6.3 5.1 4.8 3.6 3.1

4.5 4.3 4.8 5.2 5.9 7.0 6.6 6.5 6.0 8.9 5.5 3.6 3.0

104

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

part-time work increases overall work effort, additional tax revenues improve the financial health of Social Security, and any delay in taking up Social Security improves retirement security in later years. To evaluate this argument, we compare the current situation to a situation where the option to work part time is effectively eliminated and the respondents are forced to choose only between full-time work and full retirement. The figures in this table are the differences between the percentage of individuals in the stated work/retirement category in this simulation and the comparable percentage of individuals in the base simulation. For instance, the first entry in the table, 1.9%, means that at age 55, the number of husbands working full time would be 1.9 percentage points higher if part-time work were eliminated as a possibility. Note that the sum of the two columns for each gender is the percent of individuals partially retired in Table 2. That is, of the 3.4% of husbands who were partially retired in the base simulation, if partial retirement were eliminated 1.9% of them would switch to full-time work and 1.5% of them would switch to full retirement. In general, eliminating partial retirement would lead to increases both in the number working full time, which would increase overall work effort, and in the number fully retired, which would reduce overall work effort. For husbands, the increase in full-time work outweighs the increase in full retirement up until about age 62, which means that up until that age overall work effort would increase if partial retirement were eliminated. After age 62, for the husbands eliminating partial retirement would cause the increase in full retirement to outweigh the increase in full-time work. This implies a reduction in overall work effort for this group. For the entire 55
67 age range illustrated in the table, the increase in full retirement outweighs the increase in full-time work, but only by a small amount. Making partial retirement more generally available or more attractive, of course, would have the opposite effect. Husbands would slightly increase their overall work effort as individuals continued to work in partial retirement beyond the date that they would otherwise retire. This effect, however, is largely offset because the increased availability of partial retirement induces individuals to leave full-time work for partial retirement earlier than they would have left full-time work in the absence of partial retirement. The pattern for the wives is similar to that for the husbands. Thus these simulations suggest that the effects of increasing partial retirement opportunities will be offset by the decline in full-time work, so that increasing partial retirement will not substantially increase overall work. This also means that increasing partial retirement opportunities will not help to improve Social Security’s finances (Table 8).

105

Integrating Retirement Models

Table 8. Age

Simulated Effect of Eliminating Partial Retirement. Change in Percent

In Full-Time Work

Change in Percent

Fully Retired

In Full-Time Work

1.5 1.7 2.0 1.9 2.5 3.2 3.7 6.9 8.3 10.0 11.4 12.7 14.1

3.5 4.2 5.3 6.3 7.3 8.0 8.0 6.4 6.1 5.1 4.4 3.6 2.6

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

1.9 2.4 3.4 4.2 5.0 6.1 7.0 7.8 7.9 7.3 6.9 5.9 4.9

Fully Retired

Wives 3.4 3.9 4.3 4.8 5.5 6.0 7.0 10.2 10.0 9.8 9.1 8.1 7.6

Another simulation investigates the effect of the major trend toward increased women’s labor force participation. To shed some light on this subject, we examine a simulation in which the wife is assumed to be out of the labor force and examine the impact on the husband’s labor force activity. Having the wife out of the labor force has two offsetting effects. First, the loss of income from the wife’s work leads to an increase in the marginal utility of income, which in turn should induce the husband to retire later. Offsetting this is the fact that, for some households, the fact that the wife is out of the labor force increases the value of leisure for the husband, at least if the husband values having the spouse present during retirement. This effect should induce the husbands to retire earlier. Table 9 presents the results of simulations investigating these offsetting forces. The figures in the top part of the table pertain to a simulation in which the husband’s leisure preferences are adjusted upward to the level they would be if the wife were always retired. This simulation in essence looks at the effect operating through the husband’s leisure preferences, and the figures in the table are differences between this simulation and the base simulation. The effect operating through the husband’s leisure preferences is substantial, amounting to a reduction in full-time work of around 8
9 percentage points in the late 1950s. The bottom half of the table gives the full effect on the husband’s work effort of having the wife out of the labor force, including the lost income. Excluding the wife from the labor force

106

Table 9.

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

Simulated Effect on Husbands Retirement Status of Eliminating Wife’s Labor Force Participation. Change in Percent

Age

In Full-Time Work

Partially Retired

Fully Retired

Effect through leisure preferences 55 −6.9 56 −7.7 57 −8.2 58 −8.6 59 −8.6 60 −8.3 61 −7.9 62 −6.7 63 −5.5 64 −4.5 65 −3.1 66 −2.4 67 −1.8

1.2 1.2 1.1 0.9 0.6 0.4 0.0 −1.2 −1.5 −1.6 −1.7 −1.5 −1.5

5.7 6.5 7.1 7.7 8.0 7.9 7.8 7.9 7.0 6.1 4.8 3.9 3.3

Total effect 55 56 57 58 59 60 61 62 63 64 65 66 67

−2.4 −2.5 −3.0 −2.8 −2.5 −2.1 −1.5 3.5 4.0 4.5 4.3 4.6 4.8

−5.5 −6.7 −8.1 −9.5 −10.3 −11.0 −11.9 −13.2 −12.9 −12.1 −10.4 −9.6 −8.9

7.8 9.3 11.1 12.4 12.8 13.1 13.3 9.7 8.9 7.6 6.0 5.0 4.1

leads to a substantial increase in full-time work by the husband, implying that the lost income effect much more than offsets the effect operating through the husband’s leisure preferences. The increased full-time work effort by the husband amounts to over 13 percentage points in the early 1960s, and the percentage of the sample being fully retired drops by a similar amount, again reaching around 13% in the early 1960s. During the 1950s and early 1960s, the percentage of the sample who are partially retired is reduced, but after that age partial retirement is increased by the exclusion of women from the labor force. To put it another way, in the

107

Integrating Retirement Models

early years full-time work is increased substantially at the expense of both partial retirement and full retirement, but in later years full retirement is reduced substantially and both full-time work and partial retirement are increased. In any case, it seems clear that all else being equal, the flow of wives into the labor force in the last few decades has probably reduced the amount of work that the husbands would have done otherwise. A final set of simulations aims to assess the degree to which various factors influence joint retirement. These results are contained in Table 10. The first column of simulated results simply repeats the central rows of the base simulation in Table 5 in which 13.5% of couples retire in the same year. The next column gives the results of a simulation in which ρɛ = 0, that is, in which the correlation between the retirement preferences ɛh and ɛw are constrained to be zero. Joint retirement falls in this case, but not by much, only to 12.2%. The following column omits the influence of having the spouse retired on retirement preferences by setting ηh = ηw = 1. The lack of the spouse retired influence has considerably more impact in this case, reducing joint retirement of 7.7%. The final column gives the results of both of these changes together, and the implication is that the spouse retired effect and the correlated retirement preference have a more or less additive effect, with joint retirement in this case at 7.2%. It is notable that even without any correlation between retirement preferences and without any influence of one spouse being retired on the value of retirement of the other spouse, there is still a notable spike of couples retiring together. The remaining effect is due to the nature of the strategic decision process as discussed in the section “The Decision Process.” Table 10.

Husband retires first

Wife retires first

Analysis of Sources of Joint Retirement.

Difference in Retirement Dates

Base Simulation

ρɛ = 0

ηh = ηw = 1

ηh = ηw = 1 and ρɛ = 0

5 4 3 2 1 0 1 2 3 4 5

3.5 3.7 3.6 3.6 3.5 13.5 2.2 2.8 2.9 3.2 3.5

3.2 3.4 3.3 3.2 3.1 12.2 2.0 2.6 2.7 3.0 3.3

2.8 2.9 2.7 2.6 2.4 7.7 1.8 2.4 2.7 2.9 3.3

2.5 2.5 2.4 2.3 2.1 7.2 1.6 2.2 2.4 2.7 3.1

108

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

CONCLUSIONS This paper has integrated many features of retirement models into a single framework. This has made it possible to utilize the full set of labor market information provided by the HRS, including survey responses, pension plan descriptions, and Social Security earnings data provided individually for husbands and wives. The integrated model is much richer than previously specified models of family retirement, allowing each spouse to retire and unretire, transitioning among the states of full-time work, partial retirement, and full retirement. It explains in much greater detail the effects of interdependence in the decisions made by each spouse, including clustering of retirements by husbands and wives, while at the same time allowing for forward looking behavior, explaining saving at the family level, incorporating the nonlinear budget constraints from still dominant defined benefit pensions and Social Security, and allowing exogenous shocks to asset returns. Increasing the richness of the model allows us to address phenomena that otherwise cannot be explained by conventional models of retirement. At the family level we are able to isolate the key role of heterogeneity in time preference, allowing the model to explain the wide differences in wealth accumulated by families with similar earnings opportunities. The retirement hazard exhibits the important spike in retirement at age 62 in the face of an actuarially fair Social Security system, captures the extent of partial retirement by each spouse, reproduces the flow from states of greater to lesser retirement, and relates each of the flows for one spouse to the decisions made by the other. We also allow each spouse to have heterogeneous preferences for both full-time and part-time work. The theoretical discussion increases understanding of the wide variety of situations that families face when approaching the retirement decision. It illustrates how choices focused on one spouse’s welfare take account of the welfare and independent reactions of their mate, incorporating the roles of different preferences and different market opportunities. Allowing for the variety of circumstances facing different families, the theoretical framework incorporates the many different situations facing different couples into a unified framework, and directly into the estimation. We also have shown why, when both spouses realize that their own choices may affect the choices of the other spouse, a solution method based on strategic decision making is superior to a method based on a Nash equilibrium, enabling our model to provide plausible behavioral predictions when the Nash equilibrium criteria fall silent.

Integrating Retirement Models

109

Simpler specifications cannot simultaneously explain heterogeneity in wealth; liquidity preference and retirement spikes when benefits become available; the high rate of return from states of lesser work to states of greater work despite the assumptions of forward looking behavior, and in the absence of changes either in preferences or in market opportunities; and the different sequence of retirements by husbands and wives. While preserving key findings, such as the conclusion from previous work that the husband’s utility of retirement is much more sensitive to the presence of the wife than the other way around, the integrated model can explain the key features of retirement outcomes that a simple specification of the family model fails to explain, including the ability to simulate with accuracy the spike in retirements at age 62. On the policy side, the model finds that increasing the early entitlement age under Social Security will have a substantial effect on retirements of both husbands and wives, moving the retirement hazard from age 62 to 64. Another simulation suggests that policies designed to promote partial retirement opportunities are not likely to succeed. Increasing partial retirement opportunities would lead to roughly offsetting effects, on the one hand reducing the number fully retired, which would increase overall work effort, while on the other hand reducing the number working full-time, which would decrease overall work effort. Another interesting result from counter factual experiments is our finding that the increased labor force participation of wives in the last few decades has probably caused a substantial reduction in the labor supply of their husbands. Although a great deal of effort is required to estimate and simulate with a structural model that incorporates many dimensions of labor market behavior, analyzes that behavior independently for each spouse, incorporates a full range of assets, and includes the sharply nonlinear incentives from pensions and Social Security, it appears to be worth the effort. One reward is an increased understanding of behavior, providing insight into many dimensions of retirement behavior, and behavior at the family level, that is not otherwise available from more simplified approaches. Another reward is a clearer picture of the likely effects of events, where simplified approaches either fall silent or, worse, may provide misleading guidance for policies. The implications of policies such as increasing the Social Security early entitlement age or changing opportunities for part-time work by the retirement age population simply cannot be understood when behavior must be analyzed with a less structural approach.

110

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

NOTES 1. Gustman and Steinmeier (1984) and Maestas (2010) describe the various flows among retirement states. Berkovec and Stern (1991) included reverse flows in their early structural analysis. 2. Couples coordinate their retirement despite age differences and differences in incentives to retire (Gustman & Steinmeier, 2000, 2004; Hurd, 1990; Maestas, 2001; Michaud, 2003). A general overview of bargaining within the family is provided by Lundberg (1999). Bourguignon, Browning, and Chiappori (2006) discuss decision making at the family level, including the inconsistencies between the data and parameters estimated for a unified model in which a single utility function is used to describe the consumption and labor supply of the family. Cherchye, De Rock, and Vermeulen (2009) provide a nonparametric analysis of the determinants of household consumption that focuses on the problems of a unitary model in analyzing household consumption. That study does not include an analysis of the determinants of labor supply. 3. It would be preferable to allow time preference to have separate values for each of the two spouses. However, allowing for separate time preference parameters for each spouse would introduce enough additional complexity to preclude our being able to estimate the model. 4. In an alternative formulation in which spouses do anticipate potential changes in the value leisure after retirement, many individuals with initially low values of leisure preference retire very early, and then return to work, in order to increase the probability that their later leisure would be more valuable. This behavior does not appear to be very realistic, and hence we reverted to the assumption that the spouses do not anticipate potential changes in the value of leisure after retirement. 5. Nominal pensions are assumed to increase by only 38% of the inflation rate. This implies that the real values decline by 62% of the inflation rate. 6. An unpublished appendix describes sample restrictions and the construction of the variables used in this study. The appendix is available from the authors on request. 7. Unpublished Appendix 2 explores issues encountered in estimating time preference from information on asset levels and retirement. The appendix is available from the authors on request. 8. If the observed moments do not match the moments generated by the model very well, the q value will be high and the probability value will be low, indicating that the model is not doing a very good job of explaining the observed data. The degrees of freedom for the χ2 distribution is equal to the number of moments minus the number of parameters estimated, which for the current estimation is 70. 9. Some studies, notably Keane and Wolpin (1997), use approximations to simplify the solution of the backward induction process. Their model assumes that each potential outcome choice has exactly one error term associated with it, an assumption that is crucial to their simplification. Our model, in contrast, has error terms that affect all of the outcome choices in relatively complex ways, so we cannot use Keane and Wolpin’s approach to simplify the calculations.

111

Integrating Retirement Models

10. Due to relatively small sample sizes at specific ages, the standard errors for observed “pseudo” retirements are approximately 3
4%. 11. For those with no assets, the exact time preference rate is indeterminate; we only know that it is sufficiently high that no accumulation takes place. We assume that they will continue to consume all their resources for the kinds of changes in their opportunity sets that are considered in this paper.

ACKNOWLEDGMENTS This paper was supported by a grant from the National Institute on Aging, 5 R01 AG024337, through the National Bureau of Economic Research. All opinions and views expressed herein are those of the authors and do not necessarily reflect the views or opinions of either the National Bureau of Economic Research or of the National Institute on Aging.

REFERENCES Berkovec, J. C., & Stern, S. (1991). Job exit behavior of older men. Econometrica, 59(1), 189
210. Bourguignon, F., Browning, M., & Chiappori, P.-A. (2006). Efficient Intra-household allocations and distribution factors: Implications and identification. Working Paper No. 2006
02. Centre for Applied Microeconometrics, Department of Economics, University of Copenhagen. Cherchye, L., De Rock, B., & Vermeulen, F. (2009). Opening the black box of intrahousehold decision making: Theory and nonparametric empirical tests of general collective consumption Models. Journal of Political Economy, 117(6), 1074
1104. Gustman, A. L., & Steinmeier, T. L. (1984). Modeling the retirement process for policy evaluation and research. Monthly Labor Review, 107(7), 26
33. Gustman, A. L., & Steinmeier, T. L. (2000). Retirement in a family context: A structural model for husbands and wives. Journal of Labor Economics, 18(3), 503
545. Gustman, A. L., & Steinmeier, T. L. (2004). Social security, pensions and retirement behavior within the family. Journal of Applied Econometrics, 19(6), 723
738. Gustman, A. L., & Steinmeier, T. L. (2009). Integrating retirement models. NBER Working Paper No. 15607. Cambridge, MA. Hurd, M. D. (1990). The joint retirement decision of husbands and wives. In D. A. Wise (Ed.), Issues in the economics of aging (pp. 231
254). Chicago, IL: University of Chicago Press. Ibbotson Associates. (2002). Valuation edition 2002 yearbook. Chicago, IL: Ibbotson Associates. Keane, M. P., & Wolpin, K. (1997). The career decisions of young men. Journal of Political Economy, 105(3), 473
522.

112

ALAN L. GUSTMAN AND THOMAS L. STEINMEIER

Lundberg, S. (1999). Family bargaining and retirement behavior. In H. Aaron (Ed.), Behavioral dimensions of retirement economics (pp. 253
272). New York, NY: Russell Sage/Brookings. Maestas, N. (2001). Labor, love & leisure: Complementarity and the timing of retirement by working couples. Berkeley, CA: Xerox. Maestas, N. (2010). Back to work: Expectations and realizations of work after retirement. Journal of Human Resources, 45(3), 718
748. Michaud, P.-C. (2003). Joint labour supply dynamics of older couples. IZA Discussion Paper 832.

THE ROLE OF DEGREE ATTAINMENT IN THE DIFFERENTIAL IMPACT OF JOB CORPS ON ADOLESCENTS AND YOUNG ADULTS Maria Bampasidou, Carlos A. Flores, Alfonso Flores-Lagunes and Daniel J. Parisian ABSTRACT Job Corps is the United State’s largest and most comprehensive training program for disadvantaged youth aged 16
24 years old. A randomized social experiment concluded that, on average, individuals benefited from the program in the form of higher weekly earnings and employment prospects. At the same time, “young adults” (ages 20
24) realized much higher impacts relative to “adolescents” (ages 16
19). Employing recent nonparametric bounds for causal mediation, we investigate whether these two groups’ disparate effects correspond to them benefiting differentially from distinct aspects of Job Corps, with a particular focus on the attainment of a degree (GED, high school, or vocational). We find that, for young adults, the part of the total effect of Job Corps on

Factors Affecting Worker Well-Being: The Impact of Change in the Labor Market Research in Labor Economics, Volume 40, 113
156 Copyright r 2014 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-912120140000040004

113

114

MARIA BAMPASIDOU ET AL.

earnings (employment) that is due to attaining a degree within the program is at most 41% (32%) of the total effect, whereas for adolescents that part can account for up to 87% (100%) of the total effect. We also find evidence that the magnitude of the part of the effect of Job Corps on the outcomes that works through components of Job Corps other than degree attainment (e.g., social skills, job placement, residential services) is likely higher for young adults than for adolescents. That those other components likely play a more important role for young adults has policy implications for more effectively servicing participants. More generally, our results illustrate how researchers can learn about particular mechanisms of an intervention. Keywords: Job Corps training program; degree attainment; causal mediation; nonparametric bounds JEL classifications: I38; I21; C14

INTRODUCTION The United States has a long tradition of implementing governmentsponsored education and training programs that have the goal of improving the labor market prospects of their participants. Most of the extant literature evaluating the effectiveness of those programs concentrate on adults, and typically have found mixed results on their effectiveness (see the reviews and references in Heckman, LaLonde, & Smith, 1999; LaLonde, 2003). A growing but smaller body of literature concentrates on the evaluation of education and training programs for youth (e.g., LaLonde, 2003; Schochet, Burghardt, & Glazerman, 2001). A good number of the observational studies concentrating on youth also report mixed results about the effectiveness of these programs (LaLonde, 2003; Mallar, Kerachslcy, Thornton, & Long, 1982). The possibility that youths do not exhibit returns from these programs is touted as disappointing, considering that low-income youth without a high school diploma and marketable skills have limited options for increasing their human capital and employability skills. However, recent evidence pertaining to the Job Corps (JC) training program for disadvantaged youth has documented positive effects on their labor market outcomes (e.g., Blanco, Flores, & Flores-Lagunes, 2013b; Chen & Flores, 2014; Flores-Lagunes, Gonzalez, & Neumann, 2010; Flores, Flores-Lagunes, Gonzalez, & Neumann, 2012; Lee, 2009; Schochet et al., 2001).1

Degree Attainment in the Differential Impact of Job Corps

115

JC is the largest program for disadvantaged youths between the ages of 16 and 24. JC is a comprehensive intervention that emphasizes human capital accumulation and market skills enhancement through academic and vocational instruction as well as supportive services such as job placement services, counseling, and health services, primarily within its residential centers. The typical participant attends the program for an average of eight months, resides in the training center, and receives an average of 1,100 hours of vocational and academic instruction, equivalent to about one year in high school (Schochet et al., 2001). The recent evidence on the effectiveness of JC is based on the United States Congress-mandated evaluation of the program: the National Job Corps Study (NJCS). The NJCS conducted a nationwide experimental evaluation and concentrated on assessing JC’s effectiveness across various dimensions, including participants’ labor market prospects, educational achievements, health outcomes, and criminal behavior. The findings of the NJCS highlighted the beneficial role of the program, reporting positive and statistically significant average effects four years after randomization on participants’ earnings (an additional $25.2 per week) and on the probability of being employed (a 3.3 percentage-point difference) (Schochet et al., 2001). Importantly, the NJCS reported that individuals randomly assigned to participate in JC (treatment group) observed a higher and statistically significant probability of attaining an academic or vocational degree (21%) relative to those assigned to the control group. The positive impacts of JC as reported in Schochet et al. (2001) have been attributed mostly to three main features: (i) program curriculum, (ii) residential component (which reduces the potentially adverse influence of the neighborhoods participants lived), and (iii) administration (since JC is administered at the federal level, its curriculum and services offered are relatively uniform across centers).2 The original NJCS findings in Schochet et al. (2001) also uncovered a number of important heterogeneities in the effectiveness of JC. One important heterogeneity was found along the age dimension: the impact of JC differed markedly between “adolescents” (aged 16
19) and “young adults” (aged 20
24).3 The latter group experienced about three times the effect on weekly earnings in quarter 16 after randomization relative to the former group ($38 vs. $13 per week).4 Schochet et al. (2008) documented that these different effects are consistent with the 10% higher length of enrollment and accompanying larger gains in human capital accumulation (through larger academic and vocational instruction) by the older group. However, Blanco, Flores, and Flores-Lagunes (2013a) estimate impacts of JC on

116

MARIA BAMPASIDOU ET AL.

participants’ wages, which presumably better reflect human capital accumulation (relative to weekly earnings), on these two groups. They find that adolescents exhibit positive and significant effects throughout the conditional distribution of their wages (quantile treatment effects), while young adults only exhibit positive and significant effects in the middle of their conditional distribution; and that the effects on wages are not as high as it would be predicted by a higher human capital accumulation by the older group. Their results suggest that a higher human capital accumulation (as conjectured by Schochet et al., 2008) may not be the whole story behind the disparate effects on earnings on these two groups. Given that age is an observed individual characteristic that can presumably be employed in better tailoring JC services to participants (relative to, for example, race), it is important to further analyze the factors that may be behind the disparate effects between those two age groups. The present work aims to further analyze this issue by concentrating on the beneficial role of attaining a high school, General Education Development (GED), or vocational degree within JC on participants’ future weekly earnings and employment, relative to other services provided by JC (e.g., counseling, job placement, and social skills training). In particular, we analyze what part of the average treatment effect (ATE) of randomization into JC on labor market outcomes is due to attaining such a degree, and what part of the ATE is due to all the other channels through which randomization into JC affects labor market outcomes. Following Flores and Flores-Lagunes (2009, 2010), we refer to these two effects, respectively, as the mechanism average treatment effect or MATE (where the mechanism is the degree attainment), and the net average treatment effect or NATE (where the effect is net of the degree-attainment mechanism). These two effects provide an intuitive decomposition of the ATE that enables learning about how randomization into JC affects labor market outcomes.5 By analyzing MATE and NATE for each age group of interest, we can learn whether the importance of attaining a degree as a mechanism for the effect of randomization into JC (as well as the importance of the bundle of all other channels through which random assignment to JC works) varies between them. If this were to be the case, light could be shed on whether the differential role played by degree attainment as a mechanism of the effect of randomization into JC between the two age groups is a contributing factor to the overall disparate impacts on their labor market outcomes. In addition, this analysis could inform administrators on better ways to serve these two age groups. Estimation of mechanism (also called indirect or mediator) effects is not straightforward, even in the presence of a randomly assigned treatment.

Degree Attainment in the Differential Impact of Job Corps

117

The reason is that the realization of the variable representing the mechanism occurs after the treatment and it is thus impacted by it. In our context, while randomization into JC guarantees that the characteristics of the treatment and control groups are balanced, the characteristics of individuals who attain a degree differ systematically from the characteristics of those who do not attain a degree (in part because degree attainment is affected by randomization into JC). Recently, there has been interest in the economics literature on methods to estimate causal mechanism effects (e.g., Flores & Flores-Lagunes, 2009; Huber, 2013).6 Most of the available methods to estimate MATE and NATE rely on the strong assumption that, conditional on observables, the value of the mechanism variable is as if it had been randomly assigned. This assumption is very strong in our context because of the high selection that exists into who attains a degree. To avoid imposing this potentially untenable assumption, we follow an alternative strategy and partially identify (bound) the MATE and NATE of attaining a degree (relative to all other mechanisms) on labor market outcomes under (presumably) more tenable assumptions. More specifically, we employ the bounds on MATE and NATE proposed by Flores and Flores-Lagunes (2010), which rely on weak-monotonicity assumptions. Our approach illustrates the existing trade-off between the plausibility of the assumptions and how much we can learn about our parameters of interest from the available data: we employ methods that rely on weaker assumptions at the cost of forgoing estimating a single value of the effect (i.e., point identification) and estimating instead a range of values where the true effect lies (i.e., bounds). Our results highlight in general the importance of the remedial education and vocational training offered within JC. The estimated bounds for the sample combining the two age groups suggest that obtaining a degree accounts for up to 80% (100%) of the total effect of randomization into JC on earnings (employment). Accounting for sampling error, however, we cannot rule out that attaining a degree accounts for the full effect. These estimates mask the heterogeneity present for the two age groups of interest. For adolescents, we cannot rule out that attaining a degree accounts for all of the effect on the labor market outcomes (i.e., MATE = ATE). For the group of young adults, the role played by channels other than the attainment of a degree (i.e., NATE) is larger than zero. By implication, for this group, attaining a degree (i.e., MATE) cannot account for the total effect of randomization into JC on both earnings and employment (i.e., MATE < ATE). We also find suggestive
but not statistically significant
evidence that the magnitude of the mediating effect of channels

118

MARIA BAMPASIDOU ET AL.

other than degree attainment on labor market outcomes (NATE) is larger for young adults than for adolescents. Finally, based on other estimates within our econometric framework, we offer a characterization of these two age groups of youths that could aid in improving JC participants’ tracking and the menu of services offered to them. These results provide novel information about the program effect heterogeneity within youths, a group that has been relatively scantly evaluated in the literature. The remainder of the paper is organized as follows. The next section provides a brief description of JC and the NJCS, and presents some descriptive statistics for the data employed. In the section “Econometric Framework,” we define our parameters of interest and describe the econometric methods employed. The section following this presents and discusses the results. The last section concludes.

THE JOB CORPS PROGRAM AND DATA Job Corps JC was established by the Economic Opportunity Act of 1964. It operates under the provisions of the Workforce Investment Act of 1998 and is administered by the Department of Labor through a national office and nine regional offices. Every year, JC accepts about 60,000 new participants in its approximately 120 centers located across the United States (in both urban and rural areas), with an average $14,000 cost per participant (McConnell & Glazerman, 2001).7 Applicants must meet the following criteria in order to be considered eligible for JC (Schochet, 1998): (1) be of age 16
24; (2) have registered with the selective service board if aged 18 or older; (3) have parental consent; (4) be a legal U.S. resident; (5) be economically disadvantaged;8 (6) need additional education, training, or job skills; (7) live in a disruptive environment; (8) have a clean health history; (9) be free of serious behavioral problems; (10) have an adequate child care plan; and, (11) possess the capability and aspirations to benefit from JC. These criteria are assessed by counselors within the Outreach and Admissions recruiting offices. The majority of JC centers are operated by private contractors, and around one-quarter are operated by the U.S. Department of Agriculture and U.S. Department of Interior. Center operations involve vocational training, academic education, residential living (nonresidential participants

Degree Attainment in the Differential Impact of Job Corps

119

are only about 12%), health care, and supportive services such as counseling, social skills training, health education, and recreation. JC provides an intensive education curriculum offering both academic classroom instruction and vocational skills training
usually leading to degrees or certificates (our mechanism of interest). Academic education emphasizes remedial education (reading, math, and writing skills) and a GED program of high school equivalency. Vocational training areas may vary by center but typically include, among others, business and clerical, health, culinary arts and cosmetology, construction, and building and apartment maintenance. The average duration of the program is eight months (the maximum is 24 months), and is characterized by an open-exit educational philosophy where instruction is individualized and self-paced. Lastly, placement agencies within JC help participants find jobs in occupations related to their training by providing assistance with resume writing and interviewing, as well as services for job placement and referral. Usually, placement activities are performed by state employment offices, private contractors, and sometimes by the operational centers. Placement agencies are also responsible of the task of distributing a stipend (as per the JC program) that students receive after leaving JC.

National Job Corps Study The U.S. Department of Labor funded the NJCS in 1993, which was conducted by Mathematica Policy Research, Inc. (MPR). Designed to evaluate the effectiveness of the program, the NJCS was the first nationally representative experimental evaluation of a government-sponsored education and training program for disadvantaged youths (Schochet et al., 2008). The sample intake occurred between November 1994 and February 1996, in which the total eligible pool of applicants (N = 80,883) were randomly assignment into a control group (N = 5,977), a treatment group (N = 9,409), and a non-research group (Schochet et al., 2001). Individuals assigned to the treatment group were eligible to enroll in JC, while individuals assigned to the control group were denied access to the program for three consecutive years (they were eligible to apply to other training or educational programs, though). Upon notification of random assignment from MPR, Outreach and Admissions agencies assigned individuals to a center within a month’s period, and the individuals who enrolled in the centers typically did so within one to four weeks after random assignment.

120

MARIA BAMPASIDOU ET AL.

Given that random assignment occurred after the youths were determined eligible to participate in JC and not after they enrolled, there exists noncompliance. That is, the treatment group includes both youths that enrolled in the training program (about 73%) and those that decided not to enroll although they were eligible; while among those in the control group around 1.4% of individuals participated in JC prior to the end of the threeyear embargo period. In our analysis below, we will concentrate on intention-to-treat effects that represent the effect of the availability of the JC program (i.e., random assignment) on future labor market outcomes (see note 5). Following random assignment, a baseline interview was conducted for the 15,386 individuals in the NJCS sample and follow-up interviews took place at three subsequent time periods: 12, 30, and 48 months.9 The main results in the original NJCS report (Schochet et al., 2001), based on differences-in-means estimates that accounted for noncompliance as in Angrist, Imbens, and Rubin (1996), documented significant positive effects of JC at the beginning of the third year after randomization, which persisted through the end of the 48-month follow-up period. Specifically, NJCS reported that JC generated positive earning impacts of around $25.2 on weekly earnings 16 quarters after randomization, and around 3.3 percentage points on employment probability. These effects represent the average effects of JC for the individuals who complied with their treatment assignment. The same report also analyzed and documented the effects of JC on the earnings and employment rate for different subgroups. For the current paper, the analysis by age groups is of special interest. Schochet et al. (2001) documented significantly higher effects for older youth (aged 20
24) relative to younger youth (aged either 16
17 or 18
19).10 For instance, in terms of yearly earnings, the effects on older youth were more than twice those of any of the other two younger groups. Schochet et al. (2008) attributed the different effects of JC on the age groups to the older group’s larger gains in academic and vocational training, which is consistent with a higher human capital accumulation relative to the younger group. They also mentioned that older participants appeared more highly motivated and well behaved. The original NJCS report also documents the impact of JC on educational achievements (high school and GED diploma attainment) and training outcomes (vocational degree attainment). This is relevant since JC serves primarily youths with no high school credential; specifically, around 80% of the participants do not have a high school diploma or GED credential prior to applying to the JC program. Schochet et al. (2001) reported notable differences between the control and treatment groups both in terms

Degree Attainment in the Differential Impact of Job Corps

121

of participation rates in education and training programs, and in the number of certificates awarded. These differences are similar by age groups (computed using our sample described below) in some respects, but not in others. Particularly, nearly 93% of the treatment group engaged in education or training compared to 72% of the control group (as mentioned earlier, the embargo from JC did not imply an embargo from other training or education programs). The corresponding statistics are 95% (91%) for the adolescents’ (young adults’) treatment group and 79% (69%) for the adolescents’ (young adults’) control group. Given these differences, educational attainment is a plausible mechanism through which randomization into JC affected the labor market outcomes of both adolescents and young adults. In terms of the kinds of degrees attained, 35% of those in the control group earned a high school or GED degree, while 49% of those in the treatment group did; while 18% of those in the control group earned vocational certificates compared to 44% of those in the treatment group. However, there is variation by age group in the kind of degree attained. In the control group, 37% of adolescents and 26% of young adults earned high school or GED degrees; while in the treatment group 50% of adolescents and 45% of young adults earned the same degrees. In contrast, in terms of attainment of vocational degrees, in the control group 15% of adolescents and 33% of young adults earned them; while in the treatment group 40% of adolescents and 58% of young adults attained vocational degrees. Therefore, there is evidence that adolescents tend to earn high school and GED degrees at a higher rate than vocational degrees, relative to the group of young adults.

Data and Descriptive Statistics We employ public-use data from the NJCS. Our two outcome measures are weekly earnings and employment during quarter 16 after random assignment. Our mechanism variable of interest is attaining a degree, represented by a binary variable that takes the value of one if the individual attained a high school, GED, or vocational certificate between random assignment and the time of the last follow-up interview, and zero otherwise. We use this mechanism variable to analyze the causal importance of earning a degree in mediating the effect of being allowed to enroll in JC (our treatment variable) on participants’ future labor market outcomes. Our sample consists of individuals with no missing values on key baseline variables such as age, the outcomes of interest during quarter 16, and the degree

122

MARIA BAMPASIDOU ET AL.

attainment indicator. The resulting sample contains 8,020 individuals, of which 5,045 were randomly assigned to the treatment group and 2,975 were assigned to the control group.11 From the 8,020 individuals in our combined sample, 6,295 are adolescents and 1,725 are young adults. Table 1 presents the average of selected baseline characteristics for the combined sample and our two age groups, as well as the difference between Table 1.

Summary Statistics of Selected Baseline Variables.

Variable

Selected demographic variables Female

Combined Adolescents

Young Adults

Difference Young Adults − Adolescents

0.433 (0.006) 18.480 (0.023) 0.250 (0.005) 0.502 (0.006) 0.772 (0.004) 0.056 (0.003) 0.184 (0.004)

0.422 (0.006) 17.637 (0.014) 0.256 (0.006) 0.503 (0.006) 0.174 (0.005) 0.040 (0.002) 0.129 (0.004)

0.479 (0.012) 21.738 (0.032) 0.226 (0.010) 0.495 (0.012) 0.188 (0.009) 0.116 (0.008) 0.397 (0.012)

0.057*** (0.014) 4.101**** (0.035) −0.030*** (0.011) −0.009 (0.014) 0.014 (0.011) 0.076*** (0.008) 0.268*** (0.013)

Selected education and crime variables HS/GED/VOC degree at 0.112 baseline (0.004) Highest grade completed at 9.782 baseline (0.016) Ever been arrested by 0.272 baseline (0.005)

0.062 (0.003) 9.562 (0.016) 0.285 (0.006)

0.308 (0.011) 10.634 (0.036) 0.223 (0.010)

0.246*** (0.012) 1.073*** (0.040) −0.061*** (0.012)

0.192 (0.004) 102.996 (1.237)

0.184 (0.005) 92.999 (1.326)

0.222 (0.010) 141.636 (2.958)

0.039*** (0.011) 48.637*** (3.241)

8,020

6,295

1,725

Age White Black Hispanic Married Has child/children

Selected labor market variables Employed at baseline Weekly earnings at baseline Observations

Note: Standard errors in parenthesis. *** denotes statistical significance at the 99% confidence level.

Degree Attainment in the Differential Impact of Job Corps

123

the average baseline characteristics of young adults and adolescents. The baseline variables include demographic characteristics, education and crime-related variables, and employment and earnings information. The first column shows that the average JC applicant is an African-American male about 18.5 years old, is not married and has no children, is not a high school graduate (only 11% are), is not employed, and has about $100 in weekly earnings. Additionally, about 27% of eligible applicants have been arrested by baseline. Table 1 shows notable and statistically significant differences between the average baseline characteristics of adolescents and young adults. As compared to adolescents, young adults are less likely to be male, white, ever been arrested by baseline, and be non-employed at baseline. In addition, young adults are considerably more likely to be married, have children, and have about 50% higher weekly earnings than adolescents. In terms of degree attainment, young adults have earned a high school, GED, or vocational degree (by baseline) at five times the rate of adolescents; they also have, on average, a little more than a full year of schooling.

ECONOMETRIC FRAMEWORK Estimands of Interest Our goal is to analyze and compare, for our two age groups of interest, the part of the effect of random assignment into JC on earnings and employment that works through the attainment of a degree, as well as the part of the effect that works through all other channels (job search assistance, counseling, health services, etc.). In this section, we define causal parameters of interest that allow for such decomposition of the effect of interest. In our setting, random assignment into JC is the treatment (T). We denote with S the binary indicator for degree attainment (within JC or not), which is the mechanism of interest through which T impacts the labor market outcomes (denoted by Y). We rely on the potential outcome framework (Neyman, 1923; Rubin, 1974) to define our estimands of interest. Each individual i is assigned to a treatment group (Ti = 1) or to a control group (Ti = 0). Since the mechanism variable S is affected by the treatment, we denote its potential values by Si (t), where t ∈ {0,1}. Thus, the value Si (1) is the potential value of the degree indicator for individual i if she were assigned to the treatment

124

MARIA BAMPASIDOU ET AL.

group, and Si (0) is the potential value of the degree indicator if she were assigned to the control group. Define the “composite” potential outcome Yi (t,s), where t corresponds to one of the treatment values (t ∈ {0,1}) and s refers to one of the potential values of the mechanism variable (s ∈ {Si (1), Si (0)}). Note that there are four composite potential outcomes: Yi (1,Si (1)) ≡ Yi (1) (outcome under exposure to the treatment); Yi (0, Si (0)) ≡ Yi (0) (outcome under no exposure to the treatment);12 Yi (0,Si (1)) (outcome under no exposure to the treatment but having the value of the mechanism variable individual i would have under treatment, Si (1)); and, Yi (1,Si (0)) (outcome under exposure to the treatment but “blocking” the effect of the mechanism variable by keeping it fixed at its value under control, Si (0)). The last two potential outcomes are, in principle, never observed in the data, and thus are entirely counterfactual or hypothetical (see, e.g., Rubin, 1990, and the related discussions in Flores & Flores-Lagunes, 2010, 2013). To define ATEs under this framework, we adopt the customary stable unit treatment value assumption, which implies that there is no interference between individuals and that there are no different versions of the treatment (Rubin, 1980, 1990). Therefore, we can define the population average treatment effect as ATE = E [Y(1) − Y(0)], where we omit the subscript i hereafter unless necessary. Using the potential outcome Y(1,S(0)) defined above, which represents the outcome individual i would have if assigned to T but net of the degree-attainment mechanism (since treatment is received but degree attainment is held fixed at its value under control), we can formally decompose ATE into a mechanism average treatment effect or MATE (where the mechanism is the degree attainment), and a net average treatment effect or NATE (net of the degree attainment mechanism), as:13 ATE = E½Yð1Þ − Yð1;Sð0ÞÞ þ E½Yð1;Sð0ÞÞ − Yð0Þ

ð1Þ

where MATE is defined as MATE = E½Yð1Þ − Yð1;Sð0ÞÞ

ð2Þ

NATE = E½Yð1;Sð0ÞÞ − Yð0Þ

ð3Þ

and NATE as

This decomposition is not new: NATE and MATE are also known as the average pure direct and indirect effects (Robins & Greenland, 1992), or as the average natural direct and indirect effects (Pearl, 2001). MATE

Degree Attainment in the Differential Impact of Job Corps

125

captures the part of the effect of T on the outcome Y that is due to attaining a degree because of T, since S changes from S(0) to S(1). All other ways T may affect the outcome are held constant because the value of the treatment is fixed at 1. Conversely, NATE captures the part of the effect of T on the outcome that is due to all other possible channels of T, since degree attainment is held constant at the level S(0). In other words, the average effect of all other services that are provided when the JC program is available, apart from degree attainment, are captured by NATE. The definitions of these estimands imply that when all the effect of the treatment on the outcome works through the mechanism (degree attainment), then MATE = ATE, and when none of the effect works through the mechanism then MATE = 0. Therefore, by learning about MATE and NATE for each one of the two age groups, we can gauge whether the differential effects of JC on the outcomes (ATEs). For these two groups seem to be linked to the relative importance of degree attainment (vs. other services or channels) in mediating their corresponding effects. The observed data consists, for each person, of realized values of the random assignment (T), degree attainment (S), and labor market outcomes (Y). The last two realizations are equal to S = TS(1) + (1 − T)S(0) and Y = TY (1) + (1 − T)Y(0). While randomization of T allows point identification of the ATE, it does not allow point identification of NATE and MATE in a straightforward way because the realized value of S is likely affected by T. Intuitively, random assignment into JC does not imply random assignment into degree attainment. Hence, individuals who attain a degree and those who do not attain a degree differ systematically in observed and unobserved characteristics, implying that the simple difference of average outcomes by degree attainment status does not yield a causal effect. The difficulty in estimating mediation effects (such as MATE and NATE) has been recognized before in the literature (e.g., Robins & Greenland, 1992; Rubin, 2004). Most available methods to point identify MATE and NATE rely on strong unconfoundedness assumptions that are hard to justify in our empirical setting, such as the mechanism variable being exogenous after conditioning on observed individual characteristics (e.g., Flores & Flores-Lagunes, 2009; Huber, 2013; Imai et al., 2010; Pearl, 2001; Petersen et al., 2006; Robins & Greenland, 1992). In this paper, we abstract from point identification and its corresponding strong unconfoundedness assumptions, and instead concentrate on estimating nonparametric bounds on the estimands of interest by imposing more plausible assumptions. In particular, we employ the nonparametric bounds on MATE and NATE proposed by Flores and Flores-Lagunes

126

MARIA BAMPASIDOU ET AL.

(2010, hereafter FF-L).14 Advantages of these bounds are that they allow unrestricted heterogenous treatment effects, are nonparametric in nature, do not require an outcome with bounded support, and employ relatively mild assumptions in the form of weak-monotonicity restrictions on potential outcomes within or across particular subpopulations. In the following subsection, we state the assumptions and discuss their intuition and plausibility in our context.

Nonparametric Bounds on the Estimands of Interest Consider our binary treatment (random assignment to JC) and binary mechanism (degree attainment indicator), and partition individuals into groups such that, within each group, all individuals have the same vector of potential values of the degree attainment indicator {S(0) = s0, S(1) = s1} with s0, s1 ∈ {0,1}. These groups are called principal strata in the terminology of Frangakis and Rubin (2002). Given that within principal strata the treatment assignment does not affect the mechanism, we can define valid causal effects within strata (called principal effects) as the comparison of mean potential outcomes within strata. In the present context, there are four principal strata:15 {S i (0) = 0, Si (1) = 0}, {Si (0) = 0, Si (1) = 1}, {Si (0) = 1, Si (1) = 0}, {Si (0) = 1, Si (1) = 1}. These strata define the following categories of individuals: the not affected at 0 (n0), the affected positively (ap), the affected negatively (an), and the not affected at 1 (n1), respectively. The ap stratum is composed of individuals who would attain a degree only if assigned to participate in JC, while the individuals in the an stratum attain a degree only if assigned to the control group. Individuals in the n1 (n0) stratum would always (never) attain a degree whether or not they are assigned to participate in JC. The intuition behind the general approach followed by FF-L (2010) to bound MATE and NATE is to first bound the mechanism and net average effects of each one of these strata, and then combine these stratumlevel bounds to construct bounds for the population MATE and NATE. Define “local” estimands for the mechanism average effect for a given stratum as: LMATEk = E[Y(1,S(1)) − Y(1,S(0))|k], for k = {n0,n1,ap,an}, each of which has a causal interpretation. Thus, MATE can be written as: MATE = πn0LMATEn0 +πn1LMATEn1 + πapLMATEap + πanLMATEan, where πk represents the population proportion of each stratum k. Similarly, we define the local net average effect of each stratum as: LNATEk =E[Y(1,S(0)) − Y(0,S(0))|k], for k = {n0, n1, ap, an}, and

127

Degree Attainment in the Differential Impact of Job Corps

write NATE as: NATE = πn0LNATEn0 +πn1LNATEn1 + πapLNATEap + πanLNATEan.16 To derive nonparametric bounds on MATE and NATE, FF-L (2010) adopt a number of assumptions in order to point or partially identify the various expectations that appear in the definition of the LMATEs and LNATEs. Those assumptions address two fundamental problems in the identification of MATE and NATE. First, strata membership is unobserved. For each individual we observe the realized treatment (T) and the realized value of the degree attainment (S). The relation between the strata and the observed T and S is given by the following table: Table 2.

Principal Strata and Observed Groups. T

S

0 1

0

1

ap, n0 n1, an

an, n0 n1, ap

Table 2 illustrates that it is not possible to determine what stratum each individual belongs to because each cell in the table is a mixture of two strata. The second fundamental problem to identify MATE and NATE is that the data contain no information on the potential outcome Y(1,S(0)) for the ap and an strata, since this is an entirely counterfactual outcome (see, e.g., FF-L, 2010).17 The next two subsections discuss the assumptions considered in FF-L (2010) to address these two fundamental problems and bound MATE and NATE. The section on basic assumptions addresses the problem that strata membership is unobserved. These assumptions one if which rules out the an stratum allow point identification of the stratum proportions, and point or partial identification of all the terms in the LMATEs and LNATEs of the relevant strata, except for the term E[Y(1,S(0))|ap]. The assumptions added in the section “Weak-Monotonicity Assumptions on Average Potential Outcomes” allow partial identification of E[Y(1,S(0))|ap] and thus MATE and NATE, and also tighten the stratum-level bounds discussed in the section “Basic Assumptions.” Basic Assumptions The first assumption is that the treatment is randomly assigned, which is innocuous in our context given that T is the random assignment into JC in

128

MARIA BAMPASIDOU ET AL.

the NJCS. It can be written as below, which showcases the implication that T is independent of potential outcomes and potential values of the mechanism: Assumption 1. (Random Treatment Assignment) Yð1Þ;Yð0Þ;Yð1;Sð0ÞÞ;Sð1Þ;Sð0Þ⊥T Assumption 1 allows point identification of expectations E[Y(1)], E[Y(0)], E[S(1)], and E[S(0)], and thus the point identification of ATE, but not of MATE (or NATE) since E[Y(1,S(0))] is not point identified. The second assumption is a monotonicity assumption on the individuallevel effect of the treatment on the mechanism variable. Monotonicity assumptions have been widely used in the causal inference literature in different contexts (e.g., Angrist et al., 1996; Bhattacharya, Shaikh, & Vytlacil, 2008; Chen & Flores, 2014; Imbens & Angrist, 1994; Lechner & Melly, 2010; Lee, 2009; Manski & Pepper, 2000; Zhang, Rubin, & Mealli, 2008). Assumption 2. (Individual-Level Monotonicity of T on S) Si ð1Þ ≥ Si ð0Þ;∀i Assumption 2 postulates that the effect of the treatment (randomization into JC) on the mechanism variable (degree attainment) is non-negative for every individual, ruling out the an stratum. This is a potentially strong assumption that excludes individuals who would not acquire a degree if randomized into JC but would if randomized out of JC. Intuitively, knowing that JC facilitates the attainment of a degree through the academic and vocational training offered to participants makes this assumption plausible.18 Using Assumption 1 and Assumption 2, the population proportions of each stratum can be point identified. Denoting ps|t ≡ Pr(S = s|T = t) we have that (refer to Table 2 after deleting the an stratum), πan = 0, πn0 = p0|1, πn1 = p1|0, and πap = p1|1 − p1|0 = p0|0 − p0|1. The combination of the same assumptions also allows the point identification of E[Y(1)|n0] = E[Y|T = 1, S = 0] and E[Y(0)|n1] = E[Y|T = 0,S = 1]. Moreover, under these two assumptions LNATEn0 and LNATEn1 are partially identified (bounded). While neither E[Y(0)|n0] nor E[Y(1)|n1] are point identified because the observed groups of individuals with (T, S) = (0, 0) and (T, S) = (1, 1) are

Degree Attainment in the Differential Impact of Job Corps

129

still a mixture of two strata, those expectations can be bounded employing “trimming” bounds (e.g., Blanco et al., 2013b; Chen & Flores, 2014; FF-L, 2010; Lee, 2009; Zhang et al., 2008). To illustrate how these trimming bounds are constructed, take the group of individuals with (T,S) = (0,0), whose average outcome is given by (see Table 2): E½YjT = 0; S = 0 =

π n0 π ap E½Y ð0Þjn0 þ E½Yð0Þjap π n0 þ π ap π n0 þ π ap

ð4Þ

Given that all πs are identified, E[Y(0)|n0] can be bounded from above (below) by the expected value of Y for the (πn0)/(πn0 + πap) fraction of the largest (lowest) values of Y for those in the group (T,S) = (0,0). Similar application of trimming bounds provide partial identification of E[Y(0)|ap]; and of E[Y(1)|n1] and E[Y(1)|ap] using the observed group (T,S) = (1,1). We summarize in the following proposition the partial identification results under the current two assumptions. We do not list the actual expressions of the upper and lower bounds on the different estimands to save space, but instead we refer the reader to FF-L (2010), where the proofs are also presented. For ease of reference, the proposition numbers correspond to those in FF-L (2010). Proposition 1. If Assumptions 1 and 2 hold, then nonparametric bounds can be constructed on LNATEn0, LNATEn1, E[Y(0)|ap], and E[Y(1)|ap]. (See FF-L (2010) for the expressions of the bounds). Note that thus far partial identification of MATE and NATE is not yet possible. To see why, note that, under the current assumptions, MATE = πapLMATEap = πap{E[Y(1)|ap] − E[Y(1,S(0))|ap]}.19 E[Y(1)|ap] is partially identified in Proposition 1, but the observed data do not contain information on E[Y(1,S(0))|ap]. This latter expectation is also part of LNATEap and thus of NATE. Therefore, without additional assumptions, it is not possible to partially identify MATE and NATE. Weak-Monotonicity Assumptions on Average Potential Outcomes We next discuss two sets of weak-monotonicity assumptions that allow partial identification of MATE and NATE. The main intuition behind these assumptions is that they relate the unidentified term E[Y(1,S(0))|ap] to other average potential outcomes that are point or partially identified under Assumptions 1 and 2. The first set of assumptions accomplishes

130

MARIA BAMPASIDOU ET AL.

this goal by postulating weak monotonicity of average potential outcomes within strata: Assumption 3. (Weak Monotonicity of Average Potential Outcomes Within Strata) (i) E[Y(1,S(1))|ap] ≥ E[Y(1,S(0))|ap], (ii) E½Yð1;Sð0ÞÞjk ≥ E½Yð0Þjk, for k = n0; n1; ap. It is clear that Assumption 3 directly provides upper and lower bounds on E[Y(1,S(0))|ap]. In particular, Assumption 3(i) implies that LMATEap ≥ 0 and thus MATE ≥ 0; whereas the combination of Assumption 2 and Assumption 3(i) implies that the attainment of a degree has a non-negative average effect on future employment and earnings. The latter implication is supported by theoretical models of human capital. Assumption 3(ii) provides a lower bound for NATE equal to zero and implies that the combination of all other mechanisms (such as job-assistance and social skills training) has a non-negative average effect on future employment and earnings. Since JC is a program that provides a bundle of services aiming at improving future labor market outcomes, we argue that this assumption is likely to be satisfied. Importantly, we emphasize that Assumption 3 pertains to average outcomes, which permits that certain individuals do not satisfy the weak-monotonicity conditions, as long as the average over individuals belonging to the corresponding strata do.20 The following proposition states that bounds can be constructed on all estimands of interest when Assumption 3 is employed. Proposition 2. If Assumptions 1, 2, and 3 hold, then nonparametric bounds can be constructed on LNATEn0, LNATEn1, LNATEap, LMATEap, NATE, and MATE (See FF-L (2010) for the expressions of the bounds). In particular, under Proposition 2, MATE is bounded from below by zero and, depending on how much information is contained in the data, it can be bounded from above by a quantity that is strictly smaller than the ATE of the treatment on the outcome. The second set of assumptions relates E[Y(1,S(0))|ap] to other point or partially identified average potential outcomes by postulating weak monotonicity of average potential outcomes across strata. Intuitively, this set of assumptions formalizes the notion that some strata are likely to have more

Degree Attainment in the Differential Impact of Job Corps

131

favorable characteristics than others, resulting in them exhibiting higher average potential outcomes: Assumption 4. (Weak Monotonicity of Average Potential Outcomes Across Strata) (i) E[Y(1,S(0))|ap] ≥ E[Y(1)|n0], (ii) E[Y(1)|n1] ≥ E[Y(1,S(0))|ap], (iii) E[Y(0)|ap)] ≥ E[Y(0)|n0], (iv) E[Y(0)|n1] ≥ E[Y(0)|ap], (v) E[Y(1)| ap] ≥ E[Y(1)|n0], and (vi) E[Y(1)|n1] ≥ E[Y(1)|ap]. Clearly, Assumptions 4(i) and 4(ii) are sufficient for bounding the term E[Y(1,S(0))|ap] and thus all the estimands of interest. Assumptions 4(iii) to 4(vi) serve the purpose of further tightening the bounds on the estimands of interest. In addition, combining Assumptions 1, 2, and 4 results in two testable implications that can be used to “falsify” the maintained assumptions:21 (i) E[Y|T = 0, S = 1] ≥ E[Y|T = 0, S = 0], and (ii) E[Y|T = 1, S = 1] ≥ E[Y|T = 1, S = 0]. In our case, these inequalities state that, within each treatment group, the average observed outcome of those individuals who attained a degree is no less than that of those individuals who do not attain a degree. As documented in the next section, these two testable implications are satisfied for each of the three samples under study. In our context, we postulate that individuals belonging in the (n1) stratum, i.e. individuals who always attain a degree irrespective of their treatment assignment to JC, are likely to, on average, fare better in the labor market under both treatment arms relative to individuals in the (ap) stratum, who attain a degree only if they are randomly assigned to enter JC. In turn, we postulate that individuals in the (ap) stratum perform better in the labor market under both treatment arms, on average, relative to individuals in the n0 stratum, which is comprised of individuals who irrespective of their treatment assignment do not obtain a degree. As with Assumption 3, the focus on average potential outcomes by strata makes Assumption 4 weaker relative to an analogous assumption that were to focus on individual-level potential outcomes. It is possible to obtain indirect evidence about the plausibility of Assumption 4 by comparing relevant average baseline characteristics (e.g., pre-treatment outcomes) of the different strata (e.g., FF-L, 2010, 2013). If these comparisons contradict the ranking of average labor market outcomes by strata postulated above, then Assumption 4 is less likely to hold. Table 3 compares several average baseline characteristics of the three strata for the combined sample, and for each one of our two age groups.

132

MARIA BAMPASIDOU ET AL.

The last two rows of each panel of Table 3 show, for each stratum (first three columns), the average of the pre-treatment outcomes (weekly earnings and employment at baseline), which are likely to be highly correlated with our actual outcomes measured during quarter 16 after random assignment. The last three columns present tests of differences between the strata averages. For the three samples considered, we obtain that when comparing the n1 and ap strata to the n0 stratum, the former two strata have higher weekly earnings and are more likely to be employed at baseline than the n0 stratum, with most differences being statistically significant. When comparing the mean of the pre-treatment outcomes at baseline for the n1 and ap strata, we observe that most of the differences are relatively close to each other and none is statistically significant. Thus, we conclude that the data do not provide indirect evidence against Assumption 4.22 The following proposition states that bounds can be constructed on all estimands of interest when Assumption 4 is combined with Assumptions 1 and 2: Proposition 3. If Assumptions 1, 2, and 4 hold, then nonparametric bounds can be constructed on LNATEn0, LNATEn1, LNATEap, LMATEap, NATE, and MATE (See FF-L (2010) for the expressions of the bounds). Finally, the combination of all the previous assumptions yields tighter bounds, and also an additional testable implication: E[Y|T = 1,S = 1] ≥ E[Y| T = 0, S = 0], which states that the average observed outcome of individuals who were randomly assigned to participate in JC and attained a degree is no less than that of those control individuals who did not attain a degree. This testable implication is satisfied by each of the three samples under study, as will be discussed in the next section. The following proposition completes the sets of nonparametric bounds that will be employed in our empirical analysis: Proposition 4. If Assumptions 1, 2, 3, and 4 hold, then nonparametric bounds can be constructed on LNATEn0, LNATEn1, LNATEap, LMATEap, NATE, and MATE. Additionally, these bounds are tighter than those derived in Proposition 3 (See FF-L (2010) for the expressions of the bounds). Estimation and Statistical Inference for the Nonparametric Bounds The bounds discussed above involve minimum (min) and maximum (max) operators,23 which is problematic for standard estimation and inference procedures. Plug-in estimates of this type of bounds tend to be narrower

Degree Attainment in the Differential Impact of Job Corps

133

than the true bounds because of the concavity and convexity of the min and max functions, respectively, and the asymptotic distribution of the bound estimators is usually unavailable. Moreover, Hirano and Porter (2012) showed that there are no locally asymptotically unbiased estimators and no regular estimators for parameters that involve min or max operators. One available methodology to obtain valid estimates and perform statistical inference in such cases was proposed by Chernozhukov, Lee, and Rosen (2013, hereafter CLR). We employ this methodology in the next section to obtain confidence intervals for the true parameter value and halfmedian unbiased estimates for our lower and upper bounds.24 For more information about the CLR methodology, we refer the reader to CLR (2013), and to Flores and Flores-Lagunes (2013) for implementation details in the current setting. For bounds not involving min or max operators, we compute plug-in estimates of the bounds and Imbens and Manski (2004) confidence intervals.

RESULTS In this section we discuss the results, which are presented in Tables 4
6. Each table corresponds to the combined sample, adolescents, and young adults, respectively, and contains a number of panels presenting different aspects of the estimation framework discussed in the previous section. The columns in these tables correspond to the two different outcomes: weekly earnings and employment during quarter 16 after random assignment. Recall that the mechanism of interest (S) is degree attainment, defined as attaining a high school, GED, or vocational certificate during the time of the study. Panels A to D in the tables present point estimates of relevant quantities in our econometric framework. The estimated bounds are presented in panels E and F of each table. These panels correspond to estimated bounds for MATE and NATE under Propositions 3 and 4, respectively. To save space, we do not report or discuss estimated bounds under Propositions 1 or 2, since they turn out to be wide and uninformative.25 Note that Proposition 3 does not rely on Assumption 3, which has the potentially unattractive feature of imposing the direction of the sign on MATE and NATE, whereas Proposition 4 adds that assumption to the others. Recall that the estimated bounds on MATE provide information about the relative importance of attaining a degree in mediating the total effect of the treatment (the ATEs) on the corresponding outcome. In turn, NATE is informative about the relative importance of all other components

Variable Combined sample Female Age White Black Hispanic Married Has child/children

Highest grade completed Ever been arrested Weekly earnings at baseline Employed at baseline

n0

ap

n1

n0 − ap

ap − n1

n0 − n1

0.417*** (0.012) 18.281*** (0.048) 0.207*** (0.010) 0.549*** (0.012) 0.175*** (0.009) 0.063*** (0.006) 0.214*** (0.010) 0.009*** (0.002) 9.350*** (0.030) 0.290*** (0.011) 86.26*** (2.569) 0.153*** (0.009)

0.502*** (0.012) 18.966*** (0.241) 0.257*** (0.011) 0.503*** (0.012) 0.168*** (0.009) 0.040*** (0.005) 0.186*** (0.010) 0.202*** (0.010) 9.989*** (0.164) 0.201*** (0.010) 117.73*** (12.955) 0.205*** (0.010)

0.393*** (0.013) 18.534*** (0.055) 0.268*** (0.012) 0.478*** (0.014) 0.183*** (0.011) 0.061*** (0.007) 0.187*** (0.011) 0.167*** (0.010) 10.078*** (0.037) 0.287*** (0.012) 109.74*** (3.089) 0.216*** (0.011)

−0.085*** (0.017) −0.685*** (0.246) −0.050*** (0.014) 0.046*** (0.017) 0.007 (0.013) 0.023*** (0.008) 0.028** (0.014) −0.193*** (0.010) −0.639*** (0.167) 0.089*** (0.015) −31.47** (13.207) −0.052*** (0.013)

0.108*** (0.018) 0.432* (0.247) −0.011 (0.016) 0.026 (0.018) −0.015 (0.014) −0.021** (0.008) −0.001 (0.014) 0.035** (0.014) −0.089 (0.168) −0.086*** (0.016) 7.994 (13.318) −0.01 (0.015)

0.023 (0.013) −0.253*** (0.053) −0.061*** (0.011) 0.071*** (0.013) −0.008 (0.010) 0.002 (0.006) 0.027** (0.011) −0.158*** (0.004) −0.728*** (0.034) 0.003 (0.012) −23.47*** (2.852) −0.063*** (0.010)

MARIA BAMPASIDOU ET AL.

HS/GED/VOC degree

Average Pre-Treatment Characteristics by Strata.

134

Table 3.

Age White Black Hispanic Married Has child/children HS/GED/VOC degree Highest grade completed Ever been arrested Weekly earnings at baseline Employed at baseline Young adults Female Age White

0.528*** (0.014) 17.882*** (0.158) 0.244*** (0.012) 0.511*** (0.014) 0.168*** (0.011) 0.032*** (0.005) 0.146*** (0.010) 0.122*** (0.009) 9.678*** (0.178) 0.183*** (0.011) 102.04*** (14.776) 0.198*** (0.011)

0.364*** (0.015) 17.682*** (0.034) 0.290** (0.140) 0.469*** (0.150) 0.179*** (0.012) 0.039*** (0.006) 0.118*** (0.010) 0.091*** (0.009) 9.823*** (0.038) 0.317*** (0.014) 100.71*** (3.295) 0.206*** (0.012)

−0.136*** (0.019) −0.391** (0.161) −0.034** (0.016) 0.043** (0.019) 0.008 (0.015) 0.016** (0.008) 0.008 (0.014) −0.118*** (0.009) −0.458** (0.181) 0.116*** (0.016) −22.13 (15.036) −0.047*** (0.015)

0.164*** (0.020) 0.2 (0.162) −0.046*** (0.018) 0.042*** (0.021) −0.012 (0.016) −0.007 (0.008) 0.028** (0.014) 0.031** (0.013) −0.144 (0.182) −0.134*** (0.018) 1.33 (15.139) −0.008 (0.017)

0.523*** (0.028) 21.696*** (0.074) 0.196*** (0.022)

0.398*** (0.025) 21.885*** (0.301) 0.325*** (0.024)

0.504*** (0.030) 21.75*** (0.081) 0.187*** (0.023)

0.124*** (0.037) −0.188 (0.310) −0.130*** (0.032)

−0.105*** (0.039) 0.135 (0.312) 0.138*** (0.033)

0.028 (0.014) −0.191*** (0.033) −0.080*** (0.012) 0.085*** (0.015) −0.004 (0.011) 0.008 (0.006) 0.036*** (0.011) −0.086*** (0.004) −0.603*** (0.034) −0.018*** (0.013) −20.80*** (3.085) −0.055*** (0.011) 0.019 (0.030) −0.053 (0.081) 0.009 (0.024)

135

0.392*** (0.013) 17.491*** (0.030) 0.210*** (0.011) 0.554*** (0.013) 0.176*** (0.010) 0.047*** (0.006) 0.154*** (0.010) 0.004** (0.002) 9.220*** (0.030) 0.299*** (0.012) 79.91*** (2.784) 0.151*** (0.010)

Degree Attainment in the Differential Impact of Job Corps

Adolescents Female

Variable Black Hispanic Married Has child/children HS/GED/VOC degree Highest grade completed Ever been arrested Weekly earnings at baseline

n0

ap

n1

n0 − ap

ap − n1

n0 − n1

0.526*** (0.028) 0.174*** (0.021) 0.131*** (0.019) 0.474*** (0.028) 0.031*** (0.010) 9.911*** (0.085) 0.251*** (0.024) 113.72*** (6.283) 0.162*** (0.020)

0.466*** (0.025) 0.165*** (0.019) 0.049*** (0.011) 0.264*** (0.022) 0.381*** (0.025) 10.771*** (0.341) 0.286*** (0.023) 161.98*** (27.166) 0.219*** (0.021)

0.511*** (0.030) 0.198*** (0.024) 0.144*** (0.021) 0.446*** (0.030) 0.457*** (0.030) 11.040*** (0.082) 0.173*** (0.023) 143.80*** (7.595) 0.252*** (0.026)

0.06 (0.037) 0.009 (0.028) 0.083*** (0.022) 0.21*** (0.035) −0.351*** (0.026) −0.859** (0.352) −0.035 (0.033) −48.26* (27.883) −0.057** (0.029)

−0.045 (0.039) −0.033 (0.030) −0.095*** (0.024) −0.182*** (0.037) −0.076* (0.039) −0.269 (0.351) 0.113*** (0.032) 18.182 (28.208) −0.032 (0.033)

0.015 (0.030) −0.024 (0.023) −0.012 (0.020) 0.028 (0.030) −0.426*** (0.015) −1.128*** (0.092) 0.078** (0.026) −30.07*** (6.934) −0.09*** (0.023)

Note: Standard errors in parenthesis. *, **, *** denote statistical significance at the 90%, 95%, and 99% confidence levels, respectively.

MARIA BAMPASIDOU ET AL.

Employed at baseline

(Continued )

136

Table 3.

Table 4.

Point Estimates and Estimated Bounds, Combined Sample. Earnings

Panel A: Treatment effects ATE of random assignment on outcome

19.63*** (4.712)

ATE of random assignment on degree attainment

0.34*** (0.007) 0.45*** (0.009) 0.21*** (0.011)

πn1 πap

E[Y|T = 1] E[Y|S = 0] E[Y|S = 1] Panel D: Testable implications E[Y|T = 0, S = 1] − E[Y|T = 0, S = 0] E[Y|T = 1, S = 1] − E[Y|T = 1, S = 0] E[Y|T = 1, S = 1] − E[Y|T = 0, S = 0]

Panel E: Estimated bounds
Proposition 3 MATE NATE Panel F: Estimated bounds
Proposition 4 MATE NATE Observations

0.028*** (0.011) 0.21*** (0.011)

Panel B: Strata proportions πn0

Panel C: Conditional means E[Y|T = 0]

Employment

190.82*** (3.655) 210.46*** (2.955) 160.21*** (3.185) 237.26*** (3.104)

0.676*** (0.009) 0.704*** (0.006) 0.607*** (0.008) 0.760*** (0.006)

70.49*** (4.226) 82.42*** (5.851) 79.42*** (5.863)

0.153*** (0.017) 0.158*** (0.014) 0.150*** (0.015)

Earnings

Employment

LB UB −39.70 17.30 [ − 45.82, 19.90] 2.32 59.30 [ − 5.94, 70.92]

LB UB −0.084 0.033 [ − 0.093, 0.051] −0.005 0.112 [ − 0.026, 0.133]

0.00 15.64 [0.00, 21.43] 3.98 19.61 [ − 1.79, 28.43]

0.000 0.028 [0.000, 0.049] 0.000 0.028 [0.000, 0.049] 8,020

Notes: Standard errors in parentheses. The estimated bounds and their 95% confidence intervals are based on the CLR (2013) procedure. ** and *** denote statistical significance at the 95% and 99% confidence levels, respectively.

138

MARIA BAMPASIDOU ET AL.

of the JC program in mediating those same total effects. Importantly, under our assumptions (justified in the previous section), both sets of nonparametric bounds have causal interpretation.

Results for the Combined Sample The results for the combined sample, presented as a benchmark for the subsequent analysis by age groups, are shown in Table 4. From panel A, we see that the ATE of random assignment to JC on earnings is $19.63, while that on employment is 0.028. The ATE on degree attainment is 0.21. These positive and statistically significant effects of random assignment to JC are consistent with the NJCS results reported by Schochet et al. (2001). Noteworthy is the positive (and highly statistically significant) effect of random assignment on degree attainment of 21%, which is in line with the expectation that the attainment of a degree (i.e., the remedial basic education and/or vocational training) is likely to play an important role in mediating the (total) effect of random assignment to JC on future labor market outcomes. Once the individual-level monotonicity assumption of the treatment on the mechanism variable (Assumption 2) is imposed, which rules out the existence of the affected-negatively stratum, the population proportions of individuals in the remaining three strata can be estimated. Panel B of Table 4 shows that the estimated population proportions for the three remaining strata in our analysis are equal to 0.34, 0.45, and 0.21 for strata (n0), (n1), and (ap), respectively. Thus, based on these statistically significant estimates, about 79% of the combined sample belongs to strata for which the treatment does not affect the mechanism variable (n0 and n1); with 21% of the combined sample consisting of individuals who are positively affected by random assignment to JC in terms of degree attainment. The latter is the stratum for which degree attainment can play a role in mediating the effect of random assignment to JC on the labor market outcomes of interest. The strata for which the treatment does not affect the mechanism variable is made up of n0 and n1. If the attainment of degrees by youth is of interest to policymakers in charge of the JC program, then the individuals in n1 may not be of concern since they would obtain a degree regardless of treatment assignment (although it still can be useful to characterize them relative to the other strata). However, the stratum of individuals that do not attain a degree irrespective of treatment assignment (n0) is likely to

Degree Attainment in the Differential Impact of Job Corps

139

represent a group of policy concern. Looking at Table 3 that lists the average pre-treatment characteristics of each stratum, we can characterize individuals in the combined sample that belong in n0. Perhaps not surprisingly, individuals in this stratum show average pre-treatment characteristics that are traditionally related to being more disadvantaged, relative to individuals in the other two strata. Individuals in the n0 stratum are, on average, more likely to be younger, black, and have children; less likely to be white and have a degree and be employed at baseline; and they also have less years of schooling and earnings at baseline. All of these average characteristics are statistically different from the averages of n1 and ap strata. Similarly, we can characterize the stratum of the ap, a stratum that policymakers affect in terms of degree attainment by making JC available. It is interesting to observe that females are more likely to belong to the ap stratum, a difference that is statistically significant relative to the proportion of females in the other two strata. Other statistically significant average pre-treatment characteristics of the ap stratum indicate that they are more likely to be older and have a degree, not being married, and being less likely to had been arrested at baseline. Noteworthy is the evidence that the ap stratum is not more likely to be employed and have higher earnings at baseline, relative to the n1 stratum. It is also interesting to note the interplay between the average pre-treatment characteristics of female and married in the ap stratum: by implication, it is likely that non-married females belong to this strata in higher proportion. Panel C in Table 4 presents, for reference, the estimated conditional means of the outcomes by values of the treatment (T) and the mechanism (S) variables. They are all precisely estimated and conform to the expectation that those in the treatment group (T = 1) and those that attained a degree (S = 1) have better mean outcomes relative to those with T = 0 and S = 0, respectively. These estimated conditional means for the outcomes imply that randomization into JC results in an average increase in earnings (the probability of being employed) of about 10% (4%). The testable implications to falsify the set of assumptions that Proposition 3 and Proposition 4 rely on to estimate the bounds are presented in panel D. For both labor market outcomes (earnings and employment), the testable implications indicate that the data on the combined sample is consistent with the corresponding assumptions for each proposition, since they are all positive (i.e., satisfied) and highly statistically significant.26 We remind the reader that this is not hard evidence on the validity of those assumptions. Instead, this evidence merely reflects the inability to falsify those assumptions with the data.

140

MARIA BAMPASIDOU ET AL.

Panels E and F in Table 4 present the estimated bounds for both earnings and employment for the combined sample based on Proposition 3 (employing Assumptions 1, 2, and 4) and Proposition 4 (employing Assumptions 1 to 4), respectively. Looking at the estimated bounds for the earnings outcome in panel E, the estimated bounds for MATE include zero and are relatively wide. However, the estimated upper bound is $17.30, which implies that the part of the effect of randomization into JC on earnings that is due to attaining a degree is at most 88% (17.3/19.63). The corresponding confidence interval, however, indicates that the true value of MATE under Proposition 3 lies in the intervals [ − 45.82,19.90] with 95% probability. Looking at the estimated bounds on NATE in the same panel, they exclude the value of zero, which indicates that not all of the ATE of random assignment to JC on earnings is mediated through the attainment of a degree. However, the corresponding confidence interval includes zero. Turning to the employment outcome in the same panel E, the estimated bounds suggest that little can be said about the role of degree attainment as a mediator for the ATE of random assignment to JC. In particular, the estimated bounds for both MATE and NATE under Proposition 3 include zero and the ATE of random assignment on employment. Turning to the estimated bounds in panel F (under Proposition 4) for earnings, the estimated lower bound for MATE is zero, while the estimated upper bound is $15.64. These estimated bounds imply that the role of degree attainment in mediating the ATE of random assignment on earnings is at most 80% (15.64/19.63), possibly leaving room for all other channels (e.g., other services offered through JC) to explain at least 20% of the same effect. Indeed, when looking at the estimated bounds for NATE, the lower and upper bounds are [3.98,19.61]. However, while the confidence intervals corresponding to the bounds under Proposition 4 are considerable tighter relative to those on Proposition 3, they still include zero and the ATE for both MATE and NATE.27 Turning to the employment outcome in panel F, the estimated bounds are not very informative for both NATE and MATE since they go from zero to the value of the ATE. In sum, the estimated bounds for the combined sample suggest that, when considering the earnings outcome, the attainment of a degree represents, at most, 80
88% of the total ATE of random assignment to JC. By implication, the role of other channels such as other services offered through JC, would causally represent at least 12
20% of the ATE of randomization into JC. However, the 95% confidence intervals render this statement uncertain, and thus we cannot statistically rule out that MATE = ATE or

Degree Attainment in the Differential Impact of Job Corps

141

NATE = ATE. In terms of employment, little can be said about the role of degree attainment as a mediator for the ATE.

Results by Age Groups Adolescents: Youths 16
19 Years Old The results for the sample of adolescents (aged 16
19) are presented in Table 5. The point estimates in panel A indicate that adolescents experience considerably smaller effects of random assignment to JC on earnings and employment relative to the combined sample and, as shown below, relative to young adults too. This is consistent with the findings by age groups reported in the original NJCS report. For adolescents, the effect of randomization into JC on earnings is $13.37 and that on employment is 0.017. They are 32% and 39% smaller relative to the same effects for the combined sample. The effect of random assignment to JC on degree attainment, however, is not very different, amounting to 0.201 versus 0.21 for the combined sample. The estimated strata population proportions in panel B of Table 5 are similar to the ones for the combined sample, and imply that the size of the ap stratum is 20%, with the rest of the population belonging to the two strata for which the treatment does not affect degree attainment. A similar analysis based on average pre-treatment characteristics by strata can be done for adolescents by looking at their corresponding panel in Table 3. Adolescents’ average pre-treatment characteristics by strata are fairly similar to those of the combined sample. Individuals that belong in n0 are, on average, more likely to be younger (within the 16
19 range), and black; less likely to be white and have a degree and be employed at baseline; and they also have less years of schooling at baseline. All of these average characteristics are statistically different from the averages of n1 and ap strata, and are typically related to being more disadvantaged. In turn, individuals in the ap stratum are more likely to be female and less likely to ever been arrested by baseline, with both average characteristics statistically different relative to those in the other two strata. Finally, statistically significant average pre-treatment characteristics for the n1 stratum indicate that, relative to the other two strata, they are more likely to be white and (surprisingly) have ever been arrested by baseline; and less likely to be black and have children at baseline. Panel C in Table 5 indicates that adolescents show relatively similar
but smaller in magnitude
estimated conditional means of the outcomes by

142

Table 5.

MARIA BAMPASIDOU ET AL.

Point Estimates and Estimated Bounds, Adolescents Sample. Earnings

Panel A: Treatment effects ATE of random assignment on outcome

13.37** (5.263)

ATE of random assignment on degree attainment

0.36*** (0.008) 0.44*** (0.010) 0.20*** (0.013)

πn1 πap

E[Y|T = 1] E[Y|S = 0] E[Y|S = 1] Panel D: Testable implications E[Y|T = 0, S = 1] − E[Y|T = 0, S = 0] E[Y|T = 1, S = 1] − E[Y|T = 1, S = 0] E[Y|T = 1, S = 1] − E[Y|T = 0, S = 0]

Panel E: Estimated bounds
Proposition 3 MATE NATE Panel F: Estimated bounds
Proposition 4 MATE NATE Observations

0.017 (0.012) 0.201*** (0.013)

Panel B: Strata proportions πn0

Panel C: Conditional means E[Y|T = 0]

Employment

186.16*** (4.068) 199.53*** (3.315) 153.46*** (3.458) 229.79*** (3.564)

0.671*** (0.010) 0.689*** (0.008) 0.593*** (0.009) 0.756*** (0.007)

69.85*** (7.819) 82.78*** (6.471) 73.85*** (6.529)

0.160*** (0.019) 0.173*** (0.016) 0.150*** (0.016)

Earnings

Employment

LB UB −37.10 16.64 [ − 44.03, 19.55] −3.28 50.43 [ − 12.59, 63.59]

LB UB −0.085 0.035 [ − 0.095, 0.041] −0.017 0.102 [ − 0.040, 0.126]

0.00 11.63 [0.00, 18.19] 1.74 13.34 [ − 4.58, 23.32]

0.000 0.017 [0.000, 0.041] 0.000 0.017 [0.000, 0.041] 6,295

Notes: Standard errors in parentheses. The estimated bounds and their 95% confidence intervals are based on the CLR (2013) procedure. ** and *** denote statistical significance at the 95% and 99% confidence levels, respectively.

Degree Attainment in the Differential Impact of Job Corps

143

T and S as compared to combined sample. These estimated conditional means for the outcomes of adolescents imply that randomization into JC results in smaller percentage increases relative to the combined sample. They imply an increase in average earnings and probability of employment of about 7% and 3%, respectively. Like with the combined sample, adolescents also satisfy the testable implications related to the assumptions in Propositions 3 and 4, presented in panel D. Thus, the data for the sample of adolescents is consistent with the testable implications of those assumptions. Looking at the estimated bounds for earnings under Proposition 3 in panel E, we see that for both MATE and NATE they include zero. This is in slight contrast to the estimated bounds for the combined sample, which excluded zero for NATE (although none of their confidence intervals excluded zero). As for the employment outcome, the estimated bounds in panel E suggest that, as with the combined sample, little can be said about the role of degree attainment as a mediator for the ATE of random assignment to JC. The estimated bounds for both MATE and NATE include zero, and the corresponding upper bounds are considerably larger than the ATE of random assignment on employment. The estimated bounds under Proposition 4 (panel F) are tighter and more informative for earnings. The MATE is bounded from below by zero and it is at most $11.63, implying that, for adolescents, the attainment of a degree represents at most 87% of the ATE of random assignment to JC on earnings. In turn, the estimated bounds on NATE exclude zero and the upper bound is almost as large as the ATE; but the confidence intervals for these two sets of bounds include zero and the ATE. For the employment outcome, the estimated bounds in panel F for both MATE and NATE go from zero to the value of the ATE, leaving again little to say about the relative importance of degree attainment as a mediator.28 In sum, comparing the estimated bounds for adolescents with those for the combined sample, and aside from sampling error, they suggest that the upper bound on the role that the attainment of a degree represents of the ATE for earnings is proportionally more for adolescents (80% vs. 87% under Proposition 4). Still, the confidence intervals render this inconclusive. In turn, there is little that can be said about the role of attaining a degree in the ATE of random assignment on employment. Next, we present the contrasting results for young adults. Young Adults: Youths 20
24 Years Old Table 6 reports the results for the age group of young adults (aged 20
24). Also consistent with the originally reported NJCS results, panel A shows that young adults experience considerably larger effects relative to

144

Table 6.

MARIA BAMPASIDOU ET AL.

Point Estimates and Estimated Bounds, Young Adults Sample. Earnings

Panel A: Treatment effects ATE of random assignment on outcome

38.06*** (11.000)

ATE of random assignment on degree attainment

0.28*** (0.013) 0.49*** (0.021) 0.23*** (0.025)

πn1 πap

E[Y|T = 1] E[Y|S = 0] E[Y|S = 1] Panel D: Testable implications E[Y|T = 0, S = 1] − E[Y|T = 0, S = 0] E[Y|T = 1, S = 1] − E[Y|T = 1, S = 0] E[Y|T = 1, S = 1] − E[Y|T = 0, S = 0]

Panel E: Estimated bounds
Proposition 3 MATE NATE Panel F: Estimated bounds
Proposition 4 MATE NATE Observations

0.060** (0.024) 0.234*** (0.025)

Panel B: Strata proportions πn0

Panel C: Conditional means E[Y|T = 0]

Employment

210.88*** (8.684) 248.94*** (6.531) 190.65*** (7.478) 260.94*** (6.643)

0.696*** (0.020) 0.756*** (0.013) 0.672*** (0.019) 0.772*** (0.013)

69.57*** (17.002) 66.89*** (13.450) 90.64*** (13.389)

0.120*** (0.040) 0.080*** (0.029) 0.414*** (0.033)

Earnings

Employment

LB UB −42.85 21.13 [ − 56.83, 28.12] 16.85 80.58 [ − 3.66, 107.86]

LB UB −0.071 0.033 [ − 0.087, 0.047] 0.027 0.131 [ − 0.018, 0.165]

0.00 15.62 [0.00, 22.19] 22.43 38.01 [1.48, 59.00]

0.000 0.019 [0.000, 0.032] 0.042 0.060 [ − 0.006, 0.106] 1,725

Notes: Standard errors in parentheses. The estimated bounds and their 95% confidence intervals are based on the CLR (2013) procedure. ** and *** denote statistical significance at the 95% and 99% confidence levels, respectively.

Degree Attainment in the Differential Impact of Job Corps

145

adolescents. The ATE point estimate of random assignment to JC on earnings is $38.06 and that on employment is 0.060. These estimates are a striking three times larger relative to the same effects for adolescents. Relative to the combined sample, young adults experience ATEs for earnings and employment that are about twice the magnitude. Conversely, the ATE of random assignment to JC on degree attainment (0.23) is just slightly larger than that of adolescents and the combined sample (0.20 and 0.21, respectively). The estimated strata population proportions for young adults are shown in panel B. The size of the young adults ap stratum is 23%, compared to 20% for adolescents and 21% for the combined sample, and the three are not statistically different from each other. And while the estimated proportions of the strata n0 and n1 for young adults are statistically different from the corresponding ones for adolescents and for the combined sample, the overall proportion of the population whose degree attainment is not affected by random assignment to JC is very similar (77% vs. 80%) and not statistically different. This suggests that, in each of the two age groups, about the same proportion of individuals are estimated to benefit from random assignment in terms of degree attainment. Moreover, these estimates imply that among the group of adolescents there is a significantly larger proportion of individuals that will not earn a degree regardless of whether they are randomized into JC (36% vs. 28%); while a significantly higher proportion of young adults are estimated to earn such a degree irrespective of random assignment into JC (49% vs. 44%). These implications are consistent with the belief that young adults that participated in JC seemed to be more highly motivated and well behaved relative to adolescents (e.g., Schochet et al., 2008). We can also characterize the young adults’ strata by their average characteristics (presented in Table 3), as we did with the previous groups analyzed. The group of young adults turns out to have fairly different average pretreatment characteristics by strata relative to adolescents and the combined sample. Individuals that belong in n0 are, on average, less likely to have completed a degree, have lower years of schooling, and lower average earnings and employment at baseline. These average characteristics are all statistically different from the averages of young adults’ n1 and ap strata. In contrast to the n0 (strata for adolescents and the combined sample, young adults in the n0 stratum are more likely to be female, be married, and have children’ their weekly earnings at base line are also higher.” As for young adults that belong in the ap stratum (and relative to those in the other two strata), they are more likely to be white, and less likely to be female, married, and have children. These average characteristics are statistically

146

MARIA BAMPASIDOU ET AL.

different relative to those in each of the other two strata.29 Finally, the average pre-treatment characteristics for the young adults’ n1 stratum are different relative to those for adolescents and the combined sample. Some statistically significant differences indicate that young adults n1 stratum are more likely to have completed a degree by baseline, and, in contrast to adolescents, they are less likely to have ever been arrested by baseline. It is interesting to note that in the case of young adults the proportion of blacks is statistically similar in the three strata, which is another contrasting feature with respect to adolescents and the combined sample. Panel C of Table 6 presents the conditional means by T and S for earnings and employment for young adults. The patterns are similar relative to the previous two groups, with the magnitudes significantly higher. The effect of T on earnings and employment for young adults represents an 18 and 8.6% increase, respectively. Thus, the effects of randomization into JC on the outcomes for young adults are considerable larger (more than twice) in percentage terms, relative to those for adolescents. Panel D shows that the group of young adults also satisfy the testable implications related to the assumptions in Proposition 3 and Proposition 4. Thus, as with the previous two groups, the data for young adults is consistent with the testable implications of those assumptions. The estimated bounds for young adults under Proposition 3 are shown in panel E of Table 6. For the earnings outcome, we see that the estimated bounds for MATE include zero. The estimated bounds for NATE, however, exclude zero, with an estimated lower bound of $16.85 that suggests a relative importance of other channels in mediating the ATE of random assignment on earnings of at least 44% (16.85/38.06). Nevertheless, the 95% confidence interval includes zero ([−3.66,107.86]).30 Thus, we cannot statistically state that the NATE is larger than zero under the assumptions in Proposition 3
and thus that channels other than degree attainment are important in mediating the ATE of random assignment to JC. Turning to the employment outcome, for which we found it difficult before to say much about, we are now able to gather some conclusions for young adults about the importance of degree attainment as a mechanism for the ATE. The estimated bounds for MATE under Proposition 3 (panel E) include zero; however, the estimated upper bound suggests that attaining a degree accounts, at most, for 55% (0.033/0.06) of the ATE of random assignment to JC on employment. Similarly, the estimated bounds for NATE in panel E exclude zero and indicate that the importance of channels other than degree attainment is at least 45% (0.027/0.06). The confidence interval corresponding to the bonds on NATE, however, includes zero (the 90%

Degree Attainment in the Differential Impact of Job Corps

147

confidence interval, though, barely includes zero).31 In general, there is some evidence that, for young adults, randomization into JC has a positive effect on labor market outcomes through channels other than the attainment of a degree. Importantly, these results are obtained without assuming the signs of MATE and NATE (Assumption 3). The estimated bounds under Proposition 4 (panel F) reinforce the notions gathered for young adults in panel E. The estimated bounds for earnings show that MATE is bounded from below by zero and it is at most $15.62. Thus, for young adults, the attainment of a degree represents at most 41% (15.62/38.06) of the ATE of random assignment on earnings. This stands in stark contrast with the results documented above for adolescents. The estimated bounds on NATE are consistent with this, and span [22.43,38.01]. Thus, the mediating role of channels other than degree attainment is estimated to be at least 59% (22.43/38.06) of the ATE of random assignment on earnings. Importantly, accounting for sampling error, the confidence interval on NATE excludes zero with a 95% confidence, even though the sample of young adults is considerably smaller than that of adolescents. Therefore, we find statistical evidence of the importance of channels other than degree attainment (i.e., NATE > 0 and hence MATE < ATE) for young adults.32 In terms of employment as an outcome, MATE is bounded from below at zero and it is at most 0.019, representing at most 32% (0.019/0.06) of the corresponding ATE. The estimated bounds for NATE in panel F are consistent with this notion: NATE is at least 0.042 or 70% (0.042/0.06) of the ATE and at most the full ATE, while its 95% confidence interval barely includes zero. Interestingly, the 90% confidence interval for these bounds does exclude zero: [0.004,0.098]. Summing up, the partial identification evidence presented above for young adults strongly suggests that degree attainment does not completely mediate the ATE of random assignment to JC on earnings and employment, which is something that could not be ruled out for adolescents.

Discussion At the outset we reiterate that, as documented in the original NJCS report (Schochet et al., 2001), the results for the population of eligible JC applicants (i.e., our combined sample) mask important heterogeneity by age groups
in particular between the groups we call adolescents and young adults.

148

MARIA BAMPASIDOU ET AL.

In terms of our estimated nonparametric bounds for the combined sample, adolescents, and young adults, obtained under relatively weak assumptions, they point towards a number of conclusions. Perhaps the strongest one is that, for the group of young adults, the role played by channels other than the attainment of a degree (i.e., NATE) is statistically larger than zero. By implication, for this group, attaining a degree (i.e., MATE) does not mediate the total effect of random assignment on both earnings and employment (i.e., MATE < ATE). Thus, with statistical confidence, we can say that channels other than degree attainment are important in mediating the effect of random assignment to JC on the labor market outcomes of young adults. A second conclusion from our estimated nonparametric bounds is that we cannot rule out that attaining a degree accounts for all of the effect of random assignment on the labor market outcomes of the group of adolescents (i.e., MATE = ATE). That is to say, we cannot rule out that degree attainment fully mediates the ATE. This is also the case for the combined sample, likely due to the group of adolescents comprising most of it (78.5%). This contrasts with the previous conclusion concerning young adults. A third conclusion is that there is suggestive
but not statistically significant
evidence that the magnitude of the mediating effect of channels other than degree attainment on labor market outcomes is larger for young adults than for adolescents (i.e., NATEYng.Ads > NATEAdols). Comparing the estimated bounds for NATE on these two groups for the earnings outcome, we have [22.43,38.01] for young adults and [1.74,13.34] for adolescents. For employment, they are [0.042,0.60] and [0,0.017], respectively. While it is clear that the two pairs of estimated bounds do not overlap, it is true that their 95% confidence intervals do, thus tempering this conclusion. In spite of not being able to obtain a single-valued estimate of our effects of interest
a trade-off accepted for weaker assumptions
we can provide useful policy insights for JC, a program that emphasizes remedial education, vocational training, and other relevant services (e.g., counseling, job search assistance, and a residential component). To do this, we take the conclusion that degree attainment does not fully mediate the effect of randomization on young adults’ labor market conditions and combine it with other estimates from our econometric framework. This allows us to moderately characterize the set of individuals that are more likely to benefit from JC services other than degree attainment. In particular, the average pre-treatment characteristics by strata for young adults relative to adolescents are useful in performing this characterization.

Degree Attainment in the Differential Impact of Job Corps

149

Recall that the ap stratum is key to the MATE since under our assumptions this is the only stratum for which T affects S. In contrast, all three strata contribute to NATE. Our results imply that for young adults services other than degree attainment play some role in mediating the total effect. Thus, we focus on pointing out some marked differences between young adults and adolescents in the average characteristics of their ap stratum relative to the other two strata combined.33 There are three such characteristics that stand out. But before describing them, we note that, for both age groups, the differences in earnings and employment at baseline between the ap stratum and the other two strata are largely insignificant. Thus, the following average characteristics do not seem to be driven by pre-treatment labor market outcomes. The first characteristic that stands out is the proportion of females. Females are significantly in higher proportion in the ap stratum (0.53) relative to the combination of the other two strata (0.38) for adolescents. The opposite is true for young adults (0.40 for ap vs. 0.51 for the combination of the other two strata). This characteristic interacts with two related ones: young adults in the ap stratum are significantly less likely to be married and have children, whereas for adolescents these characteristics are fairly balanced between ap and the combination of the other two strata. Thus, one possibility is that the higher proportion of young adult females in the n0 and n1 strata are more likely to be married and have children. Following our results, this group may be able to benefit more from program services other than those related to degree attainment. This may be because they have more responsibilities and thus are more mature.34 At the same time, the lower proportion of young adult females that belong in the ap stratum are likely unmarried and have no children, characteristics that could facilitate their taking advantage of both degree attainment and other services within the program.35 A second characteristic that stands out is the proportion of white individuals. While this variable turns out to be close to perfectly aligned between the ap stratum and the combination of the other two strata for adolescents, for young adults the ap stratum has significantly more white individuals relative to the combination of the other two strata (0.33 vs. 0.19). This could indicate that black (and Hispanic) young adults are better at taking advantage of JC services other than degree attainment. If true, this aspect is potentially important given the substantially higher (total) effects of the program on labor market outcomes for white individuals relative to blacks and Hispanics (e.g., Flores-Lagunes et al., 2010). Finally, another

150

MARIA BAMPASIDOU ET AL.

characteristic that is contrasting is the proportion who have been arrested by baseline: while adolescents in the ap stratum are significantly less likely to have been arrested relative to the combination of the other two strata (0.18 vs. 0.31), young adults in the ap stratum are significantly more likely to have been arrested relative to the combination of the other two strata (0.29 vs. 0.22). One possible implication is that young adults that are less likely to have been arrested by baseline are more likely to benefit from JC services other than degree attainment. Potential reasons for this are unclear to us at this point. To close, we relate the results herein with previous evidence. In our view, the results above complement the evidence in Schochet et al. (2008) and Blanco et al. (2013a). Schochet et al. (2008) pointed out that the larger impacts on young adults relative to adolescents were consistent with the 10% higher length of enrollment and accompanying larger gains in human capital accumulation by the older group. Here, we find that the impact for young adults are indeed mediated by their higher human capital accumulation in the form of degree attainment. But at the same time, we document that the impacts for young adults are also mediated by channels (i.e., services) other than degree attainment. To the extent that the other services (or most of them) are not related to traditional human capital activities (i.e., classroom training)
such as job search assistance, social skills training, health services and residential services
our results suggest that non-human capital services could be especially relevant for young adults. Our results can also be related to Blanco et al. (2013a) who documented that, to the extent that human capital accumulation becomes incorporated into the individuals’ wages, the estimated bounds on the wage effects of JC for adolescents and young adults do not seem to be disparate enough to be consistent with a sole explanation based on the groups differential human capital accumulation. Our results above for young adults support this view by documenting for this group a statistically positive role of services other than those related to human capital accumulation (degree attainment) in mediating young adults’ larger effects on labor market outcomes.

CONCLUSION JC is the largest and most comprehensive training program for disadvantaged youth. A randomized evaluation of the program documented the positive effects of the availability of the program (random assignment) on

Degree Attainment in the Differential Impact of Job Corps

151

the eligible youth population. But this same evaluation documented marked differences in the effects between adolescents (aged 16
19) and young adults (aged 20
24). Those disparate effects
as much as three times larger for young adults
were documented in the original NJCS evaluation report (Schochet et al., 2001) and further analyzed in Blanco et al. (2013a). In this paper, we estimated nonparametric bounds on the MATE and the NATE to learn about the role of degree attainment in mediating the treatment effect of randomization into JC on the labor market outcomes of adolescents and young adults. Our main purpose was to analyze whether the mediating role of degree attainment on the effect of random assignment into JC is linked to the disparate effects of that treatment for those two groups. By estimating nonparametric bounds, we strove to rely on relatively weak assumptions. Our estimated nonparametric bounds indicate that, for young adults, the role played by channels other than the attainment of a degree in the mediation of the total effect of randomization into JC on labor market outcomes is greater than zero. This implies that the attainment of a degree does not fully account for that total effect. In contrast, for adolescents, we cannot rule out that degree attainment fully mediates those total effects. That those other components could play a more important role for young adults likely has policy implications for more effectively servicing participants. We frame some of those policy implications by characterizing the individuals that are more likely to benefit from components of the program other than degree attainment. Moreover, we relate the implications of our estimated nonparametric bounds to previous work that examined the disparate impacts of the availability of JC on adolescents and young adults. Our implications appear to complement those other studies by offering an explanation for why the differences in human capital accumulation between the two groups seem not to account for the entirety of the observed disparate impacts. Finally, in a more general note, the present work illustrates how researchers can learn about particular mechanisms of an intervention and derive implications from it under relatively weak assumptions.

NOTES 1. We note that Schochet, Burghardt, and McConnell (2008) cast some doubt on the effectiveness of Job Corps as they find no program effects when using state’s administrative data on individual’s earnings, as opposed to survey data (as we use here).

152

MARIA BAMPASIDOU ET AL.

2. Other studies commenting on the features of JC as contributing factors of the programs’ success include Heckman and Krueger (2002) and Schochet et al. (2008). 3. Another important heterogeneity that was tagged as the most important “failure” of JC in Schochet et al. (2001) is the lack of a program effect on Hispanics. See Flores-Lagunes et al. (2010) for a study of this particular heterogeneity. 4. These figures are based on “intention-to-treat” effects on the sample we employ. Also, relative to Schochet et al. (2001), we combine their 16
17 and 18
19 groups into a single group to simplify the analysis to be conducted below. Impacts of similar magnitude are obtained if we disaggregate adolescents into two groups. 5. We note that the ATE that we estimate and decompose below is actually an intention-to-treat effect, that is, the effect of randomization into JC on future labor market outcomes. This allows us to fully exploit the randomization conducted in the NJCS to bound MATE and NATE. 6. In disciplines other than economics (e.g., statistics and epidemiology), there is a longer tradition of developing causal mediation methods. Some references are Robins and Greenland (1992), Pearl (2001), Rubin (2004), Petersen, Sinisi, and van der Laan (2006), Imai, Keele, and Yamamoto (2010), and Mattei and Mealli (2011). 7. For program year 2008, JC enrolled more than 60,000 new students in 124 centers nationwide, at a cost of about $1.58 billion. By program year 2008 the program had served about 2.6 million youths throughout its existence (U.S. Department of Labor, 2009). 8. According to JC, a youth is categorized as economically disadvantaged if her/his family is receiving public assistance or the family income is well below the poverty level as defined by the Department of Health and Human Services (Schochet et al., 2001). 9. The response rates were fairly high and similar for the two program groups. Specifically, the response rate was 95% to the baseline interview and 90%, 79%, and 80% to the 12-, 30-, and 48-month follow-up interviews, respectively (Schochet et al., 2001). In the analysis below, we employ NJCS-supplied probability weights that account for the different sampling of some particular subpopulations into control and treatment groups and for some interview nonresponse items (see Schochet, 2001 for details on the construction of these weights). 10. As mentioned in note 4, in our analysis we combine their two younger groups into one that we call adolescents. 11. From the 15,386 individuals in the NJCS, we lose 5,587 observations that have missing observations on key baseline variables, 307 with missing values in any of the two outcomes considered, and 1,472 with missing values on degree attainment. Our final sample size is of similar magnitude to other studies using NJCS data, such as Lee (2009), Flores-Lagunes et al. (2010) and Blanco et al. (2013a). 12. Recall that individuals in the control group were permitted to enroll in other training programs, vocational or technical schools, or even in high school. Thus, we can define a potential degree attainment with value Si (0) = 1 if they succeeded in acquiring a credential and Si (0) = 0 if not. 13. It is also possible to use the potential outcome Y(0,S(1)) to perform a similar decomposition. In this paper, we focus on the decomposition based on Y(1,S(0)).

Degree Attainment in the Differential Impact of Job Corps

153

14. There is also a corresponding literature in other fields employing bounds to estimate mediation effects (e.g., Cai, Kuroki, Pearl, & Tian, 2008; Kaufman, Kaufman, MacLehose, Greenland, & Poole, 2005; Lee, 2009; Sjo¨lander, 2009). For a discussion on how those bounds are related to the ones employed here, the reader is referred to Flores and Flores-Lagunes (2010). 15. The principal strata here are analogous to the “compliance types” in Imbens and Angrist (1994) and Angrist et al. (1996). 16. It is important to note that, since for the n0 and n1 strata degree attainment is not affected by assignment to JC (i.e., S(0) = S(1)), LMATEn0 = LMATEn1 = 0 and the LNATEs of these two strata equal the local (total) average effect on the outcome, that is, LATEk = E[Y(1, S(1)) − Y(0, S(0))|k], for k = {n0, n1}. 17. Note that the data contain information on Y(1, S(0)) for the n0 and n1 strata (and thus is not an entirely counterfactual outcome for them), since for these two strata we have that S(0) = S(1), which implies Y(1, S(0)) = Y(1, S(1)) = Y(1). 18. Some indirect evidence on this intuition is obtained by considering the strong and significant average effect of random assignment into JC on degree attainment (standard errors in parenthesis): 0.21 (0.01) for our full sample, 0.20 (0.01) for adolescents, and 0.23 (0.03) for young adults. 19. Recall that LMATEn0 = LMATEn1 = 0 and Assumption 2 ruled out the an stratum 20. Importantly, other influential econometric work on partial identification imposes similar weak-monotonicity assumptions at the individual level, such as Manski (1997), Manski and Pepper (2000), and Sjo¨lander (2009). 21. In this context, if a testable implication is not satisfied, then there is certainty that some of the assumptions are violated; whereas if it is satisfied, it only indicates that the data is “consistent” with the maintained assumptions. In this regard, the testable implications cannot be interpreted as formal statistical tests of the assumptions. 22. Under Assumption 4, a couple of average baseline characteristics in Table 3 would seem to be counterintuitive. In particular, having a high school, GED or vocational degree at baseline and ever been arrested at baseline, for which the ap stratum has more favorable averages than the n1 stratum. These differences, however, are not strong enough to make the averages of the pre-treatment outcomes of the ap stratum statistically greater than those of the n1 stratum. We further discuss the rest of the average characteristics of the different strata in the section “Results.” 23. For example, in Proposition 4, the upper bound on MATE is the minimum of 16 terms, while the lower bound on NATE is the maximum of 16 terms. See Flores and Flores-Lagunes (2010, 2013) for details. 24. The half-median-unbiasedness property means that the upper bound estimator exceeds the true value of the upper bound with probability at least one half asymptotically, while the reverse holds for the lower bound. 25. We also omit discussion of the estimated bounds on the LNATEs under the different propositions. 26. Recall that the first two testable implications in that panel are derived from the assumptions in Proposition 3, while the third testable implication is derived when Assumption 3 is added in Proposition 4.

154

MARIA BAMPASIDOU ET AL.

27. We also computed 90% confidence intervals employing the CLR (2013) methodology and are available upon request from the authors. Looking at the 90% confidence intervals for NATE under Proposition 4, they are [ − 0.80, 26.87]. Thus, when considering a 90% confidence level, the confidence interval almost rules out zero. This is important because ruling out zero for NATE means that the role of channels other than degree attainment in mediating the ATE of random assignment to JC on earnings is positive. 28. Of course, relative to the corresponding bounds for employment under Proposition 3 (panel E), the bounds under Proposition 4 are more informative in the sense that they are narrower. This is not surprising as Proposition 4 imposes an additional assumption. 29. Another average characteristic that is statistically different for ap relative to each of n0 and n1 is the attainment of a degree by baseline. In the case of this variable, ap’s have higher average degrees relative to n0 but lower relative to n1. 30. The 90% confidence interval for these bounds also includes zero: [ − 1.62,103.14]. 31. The 90% confidence interval for these bounds barely includes zero: [ − 0.002,0.160]. 32. Note that comparing the upper end of the confidence interval for MATE to the ATE can be misleading for obtaining statistical evidence that MATE < ATE, as one needs to take into account the sampling error in the estimation of the ATE. Thus, we focus on whether the confidence intervals for NATE exclude zero. 33. We do not present these average characteristics and tests here to save space, but they are available from the authors upon request. 34. One way this may be possible is through the use of child care services that are available through a number of JC centers. 35. For instance, this group is more likely to use the residential component of the program by not being married nor having children.

ACKNOWLEDGMENT We thank Charles Moss and Larry Kenny for comments on an early version of this work.

REFERENCES Angrist, J., Imbens, G., & Rubin, D. (1996). Identification of causal effects using instrumental variables. Journal of the American Statistical Association, 91, 444
472. Bhattacharya, J., Shaikh, A., & Vytlacil, E. (2008). Treatment effects bounds under monotonicity assumptions: An application to swan-ganz catheterization. American Economic Review: Papers and Proceedings, 98, 351
356. Blanco, G., Flores, C. A., & Flores-Lagunes, A. (2013a). The effects of job corps training on wages of adolescents and young adults. American Economic Review: Papers & Proceedings, 103(3), 418
422.

Degree Attainment in the Differential Impact of Job Corps

155

Blanco, G., Flores, C. A., & Flores-Lagunes, A. (2013b). Bounds on average and quantile treatment effects of job corps training on wages. Journal of Human Resources, 48(3), 659
751. Cai, Z., Kuroki, M., Pearl, J., & Tian, J. (2008). Bounds on direct effects in the presence of confounded intermediate variables. Biometrics, 64, 695
701. Chen, X., & Flores, C. A. (2014). Bounds on treatment effects in the presence of sample selection and noncompliance: The wage effects of job corps. Journal of Business and Economic Statistics. (forthcoming). CLR. (2013). Intersection bounds: Estimation and inference. Econometrica, 81, 667
737. Flores, C., Flores-Lagunes, A., Gonzalez, A., & Neumann, T. (2012). Estimating the effects of length of exposure to instruction in a training program: The Case of Job Corps. The Review of Economics and Statistics, 94, 153
171. Flores, C. A., & Flores-Lagunes, A. (2009). Identification and estimation of causal mechanisms and net effects of a treatment under unconfoundedness. IZA Working Paper No. 4237. Bonn, Germany. Flores, C. A., & Flores-Lagunes, A. (2010). Nonparametric partial identification of causal net and mechanism average treatment effects. Working Paper. California Polytechnic State University, San Luis Obispo, CA. Flores, C. A., & Flores-Lagunes, A. (2013). Partial identification of local average treatment effects with and invalid instrument. Journal of Business and Economic Statistics, 31(4), 534
545. Flores-Lagunes, A., Gonzalez, A., & Neumann, T. (2010). Learning but not earning? The impact of job corps training on hispanic youth. Economic Inquiry, 48, 651
667. Frangakis, C. E., & Rubin, D. (2002). Principal stratification in causal inference. Biometrics, 58, 21
29. Heckman, J., & Krueger, A. (2002). Inequality in America. What role for human capital policies. Cambridge, MA: MIT Press. Heckman, J., LaLonde, R. J., & Smith, J. A. (1999). The economics and econometrics of active labor market programs. In O. Ashenfelter & D. Card (Eds.), Handbook of labor economics (Vol. 3, pp. 1801
1863). Amsterdam, The Netherlands: Elsevier. Hirano, K., & Porter, J. (2012). Impossibility results for nondifferentiable functionals. Econometrica, 80, 1769
1790. Huber, M. (2013). Identifying causal mechanisms (primarily) based on inverse probability weighting. Journal of Applied Econometrics. (forthcoming). doi:10.1002/jae.2341 Imai, K., Keele, L., & Yamamoto, T. (2010). Identification, inference and sensitivity analysis for causal mediation effects. Statistical Science, 25(1), 51
71. Imbens, G., & Manski, C. (2004). Confidence intervals for partially identified parameters. Econometrica, 72, 1845
1857. Imbens, G. W., & Angrist, J. D. (1994). Identification and estimation of local average treatment effects. Econometrica, 62, 467
475. Kaufman, S., Kaufman, J., MacLehose, R., Greenland, S., & Poole, C. (2005). Improved estimation of controlled direct effects in the presence of unmeasured confounding of intermediate variables. Statistics in Medicine, 24(11), 1683
1702. LaLonde, R. (2003). Employment and training programs. In R. Moffitt (Ed.), Means-tested transfer programs in the United States (pp. 517
585). Chicago, IL, University of Chicago Press for National Bureau of Economic Research. Lechner, M., & Melly, B. (2010). Partial idendification of wage effects of training programs. Working Paper. University of St. Gallen, Switzerland.

156

MARIA BAMPASIDOU ET AL.

Lee, D. (2009). Training, wages and sample selection: Estimating sharp bounds on treatment effects. Review of Economic Studies, 76, 1071
1102. Mallar, C. T., Kerachslcy, S., Thornton, C., & Long, D. (1982). Evaluation of the economic impact of the job corps program: Third follow-up report. Princeton, NJ: Mathematica Policy Research, Inc. Manski, C. F. (1997). Monotone treatment response. Econometrica, 65, 1311
1334. Manski, C. F., & Pepper, J. V. (2000). Monotone instrumental variables: With an application to the returns to schooling. Econometrica, 68(4), 997
1010. Mattei, A., & Mealli, F. (2011). Augmented designs to assess principal strata direct effects. Journal of the Royal Statistical Society, Series B, 73, 729
752. McConnell, S., & Glazerman, S. (2001). National job corps study: The benefits and costs of job corps. Princeton, NJ: Mathematica Policy Research, Inc. Neyman, J. (1923). On the application of probability theory to agricultural experiments: Essay on principles (Section 9). Translated in Statistical Science, 5, 465
480. Pearl, J. (2001). Direct and indirect effects, proceedings of the seventh conference on uncertainty in artificial intelligence (pp. 411
420). San Francisco, CA: Morgan Kaufman. Petersen, M., Sinisi, S., & van der Laan, M. (2006). Estimation of direct causal effects. Epidemiology, 17, 276
284. Robins, J., & Greenland, S. (1992). Identifiability and exchangeability for direct and indirect effects. Epidemiology, 3, 143
155. Rubin, D. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66, 688
701. Rubin, D. (1980). Discussion of ‘randomization analysis of experimental data in the fisher randomization test’ by Basu. Journal of the American Statistical Association, 75, 591
593. Rubin, D. (1990). Comment: Neyman (1923) and causal inference in experiments and observational studies. Statistical Science, 5, 472
480. Rubin, D. (2004). Direct and indirect causal effects via potential outcomes (with discussion). Scandinavian Journal of Statistics, 31, 161
198. Schochet, P. (1998). National job corps study: Characteristics of youths served by job corps. Princeton, NJ: Mathematica Policy Research, Inc. Schochet, P. (2001). National job corps study: Methodological appendixes on the impact analysis. Princeton, NJ: Mathematica Policy Research, Inc. Schochet, P., Burghardt, J., & Glazerman, S. (2001). National job corps study: The impacts of job corps on participants’ employment and related outcomes. Princeton, NJ: Mathematica Policy Research, Inc. Schochet, P., Burghardt, J., & McConnell, S. (2008). Does job corps work? Impact findings from the national job corps study. American Economic Review, 98(5), 1864
1886. Sjo¨lander, A. (2009). Bounds on natural direct effects in the presence of confounded intermediate variables. Statistics in Medicine, 28, 558
571. U.S. Department of Labor. (2009). U.S. Department of labor job corps annual report, program year 2008. Washington, DC. Zhang, J., Rubin, D., & Mealli, F. (2008). Evaluating the effects of job training programs on wages through principal stratification. In T. Fomby, R. Carter Hill, D. L. Millimet, J. A. Smith, & E. J. Vytlacil (Eds.), Modelling and evaluating treatment effects in econometrics (Vol. 21, pp. 117
145). Advances in Econometrics. Bingley, UK: Emerald Group Publishing Limited.

INSECURE, SICK AND UNHAPPY? WELL-BEING CONSEQUENCES OF TEMPORARY EMPLOYMENT CONTRACTS Vincenzo Carrieri, Cinzia Di Novi,$ Rowena Jacobs and Silvana Robone ABSTRACT This paper investigates the influences of temporary contracts along several dimensions of well-being (physical and mental health, self-assessed health and happiness) for young Italian workers. Our paper contributes to the literature exploring some new aspects of the relationship between temporary jobs and well-being in a country not frequently analysed in previous literature. We focus on the gender gap in the well-being consequences of non-permanent jobs, the influence of financial support by family in reducing well-being effects caused by temporary contracts and the interaction between gender gap and family support. We find that

$

C. Di Novi was awarded the Alan Williams Fellowship for completing the research at the Centre for Health Economics (University of York).

Factors Affecting Worker Well-Being: The Impact of Change in the Labor Market Research in Labor Economics, Volume 40, 157
193 Copyright r 2014 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-912120140000040006

157

158

VINCENZO CARRIERI ET AL.

temporary contracts are damaging in terms of psychological health and happiness mostly for young men and individuals without family economic support. On the other hand, women’s mental health is not affected by temporary contracts and they are even better off in terms of their mental health and well-being when receiving family economic support. Keywords: Health; happiness; psychological well-being; young employees; fixed-term contracts JEL classifications: I12; J08

INTRODUCTION The spread of temporary jobs is one of the most important changes in European labour markets. From 1990 to 2011, the share of temporary employment in Europe increased from 11% to 15% while, in the same period, in North American countries the share of temporary employment remained stable at around 7% (OECD, 2011). A substantial literature has tried to investigate the consequences of temporary contracts on individual well-being using several indicators such as job satisfaction, life satisfaction and health. Our paper contributes to the literature by investigating the influence of temporary contracts on the health and well-being of young Italian workers and by exploring some new aspects of this relationship. Italy is an interesting case study because the spread of temporary contracts has dramatically changed job security conditions of employees. In the last two decades, the increment of the share of temporary jobs in Italy has been one of the highest in Europe (OECD, 2011). Italy increased its share of temporary employment from around 5% of total employment in 1990 to around 13% in 2011; an increase of around 150% in the period 1990
2011, compared to 40% in Europe and around 20% in OECD countries. These figures are likely to be due, in part, to a European trend in the spread of temporary employment, and in part to some important reforms which occurred in Italy in 1997, 2001 and 2003 concerning the liberalization of the use of temporary employment (laws 196/1997, 68/2001 and 30/2003). Young workers have been affected by the spread of temporary employment more than other kinds of workers. In the period 1990
2011, the share of temporary employment of 15
24 year olds increased from 11% to around 50%. These figures are notably higher than other OECD

Well-being Consequences of Temporary Contracts

159

countries (25% in 2011) and all other European countries (40% in 2011) (OECD, 2011). Many features of the Italian liberalization process are common to the reforms occurring in other European countries such as Spain and France. However, due to certain characteristics of the Italian institutional setting, labour market and welfare system, the impact of the labour reform on Italian society has been particularly acute. First, the adoption of nonpermanent positions in Italy has been more rapid: no other country in Europe had a comparable growth rate of temporary contracts. Second, the jobs characterized by the new contractual forms are paid on average less well than traditional ones. This wage policy has been adopted by firms in many European countries (Bentolila & Dolado, 1994; Hagen, 2002), but is particularly common in Italy (Elia, 2010). Third, there is evidence that in Italy after the end of a contract, an employee with a temporary contract is more likely to be hired with another temporary contract, or to become unemployed, than to be hired with a permanent contract (Garibaldi & Pacelli, 2008; Ichino & Riphahn, 2005). Lastly, welfare safety nets for nonpermanent workers are not particularly strong in Italy for two reasons. On the one hand, in Italy there are no transfers to support workers in poverty which is much more common amongst workers with fixed-term contracts (Di Bartolomeo, Di Bartolomeo, & Pedaci, 2009). On the other hand, unemployment is more costly for individuals than in other European countries because unemployment benefits are generally very low. For these reasons we expect that recent Italian labour market changes have had an impact on young workers’ physical and psychological health, and that this impact has been particularly strong in Italy compared to other European countries. We focus on the gender gap in the health consequences of non-permanent jobs, the influence of family financial support in reducing the health consequences caused by temporary contracts and the interaction between gender gap and family support. These aspects have not previously been explored in the literature and are relevant to understanding which types of individuals more likely experience well-being effects from temporary contracts. Indeed, on the one hand, it is likely that women and men react differently to temporary contracts. Women have a greater probability of separation (change occupation, geographic location) because they place a higher value on non-market opportunities such as family responsibilities, raising children, following their partners who move job location and so on (Booth, Francesconi, & Frank, 2002; Lazear & Rosen, 1990). For this reason, their mental health could be less affected than men by a job which has a fixed term because they may know they have a higher probability of changing

160

VINCENZO CARRIERI ET AL.

jobs or exiting from the labour market at some future point. On the other hand, family economic support might be relevant for temporary workers who are generally paid less than permanent ones (Bentolila & Dolado, 1994; Elia, 2010; Hagen, 2002) and they are much more likely to be poorer than other kinds of employees (Carrieri, 2012; Di Bartolomeo et al., 2009). This is particularly important in Italy, where formal welfare safety nets for non-permanent workers are weak (OECD, 2009). While previous studies on Italy have concentrated mostly on the relationship between temporary contracts and job satisfaction (De Witte & Na¨swall, 2003; Origo & Pagani, 2009; Salvatori, 2010), we investigate the influences of non-permanent jobs on a complete battery of well-being measures: physical and mental health, self-assessed health (SAH) and happiness. The empirical investigation is performed using the cross-sectional survey ‘Health Conditions and Use of the Health Service’ of the Italian population in 2004/2005 in conjunction with the Bank of Italy’s Survey on Household Income and Wealth (SHIW), which contains information on family income and wealth. The sample consists of young people between the ages of 15 and 30. We employ a propensity score (PS) matching estimator to take into account differences in observable characteristics of workers enrolled in temporary as opposed to permanent employment, we use several matching procedures and the calibrated confounder method (Ichino, Mealli, & Nannicini, 2008) to test for potential departures from the Conditional Independence Assumption (CIA).

PREVIOUS EVIDENCE ON TEMPORARY CONTRACTS AND WELL-BEING The empirical evidence regarding the effects of temporary versus permanent contracts on health is mixed. Some studies in Europe, such as Finland, Spain, Germany and Sweden (see Gash, Mertens, & Romeu Gordo, 2007; Kivima¨ki et al., 2003; Waenerlund, Virtanen, & Hammarstrom, 2011) report that workers with fixed-term contracts have worse physical health than workers with permanent contracts. In other studies, fixed-term contracts have been shown to have no effects, or even positive ones, on health (Sverke, Gallagher, & Hellgreen, 2000; Virtanen, Kivimaki, Elovainio, Vahtera, & Ferrie, 2003). More recently, Ehlert and Schaffner (2011) have used the European Union Statistics on Income and Living Conditions (EU-SILC), a panel dataset (2004
2008) comprising 27 European

Well-being Consequences of Temporary Contracts

161

countries. In most of the countries, including Italy, employees with a permanent contract do not appear to report better health than those with a temporary contract; however repeated temporary contracts show a significant negative impact on SAH. With respect to psychological well-being, fixed-term employment is traditionally assumed to have a negative effect. Fixed-term contracts are considered stressful since they imply job insecurity (Bohle, Quinlan, & Mayhew, 2001; Burchell, 1994, 1999), not enabling workers to plan and control their lives given the short-term nature of their jobs. This traditional assumption is confirmed by several studies (Klein-Hesselink, & van Vuuren, 1999; Lasfargues et al., 1999; Quesnel-Vallee, DeHaney, & Ciampi, 2010). It is unlikely, however, that fixed-term contracts have the same impact on all workers. Characteristics at individual level, such as tolerance for ambiguity and self-monitoring, play a relevant role in influencing responses to stress and the selection process into permanent employment (Bauer & Truxillo, 2000). Therefore, it is not surprising to find studies in the literature not supporting the traditional assumption of the negative impact of fixed-term employment on psychological well-being (Cottini & Lucifora, 2010; Sverke et al., 2000). In general, recent evidence appears to suggest that workers with fixedterm contracts cannot be considered a homogeneous group when comparing their health and well-being with that of workers with permanent contracts. The relationship between fixed-term contracts and psychological well-being has been shown to be mediated by factors such as the level of employability and prospect of finding another job (Silla, Gracia, & Peiro, 2005), level of work control and choice (such as self-scheduling or gradual/partial retirement) (Joyce, Pabayo, Critchley, & Bambra, 2010) and dislike of uncertainty (Natti, Kinnunen, Makikangas, & Mauno, 2009), which may in turn be mediated by the individual’s preference for risk. Robone, Jones, and Rice (2011) investigate the influence that contractual conditions have on both SAH and psychological well-being of employees using 12 waves of the British Household Panel Survey. Their analysis reveals a negative relationship between health and psychological well-being and having a fixed-term contract, compared to having a permanent one, but these effects are much smaller for workers characterized by a high level of employability. In the literature investigating the links between labour conditions and ‘happiness’, several studies have shown a large negative effect of unemployment on individual happiness (Clark, Georgellis, & Sanfey, 2001; Ferrer-i-Carbonell & Gowdy, 2007; Frey & Stutzer, 2000; Winkelmann & Winkelmann, 1998). However, very little evidence has been reported on

162

VINCENZO CARRIERI ET AL.

the influence of contractual conditions on happiness (Dolan, Peasgood, & White, 2008). The studies of Scherer (2009) and Ponzo (2011) use, respectively, the 2004 ESS data (which involves 16 Western European countries) and three waves of the Bank of Italy’s SHIW. These studies suggest that the reported level of happiness is lower for workers with fixed-term contracts. In both Northern and Southern European countries women are more likely than men to have temporary contracts (Artazcoz et al., 2013; Bildt & Michelsen, 2002; Naldini & Jurado, 2013). However, in the literature there is mixed evidence about the existence of a gender gap in the well-being consequences of such contracts. Some studies report a stronger negative effect for women than men of temporary contracts on SAH (Bauer, Huber, Jenny, Mu¨ller, & Ha¨mmig, 2009) and psychological well-being (Bildt & Michelsen, 2002; Plaisier et al., 2007), while some studies do not find statistically different effects between women and men (Cheng & Chan, 2008; Ferrie, Shipley, Stansfeld, & Marmot, 2002). Other studies provide evidence that job insecurity is related to negative health (Robone et al., 2011) and well-being of men but not of women (De Witte, 1999). When considering the possible gender gap in the well-being consequences of non-permanent jobs, recent literature reports a still high degree of consistency in the gender division of work. Gender differences and inequalities in paid and unpaid work still persist, with females having primary responsibility for household and domestic labour and males having a primary role in paid work (Arcas, Novoa, & Artazcoz, 2013; Artazcoz, Borrell, Cortes, Escriba-Aguir, & Cascant, 2007; Berntsson, Lundberg, & Krantz, 2006; Doyal, 2000; Harryson, Novo, & Hammarstrom, 2012). This gender division of labour permeates all levels and spheres of society (Artazcoz et al., 2007). Gender differences are particularly marked in Southern European countries, which appear to be characterized by a strong ‘familialism’, with an asymmetrical gender division of work and low female participation in the labour market (Naldini & Jurado, 2013). This evidence appears to hold for Italy as well, where women’s labour participation is amongst the lowest in Europe (13% below the EU average) and the provision by the public sector of child and elderly care is very limited (Del Boca & Vuri, 2007). The role of family financial support is particularly pertinent for young workers in the Italian labour market context since it is characterized by a persistently high youth unemployment rate. A number of school-leavers and graduates are unemployed for quite some time after leaving initial education, and even those who do find a job immediately, frequently have a flexible and insecure labour market position. Both cultural and social

Well-being Consequences of Temporary Contracts

163

factors, but particularly in more recent years, economic factors, have played a key role in observing late transitions into adulthood in Italy. The lack of job security prevents any forward planning in particular amongst young people who have come to the end of a training period and for whom work represents a central factor in their lives (Lewis, Smithson, & Brannen, 1999). Due to the job insecurity and weak welfare system that does not offer financial support to young people, many young Italian workers tend to depend on their parents, even at ages at which their European counterparts have generally managed to achieve an independent lifestyle (Schro¨der, 2008). According to the Italian Statistical Office the median age for leaving home for Italian males is about 27 years old (Famiglia e soggetti sociali, 2011).

METHODS From a methodological point of view, it should be noted that workers with a temporary contract may differ substantially from workers with a permanent one in several ways; for instance, temporary workers may be selected into temporary jobs as a result of poor health conditions. This potential endogeneity problem can be corrected by matching each temporary worker (the ‘exposed/treated’) with a permanent worker (the ‘control/untreated’) on each characteristic known to be associated with employment conditions and health (Caliendo & Kopeinig, 2008). In our analysis we performed this matching by using PS matching, as formalized by Rosenbaum and Rubin (1983). PS matching has important advantages over regression-based methods. Being a non-parametric method, matching does not impose any specific linearity assumptions on the evaluated effects that are inherent in regression-based modelling. Furthermore, matching explicitly tries to find for each untreated individual a similar treated individual to evaluate the counterfactual, that is what would happen to the treatment group without the treatment. The PS matching technique produces two balanced groups, one of permanent workers and one of temporary workers: the score substitutes a collection of confounding variables with a single covariate that is a function of all the variables. The PS can be considered as a balancing score, meaning that amongst subjects with the same propensity to be exposed, treatment is conditionally independent of the covariates. By summarising the intrinsic characteristics that could generate distortions, PSs use a

164

VINCENZO CARRIERI ET AL.

matching procedure to allow for comparisons between the treated and control groups. Analytically, this method calculates an index e(X) for each worker, as a function of confounders (X) and represents the conditional probability of being a temporary worker, given all observable individual characteristics:

eðXÞ = PðI = 1 XÞ

ð1Þ

where I = 1 denotes that the individual belongs to the temporary worker group. We invoke the common support modelling option which restricts the set of data points over which the test of the balancing property is sought to those belonging to the intersection of the supports of the PS of treated and controls. The balancing test is performed in Stata 12, using the psmatch2 routine developed by Leuven and Sianesi (2003). Imposing the common support condition in the estimation of the PS may improve the internal validity of the estimates under common support (Caliendo & Kopeinig, 2008). We first compute the PS through a probit model. The estimation was carried out using the p-score program for STATA 12 developed by Becker and Ichino (2002). The dependent variable is a dummy indicator for the type of contract which equals 0 if the individual has a permanent contract and 1 if he/she has a temporary one. The reference individual in the model is male, lives in the South of Italy and is married with children. He is aged 15 years or more, is a blue-collar worker in the primary industry and has a secondary school certificate. Second the matching is carried out through algorithms which form ‘statistical twins’ that differ only in their contract status and not in other observed characteristics in order to account for self-selection. Thus, for a given PS, exposure to treatment is effectively random and treatment and control groups should on average be observationally identical. We test the robustness of our results by performing the matching with different algorithms, including the Nearest Neighbour, Radius (with caliper size 0.01) and Kernel Matching (Caliendo & Kopeinig, 2008; Imbens & Wooldridge, 2009). Sensitivity analysis was carried out on radius caliper size (0.01
0.05). Finally, health and well-being of matched individuals is then compared to estimate the average effect of working with a temporary contract instead of a permanent one. Specifically, we are interested in the average effect of the treatment on the treated (ATT), that is the difference between the health outcomes for workers with temporary jobs with respect to

Well-being Consequences of Temporary Contracts

165

the counterfactual unobservable outcome which would have prevailed for them if they had a permanent job. Analyses are carried out on the full sample and stratified by financial support categories and reported separately by gender. By stratifying the sample across gender and financial support, we allow for a better selection on observable characteristics which leads to more accurate results.

DATA The empirical investigation was performed using the cross-sectional survey ‘Health Conditions and Use of the Health Service’ of the Italian population in 2004/2005. The survey is part of the Multiscopo survey which is carried out every five years by the Italian National Institute of Statistics (ISTAT Multiscopo Survey, 2007) since 1990, using direct interviews of a sample of families living across Italy. Since the question regarding the type of employment contract of respondents was introduced in the fourth wave only, we consider only wave 2004/2005. A cross-sectional study involves the evaluation of the characteristics of individuals at the same point in time: their employment contract, and their health and happiness. This makes it impossible to disentangle the temporal sequencing of temporary contracts on individual well-being. In order to unravel the temporal sequences, panel data are needed. However panel data of this type are not available in the Italian context, and where they do exist (the Bank of Italy’s SHIW for instance), they lack information on individuals’ well-being. Thus we used the cross-sectional survey of the Italian population in 2004/2005 since it provides detailed information about respondents’ subjective and objective health status and happiness. The survey ‘Health Conditions and Use of the Health Service’ does not contain any information about family income. It was, therefore, necessary to integrate the Istituto Nazionale di Statistica (ISTAT) dataset with the Bank of Italy’s SHIW 2004, which contains information on family income and wealth. To integrate the two datasets we used a statistical matching technique. To perform this matching we made use of demographic and socio-economic information present in both the ISTAT and the Bank of Italy’s surveys. Specifically, we impute income of an individual from the SHIW to a similar individual from the Multiscopo. The matching procedure used for linking the two datasets is provided in Appendix A.

166

VINCENZO CARRIERI ET AL.

The analysis was conducted on the sub-sample of employed individuals only. We also restricted our sample to young people between the ages of 15 (the minimum legal age to start working in Italy in 2005) and 30. After deleting records with missing values, we obtained a final sample of 8,280 observations. The sample was stratified into workers with a permanent or temporary contract; 1,840 workers (22%) had temporary jobs. Our sample includes both part-time and full-time workers but we do not perform a separate analysis on the effects of temporary contracts on part-time and full-time workers since we have too few part-time workers in our sample (6%) and they are concentrated amongst females with a permanent job (80%). A definition of all variables used in our analysis is listed in Table 1. We use four dependent variables of health status: SAH, the Physical Component Score (PCS) and the Mental Component Score (MCS) from the Health Related Quality of Life instrument Short Form (SF-12), and an indicator of happiness. The following standard SAH status question was asked: ‘Would you say that in general your health is: excellent, very good, good, fair, poor’. SAH was therefore measured on a five-point scale from ‘excellent’ (score 5) to ‘poor’ (score 1) and was treated as an ordered categorical variable. The use of SAH as an indicator of health status is supported by evidence which shows a strong predictive relationship between people’s self-rating of health and mortality or morbidity (Idler & Benyamini, 1997; Kennedy, Kawachi, Glass, & Prothrow-Stith, 1998). Moreover, SAH correlates strongly with more complex health indices such as functional ability or indicators derived from health service use (Unde´n & Elofsson, 2006). The PCS and MCS were obtained from the SF-12. The SF-12 is a multipurpose short form survey with 12 questions which has been developed to provide a shorter, yet valid alternative to the full SF-36 health survey (Ware, Kosinski, & Keller, 1996). Results obtained for scales of SF-12 received a numerical score that was standardized and ranked on a scale from 0 to 100, with higher values associated with better health. Very low scores (under 20) on PCS correspond to ‘substantial limitations in taking care of oneself and in physical, social and personal activity; important physical pain; frequent tiredness; poor health’. A low value on MCS indicates ‘frequent mental trouble; important social and personal trouble due to emotional problems; poor mental health’. In our sample the PCS and MCS varied from 19.47 to 68.4 and from 8.54 to 68.7 respectively. Happiness is measured on a five-point scale by the following question: ‘Would you describe yourself as being: happy and interested in life (score 5); somewhat happy; somewhat unhappy; very unhappy; so unhappy that life is

Well-being Consequences of Temporary Contracts

Table 1.

167

Definition of Variables used in the Propensity Score Model.

Variable Name Outcome variables Self-assessed health Happiness PCS MCS Controls Age Gender Work experience Income Injuries Education Primary school

Variable Definition 1 if ‘poor’, 2 if ‘fair’, 3 if ‘good’, 4 if ‘very good’, 5 if ‘excellent’ health 1 ‘so unhappy that life is not worthwhile’ to 5 ‘happy and interested in life’ Physical Component Score
Health Related Quality of Life instrument SF-12 Mental Component Score
Health Related Quality of Life instrument SF-12 Age in years at 1st December of current wave 1 if female, 0 otherwise 1 if more than 10 years of working experience, 0 otherwise log of annual labour income (in Euros) 1 if injuries in the previous six months serious enough to limit normal activities, 0 otherwise

Lower high school High school University or postgraduate

1 if no educational certificates or if primary school certificate, 0 otherwise 1 if lower high school certificate, 0 otherwise 1 if high school certificate, 0 otherwise 1 if university or postgraduate degree, 0 otherwise

Marital status Married Previously married Never married

1 if currently married, 0 otherwise 1 if previously married, 0 otherwise 1 if never married, 0 otherwise

Family composition Living with both parents Living with a partner Living with a single parent Living alone Industry Primary industry Secondary industry Tertiary industry Occupation Business executive Supervisor White collar Blue collar

1 if living with both his/her parents, 0 otherwise 1 if living with husband/wife or cohabiting with partner, 0 otherwise 1 if living in a single-parent home, 0 otherwise 1 if living alone, 0 otherwise 1 if primary (agriculture), 0 otherwise 1 if secondary (manufacture, construction), 0 otherwise 1 if tertiary (accommodation, restaurants, transport etc.), 0 otherwise 1 if business executive, 0 otherwise 1 if supervisor and intermediate decision posts, 0 otherwise 1 if white collar, 0 otherwise 1 if blue collar, 0 otherwise

168

VINCENZO CARRIERI ET AL.

Table 1.

(Continued )

Variable Name Macro-region North Centre South Islands

Variable Definition 1 if North, 0 otherwise 1 if Centre, 0 otherwise 1 if South, 0 otherwise 1 if Islands, 0 otherwise

Stratifying variable Financial support by family 1 if receive financial support from family as main source of income, 0 otherwise

Table 2. Self-assessed health PCS MCS Happiness

Correlation Coefficients of the Health Variables. Self-Assessed Health

PCS

MCS

Happiness

1 0.4326 0.361 0.266

1 −0.0633 0.0652

1 0.3749

1

not worthwhile (score 1)’. It is treated as an ordered categorical variable. General happiness is a measure of subjective well-being which has been examined extensively in the literature (George, 2006; Veenhoven, 1996). Studies consistently show a strong relationship between happiness and both physical and psychological health (Dolan et al., 2008; Graham, 2008). Psychological health appears to be more highly correlated with happiness than physical health (see Table 2), but this is not surprising given the close correspondence between mental health, well-being and happiness. Potential confounding factors which could be associated with both health and employment conditions include age, gender, work experience, individual income, activity-limiting injuries, education, marital status, family composition, industry and occupation and geographical region. Age was modelled as a continuous variable. Work experience was measured with a binary indicator that takes the value one if respondents have more than 10 years of working experience, zero otherwise. We tested shorter intervals for work experience such as 2 or 3 years but these were not significant. Income was log-transformed to obtain a normal distribution and modelled as a continuous variable. Activity-limiting injuries is a binary variable that takes the value one if respondents reported that they

Well-being Consequences of Temporary Contracts

169

suffered injuries in the previous six months which were serious enough to limit their normal activities, zero otherwise. The International standard classification of education (Isced) was used to classify the education variable. Isced is classified into 7 levels: Isced 0 (preprimary schooling); Isced 1 (primary education); Isced 2 (lower secondary); Isced 3 (upper secondary); Isced 4 (post high school); Isced 5 (university); Isced 6 (postgraduate). In the analysis Isced levels 0, 1 and 2, 3 and 5, 6 have been grouped together, respectively. Four levels of education are therefore considered: (1) no educational certificates or primary school certificate, (2) lower high school (secondary) certificate, (3) high school graduation and (4) university degree or postgraduate. Marital status was categorized into never, currently, or previously married. Family composition was defined with reference to the family nucleus. These were defined as living with both parents, living in a single-parent home, living with a partner, or living alone. Based on the nature of products, industries were categorized into primary (agriculture), secondary (manufacture, construction), or tertiary (accommodation, restaurants, transport, communication, real estate, business services, schools, finance, insurance, data processing, law, health care). Occupations were categorized into four groups: business executive and assimilated, supervisor and intermediate decision posts, white-collar and blue-collar workers. Labour market conditions may be very different amongst Italian regions because of the economic and social dualism between the more economically developed Northern regions and the less developed Southern ones. Southern Italy is Europe’s principal empirical case study of failed modernization and is often used by researchers as a case study for corruption, crime and underdevelopment (Chubb, 1982; Davis & Marino, 2000; Micali, 2009). We control for the local labour market context using a set of geographic dummies for the macro-regions. We also consider individuals who receive financial support from their families. We do not include this as a covariate, but use it to further stratify our sample to examine differences between individuals who rely on employment versus those who rely on family maintenance as their main source of income. While we do not know the magnitude of family economic support, we are able to use this binary variable for whether they received any support at all, to further disentangle the impact of contract status on health. The following item was used to identify respondents’ main source of financial support: ‘Which one of these is your main source of income?’

170

VINCENZO CARRIERI ET AL.

Four options were made available in the Multiscopo survey for the category of workers we have considered in our paper: (1) wages or salary from employment, (2) family maintenance, (3) annuities and (4) ‘other’. We focus on groups 1 and 2 in this part of the analysis. We compute the PS through a probit model for those with employment as their main source of income and those with family maintenance as their main source of income using the same specification as described in the section ‘Methods’.1

RESULTS Table 3 presents summary statistics of the health status of workers with a permanent and temporary contract as well as the demographic and socioeconomic factors in the final sample. As shown in Table 3, women are more likely to have a temporary contract than men. Table 3 shows that being young, never married, having a university degree, having less work experience and living in the South or Island regions are characteristics associated with having a temporary contract. We also see that those on temporary contracts are more likely to receive financial support from their families. We present descriptive statistics in Table 4 by gender. Males tend to have higher levels of SAH, happiness and physical and mental well-being. They are also likely to have more work experience and higher income. Table 5 presents the descriptive statistics as stratified by those who receive financial support from their family as their main source of income. Younger females, living in the South of Italy with a higher degree of education who have never been married and live with their parents are more likely to receive financial support.

Overall Propensity Score Matching Results by Gender The covariates (not presented but available on request) for the PS estimation show that being employed on a temporary contract is positively associated with being younger and female, having a higher education and a lower income. In terms of job characteristics, working as a blue-collar worker in the primary industry and living in the Southern regions is also positively associated with a temporary job. When we split the sample

171

Well-being Consequences of Temporary Contracts

Table 3.

Descriptive Statistics of Variables in the Propensity Score Model, by Contract Type.

Variables

Full Sample

Temporary Job

Permanent Job

Mean

Std. Dev.

Mean

Std. Dev.

Mean

Std. Dev.

Outcome variables Self-assessed health Happiness PCS MCS

4.217 4.523 54.478 52.310

0.669 0.577 5.139 8.139

4.207 4.480 54.494 51.784

0.676 0.614 5.356 8.127

4.220 4.535 54.473 52.461

0.667 0.566 5.075 8.136

Controls Age Gender Work experience Income Injuries

26.424 0.4306 0.058 12031 0.97

3.489 0.495 0.235 5620 0.171

25.485 0.484 0.027 11537 0.972

3.66 0.499 0.161 5939 0.164

26.693 0.415 0.068 12172 0.9693

3.392 0.493 0.251 5517 0.173

Education Primary school Lower high school High school University or postgraduate

0.032 0.326 0.532 0.112

0.175 0.468 0.499 0.315

0.043 0.298 0.482 0.177

0.203 0.458 0.499 0.382

0.028 0.332 0.546 0.093

0.166 0.471 0.498 0.291

Marital status Married Previously married Never married

0.202 0.027 0.771

0.402 0.162 0.42

0.151 0.022 0.828

0.358 0.146 0.378

0.217 0.028 0.755

0.412 0.166 0.43

Family composition Living with both parents Living with a partner Living with a single parent Living alone

0.658 0.119 0.105 0.117

0.474 0.325 0.307 0.321

0.691 0.078 0.119 0.113

0.462 0.268 0.324 0.316

0.649 0.132 0.102 0.118

0.477 0.338 0.302 0.322

Industry Primary industry Secondary industry Tertiary industry

0.033 0.375 0.592

0.178 0.484 0.491

0.087 0.284 0.629

0.282 0.451 0.481

0.017 0.4 0.067

0.131 0.49 0.251

Occupation Business executive Supervisor White collar Blue collar

0.006 0.022 0.422 0.551

0.074 0.146 0.494 0.497

0.006 0.026 0.409 0.559

0.077 0.158 0.491 0.497

0.005 0.021 0.425 0.548

0.073 0.143 0.494 0.498

Macro-region North Centre

0.497 0.176

0.5 0.381

0.408 0.16

0.492 0.367

0.523 0.181

0.499 0.384

172

VINCENZO CARRIERI ET AL.

Table 3. Variables

(Continued )

Full Sample

Temporary Job

Permanent Job

Mean

Std. Dev.

Mean

Std. Dev.

Mean

Std. Dev.

South Islands

0.241 0.086

0.428 0.28

0.308 0.125

0.462 0.33

0.221 0.075

0.416 0.263

Stratifying variable Financial support by family

0.049

0.216

0.121

0.327

0.028

0.165

N

8,280

1,840

6,440

between male and female the results remain very similar. The common support restriction is implemented by deleting temporary workers whose PS is higher than the maximum or lower than the minimum PS of the matched permanent workers. Covariate balancing tests reported in Table 6 (a) show that the matching is effective in removing differences in observable characteristics between temporary and permanent workers. In particular, the median absolute bias is reduced by approximately 68
92% depending on the matching technique. The Pseudo R2 after matching is always close to zero, correctly suggesting that the covariates have no explanatory power in the matched samples. Table 7 presents the results for the ATT. We show results for the ATT using the three matching methods: Nearest Neighbour, Radius and Kernel Matching. Results were not sensitive to radius caliper size. Starting from the full sample and with the first estimator we find that temporary employment has a significant and negative influence on the individuals’ well-being: in particular it has a negative influence on SAH, happiness and the MCS. Concerning the interpretation of the coefficients, it should be noted that SAH is an ordered categorical variable that has five possible levels with a natural ordering (poor, fair, good, very good, excellent), scaled from 1
5. When converting a measure such as SAH to a quantitative scale, it is necessary to exercise caution in the interpretation of the difference between values. For instance, the difference between the responses excellent (1) and very good (2) may not be the same as the difference between very good (2) and good (3). For this reason, we provide only a qualitative interpretation. Happiness is also an ordered categorical variable for which we provide only a qualitative interpretation. Concerning happiness, the Kernel and Radius Propensity Score matching provide very similar results: the ATT remains negative and statistically significant.

Table 4.

Descriptive Statistics of Variables in the Propensity Score Model, by Gender.

Variables

Female

Male

Mean

Std. Dev.

Mean

Std. Dev.

Outcome variables Self-assessed health Happiness PCS MCS

4.109 4.507 54.111 50.719

0.670 0.597 5.485 8.891

4.299 4.536 54.754 53.513

0.657 0.562 4.843 7.2933

Controls Age Work experience Income Injuries

26.62 0.048 11353 0.965

3.358 0.213 5788 0.183

26.27 0.067 12543 0.973

3.579 0.249 5433 0.161

Education Primary school Lower high school High school University or postgraduate

0.018 0.240 0.582 0.159

0.132 0.427 0.493 0.367

0.042 0.388 0.494 0.076

0.201 0.487 0.5 0.264

Marital status Married Previously married Never married

0.256 0.036 0.708

0.436 0.186 0.455

0.161 0.020 0.818

0.368 0.141 0.386

Family composition Living with both parents Living with a partner Living with a single parent Living alone

0.637 0.153 0.108 0.102

0.481 0.36 0.31 0.303

0.674 0.095 0.103 0.127

0.469 0.293 0.305 0.333

Industry Primary industry Secondary industry Tertiary industry

0.029 0.221 0.750

0.168 0.415 0.433

0.036 0.491 0.473

0.185 0.499 0.499

Occupation Business executive Supervisor White collar Blue collar

0.005 0.021 0.574 0.399

0.071 0.144 0.494 0.489

0.006 0.022 0.307 0.665

0.077 0.148 0.461 0.472

Macro-region North Centre South Islands

0.543 0.185 0.202 0.068

0.498 0.389 0.402 0.253

0.463 0.168 0.269 0.099

0.499 0.374 0.444 0.299

Stratifying variable Financial support by family

0.057

0.232

0.043

0.203

N

3,565

4,715

174

Table 5.

VINCENZO CARRIERI ET AL.

Descriptive Statistics of Variables Stratified by Family Financial Support.

Variables

Financial Support

No Financial Support

Mean

Std. Dev.

Mean

Std. Dev.

23.916 0.502

3.676 0.501

26.551 0.426

3.433 0.495

Education Primary school Lower high school High school University or postgraduate

0.015 0.291 0.569 0.126

0.121 0.455 0.496 0.332

0.032 0.327 0.530 0.110

0.177 0.469 0.499 0.313

Marital status Married Previously married Never married

0.069 0.022 0.909

0.254 0.147 0.288

0.208 0.027 0.764

0.406 0.163 0.424

Family composition Living with both parents Living with a partner Living with a single parent Living alone

0.833 0.034 0.099 0.034

0.374 0.183 0.298 0.183

0.649 0.124 0.106 0.121

0.477 0.329 0.307 0.326

Macro-region North Centre South Islands

0.303 0.118 0.419 0.160

0.46 0.323 0.494 0.367

0.508 0.179 0.231 0.082

0.499 0.384 0.422 0.274

Age Gender

N

406

7,843

The SF-12 indicator provides a more readily interpretable numerical score on a scale from 0 to 100 for physical and mental health. While the ATT for PCS is not statistically significant, MCS decreases by 0.68 percentage points. The Kernel and Radius Propensity Score matching provide similar results with a negative ATT and a decrease of the mental health score of between 0.64 and 0.72 percentage points. The fact that the three mental health estimates are almost identical provides evidence of their robustness. The decrease of 0.68 percentage points corresponds to 8.4% of a standard deviation in the full sample which is around 1% from the mean. We also use formal sensitivity analyses to assess the robustness of our results to different specifications of the PS model and to possible deviations

175

Well-being Consequences of Temporary Contracts

Table 6.

Results of Covariate Balancing Tests.

Probit Pseudo R2 Before Matching

Probit Pseudo Median Bias R2 after Before Matching Matching

Median Bias after Matching

% Reduction in Median Bias

(a) Temporary job position Nearest neighbour matching Full sample 0.085 Female 0.084 Male 0.088

0.005 0.011 0.009

10.7 9.5 10.3

3.4 3.1 3.2

0.682 0.674 0.689

Radius matching Full sample Female Male

0.085 0.083 0.088

0.001 0.002 0.001

10.7 9.4 10.3

0.9 1.1 0.9

0.916 0.883 0.913

Kernel matching Full sample Female Male

0.085 0.083 0.088

0.001 0.002 0.001

10.7 9.4 10.3

0.8 0.8 0.8

0.925 0.915 0.922

(b) Job as main source of income Nearest neighbour matching Full sample 0.078 Female 0.076 Male 0.084

0.007 0.008 0.006

10.3 9.6 11.3

3.3 3.5 2.3

0.68 0.635 0.796

Radius matching Full sample Female Male

0.078 0.076 0.084

0.001 0.002 0.001

10.3 9.6 11.3

0.9 0.9 1.2

0.913 0.906 0.894

Kernel matching Full sample Female Male

0.078 0.076 0.084

0.001 0.002 0.001

10.3 9.6 11.3

0.7 1.1 0.7

0.932 0.885 0.938

(c) Family as main source of income Nearest neighbour matching Full sample 0.087 Female 0.132 Male 0.108

0.039 0.066 0.065

5.2 8.1 9

2.7 4.3 4.2

0.481 0.469 0.533

Radius matching Full sample Female Male

0.087 0.132 0.107

0.007 0.043 0.019

5.2 8.1 9.1

2.3 4 3.7

0.558 0.506 0.593

Kernel matching Full sample Female Male

0.087 0.132 0.107

0.007 0.043 0.034

5.2 8.1 9.1

1.8 3.5 2.5

0.654 0.568 0.725

176

Table 7.

Average Treatment Effects (ATT) of Propensity Score Matching under Nearest Neighbour, Radius and Kernel Methods.

(a) Self-Assessed Health (SAH) Full Sample Female

Male

(b) Physical Component Score (PCS) Full Sample

Female

Male

(c) Mental Component Score (MCS) Full Sample Female

Male

(d) Happiness Full Sample

Female

Male

−0.574 −0.625* (0.479) (0.355)

−0.068*** (0.021)

−0.056* −0.077*** (0.033) (0.028)

Radius propensity score matching (standard errors in parenthesis) −0.026 −0.014 −0.018 −0.018 0.013 −0.019 (0.018) (0.026) (0.025) (0.142) (0.216) (0.187)

−0.719*** (0.218)

−0.477 −0.654** (0.346) (0.266)

−0.054*** (0.016)

−0.030 (0.024)

−0.073*** (0.022)

Kernel propensity score matching (standard errors in parenthesis) −0.039** −0.031 −0.044* −0.150 −0.187 −0.096 (0.017) (0.028) (0.025) (0.149) (0.244) (0.174)

−0.643*** (0.217)

−0.355 −0.808*** (0.342) (0.267)

−0.050*** (0.019)

−0.029 (0.025)

−0.068*** (0.022)

*Statistical significance at the 10% level. **Statistical significance at the 5% level. ***Statistical significance at the 1% level.

VINCENZO CARRIERI ET AL.

Nearest Neighbour Propensity Score Matching (Standard Errors in Parenthesis) −0.051** −0.051 −0.054* −0.194 −0.488* −0.186 −0.680** (0.024) (0.036) (0.032) (0.178) (0.284) (0.235) (0.290)

Well-being Consequences of Temporary Contracts

177

from the CIA. In order to investigate the effect of potential departures from the CIA on our estimated ATTs, we employed the calibrated confounder method proposed by Ichino et al. (2008). Sensitivity checks reported in Appendix B show that results are robust. Consistent with previous literature, our analysis reveals a negative relationship between psychological well-being and having a temporary job (compared to having a permanent job) (see Robone et al., 2011). These effects are mainly driven by men. In fact, when analysing the temporary job influence on health outcomes and happiness in particular, splitting the sample into males and females, the effect remains strong and statistically significant for men but not for women. The results are also much stronger and more consistent for mental health (MCS) and happiness as opposed to physical health (PCS) or SAH.

Propensity Score Matching Results Stratified by Family Financial Support We further investigate the question of whether temporary contract influences on well-being, differ between young people whose main source of income is family maintenance and their counterparts who rely on their employment. All observed controls used in the PS matching analysis satisfy the balancing property again (see Table 6(b) and (c)). The first two rows of Table 8 show the ATT for the group of young workers whose main source of income is employment while the second two rows include the ATT for those who receive family maintenance as their main source of income. For the first group the ATT is similar to the baseline estimation: as before, having a temporary job compared to a permanent one seems to have adverse influences on psychological well-being and happiness, in particular for young male workers. The most interesting result arises from the group of young workers who are economically dependent on their parents. There is an asymmetry between women and men. While the effect of having a temporary contract is no longer statistically significant for men, young women who have a temporary job and who receive financial support from their family seem to enjoy good mental health compared to their permanent worker counterparts. The increase of between 3.76 and 5.645 percentage points on the MCS corresponds to between 42% and 63% of a standard deviation in the female sample or between 7.5% and 11% increase over the mean.

Average Treatment Effects (ATT) of Propensity Score Matching with and without Family Financial Support.

(a) Self-Assessed Health (SAH) Full Sample

Female

Male

(b) Physical Component Score (PCS) Full Sample

Female

Male

(c) Mental Component Score (MCS) Full Sample

Female

Nearest Neighbour Propensity Score Matching (Standard Errors in Parenthesis) Job as main source of income −0.076*** −0.050 −0.071** −0.443** −0.356 0.102 −0.608** −0.889* (0.025) (0.038) (0.033) (0.187) (0.291) (0.247) (0.306) (0.498) Family as main source of income 0.036 0.182 −0.031 0.274 −0.348 −0.804 1.364 5.645** (0.107) (0.143) (0.149) (0.739) (0.922) (0.812) (1.489) (2.275)

Kernel propensity score matching (standard errors in parenthesis) Job as main source of income −0.034** −0.029 −0.035 −0.131 −0.177 −0.067 −0.705*** −0.486 (0.017) (0.028) (0.024) (0.177) (0.205) (0.147) (0.264) (0.374) Family as main source of income −0.019 0.051 −0.106 −0.506 −0.062 −0.851 1.855* 3.961** (0.091) (0.106) (0.144) (0.552) (0.834) (0.870) (1.103) (1.589) *Statistical significance at the 10% level. **Statistical significance at the 5% level. ***Statistical significance at the 1% level.

Full Sample

Female

Male

−1.060*** (0.365)

−0.071*** (0.022)

−0.050 (0.034)

−0.075*** (0.029)

−0.625** (0.283) −0.924 (1.004)

−0.822*** (0.241) 0.560 (1.472)

0.035 (0.101)

−0.060*** (0.017) 0.098 (0.063)

−0.054*** (0.018) 0.108 (0.068)

0.224 (0.139)

−0.042* (0.025) 0.189* (0.097)

−0.034 (0.027) 0.169* (0.095)

0.188 (0.144)

−0.074*** (0.023) 0.039 (0.084)

−0.070*** (0.026) 0.063 (0.100)

VINCENZO CARRIERI ET AL.

Radius propensity score matching (standard errors in parenthesis) Job as main source of income −0.023 −0.015 −0.013 −0.012 −0.026 0.028 −0.736*** −0.569 (0.019) (0.027) (0.026) (0.149) (0.228) (0.196) (0.230) (0.365) Family as main source of income −0.047 −0.252 −0.094 −0.520 0.074 −0.725 0.663 3.760** (0.069) (0.729) (0.095) (0.508) (0.100) (0.726) (0.907) (1.510)

(d) Happiness

Male

0.219 (1.549)

178

Table 8.

Well-being Consequences of Temporary Contracts

179

This result is particularly interesting as it shows how the combined effect of preferences towards career flexibility and ‘family welfare’ may result in a completely different reaction between men and women to temporary contracts. Indeed, men who have responsibility for being the main economic provider in a family are worse off with temporary contracts but when the family provides economic support they do not seem to suffer well-being losses from having a temporary job. On the other hand, women who might have preferences towards career flexibility are not worse off with temporary contracts compared to permanent ones and they turn out to be better off when receiving family economic support.

CONCLUSIONS The spread of temporary contracts has dramatically changed job security conditions of young employees in Italy. In the last two decades, the increment of the share of temporary jobs in Italy has been one of the highest in Europe. Up to now, there has been little evidence on the relationship between atypical job status and health of Italian employees. In this paper we have tried to fill this gap investigating also new aspects of the relationship between contractual conditions and a wide range of well-being measures. Consistent with a substantial literature in this field, our analysis reveals a negative relationship between psychological well-being and having a temporary job and it shows that these effects are driven by men and are much stronger for mental than physical health. A key contribution of our results is that it reveals an interesting asymmetry between men and women in their response to job-related stress from temporary employment. These results may be due to gender-related differences in preferences towards career flexibility. This interpretation may be particularly appropriate given the well-established male work culture present in Italy, as in many other Southern European Countries (such as Greece and Spain). As argued by Akerlof and Kranton (2000), social customs and gender identity may influence the distribution of time spent on ‘market sector’ work, house work and childcare. Men tend to be focused on employed work and often have responsibility for being the main economic provider in a family; in this sense temporary jobs involve higher psychological distress and less happiness compared with permanent ones. On the other hand, women place a higher value on non-market activities such as family responsibilities and raising children (Booth et al., 2002;

180

VINCENZO CARRIERI ET AL.

Lazear & Rosen, 1990). Plagnol and Easterlin (2008) find that women’s aspirations for a happy marriage and family life are higher than those of men during early adulthood. Women’s mental and psychological health may thus be less affected than men by a job which has a fixed term because fixed-term contracts may allow women to better realize their aspirations for a happy family life which is more highly valued during early adulthood. Such differences in preferences towards career flexibility have been advanced primarily to explain the gender gap in career opportunities. The hypothesis is that workers compare wages with the value of non-market opportunities when deciding to stay in the labour market; thus women who place a higher value on non-market opportunities may decide to exit from the labour market or from their current job more easily and this would explain the better career opportunities which may exist for men especially in jobs where high investments in specific human capital is required (Booth et al., 2002; Lazear & Rosen, 1990). The analysis performed by Robone et al. (2011) with regard to a sample of UK employees also corroborates the hypothesis that asymmetries between women and men about the influence of temporary jobs on health and psychological well-being might derive from the different roles which they play within the market sector and the family structure (taking care of children is mainly a responsibility of women while men hold the main responsibility for material well-being). Traditional gender roles appear to hold not only in Italy, but also in other European countries, suggesting that the results of our analysis might be generalised beyond the Italian context. Our results further shed light on the impact of whether individuals receive financial maintenance from their families. For men who depend primarily on their employment for income, job insecurity has a marked negative influence on their psychological health. For women who receive financial maintenance from their families, job flexibility and temporary contracts have a significantly positive influence on their mental health. We also show that workers with fixed-term contracts have a lower level of happiness. These results allow our study to provide an innovative contribution since to date little evidence has been reported on the influence of contractual conditions on happiness (Dolan et al., 2008). Our results are in line with the very few other studies on this issue (Ponzo, 2011; Scherer, 2009). Our analysis provides a snapshot at an important point in time in that it allows for the impact of an intense wave of successive reforms on the Italian labour market after 1997, introducing more flexibility into the labour market. Since 2003, this wave of reform was substantially interrupted and there have not been any subsequent important interventions in

Well-being Consequences of Temporary Contracts

181

the Italian labour market. Thus, with respect to flexibility, the current labour market situation in Italy is largely comparable to the one of 2005 making the analysis relevant to the current situation. The spread of temporary employment has however grown in Italy from 2004 (37% of 15
25-year-old workers) to 2012 (50% of 15
25-year-old workers) because the new contractual types have had more time to be understood and used by employers. In this respect, our estimates of the relationship between temporary employment and well-being are likely to take into account the short-term consequences of the labour market reforms in Italy. The spread of fixed-term contracts in Italy, encouraged by important liberalization reforms, has often been justified by the need for firms to ensure partial work flexibility and as an attempt to respond to youth unemployment. However, our results may highlight some concerns around the use of fixed-term contracts both on efficiency and equity grounds. On efficiency grounds, our results, showing lower psychological well-being of temporary workers, may have a potential impact on work performance given that workers with poorer health and psychological well-being are likely to suffer more from illnesses limiting their working capacity and resulting in more sickness absence (Bartley, Sacker, & Clarke, 2004). On equity grounds, our results suggest that the spread of temporary contracts is likely to generate disparities between individuals with strong family economic support compared to those relying on employment as their main source of income, given that family maintenance is able to protect workers from the stress derived from job insecurity. These findings seem to suggest the necessity to strengthen formal welfare support for younger employees enrolled in temporary jobs. Such individuals are more exposed to the risk of being ‘working’ poor or unemployed, and informal welfare provided by family can substantially reduce the negative effects of job instability but at the expense of more inequalities between the dependents of those families who are, and those who are not, able to provide economic support. The gender gap in the health consequences caused by fixed-term contracts and its interaction with family economic support seems a promising area of future research. Our interpretation based on gender-related differences in preferences towards career flexibility may not be the only possible one. It would be interesting, for instance, to investigate other mediating factors such as whether differences in health consequences, in particular psychological well-being, may be explained by a different tolerance towards work-related stress between men and women.

182

VINCENZO CARRIERI ET AL.

This paper suffers from two important limitations: we used crosssectional data to analyse the effect of temporary contracts on health and happiness of individuals. Cross-sectional data make it difficult to identify a causal association between temporary jobs and individuals’ health status and happiness. It also makes it difficult to disentangle the temporal sequencing of temporary contracts on individual well-being. Nevertheless, we provide evidence on new aspects of the relationship between temporary jobs and well-being in Italy, not frequently analysed in previous literature. Some limitations in the use of the PS method should be also pointed out. The PS implies that the identification of the ATTs relies on the validity of the CIA, namely that the potential treatment outcomes are independent of the assignment mechanism for any given value of a vector of observable characteristics (X) (Ichino et al., 2008). In our specific case, CIA implies that selection into temporary jobs is solely based on observable variables included in the PS model. The assumption is expected to be fulfilled if all relevant variables are observable. We do not claim to have access to all variables influencing the outcome; however, we performed sensitivity analysis to take into account the potential effect of unobservable variables (see Appendix B). Our sensitivity analyses suggest that our baseline estimates are robust to deviations from the CIA assumption.

NOTE 1. Excluding groups 3 and 4 resulted in a very small loss (0.1% and 0.7% of the sample respectively).

ACKNOWLEDGEMENTS An earlier draft of this paper was presented at a Centre for Health Economics seminar at the University of York, UK. This paper was also presented at the annual meeting of the Italian Society of Economists, Matera, Italy, at the annual meeting of the Italian Society of Public Economics, Pavia, Italy and at the annual meeting of the Italian Health Economics Association, Rome, Italy. The authors wish to thank the participants for useful comments received. We also thank Giovanni Pica, Lorenzo Sacconi, Pedro Rosa Dias and Maarten Lindeboom for useful suggestions as well as two anonymous referees. The usual disclaimer applies.

Well-being Consequences of Temporary Contracts

183

REFERENCES Akerlof, G. A., & Kranton, R. E. (2000). Economics and identity. Quarterly Journal of Economics, 115, 715
753. Arcas, M., Novoa, A. M., & Artazcoz, L. (2013). Gender inequalities in the association between demands of family and domestic life and health in Spanish workers. The European Journal of Public Health, 23(5), 883
888. doi:10.1093/eurpub/cks095 Artazcoz, K., Borrell, C., Cortes, I., Escriba-Aguir, V., & Cascant, L. (2007). Occupational epidemiology and work related inequalities in health: A gender perspective for two complementary approaches to work and health research. Journal of Epidemiology & Community Health, 61(Suppl. II), ii39
ii45. Artazcoz, K., Cortes, I., Puig-Barrachina, V., Benavides, F., Escriba-Aguir, V., & Borrell, C. (2013). Combining employment and family in Europe: The role of family policies in health. The European Journal of Public Health, 1
7. doi:10.1093/eurpub/ckt170 Auld, M. C., & Sidhu, N. (2005). Schooling, cognitive ability and health. Health Economics, 14(10), 1019
1034. Bartley, M., Sacker, A., & Clarke, P. (2004). Employment status, employment conditions, and limiting illness: Prospective evidence from the British household panel survey 1991
2001. Journal of Epidemiology & Community Health, 58, 501
506. Bauer, G. F., Huber, C. A., Jenny, G. J., Mu¨ller, F., & Ha¨mmig, O. (2009). Socioeconomic status, working conditions and self-rated health in Switzerland: Explaining the gradient in men and women. International Journal of Public Health, 54, 23
30. Bauer, T. N., & Truxillo, D. M. (2000). Temp-to-permanent employees: A longitudinal study of stress and selection success. Journal of Occupational Health Psychology, 53, 37
346. Becker, S. O., & Ichino, A. (2002). Estimation of average treatment effects based on propensity scores. The Stata Journal, 2(4), 358
377. Bentolila, S., & Dolado, J. J. (1994). Labour flexibility and wages: Lessons from Spain. Economic Policy, 18, 55
99. Berntsson, L., Lundberg, U., & Krantz, G. (2006). Gender differences in work
home interplay and symptom perception among Swedish white-collar employees. Journal of Epidemiology & Community Health, 60, 1070
1076. Bildt, C., & Michelsen, H. (2002). Gender differences in the effects from working conditions on mental health: A 4-year follow-up. International Archives of Occupational and Environmental Health, 75(4), 252
258. Bohle, P., Quinlan, M., & Mayhew, C. (2001). The health and safety effects of job insecurity: An evaluation of the evidence. Economic and Labour Relations Review, 12(1), 32
60. Booth, A. L., Francesconi, M., & Frank, J. (2002). Temporary jobs: Stepping stones or dead ends? Economic Journal, 112(480), 189
213. Burchell, B. J. (1994). The effects of labour market position, job insecurity and unemployment on psychological health. In G. Duncan, C. Marsh, & C. Vogler (Eds.), Social change and the experience of unemployment (pp. 188
212). Oxford: Oxford University Press. Burchell, B. J. (1999). The unequal distribution of job insecurity, 1966
86. International Review of Applied Economics, 13(3), 437
458. Caliendo, M., & Kopeinig, S. (2008). Some practical guidance for the implementation of propensity score matching. Journal of Economic Surveys, 22, 31
72.

184

VINCENZO CARRIERI ET AL.

Carrieri, V. (2012). Working poor in Italia: Chi sono, quanti sono e quanto sono poveri? Italian Journal of Social Policy, 2(2012), 71
96. Cheng, G. L., & Chan, D. K. S. (2008). Who suffers more from job insecurity? A metaanalytic review. Applied Psychology: An International Review, 5(2), 272
303. Chubb, J. (1982). Patronage, power, and poverty in southern Italy. A tale of two cities. Cambridge: Cambridge University Press. Clark, A., Georgellis, Y., & Sanfey, P. (2001). Scarring: The pyschological impact of past unemployment. Economica, 68, 221
241. Cottini, E., & Lucifora, C. (2010). Mental health and working conditions in European countries. IZA Discussion Paper 4717. Institute for the Study of Labour.. Retrieved from http:// ftp.iza.org/dp4717.pdf D’Attoma, I., & Tassinari, G. (2011). Il lavoro a tempo determinato in Europa: Alcune evidenze statistiche. Bologna: Univerista` di Bologna, Alma Mater Digital Library. Davis, J., & Marino, J. (2000). Mezzogiorno and modernity: Southern Italy and its history, UC World History 2000. De Witte, H. (1999). Job insecurity and psychological well-being: Review of the literature and exploration of some unresolved issues. European Journal of Work and Organisational Psychology, 8, 155
177. De Witte, H., & Na¨swall, K. (2003). ‘Objective’ vs ‘Subjective’ job insecurity: Consequences of temporary work for job satisfaction and organizational commitment in four European countries. Economic and Industrial Democracy, 24(2), 149
188. Del Boca, D., & Vuri, D. (2007). The mismatch between employment and child care in Italy: The impact of rationing. Journal of Population Economics, 20, 805
832. Di Bartolomeo, A., Di Bartolomeo, G., & Pedaci, M. (2009). Chi sono i lavoratori poveri? Working Paper series 55/2009. Department of Communication, University of Teramo. Retrieved from http://wp.comunite.it/data/wp_no_55_2009.pdf Dohmen, T., Falk, A., Huffman, D., Schupp, J., Sunde, U., & Wagner, J. (2011). Individual risk attitudes: Measurement, determinants and behavioral consequences. Journal of the European Economic Association, 9(3), 522
550. Dolan, P., Peasgood, T., & White, M. (2008). Do we really know what makes us happy? A review of the economic literature on the factors associated with subjective well-being. Journal of Economic Psychology, 29, 94
122. Doyal, L. (2000). Gender equity in health: Debates and dilemmas. Social Science and Medicine, 51, 931
939. Ehlert, C., & Schaffner, S. (2011). Health effects of temporary jobs in Europe. Ruhr Economic Papers No. 295. Department of Economics, Ruhr-Universita¨t Bochum (RUB). Retrieved from http://repec.rwi-essen.de/files/REP_11_295.pdf Elia, L. (2010). Temporary/permanent workers’ wage gap: A brand-new form of wage inequality? Labour, 24, 178
200. Famiglie e Soggetti Sociali. (2011). Istat, Come cambiano le forme familiari – Anno 2009, Statistica report, 15 settembre 2011. Ferrer-i-Carbonell, A., & Gowdy, J. M. (2007). Environmental degradation and happiness. Ecological Economics, 60, 509
516. Ferrie, J. E., Shipley, M. J., Stansfeld, S. A., & Marmot, M. G. (2002). Effects of chronic job insecurity and change in job security on self reported health, minor psychiatric morbidity, physiological measures, and health related behaviours in British civil servants: The Whitehall II study. Journal of Epidemiology & Community Health, 56, 450
454.

Well-being Consequences of Temporary Contracts

185

Frey, B. S., & Stutzer, A. (2000). Happiness, economy and institutions. The Economic Journal, 110, 918
938. Garibaldi, P., & Pacelli, L. (2008). Do larger severance payments increase individual job duration? Labour Economics, 15, 215
245. Gash, V., Mertens, A., & Romeu Gordo, L. (2007). Are fixed-term jobs bad for your health? A comparison of West-Germany and Spain. European Societies, 9, 429
458. George, L. K. (2006). Perceived quality of life. In R. H. Binstock & L. K. George (Eds.), Handbook of aging and the social sciences (6th ed., pp. 320
336). San Diego, CA: Academic Press. Graham, C. (2008). Happiness and health: Lessons and questions for public policy. Health Affairs, 27, 72
87. Hagen, T. (2002). Do temporary workers receive risk premiums? Assessing the wage effects of fixed-term contracts in West Germany by a matching estimator compared with parametric approaches. Labour, 16, 667
705. Harryson, L., Novo, M., & Hammarstrom, A. (2012). Is gender inequality in the domestic sphere associated with psychological distress among women and men? Results from the Northern Swedish Cohort. Journal of Epidemiology & Community Health, 66, 271
276. Ichino, A., Mealli, F., & Nannicini, T. (2008). From temporary help jobs to permanent employment: What can we learn from matching estimators and their sensitivity? Journal of Applied Econometrics, 23, 305
327. doi:10.1002/jae.998 Ichino, A., & Riphahn, R. T. (2005). The effect of employment protection on worker effort: A comparison of absenteeism during and after probation. Journal of the European Economic Association, 3(1), 120
143. Idler, L. E., & Benyamini, Y. (1997). Self-rated health and mortality: A review of twenty-seven community studies. Journal of Health and Social Behavior, 38, 21
37. Imbens, G. W., & Wooldridge, J. M. (2009). Recent developments in the econometrics of program evaluation. Journal of Economic Literature, 47, 5
86. ISTAT Multiscopo Survey. (2007). Condizioni di salute, fattori di rischio e ricorso ai servizi sanitari. Indagine Multiscopo sulle famiglie, Condizioni di salute e ricorso ai servizi sanitari, Anni 2004
2005. Joyce, K., Pabayo, R., Critchley, J. A., & Bambra, C. (2010). Flexible working conditions and their effects on employee health and wellbeing (Review). Cochrane Database of Systematic Reviews, (2). Art. No.: CD008009. doi:10.1002/14651858.CD008009.pub2 Kennedy, B. P., Kawachi, I., Glass, R., & Prothrow-Stith, D. (1998). Income distribution, socio-economic status, and self rated health in the United States: Multilevel analysis. British Medical Journal, 317, 917
921. Kivima¨ki, M., Vahtera, J., Virtanen, M., Elovainio, M., Pentti, J., & Ferrie, J. E. (2003). Temporary employment and risk of overall and cause-specific mortality. American Journal of Epidemiology, 158(7), 663
668. doi:10.1093/aje/kwg185 Klein-Hesselink, D. J., & van Vuuren, T. (1999). Job flexibility and job insecurity: The Dutch case. European Journal of Work and Organizational Psychology, 8, 273
293. Lasfargues, G., Doniol-Shaw, G., Deriennic, F., Bardot, F., Huez, D., & Rondeau du Noyer, C. (1999). Pre´caritie´ de l’emploi et sante´: Differences hommes-femmes. Archieves des Maladies Professionnelles et de Me´decine du Travail, 8, 273
293. Lazear, E. P., & Rosen, S. (1990). Male-Female wage differentials in job ladders. Journal of Labour Economics, 8(1), 106
123.

186

VINCENZO CARRIERI ET AL.

Leuven, E., & Sianesi, B. (2003). PSMATCH2: Stata module to perform full mahalanobis and propensity score matching, common support graphing, and covariate imbalance testing. Statistical Software Components S432001. Boston College Department of Economics, revised 19 Jul 2012. Lewis, S., Smithson, J., & Brannen, J. (1999). Young Europeans’ orientations to families and work. Annals of the American Academy of Political and Social Science, 562, 83
97. Micali, S. J. (2009). Italy: The Case for Division, Pell Scholars and Senior Theses. Paper 50. Retrieved from http://escholar.salve.edu/pell_theses/50 Naldini, M., & Jurado, T. (2013). Family and welfare state reorientation in Spain and Inertia in Italy from a European perspective. Population Review, 52, 43
61. Natti, J., Kinnunen, U., Makikangas, A., & Mauno, S. (2009). Type of employment relationship and mortality: Prospective study among Finnish employees in 1984
2000. European Journal of Public Health, 19, 150
156. OECD. (2009). Employment outlook. Paris: Organisation for economic cooperation and development (OECD). Retrieved from www.oecd.org/els/social/workincentives OECD. (2011). OECD.StatExtracts. Paris: Organisation for economic cooperation and development (OECD). Retrieved from stats.oecd.org Origo, F., & Pagani, L. (2009). Flexicurity and job satisfaction in Europe: The importance of perceived and actual job stability for well-being at work. Labour Economics, 16(5), 547
555. Plagnol, A. C., & Easterlin, R. A. (2008). Aspirations, attainments, and satisfaction: Life cycle differences between American women and men. Journal of Happiness Studies, 9(4), 601
619. Plaisier, I., de Bruijn, J. G. M., de Graaf, R., ten Have, M., Beekman, A. T. F., & Penninx, B. (2007). The contribution of working conditions and social support to the onset of depressive and anxiety disorders among male and female employees. Social Science & Medicine, 64, 401
410. Ponzo, M. (2011). Status occupazionale e benessere soggettivo in Italia, QA-Rivista dell? Associazione Manlio Rossi Doria, 65, 75
106. Quesnel-Vallee, A., DeHaney, S., & Ciampi, A. (2010). Temporary work and depressive symptoms: A prospective score analysis. Social Science and Medicine, 70, 1982
1987. Robone, S., Jones, A. M., & Rice, N. (2011). Contractual conditions, working conditions, health and well-being in the British household panel survey. The European Journal of Health Economics, 12, 429
444. Rosenbaum, P. R., & Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70, 41
55. Salvatori, A. (2010). Labour contract regulations and workers’ wellbeing: International longitudinal evidence. Labour Economics, 17, 667
678. Scherer, S. (2009). The social consequences of insecure jobs. Social Indicators Research, 93, 527
547. Schro¨der, C. (2008). Economic insecurity and cohabitation strategies in Italy. MPIDR Working Paper 2008
004. Retrieved from http://www.demogr.mpg.de/papers/working/wp-2008004.pdf. SHIW. (2004). The survey on household income and wealth. Indagine campionaria “I bilanci delle famiglie italiane nell’anno 2004. Banca d’Italia. Silla, I., Gracia, F. J., & Peiro, J. M. (2005). Job insecurity and health-related outcomes among different types of temporary workers. Economic and Industrial Democracy, 26, 89
117.

Well-being Consequences of Temporary Contracts

187

Singh-Manoux, A., Ferrie, J., Lyinch, J. W., & Marmot, M. (2005). The role of cognitive ability (intelligence) in explaining the association between socioeconomic position and health: Evidence from the whitehall ii prospective cohort study. American Journal of Epidemiology, 161(9), 831
839. Sverke, M., Gallagher, D. G., & Hellgreen, J. (2000). Job stress, well-being, and work attitudes among employees with different employment contracts. In K. Isaksson, C. Hogstredt, C. Eriksson, & T. Theorell (Eds.), Health effects of the new labour market (pp. 145
167). New York, NY: Kluwer Academic/Plenum Publishers. Unde´n, A. L., & Elofsson, S. (2006). Do different factors explain self-rated health in men and women? Gender Medicine, 3(4), 295
308. Veenhoven, R. (1996). Happy life-expectancy. Social Indicators Research, 39, 1
58. Virtanen, M., Kivimaki, M., Elovainio, M., Vahtera, J., & Ferrie, J. E. (2003). From insecure to secure employment: Changes in work, health, health related behaviours, and sickness absence. Occupational Environment Medicine, 60, 948
953. Waenerlund, A. K., Virtanen, P., & Hammarstrom, A. (2011). Is temporary employment related to health status? Analysis of the Northern Swedish Cohort. Scandinavian Journal of Public Health, 39, 533
539. Ware, J. E., Jr., Kosinski, M., & Keller, S. D. (1996). A 12 item short form health survey: Construction of scales and preliminary tests of reliability and validity. Medical Care, 34, 220
233. Winkelmann, L., & Winkelmann, R. (1998). Why are the unemployed so unhappy? Evidence from panel data. Economica, 65, 1
15.

188

VINCENZO CARRIERI ET AL.

APPENDIX A We describe here how the statistical matching between the ISTAT Multiscopo dataset with the Bank of Italy’s SHIW was performed. First, two constraints need be satisfied to make matching feasible: (1) the two surveys must be random samples from the same population; (2) there must be a common set of conditioning variables. In our case, the first condition is met by design, since both the ISTAT Multiscopo 2005 and the SHIW 2004 data are representative of the Italian population. As far as the second constraint is concerned, the variables (X) common to each dataset and chosen for the process of imputation of the individual income are: age, gender, macro-region of residence (North, South, Centre and Islands), marital status, education, professional position and field of work. Since working individuals have been taken into account, the final sample is made up of 49,402 observations from the SHIW survey and 14,460 from the ISTAT Multiscopo Survey. The dataset, integrated by ISTAT-Bank of Italy was created using the PS matching technique, a statistical method which allows individuals with similar characteristics but from different datasets to be put together (Rosenbaum & Rubin, 1983). A ‘treatment’ variable (T) has been generated, or rather defined as a binary variable that takes a value 1 if the interviewee belongs to the ISTAT Multiscopo Survey and 0 if the observation belongs to the SHIW survey. The PS or rather the probability of belonging to the ISTAT Multiscopo Survey conditioned by the set of common demographic and socio-economic variables, was calculated through a probit regression of the T variable of the X set of common variables mentioned above. Once that index was obtained it was necessary to define a similarity function between the individuals of the two samples. The similarity function assigns to each individual in the ISTAT Multiscopo set a similar individual from the SHIW, according to some particular criteria. The matching was performed through the most straightforward matching estimator: the nearest neighbour matching technique. The individual from the comparison group is chosen as a matching partner for a treated individual that is closest in terms of PS. This technique selects the comparison units whose PSs are closest to the treated unit in question (Caliendo & Kopeinig, 2008). In order to obtain a more precise matching, the sample was stratified in cells according to type of occupation distinguishing between permanent and temporary jobs, and according to gender. The dataset was divided into 20 cells, for each of which a PS was calculated through a probit model. Once the matching procedure was complete, it was evaluated in terms of maintaining the income distribution, both in terms of preserving the

189

Well-being Consequences of Temporary Contracts

pre-existing income distribution as well as in terms of pre-existing relations between variables of interest. The next step was (1) the comparison between the income distribution in the integrated dataset and the pre-existing SHIW one, (2) the calculation of the correlation between income and the X vector to verify the maintenance of the sign recorded in the ‘donor set’. The differences between the common-fusion correlations in the SHIW dataset versus the fused ISTAT Multiscopo dataset were well preserved for most variables (see Tables A1 and A2). Table A1. Variables

Correlation between Income and X Vector. ISTAT Multiscopo Survey (Recipient Set)

SHIW (Donor Set)

Age Female

0.0125 −0.0493

0.1125 −0.0891

Region North Centre South Islands

0.0115 0.0027 −0.0074 −0.0123

0.0512 0.0215 −0.046 −0.0545

Education Primary school Lower high school High school University or postgraduate

−0.1452 −0.1132 0.0698 0.1799

−0.1028 −0.1293 0.0576 0.1668

0.3385 0.3029 0.3129 0.0911

0.3616 0.2533 0.3108 0.0117

−0.1051 0.0066 0.0580

−0.1273 0.0126 0.0432

Occupation Business executive Supervisor White collar Blue collar Industry Primary industry Secondary industry Tertiary industry

Table A2.

Mean Income in Two Datasets.

Income Mean (Standard Deviation in Parenthesis) ISTAT Multiscopo Survey (Recipient Set) 11652 (9504)

SHIW (Donor Set) 12493 (9357)

190

VINCENZO CARRIERI ET AL.

APPENDIX B We performed sensitivity analysis to take into account the potential effect of observable and unobservable variables on the ATTs shown in the section ‘Results’. First of all, we try a different specification of the PS model to check to what extent our ATTs are sensitive to the observable variables chosen. For instance, it might be argued that occupation and industry variables are not good pre-treatment variables but variables whose value are determined once the job is taken up. In our preferred specification, we include such variables because they are important in explaining well-being conditions. However, our results are not driven by the inclusion of these variables because the ATT excluding occupation and industry variables remains substantially unchanged. Results are not shown but available upon request. A second important identification issue might be related to the unobservables. Indeed, the identification of the ATTs relies on the validity of the CIA, namely that the potential treatment outcomes are independent of the assignment mechanism for any given value of a vector of observable characteristics (X) (Ichino et al., 2008). In our specific case, CIA implies that selection into temporary contracts is solely based on observable variables included in the PS model. This assumption seems to not be as strong as it appears, since as anticipated, job opportunities for young workers in Italy are quasi exclusively limited to the new forms of non-permanent jobs. In other words, our treatment variable seems to be mainly a direct consequence of labour market conditions rather than a free choice based on individual preferences. This hypothesis is corroborated by D’Attoma and Tassinari (2011). Based on EU Labour Force Survey data, they showed that in Italy approximately 67% of workers are enrolled in temporary contracts and stated that their choice was involuntary and related to the difficulty of finding a permanent job. This may provide support for the CIA assumption. Notwithstanding, we complemented the above reasoning with an assessment of the robustness of our estimates to possible deviations from the CIA. First of all, we considered which potential unobservable factors might represent a threat to our estimates. Our results showed negative effects of temporary contracts on health and well-being. Thus, an unobservable factor which might concern us would be one that is both positively associated with temporary contracts and positively associated with health and wellbeing. A confounder with these characteristics might drive ATT estimates towards zero and make the treatment effects non-significant. On the other

Well-being Consequences of Temporary Contracts

191

hand, a confounder which would increase the enrolment in temporary contracts (permanent contracts) and reduce (increase) health and well-being would not really be a cause for concern. In fact, in this last case the estimated ATT would be considered as a lower bound of the real effect. An important unobservable, typically associated with temporary contracts, is the worker’s cognitive ability. It may be argued that in a labour market where hiring on temporary contracts is easier, selection into permanent contracts may become even more competitive. So, only workers with the best abilities gain access to such contracts. If this is the case, temporary contracts should be gained, on average, by workers with lower cognitive ability. However, several studies have linked lower cognitive ability to a higher probability of suffering from poor health status (Auld & Sidhu, 2005; Singh-Manoux, Ferrie, Lyinch, & Marmot, 2005). Hence, workers’ ability is unlikely to be an unobservable confounder which would cause concern since the inclusion of ability as a control variable would make the ATT even more negative (neither positive nor zero). Another important unobservable may be workers’ risk aversion. Again, in the Italian labour market, unemployment levels for the young are amongst the highest in the EU. Because of the high unemployment rate and the low probability of getting a permanent job, risk-averse young entrants may be more willing to accept a temporary job as an alternative to unemployment. Risk aversion may also positively influence health status: risk-averse individuals are in general more likely to adopt healthy lifestyles and to demand preventive care (Dohmen et al., 2011). Hence, unobserved risk aversion may be a potential confounder which could cause concern. To rule out this possibility, we apply the calibrated confounder exercise proposed by Ichino et al. (2008). Briefly, the exercise simulates a ‘problematic’ confounder and re-estimates the ATT adding the simulated variable in the PS model. Following Ichino et al. (2008) we simulate a confounder variable which mimics the distribution of one important covariate, in our case, education. This choice relies on two arguments. First, education is a strong determinant of health status and it is a key factor in any occupational outcome. Second, a distribution close to the education variable might easily mimic a distribution of a potential ‘problematic’ unobservable like risk aversion. The hypothetical distribution of risk aversion in the treatment (having a temporary contract) and in the outcome (enjoying better health) is likely to be very close to the distribution of education which is positively associated with health status and with the enrolment in temporary contracts (see ‘Overall Propensity Score Matching Results by Gender’ section).

192

Table A3. (a) Self-Assessed Health (SAH) Full Sample Female

Male

Robustness Checks for the Average Treatment Effects (ATT). (b) Physical Component Score (PCS)

Full Sample

Female

Male

(c) Mental Component Score (MCS) Full Sample Female

Male

(d) Happiness Full Sample

Female

Male

−0.574 −0.625* (0.479) (0.355)

−0.068*** (0.021)

−0.056* −0.077*** (0.033) (0.028)

Sensitivity test (standard errors in parenthesis) −0.050* −0.056 −0.054 −0.274 −0.424 (0.028) (0.041) (0.036) (0.208) (0.320)

−0.161 (0.264 )

−0.525 (0.333)

−0.352 −0.716* (0.559) (0.397)

−0.063*** (0.024)

−0.052 (0.039)

−0.073** (0.032)

Sensitivity test (standard errors in parenthesis) −0.052* −0.048 −0.054 −0.291 −0.399 (0.028) (0.040) (0.038) (0.203) (0.311)

−0.117 (0.279)

−0.580 (0.359)

−0.359 −0.748** (0.543) (0.406)

−0.058** (0.024)

−0.055 (0.038)

−0.070** (0.032)

Confounder: Bachelor degree or more. *Statistical significance at the 10% level. **Statistical significance at the 5% level. ***Statistical significance at the 1% level.

VINCENZO CARRIERI ET AL.

Nearest Neighbour Propensity Score Matching (Standard Errors in Parenthesis) −0.051** −0.051 −0.054* −0.194 −0.488* −0.186 −0.680** (0.024) (0.036) (0.032) (0.178) (0.284) (0.235) (0.290)

Well-being Consequences of Temporary Contracts

193

Thus we re-estimate all the ATTs presented in ‘Overall Propensity Score Matching Results by Gender’ section adding one simulated variable with the same distribution as a dummy variable taking the value of one for individuals having a university or postgraduate degree (compared to having at most a high school certificate). We also repeat the exercise including a dummy variable which takes the value one if individuals have a high school certificate or higher (compared to having at most a lower high school certificate). Two cut-off points are useful since we have to dichotomize the education variable to simulate a sharp shock in education and the dichotomization is required to implement the method proposed by Ichino et al. (2008). We present the simulated ATTs and the baseline ATTs together in Table A3 in order to verify the robustness of the ATTs. We present just the results of overall ATTs from ‘Overall Propensity Score Matching Results by Gender’ section as sensitivity analysis for stratified samples is more problematic because we have too few treated and control individuals. Sensitivity analysis of stratified samples are substantially the same as baseline ATTs, but in some cases statistical significance is not reached as there are too few treated and control individuals. Thus, while the general robustness of ATTs to deviations from CIA is always verified, the sample sizes in some cases are too small to verify statistically. Table A3 shows that all three baseline ATTs are almost identical to the ATTs which include the simulated variable acting as education. All in all, sensitivity analyses seem to suggest that our baseline estimates are robust to exclusion of some potential pre-treatment observable variables and robust to deviations from the CIA assumption.

THE EFFECT OF LAND TITLE ON CHILD LABOR SUPPLY: EMPIRICAL EVIDENCE FROM BRAZIL Mauricio Moura and Rodrigo Bueno ABSTRACT This paper assesses the effect of property titling on child labor. Our main contribution is to investigate the potential impact of property rights on child labor supply by analyzing household response regarding the child labor force to exogenous changes in property ownership status. The causal role of legal ownership is isolated by comparing the effect of land titling using data from a unique study in two geographically close and demographically similar communities in Osasco, a town of 654,000 people in the Sao Paulo metropolitan area. Survey data were collected from households in both communities before and after the granting of land titles, with neither type knowing ex ante whether it would receive land titles. The econometric estimates, applying the Difference-in-Difference (DD) methodology and propensity score matching, suggest that land titling decreases child labor. Keywords: Property rights; land titling; child labor force JEL classifications: P14; Q15; J22; O18; O54

Factors Affecting Worker Well-Being: The Impact of Change in the Labor Market Research in Labor Economics, Volume 40, 195
222 Copyright r 2014 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-912120140000040007

195

196

MAURICIO MOURA AND RODRIGO BUENO

INTRODUCTION Many researchers have studied how property rights generate economic development (e.g., North & Thomas, 1973), but few have focused on the impact of this type of policy on child labor.1 An exception related to property rights and child labor is Field (2007), who documents the effect of transferring property protection from local communities and households to the state. She concludes that there is a reduction in the number of hours worked by children from these households. Moura and De-Losso (2010) investigate the impact of exogenous changes in formal property ownership status on child labor supply. However, because they did not address observable differences in some controls between the intervention (land titled) and comparison (non-land titled) groups, their results are very likely overestimated. To overcome such a flaw, we apply the propensity score method and make a range of important improvements, allowing us to verify if and how the results change and to make them much more reliable. The lack of a formal property rights system is more problematic for the poor, because they are constrained from using land as collateral to access credit markets (Besley, 1995; North, 1990). If such credit were available it could be invested as capital in productive projects and used to increase labor productivity and income (Demsetz, 1967; De Soto, 2000). Torstensson (1994) and Goldsmith (1995) have shown a positive relationship between economic growth and property tenure rights. The authors also recommend strengthening economic institutions to increase growth performance, encourage investment in physical and human capital, decrease macroeconomic volatility, and promote more efficient and fair distribution of economic opportunity. Property titling increasingly is considered one of the most effective public policies to benefit poor populations and encourage economic growth around the world (Baharoglu, 2002; Binswanger, Deninger, & Feder, 1995). In Asia, for example, millions of land titles have been issued in Vietnam and Cambodia, while several governments are investing in social housing in Africa (Galiani & Schargrodsky, 2010). In Latin America, Peru is the most famous example of a property-titling program; in the 1990s the government issued titles to 1.2 million urban households (Field & Torero, 2002). In 2003, the Brazilian federal government announced a massive national plan to title 750,000 families. Since its launch, this program, Papel Passado, has received US $15 million annually from the federal budget, providing titles to over 85,000 families in 49 cities in 17 Brazilian states. Its official goal is “… to increase land titles in Brazil and to promote an increase in the quality of life for the Brazilian population” by issuing land titles to families living under illegal conditions (i.e., residents illegally squatting in urban

The Effect of Land Title on Child Labor Supply

197

dwellings).2 This paper measures the impact of the Brazilian land-titling program, Papel Passado, on child labor supply, mostly among children between 10 and 17 years old. In contrast to other research on this subject, we benefit from being able to analyze a unique dataset in a context that helps us not only to isolate the causal role of land titling, but also to minimize the endogeneity problems characteristic of most studies in this field. We compare two similar neighboring communities in the Brazilian city of Osasco. The town
with around 654,000 inhabitants where almost 6,000 families live informally on urban property
is located in the metropolitan area of Sa˜o Paulo and is part of the Papel Passado program map. In one of its communities, Jardim Canaa˜, all households received land titles in 2007. In another, Jardim DR, households were scheduled to receive land titles in 2012, making it a natural comparison group. Our analysis is based on a two-stage survey conducted in Jardim Canaa˜ and Jardim DR focusing on the property rights issue. The sample consists of 326 households distributed across both neighborhoods (185 from Jardim Canaa˜ and 141 from DR). The first stage of the survey was conducted in March 2007, before titles had been issued to Jardim Canaa˜, and the second stage in August 2008, almost one-and-one-half years after the titles had been received. Those communities were municipal public land illegally invaded.3 The main consequence is that Brazilian law came to protect these new property owners. The Government relinquished post rights over the land/property after the titling execution. Other land dispute cases in Brazil are much more complex if invasion occurs on private property, since that opens a window for endless legal battles. This paper’s main finding indicates that land titling has a socially positive impact on the child labor supply, that is, children from households that received a land title decreased their weekly hours of work. The available land title impact literature lacks a precise transmission channel explaining this effect. Therefore, we discuss some transmission channels without taking a firm position. The rest of the paper is organized as follows. The section “Literature Review on Land Titling” reviews the land title literature. The section “Overview of Child Labor Force Participation” explores the economic context of child labor. The section “Land Title and Child Labor: Potential Mechanisms” provides a basic overview of the potential mechanisms to explain child labor. The section “Methodological issues, data and descriptive statistics presents the research basic methods, challenges and data”. The section “Empirical Strategy” describes the empirical strategy to measure the effects of land titling on child work, discussing the research methodology,

198

MAURICIO MOURA AND RODRIGO BUENO

including the Difference-in-Difference (DD) and propensity score techniques, and providing an overview of the data collected. The empirical results are further discussed in the section “OLS, Difference-in-Difference, and DDM: Empirical Results,” and the final section concludes.

LITERATURE REVIEW ON LAND TITLING The positive effects of land titling have been reported in several types of studies, including those on real estate values by Jimenez (1985), Alston, Libecap, and Schneider (1996), and Lanjouw and Levy (2002); studies on agricultural investment by Besley (1995), Jacoby, Li, and Rozzelle (2002), Brasselle, Gaspart, and Platteau (2002), and Do and Iyer (2003) research on credit access, labor supply, housing investment, and income by Place and Migot-Adholla (1998) and Carter and Olinto (2003). Most of the literature and the majority of policy attention to property rights focus on tenure security of rural households. According to Field and Torero (2002), this focus is due to historical interest in agricultural investment and related policies of land reform. The impact of land rights on agricultural investment is dependent on the location under analysis; Besley (1995) shows no impact in the region of Wassa but a positive effect in Angola. Similarly, Jacoby et al. (2002) estimate positive effects in China, but Brasselle et al. (2002) do not find any in Burkina Faso. The Peruvian titling program was implemented regionally in different stages in the 1990s. Field and Torero (2002) take advantage of this variability in timing and use cross-sectional data to study past and future titleholders midway through the project. They also find positive effects for the Peruvian titling program, particularly on labor supply, credit access, and housing investment. In Brazil, Andrade (2006) estimates a positive effect of land title on income using cross-sectional data from a sample of 200 families of the Comunidade do Caju, a poor urban community in the city of Rio de Janeiro. Furthermore, in Brazil Moura and De-Losso (2013) provide empirical support to show that land title can influence happiness.

OVERVIEW OF CHILD LABOR FORCE PARTICIPATION The United Nations Habitat Report (2005) estimates that in 2002, 246 million children and teenagers around the world engaged in some form of

199

The Effect of Land Title on Child Labor Supply

North

8

11.1 10.5 9.9 9.8 10.2 10.2

Brazil

%

10

14.9 14.0 13.6 13.6 11.9 11.6

12

13.8 13.1 12.4 11.3 10.3 9.6

14

11.8 12.2 11.5 10.8 10.2 9.8

16

7.9 8.6 8.4 7.9 7.9 7.6

18

14.8 15.9 14.4 13.4 12.3 11.7

work, mostly working in their own families’ businesses. While Africa has the highest ratio of child labor with respect to the work force, Asia has the highest absolute number of children working. In Brazil, the number of children between 5 and 17 years old who work is decreasing annually, as Fig. 1 shows. However, the number is still high, at 9.8% of this population or about 4.3 million people. The worst case is the Northeast region, with 11.7% of this population working, whereas the Southeast is the best case with 7.6%. Although international and local statistics are available, Edmonds (2008) claims that there is no universally agreed definition of child labor. Furthermore, many theories attempt to explain the reasons for the existence of child labor. Becker and Lewis (1973) argue that families make a cost-benefit analysis between sending their children to work or to school. Children’s labor increases the household income, but reduces study and leisure time. In terms of time allocation, there is some degree of substitution between adult and child work, which is determined by the production capacity of children and parents (see Rosenzweig, 1981). Basu and Van (1998) claim that poverty induces parents to send their children to work: since low-income households cannot afford luxury goods such as schooling and leisure, they need their children to generate income. Alternatively, Ray (1999) asserts that credit market imperfections in emerging economies cause child labor, arguing that if poor families had access to credit, then they would send their children to school rather than work given the high returns on education.

6 4 2 0 2004

Fig. 1.

2005

Northeast 2006

Southeast 2007

South 2008

Mid-West 2009

Occupation Level among the 5
17-Year-Old Population (% of Total 5
17 Population). Source: IBGE, PNAD (2009).

200

MAURICIO MOURA AND RODRIGO BUENO

In Brazil, Kassouf (2002) concludes that the probability of reducing child labor and increasing school enrollment increases with a household’s income and the parent’s level of education. She finds an inverse relationship between the mother’s level of education and a child’s labor participation, as do Bahalotra and Heady (2003) in Ghana. Further, the latter authors find that a larger family increases the probability of child labor in Pakistan, similar to the conclusion of Patrinosm and Psacharopoulos (1994) for Paraguay. In Egypt the phenomenon of “dynastic poverty traps” results in child labor increasing by one-tenth in those families whose parents worked during their childhood (see Wahba, 2002). The same conclusion holds in the work of Emerson and Souza (2003), who attribute such a result to “social norms” explaining that parents are likely to view child labor as natural if they had to work during their childhood. Some papers consider the effects of child labor outcomes. Boyden et al. (1998) focus on the real impacts of work on children’s lives and claim that some work can be positive, or at least neutral, and an important vehicle for a child’s integration into society. Emerson and Souza (2011), applying data from Brazil, show that child labor has a significant negative impact on adult earnings for male children. On the other hand, Beegle, Deheja, and Gatti (2009), using panel data from Vietnam, demonstrate that the consequences of child labor can be ambiguous; the authors found significant negative impacts on education and a higher earnings gain for young adults previously engaged in child labor. Furthermore, Dumas (2012), assessing data from Senegal, finds that children’s participation activities are associated with lower adolescent cognitive achievement.

LAND TITLE AND CHILD LABOR: POTENTIAL MECHANISMS This is an empirical paper, thus it is not in its scope to develop theoretical mechanisms explaining how land title affects child labor. We use a reduced form to highlight that land titles have an important effect on child labor and provide an indication of the magnitude of said impact. However, we find unavoidable to discuss possible channels of transmission, but do not advocate for any specific mechanism. As such, we outline possible channels, which are not necessarily mutually exclusive, through which a land registration policy could positively affect child labor supply.

The Effect of Land Title on Child Labor Supply

201

Field and Torero (2002), for example, argue that untitled households need to provide informal policing, both to deter prospective intruders from invading private properties and to participate actively in community enforcement efforts to protect neighborhood boundaries. For them, this is an important mechanism by which the lack of land titling removes adults from the labor force and creates incentives for child labor, since it leads to adults guarding their property instead of working. Consequently, they end up sending their children to work to supplement the family income. In searching the literature, we found conclusions that those without formal property rights often state that their spouses look after the home while they go to work, usually taking their children with them. That means that the shift in property protection away from informal communities and households toward formal property ownership seems to be one of the main benefits of consolidated property institutions (see more details, for example, in Cockburn, 1998). Other articles (Carter & Zegarra, 2000; Field, 2007; World Bank, 2000; among others) conclude, in general, that the lack of formal institutions exacerbates the absence of property protection provision by the state. Thus, informal or absent property rights make households and communities invest their effort, time, skills, and resources into providing tenure security. By contrast, titling allows them to move away from these activities, releasing time to work and increasing their income. Andrade (2006) observes that land property is highly respected and preserved by local courts, thus changing the dynamics between residents and local police as mentioned above. De Soto (2000) argues that land titles open up formal credit markets to people otherwise unbanked. Besley (1995) points out that better-defined rights to land affect the share of wealth that can be pledged as collateral in a credit contract. Under certain conditions, we can expect a positive effect from a land title program on access to credit due to the increment in wealth. The author claims that credit can be used to afford durable goods (a proxy for investment) or for nondurable consumption. Furthermore, Dower and Potamites (2007) also show that titles play an important ex ante role in providing information about the applicant to the potential lender. The basic concept is that the bank would rather lend to titled households, not only because the title mitigates the bank’s risk in the case of a default, but also because the title provides information about the likelihood of default or improves the borrower’s credit score. Hence, easier credit access potentially decreases child labor by stimulating investments that increase the household income.

202

MAURICIO MOURA AND RODRIGO BUENO

In addition, land title increases home equity value given that a land title program is followed by a better supply of public goods such as security, electricity, garbage collection, and provision of a sewage connection (in the case of Osasco, the implementation of those public services were not immediately part of the land title program. However, several months after the titling those services
each under a different timetable and dynamic
started to be part of the community’s routine). The supply of public goods has positive externalities, such as increasing the housing value of a neighborhood and increasing household access to credit that enables them to establish a more stable consumption pattern overtime, avoiding the need for short-term income generated by child labor.

METHODOLOGICAL ISSUES, DATA, AND DESCRIPTIVE STATISTICS Minimizing Selection Bias Osasco is part of the federal program “Papel Passado” and home to about 30,000 people (or 6,000 families) living under informal conditions. That figure amounts to almost 4.5% of the city’s total population (ANOREG, 2007). Given the fiscal resources available, the program to award land titles to all communities is scheduled to last from 2007 to 2014. Jardim Canaa˜, with 500 resident families, was the first locality to receive land titles in 2007 (all the households from Canaa˜
without exception
were eligible and received the title). Its closest neighbor, DR, a community of 450 households, received land titles by the end of 2012.4 As Behrman and Todd (1999) argue, randomization avoids selection bias in program evaluation
a bias that is generally present in nonexperimental evaluations. They also note that randomization can prevent the problem of selfselection. However, other types of bias may occur in randomized designs, such as contamination and attrition. In our sample, 95% of the first survey participants
both from Jardim Canaa˜ and DR
did not expect to be awarded land title. A question was directly asked: “Do you expect to have a deed during the next twelve months?” Ninety-five percent responded no. Indeed, respondents were unaware of Papel Passado and its meaning, thus curbing potential behavioral deviation by households included in the program.

The Effect of Land Title on Child Labor Supply

203

Contamination bias5 is also avoided here, because the comparison group residents living outside the intervention locality could not benefit from the program. Besides, no alternative formal land title program exists, and the program does not provide a dropout option (no resident of Canaa˜ could choose to be excluded from receiving a land title). After receiving the title, the household may sell the property and move away from the locality. However, by then the household would have already been affected by the program, reducing the probability of attrition bias.6 Moreover, Jardim Canaa˜ and DR share similar economic and social characteristics. They are both official neighborhoods without a clear border separating them. No one walking in the area would be able to identify which location is Canaa˜ or DR without previous local knowledge. For example, both neighborhoods are located exactly 2.5 miles from downtown Osasco, having precisely the same access to Osasco’s main economic center. This helps to ensure that the intervention group is similar in many observable and unobservable characteristics to the group that did not receive land titles.7 Additionally, this paper uses a quasi-experimental design which, as defined by Shadish, Cook, and Campbell (2002), is characteristic of a particular type of study in which researchers lack control over the allocation of interventions or other factors studied.

The Data The door-to-door survey focusing on property rights (and answered by the family head
56% female and 44% male respondents) was carried out in two stages: before and after titling. To further minimize bias, neither the questionnaire nor the interviewer provided direct information to the households about the objectives of the research. Officially, respondents were told that the study was examining general living conditions in the city of Osasco. Specific questions about child labor included: (domestic work was not considered part of the answers): (a) Do any children/teenagers contribute to the family income? How many? (= more than 5 and less than 18 years old); (b) How many hours do they work daily? (interviewers were asked to check exactly how many hours each minor was working); and (c) How many days per week do the minors work? In 2008, an additional question was asked: Regarding hours of child work, is the number of hours greater, equal, or lower than one year ago? The

204

MAURICIO MOURA AND RODRIGO BUENO

questionnaire did not include a specific question about parents taking their children to work with them. The questionnaire was administered to 326 randomly selected households and includes 39 questions.8 The format of questions and methodology closely mirrors the national statistical survey (Pesquisa Nacional de Amostra de Domicı´lios, PNAD) from the Brazilian Statistical Bureau (Instituto Brasileiro de Geografia e Estatı´stica, IBGE). It also requests information on household and individual characteristics including the social, personal, and economic benefits associated with property ownership. The researchers who conducted the door-to-door survey were not from Osasco. They first administered the questionnaires in March 2007, before titles had been issued to households of Jardim Canaa˜. The second stage was carried out with the same households in August 2008 (with 2% missing interviews) about 17 months later. The time gap between stages was designed so that all households interviewed during the first stage would have possessed the land title for at least one year by the time of the second survey.9 The study also tracked the households that moved away from both communities. In contrast to the 8% of households that moved away from Canaa˜, only 0.7% of households (1 out of 141) moved away from DR during the same period.

Descriptive Statistics Tables 1 and 2 summarize the answers of the family heads (326) in 2007 and 2008 regarding the weekly number of hours of child work. Table 1 shows that for both groups combined (intervention and comparison), the weekly hours of child work decreased between 2007 and 2008. Table 2 reports the t-test results regarding differences in means between the comparison and intervention groups before the program for the covariates in 2007. It is evident that the number of observations (households) in the intervention and comparison groups is roughly comparable. However child labor, monthly income per capita, and informal labor are significantly lower, whereas education of the family head is significantly higher in the intervention than in the comparison group. While the intervention group is more educated and engages less in informal work, it has a lower income than the comparison group. This is corroborated by Spearman correlations in Table 3 and by the wealth index displayed in Table 2. The wealth index
computed using principal

205

The Effect of Land Title on Child Labor Supply

Table 1.

Descriptive Statistics: Selected Variables, 2007
2008, Both Communities Combined.

Variable

Weekly hours of adult work Ethnicitya Genderb Mean age Marital statusc Net monthly income (currency BRLd) Number of residents Number of children (between 10 and 18) Weekly hours of child work Years of education (family head) Observations (respondents in both surveys)

Pre-Intervention 2007

Post-Intervention 2008

Mean

Std. Dev.

Mean

Std. Dev.

10.19 2.75 0.43 40.89 1.98 1,126.25 3.89 1.12 6.80 7.25 326

12.22 1.40 0.47 14.68 0.80 1,491.92 1.61 1.04 1.23 4.34 326

16.18 2.75 0.33 41.89 1.98 1,138.76 3.96 1.13 6.06 7.31 326e

14.33 1.40 0.47 14.68 0.78 1,473.35 1.62 1.07 1.20 4.33 326

Sources: The Osasco Land Title Survey and the Central Bank of Brazil. 0 = White-Caucasian (35%), 1 = Afro-Brazilian (50%), 2 = Asian (8%), 3 = African Indian (5%), 4 = Indian (2%). b 0 = male, 1 = female. c 0 = single, 1 = married, 2 = widowed, 3 = divorced. d Currency exchange rate on 31 December, 2008: 1 USD = 1.75 BRL (Brazilian Reais) and excludes child work income. e It includes residents that moved away. a

component analysis (PCA)
summarizes the stock of durable goods owned by the households including TVs, radios, cars, washing machines, refrigerators, and freezers. In order to explain the results, recall that households with higher levels of education tend to have more access to formal jobs in Osasco (Zylberstajn & Neto, 1999) and naturally have disposable income mostly based on wages. Therefore, they tend to receive additional savings (FGTS and public pensions) and significant perquisites that are not reflected in their direct cash salaries, as is customary for formal employment in Brazil.10 On the other hand, informal workers do not receive these benefits, relying instead on cash income to compensate for the lack of perquisites, and in most cases they do not pay income tax and/or contribute to a public pension.11 Furthermore, Table 3 shows a positive correlation between monthly income per capita and children’s hours worked per week (0.11). This is

206

Table 2.

MAURICIO MOURA AND RODRIGO BUENO

t-Tests and z-scores for the Differences in Means for Covariates, 2007 (N = 326). Mean Mean Comparison (A) Treatment (B)

Gender (=1 if female, =0 if not) Ethnicity (=1 if African-Brazilian, = 0 if not) Marital status (=1 if married, =0 if not) Mean age (family head) Weekly hours of adult work Weekly hours of child labor (between 10 and 18 years old) Children working (=1 if yes, =0 if not) Years of education (family head) Net monthly income (currency BRLa) per capitab Wealth index Informal sector workerc (=1 if informal, =0 if not) Access to credit (=1 if yes, =0 if not) Number of residents Number of children (between 5 and 17) Observations (households)

Test: A
B = 0 p-value

0.31 0.69

0.34 0.64

0.48 0.43

0.61 42.60 10.10 7.55

0.65 39.40 10.40 4.60

0.52 0.06* 0.81 0.00***

0.34 5.00 553.10

0.14 9.00 255.80

0.00*** 0.00*** 0.00***

1.12 0.94

−0.94 0.65

0.00*** 0.00***

0.44 3.88 1.41 185

0.45 3.91 1.48 141

0.88 0.58 0.49

Source: The Osasco Land Title Survey and the Central Bank of Brazil. * and *** rejection of the null hypothesis at 10% and 1% respectively. a Currency exchange rate on 31 December, 2008, 1 USD = 1.75 BRL (Brazilian Reais). b Monthly income per capita is calculated dividing monthly income
without child work income
by the number of residents. c Defined as an individual without an active work card or a small business license.

Table 3.

Spearman Correlation, 2007 (N = 326). Years of Informality Education

Years of education Informality Net monthly income per capita Child weekly hours worked (between 10 and 18 years old)

1 −0.14** −0.21*** 0.11**

1 0.21*** 0.08

Net Monthly Income per Capita

Child Weekly Hours Worked

1 0.11*

1

Source: The Osasco Land Title Survey. *, **, *** rejection of the null hypothesis at 10%, 5%, and 1%, respectively.

The Effect of Land Title on Child Labor Supply

207

consistent with our underlying hypothesis that families financially rely on the number of hours worked by children to supplement monthly income.

EMPIRICAL STRATEGY The previous section discussed statistically significant differences between the intervention and comparison groups of our sample. These differences can be attributed to the randomization at the community level rather than at the household level. Along the same lines, Skoufias (2001) and Berhman and Todd (1999) observe that even if the comparison and intervention groups have similar characteristics at the community level, they are not fully comparable at the household level. These authors suggest the use of control variables rather than estimating the program impact only through mean tests. We follow this strategy and adopt two econometric procedures to address the differences in observable characteristics. The first is to include control variables and use the OLS and DD techniques. The second procedure is to use these same methods combined with a propensity score technique to select a more balanced sample. The following subsections explain in detail each of the empirical methods used.

OLS Regression Definition A standard OLS procedure is first used to investigate the mean effect of land titling through the following regression equation: Hi = α þ δ titlei þ βXi þ ɛi

ð1Þ

where Hi represents the number of hours worked weekly by children (between 10 and 17 years old) and Xi is a control vector of socio-economic variables that includes gender; ethnicity; marital status; years of schooling; number of members in the household; age; age2; a dummy for informal work; household assets; and monthly income per capita. These variables are common covariates for land title and are usually used to evaluate this type of program. The variable titlei is a dummy equals 1 if the household participates in the titling program, and 0 otherwise. Since all households in Canaa˜ received titles in 2007, the parameter of interest is the average treatment effect on the intervention (treated) group (ATT).

208

MAURICIO MOURA AND RODRIGO BUENO

The Difference-in-Difference Methodology: Definition The main econometric method applied is the DD estimator, which involves comparing the difference in outcomes before and after the intervention in the affected group (treated) and in the unaffected group (comparison) (see Bertrand, Duflo, & Mullainathan, 2004; Imbens & Wooldridge, 2008). This method is estimated with the following regression model: 0 γ þ μi;s;t Hi;s;t = β0 þ β1 treats;t þ β2 yeart þ αDD ðtreats;t × yeart Þ þ Xi;s;t

ð2Þ

where Hi,s,t is the number of hours worked weekly by child i in community s at time t; treats,t is a dummy variable equal to 1 if the household resides in the treated community (s = 1) and 0 otherwise; yeart is a dummy variable equal to 0 in 2007 (baseline period) and to 1 in 2008; Xi,s,t is a vector of observable characteristics of household i in community s, changing through time; and μi,s,t denotes the error term, assumed to be independent of Xi,s,t and yeart (see Imbens & Wooldridge, 2008; Meyer, 1995).12 Our parameter of interest is coefficient αDD, which identifies the intervention’s effect on the group (see Appendix A).

OLS, DIFFERENCE-IN-DIFFERENCE, AND DDM: EMPIRICAL RESULTS This section presents the results of the empirical analysis proposed previously. We start by discussing the findings (applying a subsample of the total 182 respondents out of 326). Table 4 shows that the average weekly hours of child work decreased in titled households and increased in nontitled households. DD calculates the difference between “after” and “before” values of the mean outcomes for each intervention and comparison group. The difference between mean differences is the impact estimate. In Table 4, the impact estimate for children’s weekly work hours is −3.55 hours.

209

The Effect of Land Title on Child Labor Supply

Table 4.

Average Weekly Child Hours Worked (Households with Children between 10 and 17 Years Old) (N = 182). Titled (Canaa˜) Non-Titled (DR) Difference Δ

1. Child work hours (before) 2. Child work hours (after) 3. Change in mean weekly child work hours

4.60 1.80 −2.80

7.55 8.30 0.75

2.95** 6.50** −3.55

**Significant at 5%.

The mean effects of the program are analyzed in Table 5, which presents OLS estimates for four regression models using data from 2008. The results support the claim that land titles decrease child labor as measured by the number of hours worked. Model OLS-Naı¨ ve excludes the control variables and estimates a negative and significant coefficient (−4.28) for the dummy, land title. The control variables are included in all other models and a significant coefficient is also obtained for land title. The estimates for the control variables also have the expected results. Applying a propensity score matching, which plays an important role in balancing study groups, based on Rosenbaum and Rubin (1984), a balanced model was also tested with a significant coefficient of (−3.41), in line with the analysis in Table 4. A higher educational level (years of education) of the household head along with easier access to credit leads to a decrease in the number of weekly hours worked by children. Furthermore, the main results remain the same with the inclusion of the variable hours of adult work. Notice that gender increases labor work of children. This happens most likely because women must be at home for more hours of the day. Therefore, the need for child labor is more important. Table 6 presents the DD estimations for the years 2007
2008 (pre- and post-titling). The results corroborate the previous findings that land titling decreases the average number of hours worked by children. All columns display a negative and statistically significant coefficient for the interaction term between the dummy for land title and the year of the program (2008). Likewise, most of the previous significant control variables in the OLS estimation remain significant when we use the DD method. The second and third columns show a similar impact, suggesting that conditioning the sample on the common support does not influence the point estimate of the interaction term. Again, the inclusion of the variable hours of adult work has not changed the overall outcome. Moura and De-Losso (2010), using

210

Table 5.

MAURICIO MOURA AND RODRIGO BUENO

OLS Estimates for Land Titling Impact on Child Labor (PostProgram), 2008 OLS (Eq. (1)).

Variables

Model 1 OLS Naı¨ ve

Model 2 OLS

Model 3 OLS balanced sample

Land title

−4.28*** (0.95)

−3.62*** (1.02) 1.29*** (1.03) −1.86 (1.54) 0.18 (0.52) 0.82 (0.56) −1.30* (0.05) −1.78** (0.21) −0.00 (0.00) −2.34** (1.71) 0.23 (0.66) 0.58 (1.82) 0.39 (1.28) 3.99** (5.18)

−3.41*** (0.94) 4.10*** (1.71) −1.27 (1.65) 0.31 (0.44) 0.03 (0.02) −0.07 (0.06) −2.02** (0.16) −0.00 (0.00) −2.89** (1.38) 0.36 (0.72) 1.48 (1.86) 0.47 (1.66) 5.57** (4.08)

0.16 182

0.11 136

Gender (=1 if female) Ethnicity (= 1 if African-Brazilian) Marital status (= 1 if married) Age Age2 Years of schooling (family head) Net monthly income (currency BRLa) per capitab Access to credit (=1 if have) Wealth indexc Informal worker Number of children (10 and 18 years old) Constant

3.89*** (1.12)

Sigma Log-likelihood Pseudo-R2/R2 Observations

0.09 182

Note: Robust standard errors are in parentheses. Currency exchange rate on 31 December, 2008: 1 USD = 1.75 BRL (Brazilian Reais). b Monthly income per capita is calculated by dividing monthly income by the number of residents. c Wealth index summarizes the total value of durable goods and the principal component analysis (PCA) was applied. Imbens and Wooldridge (2008) state that the PCA is a useful technique when explanatory variables are closely related.*, **, *** statistically significant at 10%, 5%, and 1%, respectively. a

211

The Effect of Land Title on Child Labor Supply

Table 6.

Difference-in-Difference: Land Title Impact on Child Labor (2007
2008) DD (Eq. (2)).

Variables

Weekly Hours Worked (DD Naı¨ ve)

Weekly Hours Worked (DD Unbalanced)

Weekly Hours Worked (DD Balanced)

Land title

−0.22 (1.21) −4.06*** (0.94) 0.90*** (0.28)

−0.16 (1.82) −3.07*** (1.11) 0.85*** (0.37) 1.95*** (1.45) −0.72 (1.11) −0.29 (1.32) −0.03 (0.02) −0.38** (0.04) −0.28** (0.18) −0.83 (0.00)

−0.21 (1.86) −3.12*** (1.13) 0.28*** (0.29) 1.71*** (1.42) −0.00 (1.12) −0.42 (1.34) −0.08 (0.03) −0.77* (0.04) −0.29** (0.19) 0.00 (0.00)

−1.33* (1.16) 0.49 (0.43) 0.007*** (1.21) 0.003*** (1.11) 2.21*** (1.71)

−1.25 (1.14) 0.75 (0.59) 0.70 (1.33) 0.22 (1.29) 2.48*** (1.09)

Land × year (DD) Year Gender (=1 if female) Ethnicity (=1 if African-Brazilian) Marital status (=1 if married) Age Age2 Years of schooling (head of household) Monthly income (currency BRLa) per capitab Access to credit Wealth indexc Worker type (=1 if informal) Number of children (10 and 17 years old) Constant R2 Observations

3.05*** (1.06) 0.09 364

0.12 364

0.14 272

Note: Robust standard errors are in parentheses. a Currency exchange rate on 12/31/2008: 1 USD = 1.75 BRL (Brazilian Reais). b Monthly income per capita is calculated by dividing monthly income by the number of residents. c Wealth index summarizes the total value of durable goods, and the PCA was applied. Imbens and Wooldridge (2008) state that the PCA is a useful technique when explanatory variables are closely related.*, **, *** statistically significant at 10%, 5%, and 1%, respectively.

212

MAURICIO MOURA AND RODRIGO BUENO

the same data set, show that land title has a positive effect on labor supply and income. The difference-in-differences-matching (DDM) estimates are shown in Table 7. The main advantage of this estimator is that it does not impose any functional form a priori and uses all observations of the comparison group to identify the parameter of interest (see Heckman, Ichimura, & Todd, 1997; Smith & Todd, 2005, for a detailed discussion of this estimator). Since there is a trade-off between bias and variance, we provide estimates for three different bandwidths. The lower bandwidth in the kernel estimator represents the bias while the higher bandwidth captures the variance. The first column of Table 7 shows that for the smallest bandwidth the point estimate is very similar to the DD naı¨ ve (see Table 6). This gives additional support for the quality of the data. As the bandwidth is enlarged, the point estimates become smaller in absolute value, as does the variance. Therefore, despite the relatively high variance in the first column, we believe that the average treatment effect on the intervention group must be around four hours/week (or four fewer hours of weekly child work). The overall magnitude interpretation allows different perspectives. First of all, assuming an average of four hours per week without engagement on child labor represents a 60% work hour reduction compared to pre-title context. As Field (2007) has shown child labor reduction has the potential to increase adult work hours, and as the magnitude increases (less hours of child work), greater potential to occur such transmission. Third, from a more simplistic perspective, the particular magnitude found The magnitude interpretation depends on the perspective: it could be marginal considering that four hours per week represents basically 48 minutes a day (assuming

Table 7.

Difference-in-Difference Matching Estimates, 2007
2008 Estimator (Eq. (A.3)).

Variables

Land title Observations

Bandwidths of Kernel Estimator (0.01)

(0.05)

(0.10)

−4.03*** (2.62)

−3.85*** (2.25)

−3.56*** (2.24)

364

364

364

Note: Standard errors (in parenthesis) are computed using a bootstrap with 100 replications. *** statistically significant at 1%, respectively.

The Effect of Land Title on Child Labor Supply

213

children work only on business days), or 24 minutes in the morning and 24 minutes in the afternoon. Last but not least, further research should focus on work balance (domestic x outside work) to check if those four hours were not, in practice, reallocated through domestic activities.

CONCLUSION Based on Field’s model (2007), which rationalizes the effect of land title on child labor force participation, we present evidence regarding the effect of property rights on child labor using data from a developing country. Several existing studies, such as Field (2007) and Andrade (2006), suggest that property rights affect income, credit, and labor supply. However, no current studies compare data from before and after land titling, which is the major advantage of our paper. Our empirical results support the finding that property titling decreases child labor force participation by about six hours per week. Moreover, we conclude that titling programs appear to have a different effect in terms of employment. While most welfare programs remove adult households from the labor force due to a positive income effect, this type of program induces adults to work more and children to work less. It is important to understand the process through which economic outcomes respond to land titles, mostly because governments from developing countries may be able to use these findings to reduce urban informality regarding ownership. These results may also help to explain labor market frictions, particularly in developing countries
a subject studied by Goldsmith (1995). High levels of residential informality coupled with informal property protection act as an obstacle to adjustments in the labor market. Therefore, public policies involving land titling could potentially play an important role in economic growth. This analysis offers various possibilities for further research. For example, the analysis of the increase in labor force participation due to land titling and its effects on the income and utility of households could be investigated. At this stage, the increase in income gained by households through increasing their labor force participation is unclear. The distributional impact of land titling could also be investigated, not only regarding labor supply and child labor but also other economic variables, such as access to credit and fertility. This would improve the assessment of the

214

MAURICIO MOURA AND RODRIGO BUENO

impacts of such programs on the lives of millions of households living in urban squatter communities in developing countries around the world.

NOTES 1. Defined in this paper as child work in paid activities outside of the child’s home (see Edmonds & Schady, 2009). 2. See Associac¸a˜o dos Nota´rios e Registradores do Brasil
ANOREG [15]. The quotation is freely translated from Portuguese by the authors. 3. This particular land title program operates only under public invaded areas. 4. Skoufias (2001), for example, uses a similar experiment to evaluate the income transfer initiative PROGRESA in Mexico. In that program some localities were randomly selected for participation (intervention localities) while the rest received land titles later (control localities). Such an assignment at the local level has the benefit of minimizing the chances of spillover effects between intervention and control study participants (groups/units) in the same area. 5. Contamination bias occurs if members of the (randomized-out) control group seek and receive alternative forms of treatment. This is usually a problem only when there are close substitutes for the intervention. 6. Attrition bias occurs if some members of the intervention group dropout of the program. If the purpose of the evaluation is to estimate the effect of receiving an intervention (for example, the effect of taking a drug over a period of time), then attrition bias can pose a major problem. It is not usually random and can compromise the benefits of randomization. 7. Rubin and Thomas (2000) indicate that estimates based on full (unmatched) samples are generally more biased and less robust to mis-specification of the regression function than those based on matched samples. 8. The questionnaire is available upon request. 9. The 2nd Osasco Office of Registration (2.º Carto´rio de Osasco) provided us with the exact date that each household received its property title after being formally authorized by the Osasco’s City Hall. 10. For example, a formal Brazilian employee usually receives a health care plan for the whole family, subsidized transportation, and a meal plan. Furthermore, formal employees have FGTS
a compulsory savings account under Brazilian Labor Law. FGTS is the Fundo de Garantia por Tempo de Servic¸o
Guarantee Fund for Time Service. Under the FGTS, employers deposit 1/12 of the worker’s pay in a restricted bank account whose balance is withdrawn and given to the worker if and when (s)he is fired without a just reason other than to decrease costs. 11. In our data sample, 233 households had informal sector workers. They represented 92% of the workers in the control group and 64% in the intervention group (see Table 2). 12. Once all households of the intervention group receive titles, s and yeart will be the same. Thus, from now on, the subscript s is omitted for the sake of simplicity.

The Effect of Land Title on Child Labor Supply

215

13. Smith andPTodd (2005), for instance, define the weighting function as GððPC − PT Þ=bn Þ= k ∈ C GððPk − PT Þ=bn Þ, where Gð⋅Þ is the kernel function and bn is the bandwidth, P stands for propensity score, C for control group, and T for treatment group.

REFERENCES Alston, L., Libecap, G., & Schneider, R. (1996). The determinants and impact of property rights: Land titles on the brazilian frontier. Journal of Law, Economics and Organization, 12, 25
61. Andrade, M. T. (2006). Direitos de propriedade e renda pessoal: Um estudo de caso das comunidades do caju. Revista do BNDES, 13(26), 261
274. Angelucci, M., & Attanasio, O. (2009). Oportunidades: Program effect on consumption, low participation, and methodological issues. Economic Development and Cultural Change, 57(3), 479
506. Associac¸a˜o dos Nota´rios e Registradores do Estado de Sa˜o Paulo
ANOREG. (2007). Sistema de biblioteca. Cartilha dos registros Pu´blicos (pp. 3
5). Sa˜o Paulo: Versa˜o III. Retrieved from http://www.anoreg.org.br/. Accessed on September 28, 2008. Baharoglu, D. (2002, July). World Bank experience in land management and the debate on tenure security. World Bank Housing Research Background
Land Management Paper. Basu, K., & Van, P. H. (1998). The economics of child labor. American Economic Review, 88(3), 412
427. Becker, G., & Lewis, H. G. (1973). On the interaction between the quantity and quality of children. Journal of Political Economy, 81(2), S279
S288. Beegle, K., Deheja, R., & Gatti, R. (2009). Why should we care about child labor? The education, labor market and health consequences of child labor. Journal of Human Resources, 44(4), 3
18. Behrman, J. R., & Todd, P. (1999). Randomness in the experimental samples of PROGRESA: Report submitted to PROGRESA. International Food Policy Research, Washington, DC, Mimeo. Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How much should we trust differences-indifferences estimates? The Quarterly Journal of Economics, 119(1), 249
275. Besley, T. (1995). Property rights and investment incentives: Theory and evidence from Ghana. The Journal of Political Economy, 103(5), 903
937. Bahalotra, S. R., & Heady, C. (2003). Child farm labor: The wealth paradox. The World Bank Economic Review, 17(2), 197
227. Binswanger, H., Deninger, K., & Feder, G. (1995). Power, distortions, revolt, and reform in agricultural land relations. Handbook of development economics (1st ed., Vol. 3, pp. 2659
2772). Cambridge, MA: Elsevier. Boyden, J., Ling, B., & Myers, W. (1998). What works for working children. Stockholm: Radda Barnen. Boyden, J., Ling, B., & Myers, W. (1999). What works for working children? Foreign Affairs, 78(May
June).

216

MAURICIO MOURA AND RODRIGO BUENO

Brasselle, A., Gaspart, F., & Platteau, J. (2002). Land tenure security and investment incentives: Puzzling evidence from burkina faso. Journal of Development Economics, 67(2), 373
418. Carter, M., & Olinto, P. (2003). Getting institutions right for whom? Credit constraints and the impact of property rights on the quantify and composition of investment. American Journal of Agricultural Economics, 85, 173
186. Carter, M., & Zegarra, E. (2000). Land markets and the persistence of rural poverty in latin America: Conceptual issues, evidence and policies in the post-liberalization era. In R. Lopez & A. Valdes (Eds.), Rural poverty in latin America: Analytics, new empirical evidence and policy. Basingstoke: MacMillan Press. Cockburn, J. A. (1998). Regularization of urban land in Peru. Land lines. Cambridge, MA: Lincoln Institute of Land Policy. De Soto, H. (2000). O Miste´rio do capital. Rio de Janeiro: Editora Record. Demsetz, H. (1967). Toward a theory of property rights. American Economic Review, 57(2), 347
359. Do, Q. T., & Iyer, L. (2003). Land rights and economic development: Evidence from Vietnam. Policy Research Working Paper Series No. 3120. The World Bank. Dower, P., & Potamites, E. (2007). Signaling credit-worthiness: Land titles, banking practices and access to formal credit in Indonesia. Working Paper 0155. Department of Economics, New York University. Dumas, C. (2012). Does work impede child learning? The case of senegal. Economic Development and Cultural Change, 60(4), 773
793. Edmonds, E. (2008) Defining child labour: A review of definitions of child labour in policy research. International Programme on the Elimination of Child Labor, Geneva, pp. 21. Edmonds, E., & Schady, N. (2009). Poverty alleviation and child labor. Working Paper No. 15345. National Bureau of Economic Research, Cambridge, MA. Emerson, P. M., & Souza, A. P. (2003). Is there a child labor trap? inter-generational persistence of child labor in Brazil. Economic Development and Cultural Change, 51(2), 375
398. Emerson, P. M., & Souza, A. P. (2011). Is child labor harmful? The impact of working earlier in life on adult earnings. Economic Development and Cultural Change, 59(2), 345
385. Field, E. (2007). Entitled to work: Urban property rights and labor supply in peru. The Quarterly Journal of Economics, 122(4), 1561
1602. Field, E., & Torero, M. (2002). Do property titles increase credit access among the urban poor? Evidence from Peru. Working Paper No. 223. Research Program in Development Studies, Princeton University. Galiani, S., & Schargrodsky, E. (2010). Property rights for the poor: Effects of land titling. Journal of Public Economics, 94(9
10),700
729. Goldsmith, A. A. (1995). Democracy, property rights and economic growth. Journal of Development Studies, 32(2), 157
174. Heckman, J., Ichimura, H., & Todd, P. E. (1997). Matching as an econometric evaluation estimator: Evidence from evaluating a job training program. Review of Economic Studies, 64, 605
654. Imbens, G. W., & Wooldridge, J. M. (2008). Recent developments in the econometrics of program evaluation. NBER Working Paper No. W14251. Jacoby, H. G., Li, G., & Rozzelle, S. (2002). Hazards of expropriation: Tenure insecurity and investment in rural China. American Economic Review, 92(5), 1420
1447.

The Effect of Land Title on Child Labor Supply

217

Jimenez, E. (1985). Urban squatting and community organization in developing countries. Journal of Public Economics, 27, 69
92. Kassouf, A. L. (2002). Aspectos So´cio-Econoˆmicos do Trabalho Infantil no Brasil. Ministe´rio da Justic¸a, Secretaria de Estado dos Direitos Humanos, Brası´ lia, p. 123. Lanjouw, J. O., & Levy, P. (2002). Untitled: A study of formal and informal property rights in urban ecuador. The Economic Journal, 112(482), 986
1019. Meyer, B. D. (1995). Natural and quasi-experiments in economics. Journal of Business and Economic Statistics, 13(2), 151
161. Moura, M. J. S. B., & De-Losso, R. (2010). Some notes on how land title affects child labor. Revista Economia, 11(2), 357
382. Moura, M. J. S. B., & De-Losso, R. (2013). Land title program in Brazil: Are there any changes to happiness? Journal of Socio-Economics, III, 196
203. North, D. C. (1990). Institutions, institutional change and economic performance. Cambridge, MA: Cambridge University Press. North, D. C., & Thomas, R. P. (1973). The rise of the western world: A new economic history. Cambridge, MA: Cambridge University Press. Patrinos, H. A., & Psacharopoulos, G. (1994). Educational performance and child labor in paraguay. International Journal of Educational Development, 15, 47
60. Place, F., & Migot-Adholla, S. (1998). The economic effects of land registration for smallholder farms in Kenya: Evidence from Nyeri and Kakamega districts. Land Economics, 74(3), 360
373. PNAD – Sı´ ntese dos Indicadores. (2008). Microdados. Rio de Janeiro: IBGE. Retrieved from http://www.ibge.gov.br/home/estatistica/populacao/trabalhoerendimento/pnad2008/ sintesepnad2008.pdf. Accessed on September 28, 2013. PNAD
Sı´ ntese dos Indicadores. (2009). Microdados. Rio de Janeiro: IBGE. Retrieved from http://www.ibge.gov.br/home/estatistica/populacao/trabalhoerendimento/pnad2007/ sintesepnad2007.pdf. Accessed on September 8, 2010. Ray, R. (1999). Child labor, child schooling and their interaction with adult labor. The World Bank Economic Review, 14(2), 347
367. Rosenbaum, P. R., & Rubin, D. B. (1984). Reducing bias in observational studies using subclassification on the propensity score. Journal of the American Statistical Association, 79, 516
524. Rosenzweig, M. (1981). Household and non-household activities of youths: Issues of modelling, data and estimation strategies. In G. Rodgers & G. Standing (Eds.), Child work, poverty and underdevelopment. Geneva: ILO. Rubin, D., & Herzog, T. N. (2000). Combining propensity score matching with additional adjustments for prognostic covariates. Journal of the American Statistical Association, 95, 573–585. Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002). Experimental and quasi-experimental designs for generalized causal inference. New York, NY: Houghton Mifflin Company. Skoufias, E. (2001). PROGRESA and its impacts on the human capital and welfare of households in rural Mexico: A synthesis of the results of an evaluation. Washington, DC: IFPRI International Food Policy Research Institute, International Food Policy Research Institute. Smith, J. A., & Todd, P. E. (2005). Does matching overcome lalonde’s critique of nonexperimental estimators? Journal of Econometrics, 125, 305
353. Torstensson, J. (1994). Property rights and economic growth: An empirical study. Kiklos, 47(2), 231
247.

218

MAURICIO MOURA AND RODRIGO BUENO

United Nations (2005). Habitat report. New York, NY: United Nations. Wahba, J. (2002). The influence of market wages and parental history on child labor and schooling in Egypt. Working Paper 6829. Department of Economics, University of Southampton. World Bank. (2000, February). Urban poor gain access to property market. World Bank Development New Archives, Peru. Zylberstajn, H., & Neto, G. (1999). As teorias de desemprego e as polı´ ticas pu´blicas de emprego. Estudos Econoˆmicos, 29(1), 129
149.

219

The Effect of Land Title on Child Labor Supply

APPENDIX A The causal effect identification on the outcomes’ variables relies on the following assumptions: (i) Selection for the intervention does not depend on unobservable individual and community characteristics changing over time; (ii) Differences between the intervention and comparison groups would be the same in the absence of the program (i.e., there is a time invariant common effect); and (iii) The intervention does not affect child labor in households living in the neighboring areas; therefore, no spillover effects are present. .

These assumptions imply that Eðui;s;t jtreat; year; XÞ = Eðui;s;t Þ = 0

ðA:1Þ

and  9 8 EðHi;s;t jtreat = 1; year = 2008; XÞ − > > > −> = < EðH jtreat = 1; year = 2007; XÞ i;s;t   = ðβ2 þ αDD Þ − ðβ2 Þ = αDD ðA:2Þ > > EðH jtreat = 0; year = 2008; XÞ − i;s;t > > ; : EðHi;s;t jtreat = 0; year = 2007; XÞ The main problem regarding this subject is a possible selfselection, also known as an anticipation problem. This certainly would be an issue if households decided to make their children work fewer hours given the expectation of receiving a land title. Regarding assumption (ii), we use control variables to account for differences between the two groups in the baseline year (2007). In addition, the fixed effect estimator is applied to check for the robustness of the results given that the dependent variable may differ across groups but remain invariant through time. We also estimate the DDM estimator suggested by Heckman et al. (1997) and employed by, among others, Smith and Todd (2005) and Angelucci and Attanasio (2009). The main advantage of such an estimator is that it does not impose any functional form on the regression model and uses a kernel function to weight the subsample of the comparison group. The aim of the weighting function is to construct a counterfactual mean effect based on the distance between the propensity score of each

220

MAURICIO MOURA AND RODRIGO BUENO

comparison group observation and the intervention group observation.13 The identification of the ATT requires that EðΔHC jP; Land = 1Þ = EðΔHC jP; Land = 0Þ where P refers to propensity score and C stands for comparison group. Under this condition, the DDM estimator for the ATT is given by: 2 3 X 1 X 4 Hi;1;1 − Hi;1;0 − wði; jÞðHi;0;1 − Hi;0;0 Þ5 αDDM = n1 i ∈ T∩Sp j ∈ C∩Sp

ðA:3Þ

where n1is the subsample of treated units, wð⋅Þ is the weighting function, and the subsets below the summations indicate that the estimate is computed in the common support, Sp (see Smith & Todd, 2005) (P stands for propensity score, C for comparison group, and T for treatment group).

221

The Effect of Land Title on Child Labor Supply

APPENDIX B Table B1. Rural Column % Urban Column % Total %

Land Titling Brazil 2007.

Untitled

Titled

Total

2,014,497 27.43 5,328,763 72.57 7,343,260 100.00

19,989,515 16.05 104,540,315 83.95 124,529,830 100.00

22,004,012 16.69 109,869,078 83.31 131,873,090 100.00

Source: PNAD (2008), Brazil.

Table B2.

Propensity Score: Logit Estimates for the Selection of the Treatment Group, 2007.

Variables

Gender (=1 if female) Ethnicity (=1 if African-Brazilian) Marital status (=1 if married) Age Weekly hours of adult work Weekly hours of child work (5
17 years old) Years of education (head) Net monthly income per capita TV (=1 if have) DVD (=1 if have) Radio (=1 if have)

Dummy = 1 if a Household Lives in the Intervention Area (Canaa˜)

Dummy = 1 if a Household Lives in the Intervention Area (Canaa˜)

(Unmatched Sample)

(Matched Sample)

0.42 (0.48) 0.04 (0.45) 0.58 (0.47) −0.03* (0.01) 0.02 (0.01) −0.03 (0.02) 0.14*** (0.05) −0.01** (0.00) −1.48** (0.69) −0.64 (0.53) −1.68*** (0.50)

0.11 (0.51) 0.02 (0.45) 0.35 (0.49) −0.01 (0.01) 0.01 (0.01) −0.01 (0.02) 0.05 (0.08) −0.01 (0.00) −0.68 (0.85) −0.29 (0.58) −0.60 (0.84)

222

MAURICIO MOURA AND RODRIGO BUENO

Table B2. Variables

Car (=1 if have) Washing machine (=1 if have) Refrigerator (=1 if have) Informal worker Credit Constant Pseudo-R2 Prob > Chi2(16) Observations

(Continued )

Dummy = 1 if a Household Lives in the Intervention Area (Canaa˜)

Dummy = 1 if a Household Lives in the Intervention Area (Canaa˜)

(Unmatched Sample)

(Matched Sample)

−0.28 (0.45) 2.19*** (0.65) −6.07*** (1.07) −1.73*** (0.62) −0.17 (0.43) 8.18***

−0.09 (0.48) 1.06 (0.92) −2.76 (2.15) −0.75 (0.85) −0.03 (0.45) 1.87

(1.62) 0.62 0.00 305

*, **,*** Statistically significant at 10%, 5%, and 1%, respectively.

(4.09) 0.63 1.00 288

THE CHANGING TIME USE OF U.S. WELFARE RECIPIENTS BETWEEN 1992 AND 2005 Marie Connolly ABSTRACT This paper looks at the changes in the time allocation of welfare recipients in the United States following the 1996 welfare reform and other changes in their economic environment. Time use is a major determinant of well-being, and for policymakers to understand the broad influences that their policies can have on a population they ought to consider changes in all activities, not simply paid work. While an increase in market work of the welfare population has been well documented, little is known on the evolution of the balance of their time. Using the Current Population Survey to model the propensity to receive welfare, together with a multiple imputation procedure, I replicate previous difference-indifferences estimates that found an increase in child care and a decline in nonmarket work. However when additional data sources are used, I find that time spent providing child care does not increase. This is especially relevant as welfare recipients are overwhelmingly poor single mothers and the welfare reform increased time at work with ambiguous effects on time spent with children. I also find that time at work follows business

Factors Affecting Worker Well-Being: The Impact of Change in the Labor Market Research in Labor Economics, Volume 40, 223
255 Copyright r 2014 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-912120140000040008

223

224

MARIE CONNOLLY

cycles, with dramatic increases in work time throughout the strong economy of the late 1990s, accompanied by less time in leisure activities. Keywords: Welfare reform; time use; child care; leisure JEL classifications: J22; J13; I38

INTRODUCTION In 1996, the United States Congress passed the Personal Responsibility and Work Opportunity Reconciliation Act (PRWORA), establishing key elements of the welfare reform of the mid 1990s. The Act brought greater program authority to the states, as well as changes in financing
the block grant TANF (Temporary Assistance for Needy Families) replaced the AFDC (Aid to Families with Dependent Children)
and introduced ongoing work requirements, incentives to reduce nonmarital births, a fiveyear maximum time limit, and limits on eligibility for Food Stamps and Supplemental Security Income among certain populations (Blank, 2002). The general idea behind the reform was to give welfare recipients incentives to go back to work, and eventually to get off the welfare rolls. A substantial body of research has been devoted to the study of the welfare reform and its various aspects, documenting among others the sharp fall in welfare receipt and the increase in work. As often noted, identifying the impact of the welfare reform per se has proven difficult due to other changes affecting the poor, such as expansions in the Earned Income Tax Credit (EITC) of the late 1980s to mid-1990s and the implementation of welfare waivers by up to 43 states by 1996 (Meyer & Rosenbaum, 2001). Up to now however, little has been done to look at how the time allocation of the welfare-receiving population has evolved over the past two decades, and specifically at their time outside of the workplace, as impacts on paid work has received more attention (Meyer & Rosenbaum, 2001). Time use is a major input to an individual’s well-being (Krueger, 2009), so any effect a policy has on time use should be of interest to policymakers when considering the broad range of a policy’s effects. Given that the bulk of welfare recipients consists of poor single mothers, the changes in time devoted to child care could influence children well-being, yet little is known about these changes. This is no doubt due to the limited number of largescale time-use surveys in the United States up to now. A new survey, the American Time Use Survey (ATUS) conducted by the Bureau of Labor

The Changing Time Use of U.S. Welfare Recipients

225

Statistics since 2003, filled the void with an annual, year-round, nationally representative survey collecting over 13,000 time diaries each year. Two major elements of the reform are susceptible to bring changes to the time allocation of welfare recipients. Ongoing work requirements dictated that by 2002, at least 50% of all recipient families and 90% of two-parent families had to be working or in work preparation programs. The obvious effect of this requirement is to increase work time. For families with preschool-age children, this work increase implies that some form of child care arrangement must be found if mothers previously providing child care now need to work. The effect on time spent in child care is thus expected to be negative for mothers of young children. However note that child care can be divided into primary and secondary care: primary care is time actively directed towards the child, like feeding a child or reading to a child; secondary care is more passive, like watching over a playing child while cooking dinner. Thus welfare reform and other policy changes from the 1990s may not have a direct impact on primary child care, but could reduce the time children spend in contact with their parents during secondary child care. The second element of the welfare reform that can impact time use is the five-year time limit. The PRWORA set a lifetime limit of 60 months of welfare benefits. States were allowed to make some exceptions to these limits. A limit on the receipt of welfare would have a similar effect to a work requirement by inducing welfare recipients to find work more quickly in order to avoid a halt to their payments. However, the time limits are not likely to have an effect in this study given that the latest year of data is 2005 (8 or 9 years after the implementation of the reform), leaving little time for beneficiaries to hit their five-year lifetime limit. This paper combines data from previous smaller-scale time-use surveys conducted by the University of Maryland’s Survey Research Center (namely, the National Time Use Survey (NTUS), the Family Interaction, Social Capital, and Trends in Time Use Survey (FISCT), and the National Survey of Parents (NSP)) with the more recent ATUS to track the trends in Americans’ time use. Previous work by Aguiar and Hurst (2007) measured the trends in leisure since 1965 for all Americans, but did not focus on the welfare-receiving population. A paper by Meyer and Sullivan (2008) explored the changes in income and consumption before and after the welfare reform using a variety of data sources, including time-use surveys. Meyer and Sullivan used the NTUS (1992
1994) and the ATUS (2003) as pre- and post-period datasets, respectively, comparing single mothers to single women without children and to married women. They explained that their analysis focused on single mother headed families because they

226

MARIE CONNOLLY

comprise a large fraction of poor children in America, and because “selecting the sample based on demographic characteristics is preferable to restricting attention to families that report limited resources, because the latter approach will cause the sample to depend too much on the specific method used to measure income and/or consumption in each dataset.” (Meyer & Sullivan, 2008, p. 2224) They found that “the increase in time spent in market work is associated with declines in nonmarket work rather than declines in nonwork time. There is evidence of less time spent in food preparation, housework and shopping.” (Meyer & Sullivan, 2008, pp. 2234
2235) Their non-work time category included child care, which may have been the biggest driver of the increase in non-work time. My analysis improves on Meyer and Sullivan’s work in two ways. First, by using a propensity score as a treatment indicator in the difference-indifferences analysis. This provides an alternative way of defining the control and treatment groups, thus serving as a check on their methodology. Second, and more importantly, by using additional data sources, enabling us to investigate what happened between 1994 and 2003. Specifically, I used the Current Population Survey (CPS) March Supplements from the prewelfare reform period to model the propensity to receive welfare under the old regime, and then estimated a difference-in-differences model to compare the time use of individuals more or less likely to receive welfare, before and after the reform. A multiple imputation (Rubin, 1987) technique was used to compute the correct standard errors, treating the predicted propensity scores as missing information from the surveys, or non-response. A drawback of the difference-in-differences approach is that I cannot be sure that I purely identify the effect of the reform, as other changes concurrently happening to the welfare population will also be captured. I acknowledge this weakness, and can thus only say that I identify changes that concerned the likely welfare beneficiaries over the 1990s and 2000s. When using the same surveys as Meyer and Sullivan, the NTUS and the ATUS, similar results were obtained, with less time spent in nonmarket work, more time in child care activities, and a non significant coefficient for time spent working. The results no longer match up as closely when additional, smaller-scale time-use surveys conducted between the NTUS and ATUS were used. Using those, I found an increase in work time for wouldbe welfare recipients after the reform, mainly driven by an increase in hours worked during the strong economic conditions of the late 1990s. An analysis with CPS data supports this claim. Another interesting result is that there is now no change in nonmarket work time, and that the increase in time spent in child care is much smaller and not statistically different from

The Changing Time Use of U.S. Welfare Recipients

227

zero. The data show that the increase in time spent providing child care started before 1996. For proponents of the welfare reform, it is interesting to note that time spent with children did not decrease when work time increased, contrary to alarmist claims that predicted that children of welfare recipients going back to work would be left unsupervised or not receive appropriate care from their mothers. The remainder of this paper is organized as follows. The next second section presents the data sources used and some of the data issues that arose during the analysis while the section following this explains the empirical strategy. Findings are discussed next, and the conclusion appears after this. Details on the multiple imputation procedure can be found in the appendix.

DATA Five time-use surveys are used in this paper: two phases of the NTUS, the first (and more substantial) wave conducted for the Environmental Protection Agency (EPA) from 1992 to 1994 (referred to as NTUS in the text), and the second, conducted for the Electric Power Research Institute (EPRI) in 1994 and 1995 (referred to as EPRI in the text); the Family Interaction, Social Capital, and Trends in Time Use (FISCT) survey, in 1998 and 1999, the NSP, in 2000 and 2001, and the ATUS, which started in 2003, and whose 2003
2005 data are used here.1 Table 1 lists these surveys and gives some information about the sample. While the ATUS contains rich and detailed information on demographic and socioeconomic indicators, the earlier ones only have limited information. In particular, the NTUS does not contain marital status information, but only variables on household composition that give the number of adults in the household and the age of the youngest child. Appendix Table A2 contains the summary statistics of the main variables used from the time-use surveys. Note the differences between the NSP and other surveys, the former containing only parents. I do not think that the findings suffer from this difference in the universe sampled, as the analysis was also performed without the NSP and the broad conclusions remained unchanged. All the surveys come with sample weights, which were first rescaled to sum to one within any given survey, following Aguiar and Hurst’s (2007) methodology. All regressions were estimated using weighted data, to ensure that the surveys provided nationally representative estimates, and that each data source was equally

228

MARIE CONNOLLY

weighted. Doing so, I implicitly assume that all surveys are of equal quality with respect to non-response and representativeness. Table 1 reports the response rates, all within the 56
65% range, and all surveys were designed to provide representative estimates. Moreover, the first four surveys are very comparable and were conducted by the same set of principal investigators,2 providing continuity across the data sets and lending credence to the assumption of equal quality. Without equal weighting, the ATUS would dominate other surveys due to its much larger sample sizes. The total number of minutes in a day spent in different time-use categories were computed from the surveys’ time diaries, following Aguiar and Hurst’s (2007) activity classification. It is important to note that the first four surveys used the exact same activity classification, thus enabling us to meaningfully infer trends from these datasets once assembled together. Only the more recent ATUS used a different classification of activities, but it contains more detailed categories than the previous surveys. Given that broad time use categories are used and that all surveys report time use in the same unit of measurement (minutes), we can compare time use through time reasonably and confidently. Appendix Table A1 shows the different time-use variables and which activities they consist of.3 Appendix Table A3 presents the mean time spent in the main time-use activities for the different surveys, for all days of the week pooled together. Once again, the values for the NSP are not quite in line with those from the other surveys, since all respondents are parents. It is important to note that, for all surveys, the activities reported refer to the primary activity, that is, the main activity, or the one reported by the respondent. Thus, if somebody is multitasking, for example, the person is eating in front of the TV while talking to her spouse and watching the baby, what will appear in the time-use survey depends on what the respondent mentions as a primary activity
it could be eating, watching TV, child care, or socializing. This may pose a problem for the study of certain activities that are more prone to multitasking, such as child care and eating, but only as long as welfare recipients systematically changed the reporting of their main activities differently from non-recipients since the estimations are based on a difference-in-differences methodology. The only secondary activity collected in the ATUS pertains to child care. Unfortunately, no comparable measures of secondary child care exist in the earlier NTUS, and so the comparison of time spent in secondary child care between the pre- and the post-period is not possible. Additionally, Egerton, Fisher, and Gershuny (2005) and Aguiar and Hurst (2007) documented an increase in the reported time in child care between the 1992 and 1994 NTUS and the 2003 ATUS, arguing that part of the increase might be due to an

Survey and Dates of Collection National Time Use Survey (NTUS), Phase I: conducted for the Environmental Protection Agency (EPA), September 1992 to October 1994

National Time Use Survey (NTUS), Phase II: conducted for the Electric Power Research Institute (EPRI), July 1994 to July 1995 Family Interaction, Social Capital, and Trends in Time Use (FISCT), March 1998 to December 1999 National Survey of Parents (NSP), May 1999 to June 2000 American Time Use Survey (ATUS), January 2003 to December 2005

Time-Use Data Sources. Sample and Response Rates

Total Sample Size

Analysis Sample Sizea

Random-digit telephone survey, one adult 7,514 adults selected using the “Next Birthday” method, that is, the person whose birthday was the next to be celebrated was chosen as the respondent. Conducted by the University of Maryland’s Survey Research Center, using CATI software. Response rate: 63.0%. Same as NTUS Phase I. Response rate: 64.6%. 1,200 adults

573

Same as NTUS. Response rate: 55.5%.

1,151 adults

491

Same as NTUS. Response rate: 63.5%.

1,199 parents

572

Households that have completed their final month in sample for the CPS are randomly selected, and then one individual aged 15 or above is randomly selected, to whom a day of the week is randomly assigned. Conducted over the phone by the Bureau of Labor Statistics, using CATI software. Response rates: 57.8% in 2003, 57.3% in 2004 and 56.6% in 2005.

20,720 in 2003, 13,973 in 2004, 13,038 in 2005

19,293 in total

2,941

The Changing Time Use of U.S. Welfare Recipients

Table 1.

a

229

The analysis sample contains all women aged 18
59 years, inclusively. Note: All surveys are nationally representative, except for the NSP whose universe is parents living with at least one of their own children under 18.

230

MARIE CONNOLLY

undersampling of households with children in the NTUS, or to a change in the coding of activities. Whatever the reason and as mentioned about the multitasking issue, because my analysis depends on a difference-indifferences framework, my results will only suffer from this possible bias if it affects differentially the population who receives welfare benefits and the one that does not, which I do not believe is the case. The CPS March Supplements, known as the Annual Demographic Files until 2002 and as the Annual Social and Economic Supplement since 2003, are used to model the propensity to receive welfare benefits.4 The Supplements from 1992 to 1995 inclusively are used as the pre-welfare reform data. The welfare recipients are identified as individuals reporting an income from public assistance or welfare. Fig. 1 shows the sharp decline of the welfare rolls in the 1990s. The peak in welfare receipts was in 1994, with just over 6% of women receiving welfare, and just over 1% of men. The mid- to late 1990s saw a big decline in the proportion of adults on the rolls, stabilizing around 2001
2002 to just under 2% for women, and 0.3% for men, cutting the rolls to a third of their 1994 size. Overall, the proportion went from 3.7% in 1994 to 1% in 2005. While the PRWORA was only passed in 1996, some of the decline actually started before that date. The welfare rolls fluctuate with the economy, and Proportion of adults receiving welfare benefits 0.06 Date PRWORA passed 0.05 0.04 0.03 0.02 0.01 0 1992

1994

1996

Women

Fig. 1.

1998 2000 Year Men

2002

2004

All

Welfare Rolls Declined Sharply in the 1990s. Source: Author’s calculations using the CPS March Supplements.

231

The Changing Time Use of U.S. Welfare Recipients

Average welfare benefits for welfare recipients

Nominal dollars

4000

3500

3000

2500 Date PRWORA passed 2000 1992

1994

1996 Women

Fig. 2.

1998 Year

2000 Men

2002

2004 All

Average Welfare Benefits Peaked in 1996. Source: Author’s calculations using the CPS March Supplements.

many states had implemented various waiver programs. The average annual welfare benefits peaked in 1996, as shown in Fig. 2. They declined from a peak of about $3,700 for women (slightly more for men) to just under $3,000 for both men and women in 2005 (all values in nominal terms). As will be explained in the next section, my empirical strategy will take into account the effect of the waivers as well as the economic conditions. In order to do so, I supplemented the time-use data with information on waivers (monthly data on whether a state had implemented a waiver up to that point, information found in Crouse (1999) and Grogger and Karoly (2005)), monthly unemployment rates by state (taken from Local Area Unemployment Statistics on the Bureau of Labor Statistics’ website) as well as annual per capita real GDP growth by state (percent change from the preceding period, from the Bureau of Economic Analysis’ website).

EMPIRICAL STRATEGY To look at the changes in time allocation following the welfare reform, I want to compare welfare recipients from the pre-period to individuals who

232

MARIE CONNOLLY

would have been welfare recipients absent the reform in the post-period. It can be argued, as in Hubbard, Skinner, and Zeldes (1995), that given that a program like welfare affects not only welfare recipients themselves but also other individuals eligible for or nearly eligible for welfare, a policy’s impact should be evaluated without conditioning on welfare receipt. Hence one could focus on single mothers versus single women without children or married mothers, as done by Meyer and Sullivan (2008). I present here an alternative approach where the propensity to receive welfare benefits is predicted and used in the context of a difference-in-differences model. A propensity score measuring the likelihood of an individual to receive welfare is used instead of a treatment indicator, giving some amount of weight to people who are not actually receiving benefits but who are at risk of being so. This approach has the benefit of using a number of variables to create the score, compared to the one dimension used by Meyer and Sullivan (2008), being a single mother. Specifically, CPS Supplements were used to model the propensity to receive welfare during the period preceding the reform. The following probit equation was estimated on the pre-period data (1992
1995). I only use pre-reform years in this model since my goal is not to identify individuals who do receive welfare benefits, but those who would have been likely to receive assistance before the reform occurred.  ð1Þ Prðreceiving welfarejXi Þ = Φ β0p Xi where Xi is a vector of control variables for individual i. The β^ p estimated were used to construct propensity scores for each individual in the time-use surveys, which will serve as the treatment indicator, or propensity of treatment. Note that constructing the scores is akin to performing a multiple imputation, so that a Rubin (1987) procedure was followed to compute correct estimates and variances. The details can be found  0 in the appendix. Estimated propensities, or p^i , were computed as p^i = Φ β^ p Xi . Those estimated propensities became the treatment variable in the following difference-in-differences equation, Specification I, estimated by ordinary least squares, weighted using the sampling weights: timeit = α þ Xi0 β þ γ 1 p^i þ γ 2 POSTt þ γ 3 p^i  POSTt þ δ1 waiverst þ δ2 p^i  waiverst þ ɛ it ;

ð2Þ

where timeit is the amount of time individual i spent in a given activity at time t, POSTt indicates the time period after the reform, p^i  POSTt is an

The Changing Time Use of U.S. Welfare Recipients

233

interaction term capturing the difference-in-differences element, waiverst indicates whether the individual lived in a state s where a welfare waiver was implemented before time t and p^i  waiverst is an interaction term between the estimated propensity and the waiver dummy. Total time in a day is fixed, such that the sum of time-use variables must also be fixed. The time-use category “other” ensures that total time is fixed by absorbing any remaining differences. Time-use decisions are made jointly, which implies that more time spent working necessarily means less time spent doing something else. Each time-use category is the subject of a separate estimation. However, given that the set of predictor variables is the same for all regressions a methodology based on ordinary least squares will be efficient. In Eq. (2), γ3 is the parameter of interest showing the post-reform time-use changes. To take into account the fact that many states had implemented waivers prior to 1997, I controlled for whether a waiver had been implemented and allowed for the waivers’ differential effects by propensity to receive welfare using the interaction term. That way, the identification of γ3 is cleaner since it is against a pre-reform and pre-waiver baseline. Note that I do not need to control for varying implementation dates as all states had implemented the reform by March 1998, when my first post-reform data were collected.5 Specification I (Eq. (2)) is estimated not only using the NTUS and ATUS as Meyer and Sullivan (2008) did, but also using the additional time-use datasets for the period between the NTUS (1992
1994) and ATUS (2003
2005), namely the EPRI from 1994 and 1995 (pre-reform), the FISCT from 1998 and 1999 (post-reform), and the NSP from 1999 and 2000 (post-reform). Using the extra datasets, the analysis can be expanded to investigate the progression between 1994 and 2003. This second specification allows for a more flexible setup, introducing dummies for each dataset and interactions of the dummies with the estimated propensities p^i . The parameters of interest are γ 03 , γ 05 , γ 07 , and γ 09 . The following equation was used to estimate Specification II, where again the effect of the waivers is netted out: timeit = α þ Xi0 β þ γ 01 p^i þ γ 02 EPRIt þ γ 03 p^i  EPRIt þ γ 04 FISCTt þ γ 05 p^i  FISCTt þ γ 06 NSPt þ γ 07 p^i  NSPt þ γ 08 ATUSt þ γ 09 p^i  ATUSt þ δ1 waiverst þ δ2 p^i  waiverst þ ɛ it

ð3Þ

In both estimated equations, I introduced controls for monthly state unemployment rate and annual state per capita GDP growth. The goal is to provide some control for the state of the economy. As described below

234

MARIE CONNOLLY

in “Paid Work Time” section, the PRWORA and other policy changes of the 1990s happened at a time of strong economic growth and low unemployment. It is thus important to ask whether the changes observed following the reform were due to the reform itself, to the macroeconomic conditions, or to both combined. Herbst (2008) is one of the few that tackled the issue head on by trying to exploiting policy variation across mothers within a given state and year. Bartik and Eberts (1999) and Figlio and Ziliak (1999) both controlled for the differential impact of welfare waivers by the level of the unemployment rate, while Hofferth, Stanhope, and Harris (2002) allowed for the state median income to affect the effect of work requirements welfare participation. Most other studies do not explicitly take this issue into account. In this study I simply introduce time as a linear function of local unemployment rates and GDP growth. A more careful analysis could investigate different functional forms and interactions. Throughout the analysis, the sample was restricted to individuals aged 18
59. The analysis was performed with both sexes pooled and with women only. However the findings did not change significantly either way, so for brevity the paper only presents results for women. Women comprise the majority of welfare recipients and are the focus of the analysis.

FINDINGS This section presents my findings, starting with a replication of Meyer and Sullivan’s (2008) analysis. Main estimates will be presented after, and finally an analysis by presence of preschool-age children will be offered.

Replication Results I propose two improvements over Meyer and Sullivan’s (2008) methodology: additional data sources and a methodological change, identifying treatment using a propensity score rather than based solely on the presence of children. To distinguish the contribution of both changes, I first performed a replication exercise. Table 2 presents estimates computed using Meyer and Sullivan’s methodology, that is, a simple difference-in-differences model with no covariates, where treatment is identified as the presence of children. In this model, the mean time spent in various time uses is compared between the NTUS (1992
1994) and the ATUS (2003) (pre-post), between

235

The Changing Time Use of U.S. Welfare Recipients

Table 2. Replication of Meyer and Sullivan’s Difference-in-Differences Estimates (Minutes per Day). DinD by survey (Baseline 1992
1994) Dependent Time-Use Variable Work Nonmarket work Education Leisure TV (part of leisure) ESPb Child care Own medical care

Meyer and Sullivana (2008) (1)

DinD NTUSATUS (2)

DinD All Surveys (3)

28.4 (28.9) −44.8 (18.0) 8.7 (6.9) −28.4 (26.3)


30.0 (27.5) −38.3 (19.8) −2.7 (10.1) −26.2 (22.9) −9.6 (17.1) 3.1 (14.2) 50.4 (9.2) −10.5 (3.8) −1.3 (2.4) −4.3 (6.6)

37.4 (44.8) −20.8 (25.3) −0.8 (10.5) −6.4 (33.0) −14.4 (23.7) 16.5 (23.4) 29.7 (10.7) −9.4 (7.7) −4.1 (3.7) 20.3 (10.4)


44.8 (7.5)


Other care




Civic and religious




EPRI FISCT ATUS 1994
1995 1998
1999 2003
2005 (4)

(5)

(6)

54.4 (56.6) 18.4 (37.9) −5.1 (19.2) −9.2 (51.0) −44.7 (38.2) −35.0 (34.4) −8.7 (17.5) 6.4 (14.6) 5.6 (6.8) −26.9 (14.4)

−65.8 (77.4) 16.4 (41.0) −3.1 (13.8) 42.3 (49.8) −0.9 (35.0) −8.1 (36.7) −2.5 (14.0) −0.3 (4.4) −0.5 (1.8) 21.7 (17.5)

30.0 (27.5) −38.3 (19.8) −2.7 (10.1) −26.2 (22.9) −9.6 (17.1) 3.1 (14.2) 50.4 (9.2) −10.5 (3.8) −1.3 (2.4) −4.3 (6.6)

a Meyer and Sullivan (2008) results are reproduced from Table A2 column 10, p. 2234. The values are converted in minutes per day by multiplying by 60 and dividing by 7. b ESP refers to eating, sleeping, and personal care. Notes: Each estimate comes from a different regression where the dependent variable is the category of time use listed on the left. The estimates presented are the estimated coefficients on the interaction of variables indicating whether the respondent has children and whether the survey was conducted after the reform. Standard errors are in parentheses. No control variables are used and the sample only includes single women. Weighted using rescaled surveys’ sampling weights (see text for details). N = 5,569 when only NTUS and ATUS are used. N = 5,807 when all surveys are used. Data from the NSP were not used since all respondents are parents.

single mothers and single women without children (treated-non treated).6 Column 1 of Table 2 reproduces the figures from Meyer and Sullivan (2008, Table 7, column 10, p. 2234), converting hours into minutes to make the comparison in Table 2 easier. Note that time use is measured in minutes in all datasets, so this reconversion (back to minutes) should not affect the

236

MARIE CONNOLLY

reliability of the comparison. The only statistically significant changes in Meyer and Sullivan (column 1 of Table 2) presented evidence of less time spent in nonmarket work activities, driven by less time shopping and obtaining goods and services, and more time in child care. The coefficient on market work was positive, meaning more work was done after the reform, but the standard error was too large to reject that they are not equal to zero. In column 2 of Table 2, I reproduce as closely as possible column 1, with additional time use categories. My figures align very well with those in column 1. The slight discrepancies could come from sample selection or the classification of time uses. In column 3, I apply Meyer and Sullivan’s simple model to my extended dataset, that is, including data from the EPRI and the FISCT surveys. I am not able to use the NSP because it covers only parents, so everyone would appear as “treated.” In the remaining columns, I mimic Specification II (Eq. (3)) and show estimates by survey. Looking at column 3, it becomes clear that the additional data sources change the estimation results. Some effects completely change sign while others change in magnitude. Given the very basic methodology used in Table 2 (no control variables, no dealing with waivers, etc.) I will not interpret results any deeper and turn now to findings using the methodology discussed in the third section.

Main Findings The first step to produce the main findings of this paper was to estimate the probit model of Eq. (1). Marginal effects are reported in Table 3. As can be expected, being black, single, with children, and of low education greatly increased the propensity to receive welfare benefits. Once the predicted propensities to receive welfare were computed, the first model (Specification I) was estimated using only data from the NTUS and ATUS, making it directly comparable to the Meyer and Sullivan (2008) model. The results for γ^ 3 are presented in Table 4, columns 2a and 2b. In this table each estimate results from a separate regression. Because total time in a day is fixed at 1,440 minutes, adding up the coefficients in each column (apart for TV which is already accounted for in leisure time) gives a figure close to zero.7 As in Table 2, column 1 of Table 4 reproduces the Meyer and Sullivan (2008, Table 7, column 10, p. 2234) estimates. Columns 2a and 2b contain estimates from Eq. (2) using only NTUS and ATUS data, in order to compare the Meyer and Sullivan methodology with the one employed here. Column 2a does not include any covariates, thus reproducing more closely

237

The Changing Time Use of U.S. Welfare Recipients

Table 3.

Average Marginal Effects from the Probit Estimating the Propensity to Receive Welfare Benefits.

Variable Age/100 Age squared/10,000 Black Asian Other race Hispanic Elementary Some high school Some college College More than college Two adults Three or more adults Children present

Average Marginal Effect (dy/dx)

Delta-Method Standard Error

−0.2543 0.1027 0.0522 0.0067 0.0250 0.0129 0.0468 0.0440 −0.0251 −0.0685 −0.0973 −0.0840 −0.0985 0.0772

0.0405 0.0571 0.0015 0.0037 0.0032 0.0018 0.0023 0.0015 0.0015 0.0030 0.0065 0.0013 0.0017 0.0016

Notes: N = 177,530. Pseudo R-squared = 0.3157. Omitted groups are whites, high school graduates, with one adult in the household. Calculations are for all women aged 18
59, inclusively. State fixed effects are included (coefficients not reported). Calculated using the CPS March Supplements from 1992 to 1995. Weighted using the March Supplement individual sampling weights.

the Meyer and Sullivan methodology, whereas column 2b includes all the demographic and day-of-week covariates, as well as monthly state unemployment rate, annual state per capita GDP growth. Both columns however make use of the estimated propensities and control for the effect of the waivers. Inspection of columns 2a and 2b of Table 4 reveals findings in line with those of Meyer and Sullivan (2008) from column 1. The γ^ 3 coefficients estimated tell us that, post-reform, individuals that were more likely to receive welfare in the pre-period appear to spend significantly less time in nonmarket work (73
88 fewer minutes per day), more time in education (36
46 minutes), less time in religious and civic activities (around 50 fewer minutes per day) and more time in child care (14
58 more minutes). The change in market work time was positive, but not significant. Neither was the small increase in leisure time (0
10 more minutes, albeit with a 20- to 25-minute drop in television watching). The largest difference between columns 2a and 2b, that is, between not using covariates and using covariates, is the estimate for the effect on time spent providing child care, which goes from 14.1 without covariates to 57.6 with covariates. Meyer and Sullivan’s estimate actually falls between those two numbers, at 44.8. Whatever the

Difference-in-Differences Estimations Results (Minutes per Day).

Meyer and Sullivana (2008)

NTUS-ATUS

All Surveys 1994
1995

(1) Work Nonmarket work Education Leisure TV (part of leisure)

28.4 (28.9 −44.8 (18.0) 8.7 (6.9) −28.4 (26.3)


ESPb




Child care

44.8 (7.5)


Own medical care




Civic and religious




Spec. I

Spec. I

Spec. I

γ^ 3 ^ (post*p) (2a)

γ^ 3 ^ (post*p) (2b)

γ^ 3 ^ (post*p) (3)

24.3 (43.3) −88.3 (27.7) 46.3 (14.4) 9.7 (31.7) −19.9 (22.7) 47.8 (23.3) 14.1 (15.2) −3.5 (6.7) −2.0 (7.4) −49.8 (10.2)

13.0 (39.9) −73.1 (26.6) 36.2 (13.9) 0.2 (30.5) −25.2 (22.4) 18.8 (22.7) 57.6 (14.2) −3.1 (6.7) −0.6 (7.4) −50.9 (10.0)

132.6 (36.4) −5.5 (25.7) −38.6 (12.4) −101 (28.0) −63.8 (20.0) 5.2 (20.5) 16.5 (14.2) −5.4 (7.3) −7.5 (5.4) −7.0 (8.5)

γ^ 03 ^ (epri*p) (4) −85.1 (63.1) 29.1 (44.4) 56.4 (21.4) −63.5 (48.5) −49.8 (35.0) −14.2 (35.7) 39.1 (24.8) 24.8 (12.8) 0.7 (9.3) 13.4 (14.8)

1998
1999 1999
2000 Spec. II (Baseline 1992
1994) γ^ 05 ^ (fisct*p) (5) 126.1 (58.3) 2.8 (40.7) −24.1 (19.5) −322 (44.4) −221 (31.5) 126.0 (32.4) 67.5 (22.8) −8.5 (11.6) 2.2 (8.5) 31.0 (13.6)

γ^ 07 ^ (nsp*p) (6) 97.9 (52.5) 20.5 (37.0) −18.7 (17.7) −94.3 (40.3) −32.5 (28.6) −27.9 (29.4) 1.6 (20.6) 12.1 (10.9) −10.1 (7.8) 18.8 (12.3)

2003
2005 γ^ 09 ^ (atus*p) (7) −18.9 (60.2) −41.6 (42.4) 35.6 (20.4) −43.3 (46.0) −49.9 (32.9) 38.4 (33.7) 59.9 (23.6) −2.7 (12.1) −7.1 (8.9) −23.2 (14.1)

a Meyer and Sullivan (2008) results are reproduced from Table A2 column 10, p. 2234. The values are converted in minutes per day by multiplying by 60 and dividing by 7. b ESP refers to Eating, sleeping and personal care. Notes: Each estimate comes from a different regression where the dependent variable is the category of time use listed on the left. Standard errors computed using a Rubin (1987) multiple imputation procedure (with 10 repetitions) are in parentheses. All regressions except column 2a include controls for age, age squared, race and ethnicity, education, household composition, presence of children, state, day of the week, monthly state unemployment rate and annual state GDP growth. The regression in column 2a includes no control variables. Weighted using rescaled surveys’ sampling weights (see text for details). Specification I is Eq. (2) in the text, Specification II is Eq. (3). N = 22,194 when only NTUS and ATUS are used. N = 23,826 when all surveys are used.

MARIE CONNOLLY

Other care

238

Table 4. Dependent Time-use variable

The Changing Time Use of U.S. Welfare Recipients

239

size of the effect, the direction is clear: there appeared to be more time in child care activities after 1996, and that increase is statistically different from zero when controlling for a host of independent variables. Taken together, the results in column 2 of Table 2 and of columns 2a and 2b of Table 4 suggest that the approach of this paper using propensity scores generates similar findings to the simple difference-in-differences approach taken by Meyer and Sullivan (2008) when the same data sources are used (NTUS and ATUS). More discrepancies appear when using additional data sources. Paid Work Time Columns 3 to 7 of Table 4 contain estimates using all available datasets, which constitutes the second and most important contribution of this paper. In column 3, the difference-in-differences estimates for Specification I are presented, while those for Specification II can be found in Columns 4
7. The findings regarding time spent at work are markedly different when using the additional surveys. The γ^ 3 for time working in Specification I is now 132.6 and significant, compared with a not statistically significant 13 in column 2b. Looking at the γ 0 s from Specification II, it becomes clear that while time at work had not increased for would-be welfare recipients in the ATUS (coefficient of −18.9, largely not significant), there was a sharp increase in the 1998
1999 FISCT, and a more modest one in the 1999
2000 NSP. Why would the increase be so large in 1998
2000, yet drop afterwards? One likely answer comes from looking at the business cycles and the economic conditions of the time. It has been argued that social policies are reinforced by a strong economy (Herbst, 2008), even after controlling for average unemployment rates. Fig. 3 shows how the unemployment rate for the civilian labor force has fluctuated between 1990 and 2006. The unemployment rate peaked at above 7.5% during the 1992 recession, then steadily dropped during the expansion time of the late 1990s until the end of 2000, to a low of under 4%, to go back to 6% in 2003, and then back down. The omitted year in both Specifications I and II is 1992, so clearly the baseline was a time of high unemployment and difficult economic conditions. The welfare recipients, individuals at the bottom of the distribution in terms of education and wages, are more likely to be affected by the changing economic environment, so it is therefore unsurprising to see that their time spent at work varies with the business cycles (Bitler, Gelbach, & Hoynes, 2006). The results from Table 4 can be compared with similar calculations made using the CPS, which contains information on hours worked per week. Specifications

240

MARIE CONNOLLY

Unemployment rate, Civilian Labor Force 8

Unemployment rate, %

7.5 7 6.5 6 5.5 5 4.5 4 1990 1992 1994 1996 1998 2000 2002 2004 2006

Fig. 3.

Unemployment Peaked in 1992 and Dropped Until 2001. Source: BLS CPS Website.

I and II were estimated for hours of work using the CPS March Supplements. The coefficient on the interaction of p^i  POST, γ^ 3 , was a highly significant 9.97 hours per week, which corresponds to 85 minutes per day, lower than but comparable to the 133 minutes from Table 4. Fig. 4 graphs the coefficients on the year dummies interacted with p^i , with 1992 taken as a baseline. Until 1995, we cannot reject that the coefficients were different from 0 (from 1992), but in 1996 and after the coefficients are increasing and significant. In fact, Fig. 4 mirrors quite well Fig. 3, the unemployment rate. One difference to note with the findings of Table 4 is that using CPS data, the hours of work decline only slightly after 2000, while when using the ATUS the estimates show a marked decline between 2000 and 2003
2005. The source of this discrepancy is unclear. Other Time Uses Moving on to other time uses, column 3 of Table 4 shows us that contrary to Meyer and Sullivan (2008), nonmarket work did not decrease (estimate of four less minutes, but not significant), but leisure still saw a significant drop (again driven by less TV time), as did educational activities. The estimated coefficient on child care was much lower and lost statistical significance, going from 57.6 to 16.5. Looking at the results for Specification II

241

The Changing Time Use of U.S. Welfare Recipients

Difference−in−Differences for hours worked per week

Coefficient on phat*year (with 95% confidence bands)

15 Date PRWORA passed

10

5

0

−5 1992

1994

1996

1998 2000 Year

2002

2004

2006

Fig. 4. CPS Data Show an Increase in Hours Worked per Week for Would-Be Welfare Recipients. Source: Author’s calculations using the CPS March Supplements.

(columns 4
7), we can see that the trends summarized by Specification I are mostly confirmed. In particular, note that leisure (and TV time, the biggest component of leisure) decreased roughly in parallel with the increase in time spent at work, while no similar decline can be observed in nonmarket work. The same conclusions hold when looking only at weekdays rather than weekends (see Appendix Tables A4 and A5). The γ 03 , γ 05 , γ 07 , and γ 09 for child care show, however, an interesting pattern. The overall difference-in-differences coefficient, 16.5 in column 3, was not statistically significant, yet two out of the three post-reform surveys show a significant increase in child care. What happens is that in the EPRI (column 4), which is before the welfare reform was fully implemented, likely welfare recipients spent 39 extra minutes in child care compared to the 1992 to 1994 NTUS. This renders the Specification I estimate of column 3 not significant. Egerton et al. (2005) and Aguiar and Hurst (2006) pointed out that there appeared to be an undersampling of households with children in the NTUS. This could cause a problem when comparing trends in child care unconditional on the presence of children in the household. However in my estimations I not only use the presence of children in the construction of

242

MARIE CONNOLLY

the propensity scores but also as a control in the difference-in-differences regressions. This leads me to argue that undersampling, if present, should not be an issue. As a robustness check, the specification from column 3 of Table 4 was estimated excluding NTUS data (thus leaving only the EPRI as pre-reform data), and the estimates did not change significantly. Based on the analysis presented here, there is no evidence that likely welfarereceiving women spent less time caring for children.

Presence of Preschool-age Children The effect of a change in maternal child care on child well-being is not straightforward and depends on the quality of the substitute (formal care in an accredited center, informal care provided by relatives, etc.) as well as other parameters such as the quantity of child care received by the child and the age of the child. A sizeable literature exists on the topic (see Almond and Currie (2011), Ruhm (2004), Waldfogel (2002), Magnuson, Ruhm, and Waldfogel (2007), Gormley and Gayer (2005), Blau (1999), Hill, Waldfogel and Brooks-Gunn (2002) and Haeck, Lefebvre, and Merrigan (2013) to name but a few). Unfortunately the time-use surveys used in this study do not contain information on the type of child care a respondent’s child receives nor on its quantity. Also, as mentioned in the section “Data,” the earlier surveys do not contain information on secondary child care. My finding of a stable primary child care could be associated with an unmeasurable decline in secondary maternal child care, with its associated effect on child well-being. To shed some light on the issue, I present a final analysis in which the sample was split according to whether the respondent has children aged five and under or not. Table 5 contains three columns. In column 1 the main findings from Table 4 are reproduced. Column 2 contains the estimates of Specification I using all surveys for mothers of children aged five and under, while column 3 does the same but for women with older children or no children.8 The age of five was chosen because it coincides with entry into kindergarten and the end of the preschool years. After age five, working parents do not have to worry about finding child care arrangements the same way that they used to because the formal school system starts taking charge of the children. Thus the impact of the welfare reform, by promoting work, may be felt differently by parents of younger children because of the need to find child care. Table 5 highlights some interesting differences by age of the child. First, for mothers of older children and non-mothers, welfare reform was

243

The Changing Time Use of U.S. Welfare Recipients

Table 5. Difference-in-Differences Estimations Results by Presence of Preschool Children (Minutes per Day), All Surveys. Dependent time-use variable

All Surveys, Specification I Full sample

Work Nonmarket work Education Leisure TV (part of leisure) ESPa Child care Own medical care Other care Civic and religious N

γ^ 3 (post*p) ^ (1)

With children 5 and under γ^ 3 (post*p) ^ (2)

No child 5 and under γ^ 3 (post*p) ^ (3)

132.6 (36.4) −5.5 (25.7) −38.6 (12.4) −101 (28.0) −63.8 (20.0) 5.2 (20.5) 16.5 (14.2) −5.4 (7.3) −7.5 (5.4) −7.0 (8.5)

83.5 (54.1) 54.8 (42.7) −66.4 (19.9) −24.5 (43.2) −3.9 (29.3) −23.4 (33.8) −30.2 (29.3) 21.8 (16.9) −9.1 (10.7) −5.1 (12.5)

263.3 (51.7) −89.7 (34.7) −10.3 (17.0) −183.3 (39.5) −56.1 (28.8) 30.9 (28.3) −4.5 (14.6) −30.3 (7.4) 1.5 (6.5) 25.9 (12.3)

23,826

6,146

17,680

a

ESP refers to Eating, sleeping and personal care. Notes: Each estimate comes from a different regression where the dependent variable is the category of time use listed on the left. Standard errors computed using a Rubin (1987) multiple imputation procedure (with 10 repetitions) are in parentheses. All regressions include controls for age, age squared, race and ethnicity, education, household composition, presence of children, state, day of the week, monthly state unemployment rate and annual state GDP growth. Weighted using rescaled surveys’ sampling weights (see text for details). Specification I is Eq. (2) in the text. Column 2 is estimated using only women with children age five and under (age six and under for the NSP due to data limitations). Column 3 is estimated using women with no children age five (age six for the NSP) and under.

associated with a larger increase of market work (236 minutes), a decrease in nonmarket work (90 minutes) and a larger decrease in leisure time (183 minutes). For mothers of younger children, most of the effects are imprecisely estimated given the sample size, but a smaller increase in work time

244

MARIE CONNOLLY

can be observed (84 minutes) as well as an increase in nonmarket work (55 minutes) and a smaller decrease in leisure (25 minutes). Second, and interestingly, I find different effects on child care, albeit not statistically significant. Indeed, mothers of young children experienced a decline of 30 minutes in the time spent on primary child care, while the rest of the group saw a decline of only five minutes. These findings are consistent with a story where mothers of preschool-age children must find alternative child care arrangements for their children in order to increase work time, whereas mothers of older children and non-mothers do not face the same issue. Hence any effect of the welfare reform and other policy changes of the 1990s on the well-being of children should be differentiated by the age of the children. Our samples are unfortunately too small to precisely estimate the effect on the time use of mothers by the age of their children.

CONCLUSION The PRWORA of 1996 marked the passage of the welfare reform in the United States, bringing changes to how the federal and state governments dealt with aid to needy families. The PRWORA provided an impetus pushing welfare recipients off of the welfare rolls, and hopefully into the workplace. Over the same period of time, other policy changes affected poor families with children, such as the EITC expansions and the welfare waivers of the early 1990s. This study set out to look at the impact of the those changes on the way people allocate their time, more specifically on how the changes affected the people that were likely recipients of welfare benefits prior to the reform. To do so, the CPS was used to model who was likely to receive welfare before 1996, and then scores were imputed to individuals in time-use datasets. Using a multiple imputation procedure (Rubin, 1987), I then estimated a difference-in-differences model to compare the time use of welfare recipients and non-recipients before and after the reform. When only the NTUS (from 1992 to 1994) and the ATUS (from 2003 to 2005) were used, the results were similar to what Meyer and Sullivan (2008) found: less time spent in nonmarket work, more time spent in child care activities, and a positive but not statistically significant effect on market work. Apart from using the CPS to model the propensity to receive benefits, the main innovation of the present study is to include data from three smaller-scale time-use surveys that were conducted between the NTUS and the ATUS. Using these additional data, the results found with the NTUS and ATUS do not hold anymore. I found evidence of more time spent at

The Changing Time Use of U.S. Welfare Recipients

245

work, evidence, that is, corroborated by the CPS data on hours worked per week. However, the pattern of hours worked closely followed the overall unemployment rate, and I am concerned that what I am picking up is more an effect of the economic conditions on the people at the bottom of the income and education distributions, rather than an effect of the welfare reform measures. Needy families are more affected by the cyclical nature of the economy, and because the welfare reform happened during an expansion time, it is unclear what the exact effect of the reform is. Even if I controlled for state unemployment rates and GDP growth in my estimations, there could simply be not enough state-by-state variation in order to properly identify the effect of the reform. A better understanding of how the time use of poor people is affected by the economic conditions is needed if we want to tease out the effect of the welfare reform. I found no drop in the time spent in nonmarket activities. It appears that the increase in time spent working was largely balanced by less time in leisure activities, of which watching television is the most substantial component. I also found no evidence of more or less time spent in child care after the welfare reform, showing that children of likely welfare-receiving mothers were not affected by the reform and other changing conditions in terms of primary maternal child care. It would be interesting to find out how welfare reform changed day care support; perhaps welfare recipients being pushed to work also need to deploy more resources to take care of their children, thereby reducing the time spent in leisure activities. It is regrettable that earlier time-use surveys do not provide the same wealth of information on child care as the ATUS does. With data on secondary child care, the ATUS allows researchers to take a more detailed look at parent
child interactions, which often happen while other activities are going on. Earlier data on time use simply does not contain information on any aspect of multitasking. At the present time, only a few years of the ATUS data are available, but as time goes on, studies up to now impossible or hard to devise will become within our reach. It seems important for policymakers to understand how various policies affect the well-being of Americans. How people spend time is a key determinant of their well-being, so knowing how women’s time allocation has varied after the implementation of the welfare reform and other policy changes of the mid-1990s is a first step in understanding the influence of the various policies in place during that decade. This study suggests that time at work has increased for would-be welfare recipients, and that this time was taken away from leisure activities, not housework or child care. The net impact on well-being would depend on how the various time uses are valued (Krueger, 2009; Meyer & Sullivan, 2008), though Herbst (2012)

246

MARIE CONNOLLY

found that TANF reforms seemed to have increased the subjective wellbeing of single mothers. Most of those variations follow fluctuations in the business cycle, meaning that while well-being may have changed, it might be more a function of the current economic conditions or at least of a timing of implementation of the policies, rather than of the policies themselves, as argued in Herbst (2008).

NOTES 1. The NTUS data are available on the University of Maryland’s Scientific Research on the Internet’s website at http://www.webuse.umd.edu/. The EPRI data were kindly provided by Suzanne Bianchi and Sara Raley, and information on the survey can be found in Robinson and Godbey (1999). The FISCT (Robinson, Bianchi, & Presser, 2001) and NSP (Bianchi & Robinson, 2005) data are available on the Inter-University Consortium for Political and Social Research’s website at http://www.icpsr.umich.edu/. The ATUS data are available on the Bureau of Labor Statistics’ website at http://www.bls.gov/tus/. 2. Namely, Suzanne Bianchi and John Robinson. 3. I thank Mark Aguiar and Erik Hurst for making all their codes available on their websites. 4. The data can be downloaded on the NBER’s Data Collection webpage at http://www.nber.org/data/current-population-survey-data.html. 5. See Crouse (1999) and Grogger and Karoly (2005) for state-by-state implementation dates. 6. In their analysis, Meyer and Sullivan (2008) also present a difference-indifferences model where single mothers are compared to married mothers, finding similar results. For simplicity, I just present the results of a replication using only single women. 7. The balance of the total time comes from the “other” time-use category, which includes activities not coded elsewhere or missing answers. Only the ATUS contains positive time spent in “other” time uses, in the other surveys the coding procedure did not allow for diary entries to be missing. Average time spent in other activities is nine minutes per day in the ATUS (see Table A3). 8. In the NSP we can only identify children aged six and under so the sample includes children aged six in the NSP data.

ACKNOWLEDGMENT Funding from SSHRC and FQRSC is greatly acknowledged. Thanks to anonymous referees, seminar participants at UQAM and participants at IZA workshop on nonmarket time in economics for helpful comments. All remaining errors are my own.

The Changing Time Use of U.S. Welfare Recipients

247

REFERENCES Aguiar, M., & Hurst, E. (2006). Measuring trends in leisure: The allocation of time over five decades. NBER Working Paper No. w12,082. National Bureau of Economic Research, Cambridge, MA. Retrieved from http://papers.nber.org/papers/w12082 Aguiar, M., & Hurst, E. (2007). Measuring trends in leisure: The allocation of time over five decades. Quarterly Journal of Economics, 122(3), 969
1006. Almond, D., & Currie, J. (2011). Human capital development before age five. In O. Ashenfelter & D. Card (Eds.), Handbook of labor economics (Vol. 4A, 1315
1486). Amsterdam: Elsevier. Bartik, T., & Eberts, R. (1999). Examining the effect of industry trends and structure on welfare caseloads. In S. H. Danziger (Ed.), Economic conditions and welfare reform. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research. Bianchi, S., & Robinson, J. (2005). National survey of parents, 2000
2001 [Computer file]. ICPSR04247-v1, University of Maryland-College Park, Survey Research Center [producer], 2001. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], (2005-07-06). Bitler, M. P., Gelbach, J. B., & Hoynes, H. W. (2006). What mean impacts miss: Distributional effects of welfare reform experiments. American Economic Review, 96(4), 988
1012. Blank, R. M. (2002). Evaluating welfare reform in the United States. Journal of Economic Literature, 40(4), 1105
1166. Blau, D. (1999). The effect of child care characteristics on child development. Journal of Human Resources, 34(4), 786
822. Crouse, G. (1999). State implementation of major changes to welfare policies 1992
1998. Retrieved from http://aspe.hhs.gov/HSP/Waiver-Policies99/policy_CEA.htm. Accessed on June 6, 2013. Egerton, M., Fisher, K., & Gershuny, J. (2005). American time use 1965
2003: The construction of a historical comparative file, and consideration of its usefulness in the construction of extended national accounts for the USA. ISER Working Paper No. 200528. University of Essex, Colchester. Contributions from John P. Robinson, Anne H. Gauthier, Nuno Torres and Andreas Pollmann. Figlio, D., & Ziliak, J. (1999). Welfare reform, the business cycle, and the decline in AFDC caseloads. In S. H. Danziger (Ed.), Economic conditions and welfare reform, Kalamazoo, MI: W. E. Upjohn Institute for Employment Research. Gormley, W., & Gayer, T. (2005). Promoting school readiness in Oklahoma: An evaluation of Tulsa’s pre-k program. Journal of Human Resources, 40(3), 533
558. Grogger, J., & Karoly, L. (2005). Welfare reform: Effects of a decade of change. Cambridge, MA: Harvard University Press. Haeck, C., Lefebvre, P., & Merrigan, P. (2013, September). Canadian evidence on ten years of universal preschool policies: The good and the bad. Working Paper No. 13-34. Cirpe´e. Herbst, C. M. (2008). Do social policy reforms have different impacts on employment and welfare use as economic conditions change? Journal of Policy Analysis and Management, 27(4), 867
894. Herbst, C. M. (2012). Welfare reform and the subjective well-being of single mothers. Journal of Population Economics, 26(1), 203
238.

248

MARIE CONNOLLY

Hill, J., Waldfogel, J., & Brooks-Gunn, J. (2002). Assessing differential impacts: The effects of high-quality child care on children’s cognitive development. Journal of Policy Analysis and Management, 21(4), 601
628. Hofferth, S., Stanhope, S., & Harris, K. (2002). Exiting welfare in the 1990s: Did public policy influence recipients’ behavior? Population Research and Policy Review, 21, 433
472. Hubbard, R. G., Skinner, J., & Zeldes, S. P. (1995). Precautionary saving and social insurance. Journal of Political Economy, 103(2), 360
399. Krueger, A. B. (Ed.). (2009). Measuring the subjective well-being of nations. National accounts of time use and well-being. Chicago, IL: The University of Chicago Press. Magnuson, K., Ruhm, C., & Waldfogel, J. (2007). Does pre-kindergarten improve school preparation and performances? Economics of Education Review, 26(1), 33
51. Meyer, B. D., & Rosenbaum, D. T. (2001). Welfare, the earned income tax credit, and the labor supply of single mothers. Quarterly Journal of Economics, 116(3), 1063
1114. Meyer, B. D., & Sullivan, J. X. (2008). Changes in the consumption, income, and well-being of single mother headed families. American Economic Review, 98(5), 2221
2241. Ruhm, C. J. (2004). Parental employment and child cognitive development. Journal of Human Resources, 39(1), 155
192. Robinson, J. P., Bianchi, S. M., & Presser, S. (2001). Family interaction, social capital, and Trends in time use (FISCT), 1998
1999: [United States] [Computer file], ICPSR version. College Park, MD: University of Maryland Survey Research Center [producer], 1999. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2001. Robinson, J. P., & Godbey, G. (1999). Time for life. In D. Robert (Ed.), The surprising ways Americans use their time (2nd. ed.) (with a foreword by. R. D. Putnam). University Park, PA: The Pennsylvania State University Press. Rubin, D. B. (1987). Multiple imputation for nonresponse in surveys. Wiley series in Probability and Mathematical Statistics: Applied Probability and Statistics. New York, NY: Wiley. Waldfogel, J. (2002). Child care, women’s employment, and child outcomes. Journal of Population Economics, 15(6), 527
548.

The Changing Time Use of U.S. Welfare Recipients

249

APPENDIX Multiple Imputation Procedure Because I used a propensity score in my difference-in-differences regression (Eqs. (2 and 3), and not an actual indicator of treatment, additional variation was introduced due to the fact that the values of the score used are not observed, but rather estimated and drawn from a sampling distribution. In order to take into account the extra variance that the process brings, a Rubin multiple imputation procedure was used (Rubin, 1987). Once the probit modeling the propensity to receive welfare was estimated, and β^ p  and Var β^ p were obtained, multiple datasets were constructed, drawing m   vectors of βp from a normal distribution N β^ p ; Var β^ p , where m is the number of repetitions used. Estimated propensities, or p^i , were computed 0 as p^im = Φ β^ pm Xi . In Eq. (2), γ3 is the parameter of interest showing the post-reform timeuse changes. Following Rubin (1987), the results from each imputation were combined using the following formulas, valid for each coefficient but presented here for the coefficient of interest, γ3. The estimate of γ3 is the average of the estimated γ^ 3 (but note that to simplify the text, I will refer to γ^ 3 when presenting and discussing findings): γ 3m =

Xm

γ^ =m l = 1 3l

ðA:1Þ

The variance comes from two parts: first the average of the analytic variances, denoted U m , and second the variance between the estimates, or added variance due to imputation, denoted by Bm. They were computed and combined into the total variance, Tm, using the following formulas: Um = Bm =

Xm
l=1

Xm l=1

γ^ 3l − γ 3m

U l =m

0


 γ^ 3l − γ 3m =ðm − 1Þ

 1 Tm = U m þ 1 þ Bm m

ðA:2Þ ðA:3Þ

ðA:4Þ

250

MARIE CONNOLLY

To evaluate the extent to which the multiple imputation affected the variability of the results, the relative increase in variance rm can be calculated as follows:  1 ðA:5Þ Bm =U m rm = 1 þ m Note however that the rm found in the analysis were virtually all around 1%, which shows that the increase in variance due to the multiple imputation was minimal. They are not reported here, but the standard errors presented are all corrected using the procedure described above.

The Changing Time Use of U.S. Welfare Recipients

Table A1. Time Use Classification

251

Time Use Classifications. Examples of Activities Included

Work

Work for pay, main job (including time spent working at home); work for pay, other jobs; other work-related activities (commuting to/from work; meals/breaks at work; searching for a job; applying for unemployment benefits) Nonmarket Work Food preparation; food presentation; kitchen/food cleanup; washing/ drying clothes; ironing; dusting; vacuuming; indoor cleaning; indoor painting; grocery shopping; shopping for other goods; comparison shopping; clipping coupons; going to bank; going to post office; meeting with lawyer; going to veterinarian (excluding any time spent acquiring medical care); all other home production including: vehicle repair; outdoor repair; outdoor painting; yard work; pet care; gardening; etc. Education Taking classes for degree; personal interest courses; homework for coursework; research for coursework; etc. Leisure Sports/exercise (playing sports; attending sporting events; exercise); watching television; entertainment (going to movies and theater; listening to music; computer use for leisure); socializing (attending/hosting social events; playing games; telephone calls); reading books, magazines; personal mail; personal email; hobbies (arts and crafts; collecting; playing musical instrument) Eating Eating meals at home; eating meals away from home; etc. Sleeping Sleeping; naps Personal care Grooming; bathing; sex; going to the bathroom; etc. (excluding any time spent on own medical care) Child care Primary child care (breast feeding; rocking a child to sleep; general feeding; changing diapers; providing medical care to child; grooming child; etc.); educational child care (reading to children; teaching children; helping children with homework; attending meetings at a child’s school; etc.); recreational child care (playing games with children; playing outdoors with children; attending a child’s sporting event or dance recital; going to the zoo with children, taking walks with children; etc.) Own medical care Visiting doctor’s/dentist’s office (including time waiting); dressing wounds; taking insulin; etc. Other care Caring for others, except household children (taking to doctor’s; providing care; helping in activities) Civic and religious Religious practice/participation; fraternal organizations; volunteer work; union meetings; AA meetings; etc. Other Not coded elsewhere; missing or did not answer. Source: Adapted from Aguiar and Hurst (2006), Table A2.

252

Table A2.

MARIE CONNOLLY

Summary Statistics for Demographic and Socioeconomic Variables, by Dataset.

Variable

Dataset a

CPS NTUS EPRI FISCT NSP ATUS 1992
1995 1992
1994 1994
1995 1998
1999 1999
2000 2003
2005 Age White Black Asian Other race Hispanic Elementary Some high school High school diploma Some college College diploma More than college One adult in household Two adults in household More than two adults Presence of children Married Divorced Single N

36.8 (11.2) 0.83 (0.38) 0.13 (0.34) 0.03 (0.17) 0.01 (0.12) 0.09 (0.29) 0.04 (0.20) 0.10 (0.30) 0.35 (0.48) 0.30 (0.46) 0.15 (0.35) 0.06 (0.23) 0.15 (0.36) 0.57 (0.49) 0.28 (0.45) 0.54 (0.50) 0.59 (0.49) 0.15 (0.35) 0.24 (0.43)

37.1 (10.7) 0.79 (0.41) 0.11 (0.31) 0.02 (0.14) 0.08 (0.26) 0.09 (0.28) 0.01 (0.12) 0.06 (0.24) 0.37 (0.48) 0.26 (0.44) 0.18 (0.38) 0.11 (0.32) 0.12 (0.33) 0.59 (0.49) 0.29 (0.45) 0.42 (0.49)







39.3 (11.7) 0.83 (0.37) 0.11 (0.32) 0.01 (0.10) 0.02 (0.15) 0.07 (0.26) 0.03 (0.18) 0.11 (0.32) 0.38 (0.49) 0.28 (0.45) 0.15 (0.36) 0.04 (0.20) 0.12 (0.32) 0.59 (0.49) 0.29 (0.46) 0.55 (0.50) 0.62 (0.49)





37.5 (11.2) 0.80 (0.40) 0.15 (0.36) 0.01 (0.08) 0.04 (0.19) 0.06 (0.23) 0.02 (0.15) 0.11 (0.31) 0.33 (0.47) 0.31 (0.46) 0.17 (0.38) 0.05 (0.23) 0.12 (0.33) 0.58 (0.49) 0.30 (0.46) 0.52 (0.50) 0.63 (0.48) 0.14 (0.35) 0.21 (0.40)

34.9 (8.8) 0.74 (0.44) 0.15 (0.36) 0.04 (0.18) 0.06 (0.24) 0.13 (0.34) 0.02 (0.13) 0.10 (0.30) 0.35 (0.48) 0.30 (0.46) 0.13 (0.34) 0.09 (0.29) 0.18 (0.39) 0.65 (0.48) 0.17 (0.37) 1.00 0.00 0.69 (0.46) 0.16 (0.37) 0.13 (0.34)

38.5 (11.8) 0.81 (0.39) 0.13 (0.34) 0.03 (0.17) 0.02 (0.15) 0.13 (0.34) 0.03 (0.18) 0.09 (0.28) 0.30 (0.46) 0.20 (0.40) 0.29 (0.45) 0.09 (0.29) 0.16 (0.36) 0.57 (0.49) 0.27 (0.44) 0.52 (0.50) 0.58 (0.49) 0.13 (0.34) 0.27 (0.44)

177,530

2,941

573

491

572

19,293

Notes: See Table 1 for a description of the datasets and their acronyms. Weighted using rescaled surveys’ sampling weights (see text for details). Standard deviations are in parentheses. a The CPS data are from the March Supplements from 1992 to 1995. Weighted using the March Supplement individual sampling weights.

253

The Changing Time Use of U.S. Welfare Recipients

Table A3.

Summary Statistics for Time-use Variables, by Dataset, all Days of the Week Pooled.

Time-Use Variable (Minutes per Day) Work Nonmarket work Education Leisure TV (part of leisure) Eating Sleeping Personal care Child care Own medical care Other care Civic and religious Other N

Dataset NTUS 1992
1994

EPRI 1994
1995

FISCT 1998
1999

NSP 1999
2000

ATUS 2003
2005

227.7 (276.3) 190.7 (178.4) 20.1 (88.9) 305.2 (209.1) 139.8 (147.5) 62.4 (53.7) 498.7 (123.3) 73.6 (74.4) 36.0 (79.6) 3.4 (23.5) 5.4 (39.8) 16.7 (61.9) 0.0 (0.0)

244.6 (273.5) 193.8 (188.8) 20.8 (92.9) 288.4 (210.3) 121.5 (143.7) 56.8 (48.7) 492.7 (124.8) 65.8 (66.4) 51.3 (96.7) 4.8 (28.2) 7.3 (34.3) 13.9 (51.9) 0.0 (0.0)

261.6 (275.4) 194.9 (175.3) 17.3 (79.2) 256.6 (205.9) 106.3 (143.4) 65.7 (58.3) 482.1 (114.9) 80.5 (81.0) 60.9 (102.4) 2.7 (13.8) 2.8 (22.0) 14.8 (61.0) 0.0 (0.0)

231.1 (275.3) 223.7 (193.6) 15.9 (80.1) 241.1 (195.4) 94.5 (124.5) 54.5 (52.3) 463.5 (126.7) 68.3 (71.0) 111.1 (119.7) 10.0 (80.9) 5.4 (30.0) 15.3 (60.3) 0.0 (0.0)

224.2 (265.1) 186.7 (165.0) 20.5 (90.6) 267.6 (185.3) 127.3 (135.0) 68.2 (59.1) 513.0 (130.0) 48.1 (40.3) 61.7 (111.3) 9.2 (55.7) 13.3 (52.1) 18.0 (64.8) 9.4 (42.0)

2,941

573

491

572

19,293

Notes: See Table 1 for a description of the datasets and their acronyms. Weighted using rescaled surveys’ sampling weights (see text for details). Means are in minutes per day. Standard deviations are in parentheses.

254

Table A4.

MARIE CONNOLLY

Difference-in-Differences Estimations Results, All Surveys, Weekdays. Weekdays, N = 12,664 Spec. I

Specification II

Time-Use Variable (Minutes per Day)

γ^ 3 ^ (post*p)

γ^ 03 ^ (epri*p)

γ^ 05 ^ (fisct*p)

γ^ 07 ^ (nsp*p)

γ^ 09 ^ (atus*p)

Work

200.1 (58.2) −39.7 (36.2) −72.9 (19.6) −128.8 (39.0) −71.1 (28.7) −17.8 (29.3) 35.9 (21.1) 8.2 (12.3) 9.0 (8.1) 7.6 (9.1)

−60.6 (98.2) 85.3 (61.3) 89.1 (32.9) −102.9 (66.8) −12.6 (48.5) −94.4 (49.3) 82.9 (35.8) 10.2 (20.5) −22.6 (13.9) 14.1 (15.4)

279.0 (88.2) −76.0 (55.2) −53.0 (29.3) −391.6 (58.8) −225.8 (43.1) 68.7 (44.8) 116.0 (32.0) −16.1 (18.3) 5.8 (12.5) 68.9 (14.0)

159.8 (84.2) 37.3 (53.0) −35.1 (27.9) −99.8 (56.3) 2.0 (41.0) −104.2 (42.4) 23.0 (31.1) 32.2 (18.0) −7.1 (11.8) −6.3 (13.2)

6.0 (91.7) −19.7 (57.5) 40.2 (30.7) −73.0 (61.5) −57.2 (44.7) −18.4 (46.1) 83.3 (33.4) −15.4 (19.1) −4.9 (13.0) −2.9 (14.5)

Nonmarket work Education Leisure TV (part of leisure) ESPa Child care Own medical care Other care Civic and religious

Notes: Each estimate comes from a different regression where the dependent variable is the category of time use listed on the left. Standard errors computed using a Rubin (1987) multiple imputation procedure (with 10 repetitions) are in parentheses. Regressions include controls for age, age squared, race and ethnicity, education, household composition, presence of children, state, day of the week, monthly state unemployment rate and annual state GDP growth. Weighted using rescaled surveys’ sampling weights (see text for details). Specification I is Eq. (2) in the text, Specification II is Eq. (3). a ESP refers to Eating, sleeping and personal care.

255

The Changing Time Use of U.S. Welfare Recipients

Table A5.

Difference-in-Differences Estimations Results, All Surveys, Weekends. Weekends, N = 11,162 Spec. I

Time-Use Variable (Minutes Per Day) Work Nonmarket work Education Leisure TV (part of leisure) ESPa Child care Own medical care Other care Civic and religious

Specification II

γ^ 3 ^ (post*p)

γ^ 03 ^ (epri*p)

γ^ 05 ^ (fisct*p)

γ^ 07 ^ (nsp*p)

γ^ 09 ^ (atus*p)

33.2 (33.4) 45.2 (35.8) 13.3 (11.7) −47.7 (40.8) −36.9 (27.9) 35.4 (28.0) −4.6 (17.8) −26.5 (5.5) −42.3 (6.3) 0.2 (16.2)

−134.8 (59.6) −41.7 (63.9) −9.1 (21.1) 43.4 (72.0) −127.3 (49.3) −93.8 (50.3) −19.1 (31.9) 48.8 (9.8) 40.7 (11.3) −22.9 (29.1)

−82.8 (57.4) 180.4 (61.5) 11.2 (19.9) −213.1 (69.3) −266.3 (46.6) 227.8 (48.1) −58.7 (29.8) 0.5 (9.0) −9.1 (10.5) −56.3 (27.8)

27.4 (48.3) 23.1 (51.5) −10.1 (17.2) −55.5 (58.8) −49.2 (39.8) 44.3 (40.8) −14.8 (25.6) −0.2 (7.7) −17.0 (9.0) 3.2 (23.8)

−23.9 (57.6) −78.3 (61.4) 5.8 (20.4) 32.9 (69.8) −43.6 (47.5) 110.9 (48.4) 18.3 (30.5) 17.1 (9.2) −15.2 (10.7) −65.1 (28.1)

Notes: Each estimate comes from a different regression where the dependent variable is the category of time use listed on the left. Standard errors computed using a Rubin (1987) multiple imputation procedure (with 10 repetitions) are in parentheses. Regressions include controls for age, age squared, race and ethnicity, education, household composition, presence of children, state, day of the week, monthly state unemployment rate and annual state GDP growth. Weighted using rescaled surveys’ sampling weights (see text for details). Specification I is Eq. (2) in the text, Specification II is Eq. (3). a ESP refers to Eating, sleeping and personal care.

DOES HIGHER EDUCATION QUALITY MATTER IN THE UK? Arnaud Chevalier ABSTRACT This paper estimates the financial returns to higher education quality in the UK. To account for the selectivity of students to institution, we rely on a selection on observable assumption. We use several estimates including the Generalised Propensity Score (GPS) of Hirano and Imbens, which relies on a continuous measure of institutional quality. This highlights that the returns to quality are heterogeneous and mostly driven by high-quality institutions. Moving from an institution in the third quality quartile to a top quality institution is associated with a 7% increase in earnings. Keywords: College quality; returns to education; generalised propensity score JEL classifications: I22; J31

Factors Affecting Worker Well-Being: The Impact of Change in the Labor Market Research in Labor Economics, Volume 40, 257
292 Copyright r 2014 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-912120140000040010

257

258

ARNAUD CHEVALIER

INTRODUCTION Reports of high financial returns to higher education are often trumpeted by stake holders to promote attending higher education. A considerable amount of empirical evidence supports this claim but there are far less evidence on the variability of this return,1 and how it is affected by the quality of the educational input. In general, sorting whereby more able students are selected by more selective institutions and earn more in the labour market (Hoxby, 2009) generates a positive correlation between institutional quality and future earnings. Thus naı¨ ve estimates, assuming random allocation of students to institutions, are severely upward biased. Whether returns differ by institution is crucial to inform prospective students about investing in tertiary education. This paper provides estimates of the effect of institutional quality on early career earnings in the UK using several identification strategies, including, for the first time in this context, the Generalised Propensity Score (GPS) (Hirano & Imbens, 2004). While graduates from higher quality institutions earn more (see James, Alsalam, Conaty, & To, 1989 for early evidence) whether this is due to the quality of the institution remains debatable. Assuming that the better inputs improve educational attainment,2 the human capital model predicts that graduates from higher quality institutions will be more productive and obtain higher wages. In fact, even in the absence of increased human capital, graduates may obtain higher wages if these institutions endow them with a set of peers that improve their job prospects. However, there are reasons to believe that the positive correlation between institution quality and graduate earnings is not causal. As in a signalling model, students may attend more selective institutions to signal their greater ability to future employers.3 Obtaining unbiased estimates of the effect of higher education quality on earnings is difficult since students and institutions select each other, and the characteristics, observable or not, which affect these choices are also correlated with labour market outcomes. To account for selection, researchers have relied on adding controls for students’ ability (Brewer, Eide, & Ehrenberg, 1999), matching models (Black & Smith, 2004, 2006), instrumental variable (Long, 2008), twins/family fixed effects (Behrman, Rosenzweig, & Taubman, 1996; Lindahl & Regne´r, 2005), regression discontinuity (Hoekstra, 2009; Saavedra, 2008) or pairing students applying to the same institution (Broecke, 2012; Dale & Krueger, 2002, 2014). Here, we rely on different strategies assuming selection on observables, including propensity score matching (PSM) and GPS; that is, we assume that the

Does Higher Education Quality Matter in the UK?

259

observed match of a student to an institution is based on observable characteristics. Part of the selection process is unknown to us whether to attend tertiary education and which institution to apply to. However, conditional on applying, the centralised application system in use in the UK makes the assumption of selection on observable characteristics of the prospective student plausible. The UK is characterised by a centralised application system, whereby all prospective students complete the same standardised form, and institutions base their decision to accept or not a student only on this information.4 In the data at our disposal, we have (almost) the same information than the admission officers, which makes the assumption of selection on observable plausible.5 Compared to a PSM estimate, GPS allows us to capture the heterogeneity in the returns to quality, by estimating the effect of educational quality at all points of the quality distribution. The debate on the effect of institutions on earnings has received a large amount of attention in the US where the market for higher education is fiercely competitive but much less in the rest of the world. There are a few empirical evidences for the UK but their results are somehow ambiguous. To inform this debate, we use survey data pertaining to the 2003 cohorts of graduates, which is linked to administrative data so that background information on the students and academic performance can be added. Another difficulty in this literature is to measure quality as there is no agreement on which input matters. We measure institutional quality as the principal component of a set of education inputs: Research Assessment score, student/staff ratio, academic expenditures per student, mean entry grade and graduate prospect, which are commonly used to create league tables. We demonstrate that there is considerable heterogeneity in returns to quality, with almost no returns for below median quality institutions and large returns for attending the most prestigious institutions. This is consistent with theoretical predictions that as attendance to college increases, returns to quality increase (Hoxby, 2009). Due to this strong non-linearity, OLS estimates are biased upwards and considerably larger than those obtained by GPS. Defining quality as a discrete step function somehow helps to capture the non-linearity. The larger impact of higher education quality at the high end of the distribution is also present in quantile regressions, thus graduates with the higher earning potentials are the one who benefit the most from attending high-quality institutions. The rest of the paper is organised as follows. The section ‘Literature Review’ reviews the US and non-US literature on the financial returns to higher education quality. In the section ‘Institutional Background and Data Description’, we discuss the institutional background as well as the

260

ARNAUD CHEVALIER

data. The section ‘Empirical Strategies’ presents the estimators used in the analysis. The results are available in the section ‘Results’ and a discussion on the robustness checks and the implication of the returns is found in the section ‘Discussion’.

LITERATURE REVIEW The literature on the effect of higher education quality on earnings originated from the US. One of the first studies to account for selection of students between institutions of different quality was by Brewer et al. (1999). They conclude that even correcting for selection into the type of university attended (identified by net tuition costs), prestigious private institutions provide significantly higher financial returns compared to low-cost public institutions, but there are little returns to attending an elite public institution. Black and Smith (2004) confirm that fee differentials are in line with quality differentials but criticise the parametric approach adopted in the rest of the literature, especially the linearity assumption, and instead recommend using PSM. Their main results pertain to the wage differential between graduates from institutions in the top and bottom quarter of the quality distribution. This large quality gap makes the assumption of selection on observable potentially problematic and may not represent the typical choice of students who are more likely to arbitrate between institutions of more similar quality. They report premium of 12% for men and 7% for women for attending a better quality institution. When estimating the wage differential between students from the second and third quality quartile compared to the lowest one, the estimated quality effect was small and statistically insignificant for men and around 12% for women. However, students’ unobservable characteristics may still bias these estimates upwards; for example, if more motivated students attend more prestigious institutions and also, independently of the institution quality, earn higher wages. Dale and Krueger (2002, 2014), use information on all applications to a selection of high-quality institutions, linked to Social Security Administration records. To control for selectivity on unobservable, they compare students who applied to the same institutions but went to different colleges. They find no financial return to attending a more selective institution, maybe due to the homogeneity of institution quality in the dataset used (30 highly selective institutions only). They report substantial returns to quality for ethnic minority students, which they reckon, could stem from

Does Higher Education Quality Matter in the UK?

261

these students obtaining a network of peers which boosts their career prospect. Hoekstra (2009) uses data on all applicants at a large state university and identifies the quality premium from a regression discontinuity; that is, comparing the earnings at age 28
33 of applicants that barely made it, to those of applicants that just failed. The earning premium reaches 20% but this may have little external validity. An alternative strategy to account for unobservable characteristics is to rely on twins who attend different institutions. Behrman et al. (1996) find significant wage differentials between female twins who attended colleges that differ along various measures of quality.6 With the exception of Dale and Krueger (2002, 2014) all studies report significant premium to attending a more prestigious institutions, and that OLS estimates are positively biased by selection. It is, however, unclear how much can be extrapolated from US evidence since in other countries the market for higher education tend to be more regulated. Evidence from other countries is scarcer. Papers using administrative registries from Sweden and various identification strategies report positive effect of institutional quality on graduate earnings (de Luna & Lundin, 2014; Lindahl & Regne´r, 2005).7 However, no returns to the selectivity of the institutions are found in Australia, Finland or Japan (Birch, Li, & Miller, 2009; Nakamuro & Inui, 2013; Suhonen, 2012, respectively). Amongst British evidence, Hussain, McNally, and Telhaj (2009) exploit various graduate cohorts and a set of quality variables. When combining all measures of quality, an increase of one standard deviation in quality is associated with a wage premium ranging from 2.5% to 5.5%, increasing for the more recent cohorts. They account for selection by including measures of student’s ability. Chevalier and Conlon (2003) use PSM to estimate the effect of university quality on the earnings of three cohorts of graduates (1985, 1990 and 1995). Their measures of quality are an indicator of appurtenance to a self-selected pressure group of prestigious universities (Russell Group) and an indicator for when the institution was granted university status.8 Graduating from the most prestigious institution is associated with a wage premium ranging from 1% to 6% but there is no significant difference between the earnings of graduates at old and new institutions. There is also some evidence that the premium for attending the higher quality institutions increases for more recent cohorts suggesting that as the number of graduates expanded employers may have used institution prestige to differentiate between candidates. Chevalier and Conlon (2003) also estimate the effect of institution quality on wage growth and reports that the effect of institution

262

ARNAUD CHEVALIER

quality on earnings is stable (for the first 10 years after graduation). Broecke (2012) uses administrative data for a cohort of school leavers and their full record of university applications. He matches students who were accepted to the same institution but were one failed to gain the grades to satisfy the conditional offer and reports that a one standard deviation improvement in quality leads to a 7% increase in wages three years after graduation. Finally, Black and Smith (2006) also highlight that the estimate of quality is likely to be biased downwards by measurement error. Most studies have estimated quality by relying on the average SAT scores of attending students, but since quality is likely to be multi-dimensional this is only a proxy for the institutional quality. Instead, Black and Smith (2006) recommend using an array of quality measures and/or estimation techniques that account for measurement error (IV, GMM, Bounds).

INSTITUTIONAL BACKGROUND AND DATA DESCRIPTION The Longitudinal Destination of Leavers from Higher Education (LDLHE) was conducted in November 2006 amongst a random sample of students who graduated in the summer of 2003. The survey is conducted in two stages. First, the universe of all higher education leavers is sampled six months after graduation (75% response rate). In the second stage, a 55,900 sample is selected to take part in the longitudinal study. The response rate at this second stage is typical of postal survey (44%) and Tipping and Taylor (2007) provide evidence in favour of the representativity of the survey. Survey weights are used throughout the analysis. The LDLHE is linked to administrative data from the Higher Education Statistical Agency so that additional information on secondary schooling and university achievements, as well as family background can be added. We select first degree holders, aged 18 to 25 on graduation, non-special entry students and who are currently observed in employment, with valid earning information. Since hours of work are not available in the survey, the labour market outcome of interest is annual earnings. We thus drop part-time workers. We also drop 21 individuals with self-reported earnings above £60,000 (in 2006) three years after graduation so as to reduce measurement error. This leads to a sample of 6,986 observations. (See Table A1 for details on the sample selection).

Does Higher Education Quality Matter in the UK?

263

We supplement the dataset with measures of institution quality. There is no unambiguous measure of institution quality in the UK. All major newspapers provide annual university ranking differing in their methodology. We use information collected from the ‘Good University Guide’,9 the longest running provider. Rather than using the ranking we compute a quality measure based on the first principal component along five dimensions of quality for 113 institutions: Research Assessment score, student/staff ratio, academic expenditures per student, mean entry grade and graduate prospect, which we then normalise to a mean of 0 and a standard deviation of 1. The first three variables measure the quantity and quality of the input, the mean entry grade is a measure of the selectivity and popularity of the institution but also captures the quality of peers. This has been widely used as a measure of institution quality. The last measure of quality is the probability that graduates are in further studies or graduate employment six months after leaving university and reflects the employers/universities view on the institution quality. The variables pertain to data collected for the academic year 2000, when the students applied and the 2001 Research Assessment Exercise. The quality is computed at the level of the institution and not specifically for the subject studied. It can be argued that this is the appropriate level of aggregation to measure higher education quality for labour market related outcomes since employers may have knowledge of institutional quality but not subject specific quality. The first principal component accounts for 70% of the variation in quality. The resulting ranking appears plausible with the top five institutions (in alphabetical order) being Cambridge, Imperial College London, London School of Economics, Oxford and University College London. It also compares favourably with other measures of quality that have been used. For example, the mean quality score is significantly different for Russell group institutions (2.42), a group of the most prestigious institutions, the ‘group 1994’, a second tier of research institution (1.42) and the remaining institutions (−1.14). Indeed, the distribution of quality appears bimodal (Fig. 1) and is characterised by a long tail at the high-quality end. The lower part of the distribution is mostly formed of the new institutions which were granted university status in 1992. It is unclear what the driver of the quality differences are. UK institutions are charities whose teaching activities are funded by a block grant and tuition income. The grants are a function of the number of students weighted by subject types. In England, tuition fees are capped for home and EU students (currently at £9,000) but are uncapped for non-EU students. As such, teaching funding differs by the mix of subject offered and

ARNAUD CHEVALIER

0.1 0

0.05

Density

0.15

0.2

264

–4

–2

0 2 university quality score

4

6

kernel = epanechnikov, bandwidth = 0.6835

Fig. 1. Distribution of Institution Quality. Note: Quality is measured as the principal component of research assessment, student staff ratio, academic expenditure, mean entry grade and graduate prospect. The dash line reflects the thresholds to the first, second and third quartile of the quality distribution.

the share of non-EU students. The latter varies from 0% to 40% and may be a driver of quality differences. The other sources of incomes are research income, which also varies dramatically by institutions
the Russell group institutions, for example, accrue 75% of the research grants income;10 however the extent that this income is used towards teaching is unknown. Endowment and other type of charitable contributions are very small, representing less than 1% of the sector income and are unlikely to generate large variations in funding between institutions (Universities UK, 2011). Fig. 2 confirms that in the absence of any additional control there is a positive relationship between the institution quality and the average earnings of graduates. In fact, two clusters can be distinguished. At below median quality, there is little variation in earnings between institutions. For institutions with positive quality score, the relationship between quality and graduate earnings is much stronger. The graph thus highlights that the relationship between institutional quality and graduate earning is convex, rather than linear. OLS estimates, which rely on a linearity assumption between quality and earnings, may thus be biased. The differences in wages by quality are substantial. Moving from an institution in the bottom tiers to a top five institution is associated with 50% higher wages.

265

20005 20000 15000

Mean wage

30000

30005

Does Higher Education Quality Matter in the UK?

–4

–2

0 2 University quality

4

6

Fig. 2. University Quality and Mean Institutional Wage for Graduates. Note: The circle size represents the number of students at the institution. Mean wages are selfreported annual wages for full-time workers earning less than £60,000, three years after graduation. The line is based on a quadratic fit of university quality on earnings.

The second relationship that we need to investigate is the amount of sorting between the academic ability of students and the quality of institutions. As stated previously, if the sorting was perfect it would not be possible to identify the effect of institution quality. However, the allocation mechanism of students to institutions in the UK, offers scope for some heterogeneity in the composition of the student body by institution. The admission process to the 134 universities is centralised. From the autumn to the spring preceding their admission to university, high school pupils fill a standard form online at the Universities and Colleges Admissions Service (UCAS). Prospective students can state a maximum of six choices (institution, subject) on this application.11 Since students have imperfect information on their ability (Chevalier, Gibbon, Thorpe, Snell, & Hoskins, 2007; Furnham, 2001, for a review), the restricted number of applications means that one strategy to guarantee access to higher education is for some students to apply to institutions spanning the quality range. UCAS then send this form to all chosen institutions/departments which decide whether or not to make an offer to the prospective students. As such, all institutions decide simultaneously, based on exactly the same information. Note that at

266

ARNAUD CHEVALIER

this point, candidates have not yet sat their end of secondary education exams (usually A-levels). Instead, the decision to offer a place at an institution is based on the teacher’s prediction of grades at this high-stake exam and previous grades. The offers are typically conditional on applicants reaching a pre-determined score in their, still to come, high-stake exams.12 When all institutions have responded, the applicant has to keep only one offer, and choose another one as an insurance choice, in case she does not reach the standards required by her preferred conditional offer. In August, preceding the start of the academic year, the high-stake exam results are revealed. Conditional offers are confirmed or terminated. There is no possibility to trade-up, that is, candidates who over-performed are locked in their accepted offer and may thus be observed in institutions of lower quality compared to their peers whose high-exam performance was correctly predicted.13 Pupils who fail to achieve the requirements of their first choice are committed to their insurance choice. For those who also fail the requirement of their insurance choice, a clearing mechanism allocates candidates to institutions that still have places on their courses. About 10% of successful applicants gain access through clearing. These clearing places are allocated on a first come first serve basis
as long as academic credentials are appropriate. A candidate who was accepted in a highquality institution but marginally failed to achieve the required score may at this stage only be able to register in a lower quality institution. Institutions which under-recruited may at this stage lower the recruitment threshold, leading to lower ability students being accepted to high-quality institution. This allocation mechanism of students to institutions makes it credible that while selection is based on academic merit, it is imperfect and creates the potential for mismatch between student’s ability and institutional quality. The LDHLE is linked to administrative data so that we observe the same information as universities do when deciding whether or not to make an offer, with the exception of the entry score, for which we observe the realisation rather than the prediction.14 However, since realised scores determine whether the offer is upheld or not, they contain the relevant information to match students to the institution they attend. Average attainment differs widely between students at colleges of different quality. At the top quartile institutions the average entry score is 23.60 out of 30, 6 points more than in the third quartile and 14 points more than in below median institutions (Table 1). Fig. 3 plots the distribution of entry score by quality quartile. Clearly, the allocation of students to institution is not random, and the distribution of test scores shift to the right for every

267

Does Higher Education Quality Matter in the UK?

Table 1.

Summary Statistics by University Quality Quartile.

Variable

Quartile 1

Quartile 2

Quartile 3

Quartile 4

A-levels score missing

0.42 13.41 (6.21) 0.11

0.43 15.65 (6.03) 0.10

0.47 21.49 (5.84) 0.02

0.45 25.61 (5.03) 0.03

Parental occupation Manager Professional Associate professional Other Not reported/no occ.

0.11 0.09 0.07 0.28 0.45

0.17 0.14 0.07 0.26 0.35

0.22 0.22 0.09 0.28 0.19

0.26 0.34 0.08 0.23 0.09

0.13 0.27 0.31 0.13 0.17 0.04 0.78 20.31 (1.54)

0.10 0.24 0.44 0.12 0.09 0.06 0.84 20.35 (1.53)

0.13 0.14 0.45 0.11 0.17 0.04 0.88 19.79 (1.11)

0.20 0.08 0.62 0.09 0.02 0.03 0.86 20.00 (1.15)

20,683 (6,113) 1,658

21,005 (6,183) 2,001

22,785 (7,403) 1,762

26,655 (9,112) 1,551

Male A-levels score

Accommodation University Parents Own Other Unknown Disable Whitea Age on graduation Labour market outcomes Ln salary Observations

Note: Cells report means of variables and standard error in parentheses when the variable is continuous. The sample is weighted to account for survey structure. Only individuals currently working full time and reporting earning less than £60,000 per annum are selected in the sample. a The analysis is conducted at a more disaggregated level and ethnicity is broken down between Black, Indian, Pakistani/Bangladeshi, other Asian, mixed, other ethnicity.

quartile of institutional quality. The top quality institutions recruit almost no students with below average ability and have a more compressed distribution of student ability; this is expected since low performers would not have received an offer if correctly predicted or would have seen their conditional offer elapsed. However, a small fraction of high-ability students are found in the lowest quality institution. The sorting is thus asymmetrical which is consistent with the recruitment procedure highlighted above. While common support is universal, the support is thin in the tail of the

ARNAUD CHEVALIER

0.02 0.04 0.06 0.08

4th Quartile 2nd Quartile

3rd Quartile 1st Quartile

0

kdensity ascore

0.1

268

0

10

20

30

A-levels score

Fig. 3.

Distribution of Entry Exam Score by University Quality Quartile. Note: Distribution of entry exam score by quartile of institution quality.

distribution. These conclusions of asymmetric sorting and thin common support are similar to Black and Smith (2004) for the US. Students at institutions of different quality differ along other dimensions too. Table 1 reports the means for all variables reported by applicants on their application form. Students at the highest quality institutions have in general more favourable characteristics. For example they are four times less likely not to have a reported entry score, three times more likely to have parents in manager or professional occupations and 10 times more likely to have been educated at a private school than students at the lowest quality quartile institutions. Living with parents while studying reduces the choice of institutions that a student can apply; students at the top institutions are three times less likely to be living with their parents. The differences in student characteristics between institutions are thus consequent. The survey has only one measure of income: self-reported yearly income three years after graduation. As highlighted in Fig. 1, institution quality is associated with wage differential and students from the top quartile institutions earn 25% more than their peers at the lower end of the quality distribution. The wage gap appears especially large between third and fourth quality quartiles. Moving from quartile 1 to 2 and to 3 is associated with an average wage increase of £1,000 but moving from 3 to 4 the gap is more than three times as large. Quality effects may thus be heterogeneous.

Does Higher Education Quality Matter in the UK?

269

EMPIRICAL STRATEGIES We conduct the analysis using several empirical strategies. The first is to estimate the effect of institution quality on log earnings (ln Y) by Ordinary Least Squares. Typically, the estimated model would be of the form: lnY = β0 þ β1 Q þ β2 X1 þ β3 X2 þ ɛ

ð1Þ

where Q is a measure of university quality, X1 a set of individual characteristics on graduation, including a measure of the student pre-enrolment ability, and X2 a set of current characteristics. X2 can be considered endogenous as the university quality may affect various dimensions of the labour market attainment of graduates. In the application, X2 contains only indicator of the regional labour market conditions (126 postcode for work location). Note that standard errors are clustered at the institution level to account for possible correlations between observations. The parameter of interest is β1 which represents the increase in earnings due to an increase in university quality. As stated previously β1 may be biased, even if Eq. (1) includes all confounding factors correlated with both quality and earnings, if the relationship between quality and earnings is non-linear and if there is a lack of common support (in which case the identification is purely due to the imposed functional form). The first limitation can be eliminated by measuring quality in a non-linear form: a set of dummies or a polynomial function, for example. Eq. (1) can also be estimated by quantile regression in order to test for heterogeneity in the effect of institutional quality on earnings which have been highlighted in the descriptive statistics. So that for each quantile p of the wage distribution we can, following Koenker and Bassett (1978), define the conditional quantile of the distribution as:

Q

ðpÞ



Y ðpÞ ðpÞ ðpÞ ðpÞ ln ; X1 ; X2 = βðpÞ 0 þ β1 Q þ β 2 X1 þ β 3 X2 þ Q ðɛÞ Q

ð2Þ

and impose the restriction that QðpÞ ðɛÞ = 0 to estimate the quartile specific parameters. To test for common support, we also estimate model Eq. (1) by PSM, where, quality is measured as a binary variable. The parameter of interest is then the mean differences in wages between graduates who attended a high-quality institution compared to those who did not but who, based on

270

ARNAUD CHEVALIER

their observable characteristics, could have done so, this can thus be considered an Average Treatment on the Treated (ATT). If we define Y1 the wage if attended a high-quality institution and Y0 the wage if attended another institution, and D is an indicator of having graduated from a highquality institution then the parameter of interest is simply: 

Y1 − Y0 ATT = E D=1

ð3Þ

ATT can be estimated if E(Y0/D = 1), which is never observed, can be approximated. If the allocation of students to universities is not random, then EðY0 =D = 1Þ ≠ EðY0 =D = 0Þ. However, if it can be argued that the selection is based on observable characteristics (X) then the earnings of non-treated individuals can be used as counter-factual to approximate the unobserved earning of treated individuals if they had not been treated, that is, conditional on this set of variables X an observation can be considered randomly allocated to either the treatment (high-quality institution) or the control group ðEðY0 =X; D = 1Þ = EðY0 =X; D = 0ÞÞ. Formally, this is expressed in the Conditional Independence Assumption: Y0⊥D/X. Since we observe almost the same information as universities do when they make an offer to a student the assumption of selection on observable is plausible. The only difference in the set of observables is that universities make their offer using the results of past exams and the predicted grades at A-levels while we observe the realised A-levels grades. We do not have access to the predicted grades, nor the breakdown of the entry score by subjects. However, since the realised grade, not the predicted one, guarantees the access to higher education it is in fact the most relevant information to base our matching. Both OLS and matching produce unbiased estimates as long as the selection into institution of high quality is due to observed variables, but matching also highlights potential bias due to lack of common support. Formally, the common support can be expressed as 0 < Pr(D = 1/X) < 1; individuals with the same characteristics have a positive probability to be treated or non-treated. The probability of treatment is the propensity score which is used to match observations, rather than X, so as to reduce the dimensionality of the matching problem. Rosenbaum and Rubin (1983) show that it is equivalent to match on all the components of X or on Pr(D = 1/X). For each treated individual i, the counter-factual outcome associated with no treatment is a weighted average of the outcomes from control

Does Higher Education Quality Matter in the UK?

271

observations, where the weights are a function of the distance between the treated and control observations’ propensity scores (using Epanechnikov kernel). If no match is found within a bandwidth, the treated observation cannot be matched which highlights the lack of common support discussed above. Black and Smith (2004) estimate the effect of moving from the bottom quartile of the quality distribution to the first quartile. Only a fraction of students will consider switching from a bottom to a top quality institution, and the estimate may be of limited interest. Moreover, by considering institutions that are so different, the common support always become extremely thin. Instead, we estimate three effects: moving from the first to the second quality quartile, second to third and third to fourth. Finally, since quality is (almost) a continuous variable we extend the PSM estimator to the case of continuous treatment (Hirano & Inbems, 2004). The remaining part of this section draws heavily on this article. The intuition behind Continuous Propensity Score Matching (CPSM) is rather similar to the dichotomous case. A GPS is calculated which is the density of the treatment (T) conditional on a set of covariates X: R = r(T,X), the GPS has similar properties to the propensity score in the dichotomous case, mainly it satisfies CIA, so that: X⊥1fT = tgjrðt; XÞ. A two-step procedure is then used to remove any bias due to differences in X. First, we estimate the conditional expectation of the outcomes as a function of the treatment (T) and the GPS: β(t,r) = E[Y/T = t,R = r]. This regression function does not have a causal interpretation and is only used to calculate the dose response function (μ) in the second step. μ(t) = E[β(t,r(t,X))]. Computationally, we estimate the GPS by OLS (T=X ∼ Nðβo þ β01 X; σ 2 Þ and calculate:  1 1 0 2 ^ ^ ^ R = pffiffiffiffiffiffiffiffiffiffi exp − 2 ðT − βo − β1 XÞ 2^σ 2π σ^ 2

ð4Þ

In the second stage we estimate, again by OLS, the outcome as a function of a flexible function of the treatment and the GPS. Following Kluve, Schneider, Uhlendorff, and Zhao (2012) we estimate:   2 3 E YjT; R^ = α0 þ α1 T þ α2 T 2 þ α3 T 3 þ α4 R^ þ α5 R^ þ α6 R^ þ α7 R^ × T 2 þ α8 R^ × T þ α9 T 2 × R^

ð5Þ

Eq. (5) is then used to estimate for a given level of treatment t, the average potential outcome. The dose response function is found by computing this

272

ARNAUD CHEVALIER

last stage for all level of the treatment considered. Hirano and Imbens (2004) recommend bootstrapping to obtain standard errors of the estimates. Rather than the dose response function, we report its derivative; in our case the interpretation is the marginal return to university quality. We also compute the expected (log) earnings at all values of the treatment. As in the case of dichotomous treatment, the GPS is only valid if the matching leads to a balancing of the characteristics of the treated and untreated group, that is, akin to ex-post random allocation of treatment. To assess the balancing we follow two strategies. First, as in Hirano and Imbens (2004), we divide the institution quality variable into three terciles and test, for each variable, whether the GPS adjusted mean differs in one tercile compare to the other two. This is equivalent to testing that the conditional mean and the treatment indicator are independent (CIA) where r(t,X) is evaluated at the median value of the treatment within the tercile (t*). We follow Hirano and Imbens (2004) and test this hypothesis by blocking. For each tercile, we define five blocks defined by the quintile of r(t*,X). For each block we calculate the mean difference in X for observations (T = t) and (T ≠ t), and combine these five mean differences, weighted by the number of observations in each block, to calculate a t-statistics of the statistical differences in the mean of X between treated and untreated observations in that tercile. Imai, King, and Stuart (2008) suggest that such a test may have low power as the decrease in the t-statistics may be due to an increase in the variance rather than a reduction in the difference in means (Balance Test Fallacy). Thus, we also conduct a second balancing test. Imai and van Dyk (2004) recommend to test the balancing of the covariates by regressing each covariate on the treatment variable and the predicted treatment E[T/X] and testing for the significance of the treatment. An alternative to this test is to regress each covariate on the treatment and the GPS, at different values of the GPS (Kluve et al., 2012).

RESULTS As discussed above, the main criteria used by universities to accept applicants is their predicted score at a high-stake exam but we use the realised score instead to control for students’ ability.15 The decision to apply to universities of different quality may also be related to individual characteristics

Does Higher Education Quality Matter in the UK?

273

(sex, age, ethnicity, disability status) and other background characteristics (private school, parental occupational class, whether expect to live at home while studying and tuition fee status) which we include in the earning regression or in the matching process. All these variables are available to universities when assessing a student’s application. We additionally control for region of origin and region of the institution attended.

Discrete Measure of Quality First, we recode the measure of university quality into a discrete measure to facilitate comparisons with previous research. Quality is then measured in quartiles of the continuous quality measure with the fourth quartile representing the highest quality. Table 2 reports estimates of Eq. (1) estimated by OLS for different specifications of the covariates. In the first column, the only additional control is a set of postcode dummies to capture the characteristics of the local labour market. There are some marked institution quality effects and graduates from the top quality institutions earn 20% more than those from the lowest quality institutions. This premium is only 8.4% for graduates from the third quality quartile and no significant premium is observed for graduates from the second quality quartile. However, these estimates are biased upwards since the ability of students is correlated both with institutional quality and earnings. Thus, the second column adds a cubic function of the entry score to control for ability. The estimates on the effect of quality are reduced by at least a third, confirming that there is a large amount of sorting of students by ability. Note that for students attending a second quartile institution, controlling for attainment does not alter the returns; that is, the intake at these institutions are similar. The next column reports estimates that controls for various dimensions of the individual characteristics and socio-economic background, including region of origin and region of institution; this reduces the quality effect further, so that the premium to attending a top institution is halved compared to the original specification. The fourth column adds controls for subject of studies and degree grades since institutions of different quality may differ along these dimensions as well. Indeed there are marked differences in the distribution of subjects by quality of institution (Table A2). Medicine, all sciences and engineering degrees, law, CLASSICS, LANGUAGES, HISTORY, ECONOMICS, are more likely to be taught at higher quality institutions. While Business, COMMUNICATION, CREATIVE ARTS, IT, being more popular at

274

ARNAUD CHEVALIER

Table 2.

OLS Financial Returns to Quality.

Discrete Measure Quality 2 Quality 3 Quality 4 Location controlsa A-level score (cubic function) Gender Parental background Other socio-economic characteristicsa Subject of graduation Grade Adjusted R2 Observations Continuous measure Normalised quality score Normalised quality score2 Location controlsa A-level score (cubic function) Gender Parental background Other socio-economic characteristicsa Subject of graduation Grade Adjusted R2 Observations

(1)

(2)

(3)

(4)

0.025 (0.020) 0.084*** (0.021) 0.205*** (0.021) Yes

0.026 (0.019) 0.065*** (0.021) 0.141*** (0.025) Yes Yes

0.028* (0.017) 0.053*** (0.019) 0.093*** (0.023) Yes Yes Yes Yes Yes

0.025* (0.014) 0.046*** (0.015) 0.080*** (0.018) Yes Yes Yes Yes Yes Yes Yes

0.20

0.22

0.29

0.39

6986 0.077*** (0.007) 0.009 (0.005) Yes

0.20

0.055*** (0.009) 0.003 (0.005) Yes Yes

0.041*** (0.009) 0.007 (0.005) Yes Yes Yes Yes Yes

0.22

0.28

0.034*** (0.007) 0.009* (0.005) Yes Yes Yes Yes Yes Yes Yes 0.38

6986

Note: Standard errors reported into brackets adjusted for clustering at the institution level. ***, **, * denote statistical significance at the 1%, 5% and 10% statistical level, respectively. a Location controls are a set of 126 postcodes to account for local labour market characteristics. The other socio-economic characteristics are age, ethnicity, disability status, school type, whether expect to live at home when studying and tuition fee status, dummies for region of origin and region of institution.

lower quality institutions. Controlling for subjects (and final grades) improves the precision of the estimates and reduces the gap further, so that they now range from 2.5% to 8.0%. Despite the rich set of controls significant quality differentials remain for institution above the median

275

Does Higher Education Quality Matter in the UK?

quality, and especially for those in the top quality quartile, which suggests that differences in earnings cannot be fully explained by selection. We now estimate the institution quality effects by PSM for three different treatments: attending an institution in the fourth quality quartile rather than one in the third; a third quality quartile institution rather than a second; and a second quality quartile rather than a first. These treatments are likely to be similar to students’ alternatives when selecting universities. Propensity scores are estimated separately for each treatment. In each case, the controls are a cubic function of entry score, an indicator for having no reported score, gender and disability, dummies for age and ethnicity, proxies for financial constraints, subject dummies, and dummies for region of origin.16 The matched samples are balanced for the treated and control groups on all characteristics. The common support is almost universal, between 94% and 99% of the treated are matched to some control observations but, the support tends to be thin for high value of the propensity score
see Fig. 4 for an example when the treatment is attending a fourth

0

0.2

0.4

0.6

0.8

1

Propensity Score Untreated

Treated: On support

Treated: Off support

Fig. 4. Balancing Propensity Score
Treatment (Top Quality Quartile) and Control (Third Quality Quartile). Note: The propensity score includes an indicator of A-level score missing, a cubic in A-level score, gender, parental occupation categories, subject of degree, tuition fees status, accommodation status, disability, type of school attended, ethnicity and a set of dummies for age.

276

Table 3.

ARNAUD CHEVALIER

Propensity Score Matching Estimates of University Quality on Earnings.

Treatment Ln salary Treated Control Propensity score Matched treated Matched control ATT % treated matched % control used for matchinga Nbr of controls accounting for 50% of match
total number of controls useda Observations

Q4-Q3 10.15 10.00

Q3-Q2

Q2-Q1

9.97 9.91

9.91 9.93

10.15 10.09 0.053*** (0.018)

9.97 9.93 0.036* (0.021)

9.91 9.91 0.008 (0.021)

99% 86% 47/1,526

94% 74% 35/1,485

99% 89% 26/1,857

3,326

3,682

3,660

Note: The propensity scores are estimated by a probit with the following covariates: an indicator of A-level score missing, a cubic in A-level score, gender, parental occupation categories, subject of degree, tuition fees status, accommodation status, disability, type of school attended, ethnicity and a set of dummies for age on graduation and region of residence. The matching estimators is based on Epanechnikov kernel matching with a bandwidth of 0.01 ***, **, * denote statistical significance at the 1%, 5% and 10% statistical level, respectively. a These statistics are based on a nearest neighbour matching with a caliper of 0.01 which are available upon request.

quartile university rather than a third quartile university. Figures for the other treatment are similar. The PSM estimates are reported in Table 3. The results confirm that returns to quality are only observed for above median institutions. Wages are 5% higher for graduates from the fourth quartile over those from the third quartile. Those graduates also enjoy a premium of 4% over those who attended an institution from the second quartile. There is no difference in earnings between graduating from the bottom two quality quartiles. The OLS and matching estimates lead to the similar conclusions; the effect of institution quality on earnings is non-linear, and only exists for above the median institutions. Note that the returns to quality are substantially larger when estimated by PSM. For example, the returns to attending a fourth quartile rather than a third quartile institution is 3.4% when estimating by OLS (0.08
0.046) but 5.3% when estimated with PSM.

Does Higher Education Quality Matter in the UK?

277

Continuous Measure of Quality Rather than relying on discrete measures of quality, the analysis is now conducted using a continuous measure. The lower panel of Table 2 reports the estimates on a quadratic function of quality for various specifications.17 In the most parsimonious model, moving from one standard deviation from the mean quality increases earnings by 7.7%. However, as in the discrete case, adding further covariates considerably reduces this premium and in the most complete model this premium has been more than halved and reaches 3.4%. This estimate is doubled for an institution with a normalised quality score of 2, highlighting the nonlinearity of the returns. A specification using a log quality measure was also tested, the estimated elasticity of quality then ranges from 0.12, for the most parsimonious model, to 0.05, for the full control one. Later we investigate the non-linearity of this relationship without imposing a functional form. Relying on a central tendency measure may be misleading. We thus additionally report estimates from quantile regressions. Rather than imposing that the effect of university quality is homogenous at all level of income, quantile regressions allows it to differ for different quantile of the income distribution. One may, for example, expect that individuals with higher earning potential may gain more from attending a higher quality institution. To simplify the interpretation of these estimates those estimates are based on a linear function of quality. For each decile the estimated effects are reported graphically in Fig. 5. For the first six income deciles, the estimated effect of university quality lies at the bottom of the OLS confidence interval and average 0.046. The estimated returns to quality then jumps almost 1 percentage points to the OLS estimate for the seventh decile, and thereafter becomes a positive function of the income decile. However, only the estimate for the 9th decile (0.078) is significantly different from the OLS estimate. Thus, there is a strong heterogeneity in the returns to institution quality but only for individuals at the top of the earnings distribution. The previous estimates using a continuous measure of university quality have assumed a random allocation of students to universities. We keep assuming that the allocation is based on observable characteristics but use the Continuous Treatment Matching method proposed by Hirano and Imbens (2004) to estimate the institution quality effect on the earnings of recent graduates, so that no functional form on the relationship is assumed. We first estimate the GPS by regressing a log transformation of the quality

278

ARNAUD CHEVALIER 0.1 Quantile estimate

OLS

0.09 0.08 0.07 0.06 0.05 0.04 0.03 Q1

Q2

Q3

Q4

Q5

Q6

Q7

Q8

Q9

Fig. 5. Estimated Returns to Institution Quality: Quantile Regressions. Note: Darker dash lines indicate the 5% confidence interval for quantile estimates, grey dash lines indicate the 5% confidence interval for OLS estimate. Estimated model has the same specification than the most extensive models presented in Table 2, but where the categorical measures of institution quality have been replaced by a linear function of the normalised quality score. Regional dummies have been omitted as small number of observations in some categories prevented the model converging.

score on all the covariates used previously and computing GPS value using the normal distribution (4).18 The balancing properties of the estimated GPS are assessed in Table 4. Following Hirano and Imbens (2004), we first split the sample into three terciles with respect to the quality of the institution attended. The first three columns of Table 4, report for each covariate a t-test of the mean difference in one tercile (g) compare to the other two groups. Clearly, there are marked differences between graduates from institutions at different level of the quality distribution, highlighting that selection of students take place. We then compute the GPS at the median value of the tercile and split the group into five blocks. Within each block, we compute the t-statistics of the difference in means between treated (went to an institution of quality g) and non-treated observations. Columns 4
7 report the weighted average t-statistics for each block. After conditioning on GPS, most mean differences between groups are dramatically reduced and insignificant. There remains a small significant difference in the highest A-level score category,

279

Does Higher Education Quality Matter in the UK?

Table 4.

Balancing of the GPS Covariates
t-Statistics of Mean Difference. Unconditional

Conditional on GPS

Imai Test

Tier 1 Tier 2 Tier 3 Tier 1 Tier 2 Tier 3 Uncond. Cond. Male A-level none A-level [0,2] A-level [2,4] A-level [4,6] A-level [6,8] A-level [8,10] A-level [10,12] A-level [12,14] A-level [14,16] A-level [16,18] A-level [18,20] A-level [20,22] A-level [22,24] A-level [24,26] A-level [26,28] A-level [28,30]

4.56 22.93 6.42 7.56 4.71 10.64 9.64 7.66 5.94 5.62 0.32 6.30 9.34 10.87 12.06 15.22 18.23

4.94 1.84 1.50 1.00 2.33 1.71 0.15 1.53 3.71 0.84 4.95 8.22 4.83 3.25 3.04 5.09 13.56

0.14 20.85 4.88 6.47 6.92 8.80 9.65 9.08 9.51 6.42 5.10 1.57 4.80 7.91 15.51 21.05 33.86

0.53 1.27 0.26 0.62 0.13 0.36 0.40 0.10 0.23 0.43 1.00 0.75 0.09 0.88 1.07 1.17 1.17

1.38 0.22 0.13 0.17 0.88 0.09 0.65 0.71 0.97 0.00 0.41 0.90 0.27 0.09 0.96 0.72 1.03

0.58 1.30 0.39 0.92 0.96 1.01 0.93 0.93 0.98 0.48 0.13 0.30 0.18 0.83 0.16 0.66 2.07

3.66 25.12 6.75 7.92 8.28 10.88 11.71 10.42 10.06 7.44 4.13 1.90 6.18 9.28 14.16 20.96 40.74

0.18 1.32 0.84 0.92 0.89 0.61 0.77 0.66 1.15 0.60 2.57 3.02 3.32 2.65 1.12 0.41 10.16

Parental occupation gp1 gp2 gp3 gp4 gp5 gp6 gp7 gp8 gp9 gp10 Disability status School
FE institution School
HE institution School
independent School
unknown School
state school White Black Indian Pakistani/Bangladeshi Chinese Mixed Other

7.97 13.60 0.10 0.91 3.62 0.08 1.58 3.84 0.26 15.44 1.46 14.03 2.91 14.15 6.85 9.04 7.75 6.20 4.76 2.83 0.33 0.48 3.19

0.61 3.61 1.22 0.61 0.63 2.19 2.01 1.25 0.61 3.51 5.15 2.77 1.45 6.61 7.46 1.36 5.55 2.16 3.62 0.74 1.76 0.15 2.95

7.49 17.52 1.07 1.51 4.23 2.18 3.53 2.64 0.85 19.01 3.50 11.26 4.31 21.40 14.19 10.47 2.46 4.07 1.32 2.12 2.01 0.34 0.38

0.68 1.01 0.65 0.39 0.10 0.25 0.18 1.32 0.02 0.26 0.75 0.89 0.06 1.32 0.12 0.32 0.82 0.62 0.85 0.38 0.44 0.17 0.05

0.19 1.13 0.61 0.08 0.07 0.70 0.49 0.67 0.13 1.02 1.31 0.00 0.61 0.87 3.01 0.92 1.67 0.42 1.17 0.05 0.73 0.18 0.84

0.56 2.08 0.01 0.85 0.84 0.39 1.01 0.29 0.24 1.00 1.14 0.20 0.78 2.64 2.48 0.42 1.18 0.23 0.74 0.01 0.55 0.11 0.98

9.04 19.20 0.84 0.42 4.90 0.01 3.07 2.23 0.37 22.05 2.83 16.16 3.89 24.85 14.61 12.10 3.72 5.68 2.31 2.62 2.45 0.41 0.58

0.36 0.77 0.06 0.38 1.60 0.24 0.82 0.58 0.16 1.38 1.40 2.28 0.27 3.62 4.27 0.25 3.18 1.19 1.29 0.45 0.95 0.08 2.95

280

ARNAUD CHEVALIER

Table 4.

(Continued )

Unconditional

Conditional on GPS

Imai Test

Tier 1 Tier 2 Tier 3 Tier 1 Tier 2 Tier 3 Uncond. Cond. Age < 18 Age [18,19] Age [19
20] Age [20,21] Age [21,22] Age [22,23] Age [23,24] Age > 24 Fee status

2.98 4.49 0.49 2.56 0.29 4.13 1.43 6.84 0.23

0.39 0.85 0.94 2.17 0.47 0.31 2.21 1.94 1.80

3.44 3.71 0.41 0.50 0.75 4.41 3.56 4.93 1.50

0.33 0.12 0.12 0.30 0.46 0.29 0.18 0.92 0.88

0.29 0.39 0.39 0.00 0.14 0.33 0.50 0.32 0.47

0.70 0.52 0.99 0.38 1.41 0.65 0.00 0.44 0.55

4.93 4.58 0.22 1.43 0.01 4.08 4.15 7.28 0.05

0.85 0.86 0.99 1.56 0.98 0.10 0.20 0.24 1.23

Subject of studies Medicine 10.35 Sub. allied to medicine 0.53 Biology 3.09 Physics 4.86 Math 7.32 Engineering 3.05 Architecture 1.95 Social studies 2.85 Law 4.49 Business and administration 11.43 Communication 7.52 Linguistic 6.23 Literature 7.55 History/philosophy 7.53 Arts 8.78 Education 6.91 Other 1.50 Sport sciences 10.66 Psychology 0.91 IT 4.65 Economics 7.02 Finance/accounting 3.18 Mixed no science 3.21 Mixed half science 2.08 Mixed 100% science 4.66

6.68 0.73 1.13 1.53 2.77 1.08 1.75 3.97 1.91 4.15 2.32 0.13 1.85 2.19 1.96 6.50 0.72 3.89 0.37 3.08 0.21 2.74 1.71 0.43 0.72

17.50 1.24 2.04 3.46 10.27 4.13 3.64 0.95 6.43 15.49 5.21 6.26 9.62 9.88 10.62 0.75 2.18 6.71 1.27 7.62 7.47 5.81 1.58 2.50 5.50

1.17 0.46 0.30 0.63 0.87 0.73 0.06 0.57 0.31 0.16 0.92 1.01 1.08 0.79 0.25 1.62 0.19 1.34 0.14 0.27 1.03 0.26 0.84 1.12 0.68

1.82 0.03 0.10 0.05 0.78 0.66 0.11 0.48 0.70 1.68 0.53 0.14 0.37 0.63 0.80 1.18 0.51 0.57 0.15 0.75 0.05 1.38 0.63 0.30 0.37

2.15 0.16 0.11 0.01 1.12 0.70 0.57 0.82 0.02 1.41 0.47 0.01 0.81 0.51 1.12 1.23 0.41 0.48 0.01 0.94 0.42 1.40 0.42 1.36 1.15

17.66 2.75 3.70 3.19 10.02 5.38 3.56 0.51 7.78 17.02 7.56 6.33 9.60 10.74 12.43 3.58 3.47 9.08 2.23 6.44 8.49 6.01 2.30 0.40 7.14

2.46 2.13 1.76 1.35 0.26 0.32 1.64 1.51 0.68 0.77 1.80 0.19 0.10 0.50 3.07 0.39 0.23 2.20 0.62 0.68 0.11 0.26 1.12 0.34 1.12

Region of residence UK not specified Channel Island Isle of Man England not specified Wales

0.10 2.78 1.96 3.23 1.19

1.86 1.19 2.16 2.56 0.23

0.61 0.07 0.26 1.39 0.00

0.07 0.35 0.35 1.21 0.27

0.07 0.39 0.57 0.74 0.04

0.93 0.74 1.40 3.39 0.48

0.42 0.54 0.83 1.35 1.07

1.70 1.52 0.26 5.68 1.37

281

Does Higher Education Quality Matter in the UK?

Table 4.

(Continued )

Unconditional

Conditional on GPS

Imai Test

Tier 1 Tier 2 Tier 3 Tier 1 Tier 2 Tier 3 Uncond. Cond. Scotland Northern Ireland East Anglia East Midlands Greater London Rest of the world North North west South east South west West Midlands Yorkshire

13.03 9.69 1.47 6.29 1.04 4.22 5.81 4.05 3.66 2.20 2.51 7.68

0.58 17.97 0.19 1.90 3.64 0.04 7.35 0.01 1.99 1.67 2.53 3.86

13.99 7.67 1.67 4.46 2.44 4.23 1.13 4.06 1.78 0.61 0.11 3.97

1.44 2.02 0.40 0.77 0.54 0.18 0.58 0.20 0.47 0.35 0.45 1.09

0.65 5.26 0.13 0.33 0.94 0.25 1.60 0.24 0.60 0.45 0.65 0.84

2.30 1.17 0.07 0.91 1.15 0.57 1.11 0.07 0.76 0.14 0.04 0.71

10.78 1.61 2.35 3.66 2.44 4.25 1.73 5.18 4.12 0.53 3.03 5.46

3.47 4.39 0.29 1.58 3.07 3.07 0.57 0.64 0.62 0.09 1.29 0.69

Region of institution North east Yorkshire & Humber North west East Midlands Eastern South east London South west West Midlands Scotland Wales Northern Ireland

2.00 5.11 5.90 5.15 3.60 5.48 4.60 0.39 4.89 15.16 5.29 10.69

13.88 3.56 0.23 2.29 3.84 6.17 9.14 5.84 7.41 1.56 2.56 22.02

10.99 8.57 6.13 2.96 0.05 0.42 4.02 6.04 2.11 17.22 2.90 10.63

0.79 0.01 0.11 0.42 2.51 1.30 1.38 0.09 0.85 1.52 1.27 2.04

4.44 0.89 0.30 0.44 0.60 1.54 2.98 1.26 1.94 0.71 0.48 5.38

2.30 3.24 0.38 1.58 0.17 0.31 3.65 1.34 0.87 2.48 0.57 1.29

2.73 6.42 10.72 2.11 10.12 1.98 4.97 5.24 4.34 13.17 0.05 0.78

1.93 0.54 1.53 1.77 5.22 4.16 9.22 3.71 0.80 3.87 3.80 5.08

Note: Columns 1
3 report the value of t-test of mean differences of each covariate between observation in tier i and observations in the other two tiers of the institutional quality distribution. Columns 4
6 report the t-tests after blocking on GPS using five blocks. The mean differences are computed within each block and then the weighted average over the five blocks is reported for each tier. Column 7 reports t-test on the coefficient of treatment in a regression of covariate k. Column 8 reports the same t-test when conditioning on the predicted treatment value.

parental background, school type and for medical graduates. We fare less well in balancing on regions of origin and region of institution with nine tests being positive. Overall, we can still be pretty satisfied of the balancing, out of the 297 tests conducted only 15 are significant, which is exactly the number of false positive expected with a 95% confidence interval. Imai et al. (2008) suggest that such a blocking test may be subject to balance test fallacy. As an alternative balancing test, we also conduct the

282

ARNAUD CHEVALIER

Imai and van Dyk (2004) test. In column 7, we report for each covariate, the t-tests on the log treatment coefficient in an unconditional regression. Those are large, indicating that covariates and quality are correlated. In column 8, we report the same t-test when we condition on the predicted value of the treatment. Again we observe a large reduction in the value of the t-test but some remain above the critical value for significance. These concern the top of the ability distribution, school type, some subjects and the regional dummies. Kluve et al. (2012) recommend to run a similar test but conditioning on GPS values at the 25th, 50th and 75th percentile of the quality distribution instead. This test broadly confirms the Imai and van Dyk and is not reported here. We have conducted these three tests for various specifications, including interactions of the various controls, but the conclusions are always the same. The blocking test suggests that after conditioning the samples are balanced while the other two tests highlights some balancing issues. Eyeballing the means and standard errors in the blocking test, it does not appear that the reduction in the t-tests when blocking is due to an inflation of the standard error. Since the blocking test assess the balancing over a finer range of the common support, we are satisfied that after conditioning on GPS the sample is balanced. To test for common support, we follow the procedure highlighted in Kluve et al. (2012). We split the sample into three tercile as previously. We then evaluate the GPS at the median quality value of group g for all observations. We replicate this procedure three times, once for each tercile. Fig. 6 reports the distribution of GPS estimated at the median value of group g, for group g versus not group g. For each group, the GPS distributions overlap indicating common support. Note that the common support is thinner for the lowest and highest quality groups. Fig. 7A reports the marginal effects of (log) institution quality on earnings throughout the distribution of treatment. The quality effect is convex and positive but for a large section in the lower part of the quality distribution, not actually significantly different from 0. For institutions above the median quality, there are some positive returns to quality. Note that the return is increasing with quality, so that the highest quality institutions generate the highest return to quality. This may be better appreciated in Fig. 7B which reports the expected log earnings over the quality distribution. The expected wage differentials are only significantly larger for attending an institution in the top quartile of the quality distribution. The predicted wage differential between a graduate from the median fourth quality quartile and one from the third quartile is 7%. In Table 5, we provide evidence that these results are robust. First, we eliminate institutions and subjects for which the selection process includes

283

(a)

5

Does Higher Education Quality Matter in the UK?

Control groups

3 2 0

1

Percentage

4

Treatment Group

0

0.2

0.4

0.6

(b)

5

GPS

Control groups

3 2 0

1

Percentage

4

Treatment Group

(c)

0.2

GPS

0.4

Control groups

0

2

4

Treatment Group

Percentage

0.6

6

0

0

0.2

0.4

0.6

GPS

Fig. 6. Distribution of GPS
Evaluated at the Median Value of the Treatment for Each Quality Treatment Tercile. Note: Distribution of GPS estimated at the median university quality for the lowest GPS tercile (a), medium tercile (b) and upper tercile (c).

284

ARNAUD CHEVALIER

(a)

0.05 –0.05

0

Dose response

0.1

.15

Continuous Treatment Matching

–1.1

–0.6

–0.1

0.4

0.9

1.4

log Normalised Treatment (b)

10.4 10.2 10 9.8 9.6

Predicted Log wage

10.6

Continuous Treatment Matching

–1.1

–0.6

–0.1

0.4

0.9

1.4

log Normalised Treatment

Fig. 7. Estimated Marginal Effects of Institution Quality on Earning. Continuous Propensity Score Matching. (a) Dose Response to Log Quality. (b) Predicted Wage to Log Quality. Note: Estimates of the marginal effect of institution quality on earnings from the generalised propensity score matching. Dash lines represent the 95% confidence interval obtained from 200 replications bootstrap.

285

Does Higher Education Quality Matter in the UK?

Table 5. Excluding Institutions with Interview Process

Robustness Checks. Quality Exclude Graduate Prospect

Quality: Research, Expenditure, Student Staff Ratio

0.020 0.028 (0.016) (0.020) 0.049** 0.036 (0.020) (0.022) 0.074*** 0.063** (0.020) (0.029)

0.022 (0.019) 0.042** (0.015) 0.052*** (0.019)

0.025* (0.014) 0.048** (0.015) 0.056*** (0.019)

Propensity score matching Quality 2 vs. 0.006 Quality 1 (0.021) Quality 3 vs. 0.044** Quality 2 (0.020) Quality 4 vs. 0.039** Quality 3 (0.019)

0.013 (0.022) 0.051** (0.018) 0.033 (0.023)

0.022 (0.025) 0.023 (0.020) 0.047** (0.018)

0.016 (0.021) 0.030** (0.015) 0.033** (0.019)

Continuous measure OLS estimates Quality score 0.032*** (0.007) Quality score 0.010* square (0.006)

0.034*** 0.025** (0.009) (0.012) 0.002 0.013* (0.005) (0.008)

0.028*** (0.008) 0.008* (0.005)

0.025*** (0.008) 0.005 (0.005)

Continuous treatment matching E[W/T = 1] 9.942 (0.009) E[W/T = 2] 9.924 (0.006) E[W/T = 3] 9.976 (0.023) E[W/T = 4] 10.044 (0.032)

9.890 (0.036) 9.892 (0.023) 9.964 (0.030) 10.025 (0.040)

9.892 (0.052) 9.902 (0.022) 9.993 (0.008) 10.064 (0.009)

9.915 (0.036) 9.918 (0.015) 9.984 (0.014) 10.086 (0.011)

Discrete measure OLS estimates Quality 2 Quality 3 Quality 4

0.026* (0.014) 0.044*** (0.016) 0.065*** (0.019)

Female

Male

-0.023 (0.044) -0.019 (0.037) 0.026 (0.031)

9.956 (0.015) 9.953 (0.004) 10.024 (0.004) 10.099 (0.011)

Note: OLS estimates are based on the full specification details of which are available in Table 2. Propensity score matching is estimated using the full set of controls, details of which are found in Table 3. Continuous treatment matching is based on a propensity score estimated with the full set of parameters as explained in Fig. 5. Standard errors are based on bootstrapping with 500 replications. ***, **, * denote statistical significance at the 1%, 5% and 10% statistical level, respectively.

286

ARNAUD CHEVALIER

an interview (Oxford, Cambridge) or an additional test (Medical schools, Art schools) as this makes the assumption of selection on observable characteristics less palatable. This selection eliminates both high-quality institutions from which there is a high premium to quality, and arts graduates which typically have low earnings. As such, the quality effect remains almost identical to those obtained on the full sample. Second, we assess whether there are gender differences in the returns to institution quality. The gender wage gap reaches £2,000 or nearly 10%. However, there are no significant differences in the returns to quality by gender. In fact, the CTM predicted wages suggest that the gender wage gap may open up for highquality institution graduate. The gender wage gap in predicted wage is around 6% for graduates from the bottom three quartiles but reaches 7.5% for graduates from top quartile institution. Finally, the last two columns assess the sensitivity of our results to alternative measures of quality. First, we eliminate graduate prospects as this is measuring a post-university outcome. The second alternative is to additionally drop the research quality score, as this may not be directly related to teaching quality. The estimates tend to be somehow smaller, confirming that simpler measure of quality tend to reduce the returns to quality (Black & Smith, 2004). Overall, we are satisfied that our estimates are robust. Additionally, we have experimented with different transformation of the quality measure but the general conclusions remain the same.

DISCUSSION Using a wide array of estimators, we consistently find that there is a large amount of sorting to institution, the estimated quality effects are halved when including students’ characteristics (mostly ability). Secondly, institutions quality is associated with positive financial returns for graduates. While there are some differences on the size and the distribution of the effect, the overwhelming conclusion is that the quality effect is non-linear and accrues mostly to graduates from the highest quality institutions or with the best earnings credentials. The estimated returns to graduating from a top quartile institution as oppose to a third quartile one, range from 3.5% to 5.5% when estimated by OLS or PSM. Our favoured estimates, GPS, are even larger reaching 7% for moving from the median quality of the third quartile to the median value of the fourth quartile.

Does Higher Education Quality Matter in the UK?

287

Graduates from a third quartile institution earn on average £22,785 per annum three years after graduation. We can compute the lifetime premium that would have been associated with these students graduating from a fourth quartile quality institution instead. On average the difference in quality score between these institutions is just short of one standard deviation. Our favoured estimated effect in this quality range is 7% (CTM). Assuming a growth rate of 2% over the lifetime, a 40-year-long participation to the labour force, a discount rate of 3.5% and that the returns to quality are constant over the lifetime, the differences in the present value of the lifetime earnings at 18 between a would-be graduate at a third quartile institution and one at a fourth quartile institution reaches £40,000. Since tuition fees are (almost) identical between all institutions
at the cap value stated by the government
there is no potential arbitrage for students between a high-quality institution and a cheaper one, so £40,000 is a net gain from attending a top quality institution. The current admission system is based on predicted grades and trap applicants in their initial choice. As such an applicant whose grades were wrongly predicted and could not get access to a top quality institution suffers from a large financial penalty over her lifetime.

CONCLUSION We show that currently in the UK, there is a wage premium to university quality. This premium is non-linear. There is no significant quality effect for institutions in the bottom half of the quality distribution but graduates from the most prestigious institutions earn 7% more than graduates from institutions in the third quality quartile. These estimates are robust to an array of methods, all relying on selection on observables. Compared to other UK estimates, such as Broecke (2012) which accounts for selection on unobservable, our results are much lower, maybe because his sample is composed for higher ability individuals than the one used for this analysis. While smaller, the returns to quality still lead to some important differences in lifetime earnings. Policymakers may thus worry that the current system of fixed price between all institutions can be considered unfair towards students attending the lower quality institutions. The allocation system could also be revised to either be based on the final secondary education exam, rather than its prediction, or by allowing successful applicants to trade-up if their exam results were better than expected. A remaining question is

288

ARNAUD CHEVALIER

whether this premium stems from an increase in human capital, peers effects or signalling.

NOTES 1. The Economist (2014) suggests that some US colleges have even negative financial returns. 2. Bowman and Mehay (2002) find a positive effect of attending private and higher rated institutions on appraisal and promotion, consistent with the earnings effects estimated elsewhere. Robst (1995) also reports that graduates from higher ranked institutions are less likely to experience over-education. Smith (2013) using twins also reports that quality reduces drop-out. 3. Distinguishing between the mechanisms is relevant for policy. Only if the premium is generated by an increase in human capital does the economy gain from institutional quality, and there is then some possible justification for subsidising students to attend higher quality institution. If the returns are only due to signalling or networking, then the returns to quality are only private. The data available to us do not offer opportunities to test these hypotheses convincingly. Hershbein (2013) proposes a test of the origin of the returns to attend more selective institutions and argues that selectivity provides a positive signal to employers. Jawaguchi and Ma (2008) provide some evidence supporting human capital. 4. There are a few exception to this general rule, which we explain later. 5. Some institutions, typically Oxford and Cambridge universities, as well as medical, dentistry and veterinary schools also rely on an interview to assess the suitability of the candidates. We conduct robustness checks excluding these graduates. 6. Smith (2013) found that attending an institution with a 100 point higher median SAT score increases the probability of graduating by 5%. This estimate is similar when estimating a between twins model. 7. Lindahl and Regne´r (2005) use a between-sibling estimates to reduce the bias due to unobserved characteristics and estimate that in Sweden OLS estimates are twice as large as the within family estimates, suggesting a large bias in regressions not correcting for selectivity. 8. Eliasson (2007) uses the same dichotomy as a proxy for institution quality in Sweden. Propensity score estimates reveal no effect of institution quality on earnings four to nine years after graduation. 9. The Good University Guide is one of the providers of ranking of universities. Rather than using its ranking, we only use the raw variables which can be obtained from: http://www.thegooduniversityguide.org.uk. Forty-eight institutions not reporting all the covariates used to compute the quality score are thus excluded. They are mostly small specialised colleges and represent 876 observations. 10. www.russellgroup.ac.uk/research
accessed on 25th July 2014. 11. Students applying to medical, dentistry and veterinary schools, as well as candidates to Cambridge and Oxford universities, who can apply only to one or the other institution but not both, can only state four choices. These applications need to be received earlier on in the cycle, typically in the October of the year preceding

Does Higher Education Quality Matter in the UK?

289

their entry to higher education. Since 2008, the number of choice has been further restricted to five. There is also a slightly different application process for performing arts courses. 12. Applicants can also apply with other qualification than A-levels. UCAS creates a score to harmonise grades at different type of qualifications. For the remaining of the paper, we will refer to the score at the exams warranting entrance at the university as the entry score. 13. The possibility to continue searching after having an offer confirmed will be offered to applicants in 2014. 14. The compulsory fields on the UCAS form are name, gender, age, address, country of birth, nationality, financing, disability, ethnicity, occupational background of parental figure if under 21 and previous school attended, which we also obtain when linking the survey to administrative data. The UCAS form also includes previous qualification, references
typically for the pupils’ teacher
the student personal statement and their predicted exam results if still currently in high school. This information is not available in our dataset, and instead we have the realised score. We do not know which institutions a candidate applied to nor what their decision was. 15. Gibbons and Chevalier (2008) find small differences between teacher assessments and realised scores at age 16 especially for pupils at the extremes of the ability distribution, but these differences have no impact on subsequent pupil outcomes. Hayward, Sturdy, and James (2005) report ‘reasonable accuracy’ (+/− one grade) at A-levels and no impact of the error on higher education participation. 16. It was not possible to control for the region of institution since for some regions, there is only one institution in a given quality quartile. 17. A cubic function of institution quality was also estimated but neither the quadratic nor the cubic terms were ever statistically significant. 18. We report results when a log transformation of the quality score is used since the log transformation is closer to a normal distribution which we use to compute the GPS. Estimates based on the untransformed quality score were also computed and do not substantially differ from those presented.

ACKNOWLEDGEMENT This paper was drafted during a placement at the Department for Innovation, Universities and Skills. The views represented in this manuscripts are the author’s and do not represent the view of the DIUS. Financial support from the ESRC for their placement scheme is also gratefully acknowledged (RES-173-27-0040). I thank Stijn Broecke, Victor Lavy, Tarja Viitanen and participants at seminars at DIUS and Oxford University for comments on earlier drafts, as well as three anonymous referees and Kostas Tatsiramos for detailed comments that have substantially improved the manuscript.

290

ARNAUD CHEVALIER

REFERENCES Behrman, J., Rosenzweig, M., & Taubman, P. (1996). College choice and wages: Estimates using data on female twins. Review of Economics and Statistics, 78, 672
685. Birch, E. R., Li, I., & Miller, P. W. (2009). The influences of institution attended and field of study on graduates’ starting salaries. Australian Economic Review, 42(1), 42
63. Black, D., & Smith, J. (2004). How robust is the evidence on the effects of college quality? Evidence from matching. Journal of Econometrics, 121, 99
124. Black, D., & Smith, J. (2006). Estimating the returns to college quality with multiple proxies for quality. Journal of Labor Economics, 24, 701
729. Bowman, W., & Mehay, S. (2002). College quality and employee job performance: Evidence from naval officers. Industrial and Labor Relations Review, 55, 700
714. Brewer, D., Eide, E., & Ehrenberg, R. (1999). Does it pay to attend an elite private college? Journal of Human Resources, 33, 104
123. Broecke, S. (2012). University selectivity and earnings: Evidence from UK data on applications and admissions to university. Economics of Education Review, 31(3), 96
107. Chevalier, A., & Conlon, G. (2003). Does it pay to attend a prestigious university? London School of Economics, CEE, DP 33. Chevalier, A., Gibbon, S., Thorpe, A., Snell, M., & Hoskins, S. (2007). Students’ academic self-perception, IZA, DP 3031. Dale, S., & Krueger, A. (2002). Estimating the pay off to attending a more selective college: An application of selection on observables and unobservables. Quarterly Journal of Economics, 117, 1491
1527. Dale, S. B., & Krueger, A. B. (2014). Estimating the effects of college characteristics over the career using administrative earnings data. Journal of Human Resources, 49(2), 323
358. de Luna, X., & Lundin, M. (2014). Sensitivity analysis of the unconfoundedness assumption with an application to an evaluation of college choice effects on earnings. Journal of Applied Statistics, 1
18 (ahead-of-print). Economist. (2014). Is college worth it? The Economist, April 05, 2014. Eliasson, K. (2007). How robust is the evidence on the returns to college choice? Results using Swedish administrative data. Umea University, Mimeo. Furnham, A. (2001). Self-estimates of intelligence: Culture and gender difference in self and other estimates of both general (g) and multiple intelligences. Personality and Individual Differences, 31, 1381
1405. Gibbons, S., & Chevalier, A. (2008). Assessment and age 16 + participation. Research Papers in Education, 23(2), 113
123. Hayward, G., Sturdy, S., & James, S. (2005). Estimating the reliability of predicted grades, UCAS, UCAS report. Hershbein, B. J. (2013). Worker signals among new college graduates: The role of selectivity and GPA. Upjohn Institute, DP 13
190. Hirano, K., & Imbens, G. W. (2004). The propensity score with continuous treatments. Applied Bayesian Modeling and Causal Inference from Incomplete-Data Perspectives, 226164, 73
84. Hoekstra, M. (2009). The effect of attending the flagship state university on earnings: A discontinuity-based approach. The Review of Economics and Statistics, 91(4), 717
724.

Does Higher Education Quality Matter in the UK?

291

Hoxby, C. M. (2009). The changing selectivity of American colleges. Journal of Economic Perspectives, 23(4), 95
118. Hussain, I., McNally, S., & Telhaj, S. (2009). University quality and graduate wages in the UK? IZA, DP 4043. Imai, K., King, G., & Stuart, E. (2008). Misunderstandings between experimentalists and observationalists about causal inference. Journal of the Royal Statistical Society, series A, 171, 481
502. Imai, K., & van Dyk, D. A. (2004). Causal inference with general treatment regimes: Generalizing the propensity score. Journal of the American Statistical Association, 99, 854
866. James, E., Alsalam, N., Conaty, J. C., & To, D.-L. (1989). College quality and future earnings: Where should you send your child to college? American Economic Review, 79(2), 247
252. Jawaguchi, D., & Ma, W. (2008). The causal effect of graduating from a top university on promotion: Evidence from the university of Tokyo’s 1969 admission freeze. Economics of Education Review, 27, 184
196. Kluve, J., Schneider, H., Uhlendorff, A., & Zhao, Z. (2012). Evaluating continuous training programs using the generalized propensity score. Journal of the Royal Statistical Society, series A, 175, 587
617. Koenker, R., & Bassett, G. (1978). Regression quantiles. Econometrica, 46, 33
50. Lindahl, L., & Regne´r, H. (2005). College choice and subsequent earnings: Results using Swedish sibling data. Scandinavian Journal of Economics, 107, 437
457. Long, M. (2008). College quality and early adult outcomes. Economics of Education Review, 27, 588
602. Nakamuro, M., & Inui, T. (2013). The returns to college quality in Japan: Does your college choice affect your earnings? ESRI Discussion Paper Series. Robst, J. (1995). College quality and overeducation. Economics of Education Review, 14, 221
228. Rosenbaum, P., & Rubin, D. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70, 41
55. Saavedra, J. E. (2008). The returns to college quality: A regression discontinuity analysis. Unpublished Harvard Mimeo. Smith, J. (2013). Ova and out: Using twins to estimate the educational returns to attending a selective college. Economics of Education Review, 36, 166
180. Suhonen, T. (2012). University Choice and Subsequent Earnings in Finland: Estimates for Groups of Majors. Unpublished Mimeo, University of Jyva¨skyla¨. Tipping, S., & Taylor, R. (2007). Destination of leavers from higher education longitudinal survey 2002/3 cohort: Assessment of robustness and fitness for purpose. Higher Education Statistical Agency. Universities UK. (2011). Higher education in facts and figures.

292

ARNAUD CHEVALIER

APPENDIX Table A1.

Sample Selection.

Selection Criteria
Applied Incrementally

Number of Observations

Original sample First degree only Age on graduation [19, 25] Not special entry student Measure of institution quality Earnings non-missing FT employee Earnings < £60,000

19,979 11,866 9,850 9,738 8,500 7,508 7,007 6,986

Note: Longitudinal destination of leavers from higher education (2002/2003).

Table A2.

Distribution of Subject by Institutional Quality.

Subject

Q1

Q2

Q3

Q4

Total

Medicine and dentistry Sub. allied to medicine Biology, vet, agri. Physics Mathematics Engineering and tech. Architecture and planning Social studies Law Business and administration Communication Linguistic and classics Language and literature History and philosophy Creative arts Education Sport science Psychology IT Economics Finance & accounting Mixed no science Mixed 45
55 science Mixed 100% science Other

0.00 7.12 2.83 3.46 0.80 5.60 1.47 4.58 2.79 18.31 4.98 1.66 0.44 1.80 10.72 3.95 4.62 3.72 8.32 0.23 2.55 3.94 4.74 0.41 0.98

0.00 7.39 4.98 5.17 0.82 6.76 3.84 6.27 2.08 19.53 1.90 1.92 0.94 0.95 8.22 1.76 1.35 3.27 7.82 0.55 2.55 2.65 7.19 1.10 0.99

2.83 5.56 6.10 5.69 3.81 7.71 1.12 7.76 6.06 6.81 1.15 5.13 2.60 6.17 4.37 1.02 1.28 3.62 5.81 2.48 1.71 3.38 5.98 1.40 0.45

9.52 4.96 6.03 6.43 5.49 9.02 0.97 5.60 5.81 3.05 0.51 5.07 6.07 7.97 1.30 1.66 0.14 2.23 3.20 3.91 0.17 2.90 5.14 2.65 0.23

3.04 6.26 4.97 5.18 2.71 7.26 1.85 6.05 4.18 11.97 2.15 3.43 2.49 4.20 6.19 2.11 1.87 3.22 6.31 1.78 1.75 3.22 5.76 1.38 0.67

Observations

1783

1732

1758

1713

6986

Note: For each column, a cell reports the fraction of individuals in that subject. The total row reports the number of individuals.

BUSINESS VISITS AND THE QUEST FOR EXTERNAL KNOWLEDGE Massimiliano Tani ABSTRACT This paper analyses the role of face-to-face interactions between employees of different firms meeting during work-related visits in fostering skills and generating new productive knowledge. The analysis is based on primary data collected from 1,982 business visitors to/from Australia. The results suggest that face-to-face interacting can be an effective mechanism to enhance skill formation, as it improves the stock of useful knowledge and offers opportunities to learning to both visiting and visited parties. Keywords: External knowledge; face-to-face interactions; international business visits JEL classifications: F2; J6

INTRODUCTION Rising wage inequality between highly- and less-educated workers in many economies has prompted a new interest in the labour market for skills and the design of policies fostering them. Skills formation is analysed by a large

Factors Affecting Worker Well-Being: The Impact of Change in the Labor Market Research in Labor Economics, Volume 40, 293
324 Copyright r 2014 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0147-9121/doi:10.1108/S0147-912120140000040011

293

294

MASSIMILIANO TANI

literature studying the role of educational institutions (Becker, 1964; Card & Krueger, 1992), families (Altonji & Dunn, 1996; Cawley, Heckman, & Vytlacil, 2000) and firms, predominantly insofar as formal training is concerned (Couch, 1992; Heckman, 2000; Topel, 1999). Research on the sources of innovation, however, suggests that skill formation and learning within firms occur in a variety of ways besides Arrow’s (1962) learning-by-doing. Such variety underpins different trajectories of technological advancement which enhance productivity and growth via horizontal or vertical product differentiation, various types of yield improvements and changes in the inputs or the organisation of production (Malerba, 1992). Understanding the mechanisms through which firms contribute to skill formation acquires more prominence as globalisation and technological change further heighten the international competition for highly skilled workers. This paper analyses one possible mechanism of skill formation: namely, labour mobility in the form of face-to-face interactions between employees of different firms meeting during work-related visits. In particular, I study whether interacting through work-related travel results in the generation of new productive knowledge and a more skilled workforce (Lundvall, 1988). The underlying assumption throughout the analysis is that visiting is a choice variable with well-defined financial and time trade-offs. An academic analogy may clarify the topic of this study. It is conventional for economics’ departments to have a weekly seminar programme where visitors from other universities present their work: these visits neither change the number of researchers in the visited and visiting institutions, nor their organisational allocation or functional unit. Yet, listening and talking to speakers may enrich the stock of knowledge of presenter and attendees, spark new ideas and lead to observable outcomes, like collaborations and sharing data and/or techniques (Hamermesh, 2006). Different speakers presenting their chosen topics are also likely to have a different impact on the audience, if at all. Does having a seminar programme enhance the stock of knowledge of guest and hosting faculty, generate new ideas and other academic output relative to not having a seminar programme? Do business-related visits generate new productive knowledge and a more skilled workforce? To address this question, I analyse primary data collected from 1,982 people travelling to or from Australia1 for work. The survey asked for background information on the visitor and his/her job, the employer and the nature work-related travel. An open-ended question posed the counterfactual about what would occur if the trip did not take place. Answers explicitly referring to learning and accessing knowledge such as ‘would not

Business Visits and the Quest for External Knowledge

295

be able to apply the latest techniques in transplant immunology’ are used as a proxy for visits generating new productive knowledge. Some examples and other reasons for travelling, and copy of the questionnaire are reported in Table A1 in the Appendix. Although subjective, answers to the counterfactual capture whether short-term mobility is itself an activity that can raise the dynamic efficiency of a firm, via enhanced embodied (e.g. cognitive) skills amongst visitors and visited, and improved organisational capabilities amongst employers. Knowledge creation is regarded as a source of sustainable economic prosperity by literatures on innovation and economic growth. However, existing studies tend to be either preoccupied with the static efficiency of knowledge-producing inputs, especially labour, or the determinants of observable measures of knowledge creation, such as patents, which account for only a tiny fraction of the stock of knowledge used for producing output (Noorderhaven & Harzing, 2009). This paper uses a novel direct measure of knowledge and skill production, which is identified by the answers to the counterfactual. The empirical analysis consists of the simultaneous estimation of the determinants of knowledge-enhancing visits and the type of relationship between visiting and visited organisations. As this approach controls for the likely endogeneity linking the purpose (e.g. to gather knowledge) and the nature of the visit carried out (e.g. attend a conference), it supports the causal interpretation of the estimates obtained. The results suggest that face-to-face interacting through business visits is an effective mechanism to enhance skill formation in the workplace. Not only do work-related visits increase employees’ ability to learn and the stock of useful knowledge in an organisation, but they do not require changes in the overall organisational headcount, the job allocation of employees, and their level of education or training. Interacting via business travel appears prima facie a relatively simple mechanism through which organisations can foster skills amongst their workforce. Such mechanism can be a suitable alternative for competing internationally for skilled labour, especially for organisations and nations that are disadvantaged by resources, geographic location or other factors. The results also support that interacting can be a strategic choice to foster skills. The opportunity cost of interacting via business travel varies across visiting and visited organisations, and interacting with different counterparties does not lead to new productive knowledge at a fixed rate. Not understanding such opportunity cost risks viewing business visits as an easily disposable expenditure at time of financial difficulties. Yet, simplistic

296

MASSIMILIANO TANI

cuts to travel budgets may undermine access to external knowledge and relevant opportunities for learning and upskilling for an organisation and its employees. Moreover, at the national economy level, treating work-related visits as consumption expenditure does not recognise that they are often a form of investment to access external knowledge, and they foster skills formation. While the social benefits associated with interacting via work-related visits are indirectly recognised by the possibility to deduct work-related travel costs from taxable income, it is somewhat paradoxical that governments tax work-related mobility but promote other knowledge-producing activities like R&D with extra financial and other incentives. This status quo reflects a general lack of relevant data on business visits and their characteristics at individual and firm level. The primary data presented in this paper call for more data collection and research to understand the net social benefits of work-related travel, and the consistency of government policies aimed at raised national productivity. The rest of the paper is organised as follows. The section ‘Interactions through Business Visits and the Access to Knowledge’reviews the relevant literature. The section ‘Primary Data: The Airport Survey’presents the data. The section ‘The Determinants of Knowledge-Related IBTs Flows’discusses the empirical approach and the results. The section ‘Implications for Policy and Conclusions’concludes.

INTERACTIONS THROUGH BUSINESS VISITS AND THE ACCESS TO KNOWLEDGE It is a fact that firms develop new products and applications with knowledge that typically exists or is originally produced outside, rather than within, the successful innovator (Mansfield, 1968; March & Simon, 1958; Mueller, 1962; Rosenberg & Steinmueller, 1988). Firms can, therefore, gain an edge over their competitors by recognising useful ‘external’ knowledge and exploiting it. This ability, or absorptive capacity (Cohen & Levinthal, 1989) or dynamic capability (Teece, Pisano, & Shuen, 1997), is fostered by carrying out R&D, engaging in production, investing in advanced technical training and forming collaborations. These activities facilitate the creation of novel linkages between what is already known and the new information acquired,2 hence expanding problem-solving capabilities and skills, and the efficient absorption of new information.

Business Visits and the Quest for External Knowledge

297

‘Knowledge’, however, is not uniformly distributed in space. In its traditional characterisation, it includes disembodied features, like data and information, which make it codifiable and replicable through blueprints, as well as embodied features, which are inextricably connected with an individual’s skills and experience (Polanyi, 1966). While disembodied knowledge is almost a ‘ubiquitous’ resource equally available to each competitor, embodied knowledge is less replicable and heterogeneous, and typically ‘sticks’ to individuals and the physical spaces where they work (von Hippel, 1987). The literature has explored the role of certain categories of workers in accessing effectively external knowledge including expatriates (Collings, Scullion, & Morley, 2007), managers of subsidiaries (Riusala & Suutari, 2004), employees temporarily working for another employer in the context of a co-operation (Franco & Filson, 2000; Zellner, 2003) and business visitors (Andersen & Dalgaard, 2011; Dowrick & Rogers, 1995; Dowrick & Tani, 2011; Hovhannisyan & Keller, 2011). These studies, however, as well as those focusing on co-location as a strategy to access external knowledge (Bathelt, Malmberg, & Maskell, 2004; Florida, 2002; Howells, 2002; Torre & Rallet, 2005), suffer from the general lack of individual-level data. As a result, they tend to treat the link between labour mobility, learning and the generation of new knowledge as a black box. Most studies use observable but narrow measures of productive knowledge, such as the number of new patents or citations (e.g. Hovhannisyan & Keller, 2011), or inputs such as R&D expenditures and number of employees engaged in research activities. But patents do not account for previous attempts, past failures and other activities that nevertheless enable firms and their employees to acquire experience instrumental in turning a knowledge breakthrough into a commercially successful product or service, even after a substantial time lag. Moreover, they exclude knowledge that is not copyrighted but that still contributes positively to productivity like informal exchanges (Noorderhaven & Harzing, 2009), casual meetings or simply being located in a ‘buzzing’ place (Gertler, 2003; Storper & Venables, 2004). Other studies instead use multifactor productivity as a broader measure of knowledge and skills advancement. While useful insofar as assessing the existence of an aggregate productivity effect of work-related visits at sectoral (Dowrick & Tani, 2011) or national level (Andersen & Dalgaard, 2011), multifactor productivity does not provide information about its underlying determinants (e.g. does it increase across all or only some firms?), limiting the usefulness of this literature for policy design.

298

MASSIMILIANO TANI

Addressing the lack of relevant individual-level information covering work-related visits and the production of useful knowledge motivates the collection of primary microeconomic data, as is later discussed. The logical nexus between interacting and fostering skills and new knowledge builds on research studying why face-to-face interacting thrives despite the availability of cheaper and faster forms of communication. This literature suggests that this occurs because face-to-face interacting is the most effective form of inter-personal communication, as it makes participants decide immediately whether to trust each other (Gambetta, 1988; Storper & Venables, 2004). If mutual trust is established, then reciprocal understanding and co-operation behaviours arise, as the transaction costs and uncertainty associated with sharing or exploring new knowledge decrease. This facilitates exchanges of know-how and experiences (Amin & Cohendet, 2004; Hansen, 1999), promotes individual and organisational learning and creates ‘social capital’ and networks (Burt, 1997; Portes, 1998). Once trust is established, the range of communication means used can expand, though interacting face-to-face seems to remain the preferred means to co-ordinate interdependent organisational activities (Mu, Peng, & Love, 2008). Building upon this literature, the main contribution of this paper is the analysis of detailed individual-level information collected through a survey of almost 2,000 business visitors.

PRIMARY DATA: THE AIRPORT SURVEY The survey (in the Appendix) was carried out in November 2006 across four of Australia’s international airports: Sydney, Melbourne, Brisbane and Adelaide. These account for over 80% of the volume of international travel from and to Australia. Travellers were approached at the boarding gates after immigration and passport controls by licensed surveyors. Respondents were initially asked whether they were travelling for work-related purposes. The interview continued only if the response was affirmative. Overall, 1,016 Australian residents and 966 foreign residents returning home were interviewed. Non-response was minimal (less than 5% of those approached), and only one employee per organisation was interviewed. Age, gender, occupation and country of origin of the respondents were compared with a second random sample of departure and landing cards of business visitors during the same period, which was carried out by Australia’s then Department of Immigration and Citizenship (DIAC), and

Business Visits and the Quest for External Knowledge

299

the population of business visitors in the same year. The characteristics of the respondents of the airport survey resulted similar to those of the people sampled by DIAC and the overall population (Table A2 in the Appendix). The survey revealed that business visitors share similar personal and occupational characteristics (Table 1). They are mostly males, aged between 35 and 44, professionals or managers employed by either a multinational company or a small
medium-sized firm with less than 100 employees. The functional units most commonly engaged in business visits are production, strategy and training. Less relevant is the presence of those working in sales and marketing. Respondents are mostly specialist managers, IT professionals and scientists and engineers. There are also several health specialists working for government and NGOs, typically providing medical and other health relief to areas in less developed countries. The survey shows that a variety of employers use international business visits to interact in each of the four major categories identified (Table 2): these are visits to (i) other parts of the same organisation or affiliate, (ii) companies within the same supply chain, (iii) new customers and suppliers and (iv) conferences and trade fairs. Finding that each type of employer uses visits across these interaction types suggests that corporations, governments and NGOs contemporaneously engage in a variety of interactions with the external environment. As such, they face choices about which one to prioritise or be strategic about, as each employer has limited financial, time, and human resources. The identification of ‘knowledge’ relies on the answer to an open-ended question about the counterfactual to what would occur to the respondent’s employer if the visit did not take place (Q19). Respondents were not prompted in any way before formulating their answer; they were not told to give a single reason, though they were asked to highlight the most important consequence, or were allowed to give more than one; they were, however, invited to rank the most important reason if they gave more than one. Their answers could be classified in a handful of broad topics, as illustrated by the sample in Table A1. The author and two other researchers independently organised the responses into five mutually exclusive categories: (1) certain or potential financial losses, as in ‘we would not be able to generate enough revenues in this region’; (2) separation from the industry’s best practice and latest developments, as in ‘would no longer be efficient as we exchange ideas with hospitals’; (3) break-ups, or diminished strength, of an existing relationship with a customer or supplier, as in ‘would have a bad reputation’; (4) other effects, such as legal liabilities

300

MASSIMILIANO TANI

Table 1. Code

Summary Statistics.

Mean

Std. Dev.

Description

EFFECT_K WITHIN2

.293 .439

.455 .496

WITHIN3 WITHIN4 EMP_1 EMP_2 EMP_3 EMP_4 OWNER FNC_1 FNC_2 FNC_3 FNC_4 FNC_5 FNC_6 NR_TRIPS1 NR_TRIPS2 NR_TRIPS3 NR_TRIPS4 LENGTH_1 LENGTH_2 LENGTH_3 LENGTH_4 ONGOING GENDER EDU_1 EDU_2 EDU_3 EDU_4 AGE_1 AGE_2 AGE_3 AGE_4 DEST_1 DEST_2 DEST_3 DEST_10

.298 .362 .181 .371 .115 .324 .115 .237 .220 .204 .105 .138 .089 .127 .426 .166 .283 .366 .365 .161 .100 .200 .126 .199 .435 .255 .109 .199 .324 .304 .170 .099 .306 .195 .394

.457 .480 .385 .483 .319 .468 .319 .426 .414 .403 .306 .345 .286 .333 .495 .372 .451 .482 .481 .367 .300 .400 .333 .399 .496 .436 .312 .400 .468 .460 .376 .298 .299 .397 .489

NO_S&E SCIENTIST ENGINEER RESIDENCE

.501 .213 .287 .487

.500 .410 .452 .500

Dependent variable 1: No travel = no knowledge Dependent variable: 1 = to supply chain; 0 = within MNC Dependent variable: 1 = to new; 0 = within MNC Dependent variable: 1 = to conf./fair; 0 = within MNC Employed by government, NGO, university Employed by MNC (reference group) Employed by large domestic firm Employed by SME Company owner or self-employed Work in strategy department, or CEO Work in production Work in sales and marketing (reference) Work in co-ordination (admin, HR, finance) Work in training Work in other departments Travel internationally once a year Travel internationally 2
5 times a year (reference) Travel internationally 6
10 times a year Travel internationally more than 10 times a year Average stay per trip one day Average stay per trip 2
5 days (reference) Average stay per trip 6
10 days Average stay per trip 11 + days Trip is part of series Female Has high school degree of less Has university degree (reference) Has Masters’ degree Has PhD Age 0

0

if

y2 ≤ 0

if the visit occurs between firms with fewer formal connection than those belonging to the same organisation. Since the dependent variable of the selection equation contains four categories but the model can be estimated only when this is binary, three separate regressions are performed, whereby the dependent variable of the selection equation equals zero if the visit occurs within the same organisation and one if the visit is: (i) within the supply chain, (ii) to new clients and suppliers and (iii) to conferences and trade fairs.

Binomial Probit Estimates
Coefficients.

Within S-Chain

New Client/Supplier

304

Table 4.

Conference/Trade Fair

P > |Z|

Coeff.

Std. err

P > |Z|

Coeff.

Std. err

P > |Z|

Outcome
knowledge exchange Number of trips/year: ref: 2
5 1 −0.179 6
9 −0.318 *** 10 + −0.267***

0.154 0.118 0.102

0.243 0.007 0.009

−0.148 −0.272** −0.249**

0.154 0.137 0.114

0.338 0.047 0.029

−0.108 −0.199 −0.340***

0.121 0.130 0.112

0.371 0.125 0.002

Length trip: ref: 2
5 6
10 days 11
20 days 21 days
12 months

−0.070 −0.150 −0.387***

0.101 0.125 0.147

0.489 0.231 0.009

0.017 0.062 −0.281*

0.112 0.143 0.161

0.881 0.664 0.081

−0.070 −0.128 −0.471***

0.101 0.130 0.161

0.487 0.322 0.003

Employer: ref: MNC Gov/NGO/university 100 + employees SME ( |Z|

Coeff.

0.000

−0.275*** −0.268***

Std. err 0.067 0.062 946 229.6 −982.8 0.000

Conference/Trade Fair

P > |Z| 0.000

Coeff. 0.212*** 0.209***

Std. err 0.061 0.058 1,029 370.4 −1,101.0 0.000

P > |Z| 0.000

* = p-value < 0.1; ** = p-value < 0.05; *** = p-value < 0.01.

MASSIMILIANO TANI

Marginal Effects (Biprobit Estimations).

Within S-chain

New Client/Supplier

Conference/Trade Fair

Std. err

P > |Z|

dy/dx .152

Std. err

P > |Z|

dy/dx .483

Std. err

P > |Z|

Number of trips/year: ref: 2
5 1
.050* 6
9 −.068*** 10 + −.060***

.034 .026 .025

.100 .010 .016

−.020 −.068*** −.061**

.035 .026 .024

.573 .009 .014

−.063 −.049 −.089**

.048 .052 .046

.190 .345 .050

Length trip: ref: 2
5 6
10 days 11
20 days 21 days
12 months

.026 .030 .030

.603 .240 .000

.012 .009 −.061**

.027 .035 .031

.653 .803 .050

−.031 −.032 −.104

.040 .052 .065

.448 .532 .109

.004

.953

.002

.006

.764

−.029***

.011

.007

.035 .029 .038

.065 .084 .562

.009 .006 −.031

.032 .027 .038

.766 .820 .413

.056 .010 −.095

.048 .042 .063

.246 .803 .137

Variable Conditional Probability

dy/dx .179

−.013 −.036 −.082***

Ongoing trip −.0002 Origin/Destination: ref: New Zealand EU/North Am. .065** Asia .051* Other −.022 .011

.022

.622

.020

.022

.359

−.009

.034

.773

.105* .057 −.013 −.047

.057 .038 .030 .039

.067 .143 .654 .228

.078 .013 −.023 −.018

.056 .036 .032 .041

.167 .722 .475 .654

.072 .034 −.060 −.043

.061 .055 .054 .076

.233 .533 .262 .574

Industry: ref: Manufacturing Agric_mining Trade Transport Finance Government

−.088*** −.050* −.058** −.015 −.068**

.032 .031 .028 .040 .034

.006 .100 .039 .706 .046

−.072** −.034 −.053* −.021 −.064*

.033 .031 .029 .037 .035

.030 .274 .074 .571 .070

−.166** −.118** −.167*** −.093 −.145**

.067 .058 .053 .067 .063

.013 .042 .002 .161 .022

307

Foreign resident Employer: ref: MNC Gov/NGO/university 100 + employees SME ( |Z|

dy/dx .152

Std. err

P > |Z|

Functional area: ref: sales & marketing Strategy .053 Production .049 Co-ordination .048 Training .185***

.040 .038 .046 .064

.185 .201 .294 .004

.040 .005 .007 .135**

.038 .039 .044 .068

.289 .894 .873 .067

Field education: ref: non-S&E Scientist .056 Engineer −.003

.036 .028

.115 .911

.050 .008

.034 .028

.044

.036

.224

.001

.042 −.023 −.027

.032 .026 .323

.193 .369 .402

.025 −.054** −.019

Variable Conditional Probability

Female Age group: ref: 35
44 |Z|

.183*** .110* .151** .220***

.058 .061 .065 .065

.001 .071 .021 .001

.147 .771

.057 .054

.046 .045

.213 .232

.035

.980

.026

.049

.596

.031 .026 .032

.427 .036 .557

.031 −.067* −.116**

.047 .040 .052

.670 .100 .026

dy/dx .483

* = p-value < 0.1; ** = p-value < 0.05; *** = p-value < 0.01.

MASSIMILIANO TANI

Business Visits and the Quest for External Knowledge

309

Estimates of coefficients and marginal effects are reported in Tables 4 and 5, respectively. For each set of joint regressions, the test of independence between the two probabilities is rejected, suggesting that the likelihood of accessing knowledge is not independent of the chosen type of interaction. This result supports using the joint estimation approach undertaken. Rho is negative in the case of visits within the supply chain and to new clients and suppliers, implying that visitors in these interactions are less focused on accessing productive knowledge than visitors travelling within parts of the same organisation (rho: −.21 and −.268, respectively. Both are highly statistically significantly different from zero). The probability of accessing new relevant knowledge conditional on visiting new clients and suppliers, or a firm within the supply chain, is however not negligible being 15.2% and 17.9%, respectively. This means that accessing knowledge affects 1 in 6 of these visits. In contrast, the selection for attending conferences and trade fairs is positive (rho: .209), suggesting that these interactions attract visitors more likely to seek useful knowledge than those interacting within an organisation. Attending conferences and trade fairs is characterised by the highest conditional probability of gaining new knowledge: 48.3%. This result confirms that such events are focal references for gathering and exploring new productive knowledge. Some general patterns about the determinants of knowledge-enhancing visits emerge, even after conditioning the outcome on the type of interaction. With reference to the characteristics of travel, gathering new productive knowledge is more likely when visits are not continually taking place. The marginal effects of more frequent or longer visits on the production of useful knowledge tend to be negative and statistically significantly different from zero, implying that this knowledge-generating interaction is more common when there is some novelty to be explored between visitors and visited hosts. In contrast, the personal characteristics of travellers do not appear to matter across the various types of face-to-face interactions: the marginal effect associated with gender and the level of formal education is no different from zero (though most travellers have at least a bachelor degree), while the field of education and age seems to matter. In particular, knowledge-enhancing visits are more likely when carried out by younger travellers (less than 35 years of age), consistent with the fact that older employees face higher opportunity costs of travelling due to work (e.g. responsibilities within their organisations) or personal reasons (e.g. family responsibilities). They are also more likely when carried out by

310

MASSIMILIANO TANI

graduates from a science discipline versus those with engineering or humanities degrees, though the statistical significance of marginal effects is affected by the relatively high standard error, likely reflecting the large number of bivariate explanatory variables (p-values just over .10). The country of residence of the traveller also does not appear to affect knowledge-enhancing visits, supporting that in- and out-bound flows of visitors share similar purposes. There are instead differences between knowledge-related visits and the origin/destination of travellers. These show that visits to/from Europe and North America tend to have higher probability of leading to new productive knowledge than visits to/from other parts of the world. Firm size and government/market ownership do not appear to substantially affect knowledge-related visits, though visits carried out by government, NGOs and the tertiary sector are associated with a higher likelihood of gaining new knowledge in visits within the supply chain (+.105). This is perhaps not surprising as these trips include visits to international affiliates, like a corresponding foreign university or foreign aid partner, and are carried out by employees working in training functional units. Indeed if a visitor works in a training unit rather than in sales and marketing (the reference group), the likelihood of travelling to gain new knowledge rises significantly across all types of interactions, with marginal effects ranging from 13.5% in the case of visits to new customers and suppliers to 22% in the case of travel to conference and trade fairs. Business visitors of virtually every sector aside from manufacturing and finance are associated with a lower probability of interacting to access external knowledge, but there are differences across types of interactions. In travels within the supply chain, the sectors with the most significant reductions of knowledge-related travel relative to manufacturing arise for those employed in agriculture and mining (−.088), the public sector (education, health, other government services: −.068), transport (−.058) and utilities/retail trade (−.05). In travels to new customers and suppliers there is a marked difference only with respect to those working in agriculture and mining (−.072), and to a much lesser extent transport (−.053) and the non-market sector (−.064). No detectable difference instead emerges between manufacturing, finance and trade. In the case of conferences and trade fairs working in an industry other than manufacturing or finance is associated with a significantly lower probability (about 15
17%) of gaining new productive knowledge via short-term mobility. There seems to be no noticeable statistical difference amongst various employers in the probability of gaining new knowledge and amongst

Business Visits and the Quest for External Knowledge

311

employees of production, strategic, and co-ordination units when visits occur inter- or intra-firm. In contrast, in the case of conferences and trade fairs, positive and statistically significant marginal effects are associated with working in a functional area devoted to strategy (+.183), production (+.11), and co-ordination (+.151) relative to working in sales and marketing. These results further support that conference and trade fairs are opportunities for gaining new productive knowledge rather than occasions for purely market own goods and services. This conclusion is reinforced by the negative statistically significant marginal effects for those aged 45 and above, suggesting that conferences and trade fairs provide the strongest opportunities to gain new knowledge for those of younger age.

IMPLICATIONS FOR POLICY AND CONCLUSIONS These results support the hypothesis that work-related visits are a channel for fostering skills via accessing and developing new productive knowledge, implying that mobility offers a way to enhance skills and learning at individual and organisational level. However, the features of business visits are not easily reconcilable with the utility maximisation of the human capital model and the focus on static efficiency of the general labour-supply approach to labour movements. The knowledge-enhancing effect of visits changes the human capital content of the parties interacting, but not their headcount. The net effect of two visitors travelling to each other’s organisations for the same amount of time on the skill endowment of both employers cannot be netted out, as is done to measure the net effect of migration on the labour supply of places of origin and destination. Interacting face-toface to gather external knowledge emphasises the dynamic efficiency associated with labour mobility, and the opportunity to upgrade the learning ability of the workforce. This aspect seems however under-researched in the study of the market for skills. In addition, interacting via mobility can be a strategic choice to improve an organisation’s efficient use of human resources. This is particularly relevant for firms and countries unable to attract highly skilled workers due, for instance, to disadvantages in size and location. Yet, the lack of research on this topic obfuscates the actual productivity effects of work-related visits. As a result, at times of economic restrictions the budget for travel can become an easy target for a reduction because its opportunity cost is not clearly understood. In the academic example presented in the introduction,

312

MASSIMILIANO TANI

a department in financial difficulty may decide to eliminate a visitors’ seminar programme or reduce the funds for visits of its faculty. This may help the department’s finances in the short-term but may negatively affect the volume and quality of the output of its faculty if the lack of funding to interact persists. At the aggregate national level, the uncertainty surround the opportunity cost of interacting through business visits leaves unresolved the question of whether work-related travel should be seen as a consumption expenditure, and taxed accordingly, of a form of investment in knowledgeproduction worth of targeted public support. Though the cost of travel can be deducted from taxable income, the net social benefits of visits are unclear pending future collections of detailed information at individual and firm level. For example, it is unclear to what extent small- and mediumsized firms, which employ a third of those surveyed at the airports, have sufficient financial capacity to undertake work-related travel in their quest for external knowledge, or can do so sustainably. Public finances’ support tends to be restricted to R&D activity, but there may be scope to review other channels to generate new productive knowledge and promote skills formation, as highlighted by the primary data presented.

NOTES 1. Australia is an ideal country to carry out such studies as mandatory requirement to fill departure and landing cards enabled us to verify the bias of the primary data collected versus the population of work-related visitors to/from the same airports at the same time. 2. Of course such ability is wasted if the transfer of knowledge from the environment to a firm’s knowledge ‘core’ is inefficient. Employers, therefore, face a critical human resource issue in the selection and training of capable ‘gatekeepers’ to be placed at key intersections between the organisation and the environment as well as at the crossing between its internal functional units, such as marketing and production (Cohen & Levinthal, 1990).

ACKNOWLEDGEMENT This paper is part of a research project funded by ARC Grant No. LP 0561107 and the Department of Immigration and Citizenship (DIAC). I thank David O’Dea and the team at McGregor Tan for their superb effort in gathering the data. Thanks also to Stephane Mahuteau for valuable

Business Visits and the Quest for External Knowledge

313

comments, and Gregory Ainsworth and Matias Vaira for excellent research assistance. The usual disclaimer applies.

REFERENCES Altonji, J., & Dunn, T. (1996). Returns to education and the family. Review of Economic Studies, LXXVIII, 692
704. Amin, A., & Cohendet, P. (2004). Architectures of knowledge. Oxford: Oxford University Press. Andersen, T. B., & Dalgaard, C. J. (2011). Cross-border flows of people, technology diffusion and aggregate productivity. Journal of Economic Growth, 16(1), 1
32. Arrow, K. (1962). The economic implications of learning by doing. Review of Economic Studies, 29(3), 155
173. Bathelt, H., Malmberg, A., & Maskell, P. (2004). Clusters and knowledge: Local buzz, global pipelines and the process of knowledge creation. Progress in Human Geography, 28(1), 31
56. Becker, G. S. (1964). Human capital: A theoretical analysis with special reference to education. National Bureau for Economic Research. New York, NY: Columbia University Press. Burt, R. S. (1997). A note on social capital and network content. Social Networks, 19, 355
373. Card, D., & Krueger, A. (1992). Does school quality matter? Returns to education and the characteristics of public schools in the United States. Journal of Political Economy, 100, 1
40. Cawley, J., Heckman, J., & Vytlacil, E. (2000). Understanding the role of cognitive ability in accounting for the recent rise in the return to education. In K. Arrow, S. Bowles, & S. Durlauf (Eds.), Meritocracy and economic inequality (pp. 250
296). Princeton, NJ: Princeton University Press. Cohen, W., & Levinthal, D. (1989). Innovation and learning: The two faces of R&D. Economic Journal, 99(397), 569
596. Cohen, W., & Levinthal, D. (1990). Absorptive capacity: A new perspective on learning and innovation. Administrative Science Quarterly, 35(1), 128
152. Collings, D. G., Scullion, H., & Morley, M. J. (2007). Changing patterns of global staffing in the multinational enterprise: Challenges to the conventional expatriate assignment and emerging alternatives. Journal of World Business, 42(2), 198
213. Couch, K. (1992). New evidence on the long-term effects of employment training programs. Journal of Labor Economics, 10, 380
388. Dowrick, S., & Rogers, M. (1995). Economic approaches to innovation. Aldershot, UK: E. Elgar. Dowrick, S., & Tani, M. (2011). International business visits and the technology frontier. Economics Letters, 110(3), 209
212. Florida, R. (2002). Bohemia and economic geography. Journal of Economic Geography, 2, 55
71. Franco, A. M., & Filson, D. (2000). Knowledge diffusion through employee mobility. Staff Report 272. Federal Reserve Bank of Minneapolis. Gambetta, D. (1988). Trust. New York, NY: Basil Blackwell.

314

MASSIMILIANO TANI

Gertler, M. (2003). Tacit knowledge and the economic geography of context or the undefinable tacitness of being (there). Journal of Economic Geography, 3, 75
99. Hamermesh, D. (2006). The value of peripatetic economists: A sesqui-difference evaluation of Bob Gregory. Economic Record, 82(257), 138
149. Hansen, M. T. (1999). The search-transfer problem: The role of weak ties in sharing knowledge across organization subunits. Administrative Science Quarterly, 44(1), 82
111. Heckman, J. (2000). Policies to foster human capital. Research in Economics, 54, 3
56. Hovhannisyan, N., & Keller, W. (2011). International business travel: An engine of innovation? Working Paper No. 17100. National Bureau of Economic Research. Howells, J. (2002). Tacit knowledge, innovation and economic geography. Urban Studies, 39(5
6), 871
884. Lundvall, B. (1988). Innovation as an interactive process: From user-producer interaction to the nation system of innovation. In G. Dosi, C. Freeman, R. Nelson, G. Silverberg, & L. Soete (Eds.), Technical change and economic theory. New York, NY: Pinter Publishers. Malerba, F. (1992). Learning by firms and incremental technical change. Economic Journal, 102, 845
859. Mansfield, E. (1968). Economics of technological change. New York, NY: Norton. March, J. G., & Simon, H. A. (1958). Organizations. New York, NY: Wiley. Mu, J., Peng, G., & Love, E. (2008). Interfirm networks, social capital, and knowledge flow. Journal of Knowledge Management, 12(4), 86
100. Mueller, W. F. (1962). The origins of the basic inventions underlying dupont’s major product and process innovations, 1920 to 1950. In R. R. Nelson (Ed.), The rate and direction of inventive activity (pp. 323
358). New York, NY: Princeton University Press. Nichols, A. (2011). Causal inference for binary regression. Retrieved from http://www.stata. com/meeting/chicago11/materials/chi11_nichols.pdf. Accessed on June 30, 2014. Noorderhaven, N., & Harzing, A. W. (2009). Knowledge sharing and social interaction within MNCS. Journal of International Business Studies, 40(5), 719
741. Polanyi, M. (1966). The tacit dimension. Gloucester: Peter Smith. Portes, A. (1998). Social capital: Its origins and applications in modern sociology. Annual Review of Sociology, 24, 1
25. Riusala, K., & Suutari, V. (2004). Expatriate as knowledge transferor: Impediments to international knowledge transfers through expatriates. Proceedings of the EURAM04 conference, St Andrews. Rosenberg, N., & Steinmueller, W. E. (1988). Why are Americans such poor imitators? American Economic Review, 78(2), 229
234. Storper, M., & Venables, A. J. (2004). Buzz: Face-to-face contact and the urban economy. Journal of Economic Geography, 4(4), 351
370. Teece, D. J., Pisano, G. P., & Shuen, A. (1997). Dynamic capabilities and strategic management. Strategic Management Journal, 18, 509
533. Topel, R. (1999). Labor markets and economic growth. In O. Ashenfelter & D. Card (Eds.), Handbook of labor economics (Vol. 3, pp. 2943
2984). Amsterdam: North Holland. Torre, A., & Rallet, A. (2005). Proximity and localization. Regional Studies, 39(1), 47
59. Von Hippel, E. (1987). Cooperation between rivals: Informal know-how trading. Research Policy, 15, 285
305. Zellner, C. (2003). The economic effects of basic research: Evidence for embodied knowledge transfer via scientists’ migration. Research Policy, 32, 1881
1895.

Business Visits and the Quest for External Knowledge

315

APPENDIX: THE QUESTIONNAIRE International Business Travellers Survey
November 2006 Questionnaire # ________ Q. 1 Airport □ Adelaide □ Sydney □ Melbourne □ Brisbane Q.2 Where are you flying today ______________________________________________________________ ______________________________________________________________ Q.3 Do you live in Australia and are you travelling abroad or do you live abroad and returning home? □ Australian resident travelling overseas □ Resident abroad visiting Australia Q.4 Approximately, how many international business trips do you carry out in an average year? □1 □ 2 to 5 □ 6 to 9 □ 10 to 15 □ 16 to 20 □ More than twenty Q.5 How long do these business trips normally last? □ 1 day □ 2 to 5 days □ 6 to 10 days □ 11 to 20 days □ 21 to 30 days □ 1 to 3 months □ More than 3 months

316

MASSIMILIANO TANI

Q.6 Can you tell me what exactly your current occupation is? ______________________________________________________________ ______________________________________________________________ ______________________________________________________________ ______________________________________________________________ Q.7 Can you tell me whether your occupation is among: □ Business Owners □ Managers and Administrators □ Professionals □ Associate Professionals □ Tradespersons and Related Workers □ Advanced Clerical and Service □ Intermediate Clerical, Sales and Service Workers □ Intermediate Production and Transport Workers □ Elementary Clerical, Sales and Service Workers □ Labourers and Related Workers Q.8 What is the industry sector of your employer? □ Agriculture, hunting and forestry □ Fishing □ Mining and quarrying □ Manufacturing □ Electricity, gas and water supply □ Construction □ Wholesale and retail trade, repair of motor vehicles, motorcycles and personal household goods □ Hotels and restaurants □ Transport, storage, IT, communication, telecommunication □ 10 Financial intermediation □ Real estate, renting and business activities □ Public administration and defence, compulsory social security □ Education □ Health and social work □ Other community social and personal services □ Private households with employed persons □ Extra-territorial organisations and bodies

Business Visits and the Quest for External Knowledge

317

Q.9 What is the main function that you perform in this job? □ Strategic management, CEO, board member, owner □ Architecture, design, R&D, product development, testing, quality assurance □ Purchasing, logistics □ Sales, marketing, advertising □ Production, engineering, building, construction, maintenance □ Accounting, treasury, finance, lending, risk management, auditing □ Human resources management, administration, recruitment □ Journalism, TV, filming, media, writer, photographer □ Legal work, legislator, valuer, negotiator □ 10 Medical work, hospital, nursing, health related □ Training, education, academic and scientific research □ Other (not coded) Q.10 What is the main function that you perform in this job? Other
specify ______________________________________________________________ ______________________________________________________________ ______________________________________________________________ Q.11 How many years have you worked in this job? □ Less than one year □ 1 to 2 □ 3 to 5 □ 6 to 9 □ 10 to 15 □ More than 15 years Q.12 Which of the following best describes your employer? □ Government □ Multinational firm (branches in two or more countries) □ Large corporate (no international branches/affiliates and >100 employees) □ Medium-size and small company (