CHILDREN AND FAMILIES
EDUCATION AND THE ARTS
The RAND Corporation is a nonprofit institution that helps improve policy and
decisionmaking through research and analysis.
ENERGY AND ENVIRONMENT
HEALTH AND HEALTH CARE
INFRASTRUCTURE AND
TRANSPORTATION
This electronic document was made available from www.rand.org as a public service
of the RAND Corporation.
INTERNATIONAL AFFAIRS
LAW AND BUSINESS
Skip all front matter: Jump to Page 16
NATIONAL SECURITY
POPULATION AND AGING
PUBLIC SAFETY
SCIENCE AND TECHNOLOGY
TERRORISM AND
HOMELAND SECURITY
Support RAND
Browse Reports & Bookstore
Make a charitable contribution
For More Information
Visit RAND at www.rand.org
Explore the Pardee RAND Graduate School
View document details
Limited Electronic Distribution Rights
This document and trademark(s) contained herein are protected by law as indicated in a notice appearing
later in this work. This electronic representation of RAND intellectual property is provided for noncommercial use only. Unauthorized posting of RAND electronic documents to a non-RAND website is
prohibited. RAND electronic documents are protected under copyright law. Permission is required from
RAND to reproduce, or reuse in another form, any of our research documents for commercial use. For
information on reprint and linking permissions, please see RAND Permissions.
This product is part of the Pardee RAND Graduate School (PRGS) dissertation series.
PRGS dissertations are produced by graduate fellows of the Pardee RAND Graduate
School, the world’s leading producer of Ph.D.’s in policy analysis. The dissertation has
been supervised, reviewed, and approved by the graduate fellow’s faculty committee.
Three Essays on the
Labor Supply, Savings
and Investment Behavior
of Older Workers
Jack W. Clift
This document was submitted as a dissertation in September 2012 in partial
fulfillment of the requirements of the doctoral degree in public policy analysis
at the Pardee RAND Graduate School. The faculty committee that supervised
and approved the dissertation consisted of Julie Zissimopoulos (Chair),
Pierre-Carl Michaud, and Paul Heaton.
PARDEE RAND GRADUATE SCHOOL
The Pardee RAND Graduate School dissertation series reproduces dissertations that
have been approved by the student’s dissertation committee.
The RAND Corporation is a nonprofit institution that helps improve policy and
decisionmaking through research and analysis. RAND’s publications do not necessarily
reflect the opinions of its research clients and sponsors.
R® is a registered trademark.
All rights reserved. No part of this book may be reproduced in any form by any
electronic or mechanical means (including photocopying, recording, or information
storage and retrieval) without permission in writing from RAND.
Published 2012 by the RAND Corporation
1776 Main Street, P.O. Box 2138, Santa Monica, CA 90407-2138
1200 South Hayes Street, Arlington, VA 22202-5050
4570 Fifth Avenue, Suite 600, Pittsburgh, PA 15213-2665
RAND URL: http://www.rand.org
To order RAND documents or to obtain additional information, contact
Distribution Services: Telephone: (310) 451-7002;
Fax: (310) 451-6915; Email: order@rand.org
Acknowledgements
After a lengthy journey, I owe a huge debt of gratitude to all the people who have helped me along the road. Julie Zissimopoulos, the chair of my dissertation committee, was incredibly supportive at all stages of this process, from exploring my earliest thoughts about these topics during her PRGS class through to suggesting final edits in response to feedback at my formal dissertation defense, and stuck with me through some extremely challenging times. It has been a pleasure to work for her on RAND projects and then with her on our collaborative work, and I am very grateful that she was willing to continue to serve as my chairperson even after leaving full‐time work at RAND. Pierre‐Carl Michaud played a significant role in all aspects of my PRGS experience, as a teacher, supervisor, colleague, committee member, and teammate on the PRGS softball team. I am particularly grateful for the extensive support on my third dissertation paper, and, along with Raquel Fonseca (and their cats), for welcoming me into their home in Quebec in July 2012 for a stay that was both highly enjoyable and very productive. Paul Heaton provided me with astute feedback and comments throughout my dissertation. His thoughts on what was interesting to him “from an outside perspective” were always extremely helpful, adding direction to the research and clarity to the narrative. His contributions were exactly what I hoped for when I asked him to be my third committee member. Collectively, my committee members are responsible for much of what is good in this dissertation, and blameless for what remains of the bad. I received financial support for my dissertation work from James & Anne Rothenberg, and Kip & Mary Ann Hagopian, through the PRGS Dissertation Awards that their generous gifts made possible. I received additional support for my work from the Roybal Center for Financial Decision Making, and from National Institutes of Aging Grant P01 AG022481. My experience at RAND has also been both financed and enriched by work on numerous projects across a range of units with a distinguished group of researchers. Aside from members of my committee, I would like to give special thanks to Joanne Yoong, Angela Hung, Dick Neu, Emmett Keeler and Stijn Hoorens for their mentorship and collegiality. I thank also the administration, faculty and staff of the Pardee RAND Graduate School for their support and patience. Particular thanks are due to Rachel Swanger, whose flexibility and compassion allowed me to stay, and Ira Krinsky, through whose dedication and backing I have a good reason to leave. I have received a large amount of academic, professional and social support from friends, particularly those within the student body at PRGS and from students, faculty and fellow alumni of the UCLA Luskin School of Public Affairs. I am grateful for the part they have played in my life to date and excited to have them as my peers, friends and colleagues as I embark on my professional career. Finally, I must thank the two people without whom I would literally not be here today: my parents Parry John Clift and Patricia Ann Clift. Their unconditional love is appreciated and reciprocated, always. iii Introduction
Populations across the ‘Western’ world are ageing: in the United States and EU‐15 countries, the ratio of those aged 20‐64 to those aged 65 and over stood at 6.43 in 1950; by 2000, this ratio had decreased to 4.09; by 2050, this ratio is projected to decrease to 2.13.1 This shift in the age structure of societies is the result of demographic trends over which government policies have relatively little control, and has significant consequences both for the welfare of older individuals and for the sustainability of the public social programs that support those individuals in old age. As the demographic trends continue, government, business and private individuals must adapt to the changed circumstances. Many countries instituted their old‐age support programs at a time where the demographic structure allowed a large number of working‐age citizens to support a small number of retirees; some reforms have been forthcoming (despite political challenges), but further reforms are likely to be necessary. Businesses are adapting their benefit structures to take demographic changes into account, reducing their long‐term legacy costs by moving from Defined Benefit pension plans to Defined Contribution plans, and are still in the process of adapting to a labor force that contains a greater proportion of older workers than in previous years. Individuals in turn can no longer rely so heavily on their public and occupational pensions to sustain their standard of living in retirement, and must find ways of working more, saving more or growing their savings more rapidly if they want to enjoy a comfortable old age. In this dissertation, I provide three distinct analyses addressing labor supply, saving and investment behavior of (older) workers, in the context of the incentives and constraints they face due to employer and government policies. In the first paper, I examine labor supply flexibility and its effect on the labor supply decisions of older workers. Previous literature suggests that people would like to reduce hours of work gradually over time as they get older, but do not have the flexibility to do so in their job, and consequently may retire early rather than continue to work high hours at older ages. If greater flexibility allows individuals to stay in the labor force longer, this could increase total labor supply, helping to increase both private resources for retirement and tax revenue to support public programs. Following a sample of older Americans for 16 years from 1992 to 2008, I find that there are noticeable differences in labor outcomes between those who had flexibility over their hours in 1992 and those who were not able to adjust their hours: those with flexibility worked fewer hours in their 50s, but tended to stay in the labor force longer; the major difference between groups occurred when individuals were in their early‐mid 60s, at which time those who did not have flexibility in 1992 were much more likely to retire than those with flexibility. This work provides support for the theory that people prefer gradual retirement to more abrupt departures from the labor force, and indicates that flexibility around key retirement ages might have an impact on behavior. The overall effect on total labor supply of providing 1
Author calculations using data from United Nations (2009). World Population Prospects: The 2008 Revision v flexibility at all points of the lifecycle is ambiguous, as small reductions in hours earlier in life may offset any gains at older ages; when individuals enter our dataset, those with flexibility are already working slightly fewer hours than those without. In the second paper, we examine whether labor supply flexibility affects investment behavior. Individuals can receive higher returns (on average) on their investments if they are willing to bear more risk, which allows people to reach retirement with greater resources (on average) than if they had pursued low‐risk strategies; but the fear of suffering big losses discourages people from taking risks. Theoretical work has argued that individuals with flexibility over their labor supply over the lifecycle can bear more risk in their portfolio of investments, as they can increase their labor to offset any losses they might suffer. Using a new survey we fielded in the American Life Panel (ALP), we examine how different measures of labor supply flexibility are related to measures of risk‐taking in investments: individual participation in the stock market, and the percentage of an individual’s financial wealth held in stocks. We find no evidence that flexibility over number of hours worked per week is related to investments in stocks. We find weak evidence that other flexibility measures – an individual’s belief that they would be able to continue to work longer to make up for any negative wealth shocks, and the absence of factors that make it difficult to sustain a job into old age – may be related to greater risk‐taking in investments. These results are not robust across all specifications, and require further research for confirmation, but may indicate that flexibility at the extensive margin (ability to extend a career) may be more relevant to investment decision‐making than flexibility at the intensive margin (ability to adjust hours). In the third paper, I describe the construction and characteristics of a unique dataset with which I lay the foundations for understanding pension system incentives and how they influence work and savings behavior over the lifecycle. Public pension systems across the developed world are in need of reform, but it is important to understand how the incentives in these systems affect behavior if we are to predict the consequences of different possible reforms. Previous literature has argued that public pensions displace private savings, but with elasticity of less than 1; this suggests that possible reductions in pension benefits through reforms would be partially (but not fully) offset by increases in private saving. Using new retrospective earnings history data for five European countries, in conjunction with linked survey data describing household wealth, I construct a dataset that captures the heterogeneous pension system incentives faced, and labor supply decisions made, at each point in the lifecycle for a large group of European men. My exploratory analysis of this dataset is consistent with the hypotheses that more generous income replacement by pension plans leads to lower private wealth accumulation, and greater reward within the pension system for continued work leads to later retirement. However, these statistical associations admit of plausible alternative explanations; the work documented in this paper cannot provide definitive answers on the incentive effects of pension systems, but provides the groundwork for significant extensions of research in this field, and eventually for detailed policy simulation of pension reform.
vi Paper1:LaborMarketRigiditiesandtheLaborSupplyofOlderWorkers
Introduction
As life expectancy in the United States increases and the Baby Boomers reach normal retirement age, Social Security and Medicare face a serious threat of insolvency2. As with any mismatch of revenues and costs, there are a number of possibilities for restoring balance to the finances, but political factors make many of these solutions unpalatable: very few politicians wish to run on a platform of increased taxation and benefit cuts for the elderly, policies which place visible concentrated costs on some voting members of society but have only diffuse, abstract benefits. However, examination of lifetime work patterns suggests some policy alternatives that may be less contentious. It is possible that people at older ages are willing to work longer than they currently do, but face rigidities in the labor market that prevent them from working the amount that suits them best, and consequently drop out of the labor force prematurely. By dropping out, people reduce the size of the labor force (and, from a fiscal perspective, tax revenue), and are also more likely to require support from Medicare rather than from an employer‐based insurance plan. As a first step towards understanding the potential benefits of policies that would provide older workers with more flexibility, in this paper I examine what impact labor market rigidities have on the labor supply of older workers. 2
Population aging is projected to play the dominant role in the growth of Federal entitlement spending in the medium term (64 % of spending growth up to 2035) and a lesser – but still significant – role in long‐term spending growth (44% of spending growth up to 2080). Looking specifically at Medicare/Medicaid, population aging is projected to cause 44% (30%) of medium‐term (long‐term) expenditure growth in health‐related entitlement programs, with excess cost growth in age‐adjusted health care expenditures responsible for a majority of spending growth in these programs. Source: Box 1.2 of Congressional Budget Office (2009) The Long‐Term Budget Outlook, retrieved 8/15/2011 from http://www.cbo.gov/ftpdocs/102xx/doc10297/toc.html 1 BackgroundandLiterature
Examination of different cohorts in the Health and Retirement Study (HRS) have shown that people in the younger cohorts expect to work full‐time to older ages than people in previous cohorts (Mermin, Johnson et al. 2006), largely due to changing retirement incentives, such as a decrease in the prevalence of Defined Benefit pensions and a reduction in the number of workers receiving offers of retiree health benefits; more recent cohorts may also be able to work to older ages in more rewarding and less physically demanding jobs than older cohorts, due to their higher levels of education (Goldin and Katz 2007). At the same time as employment at older ages is increasing, a study of the changing workforce in the United States found that among workers over the age of 50, 65% of men and 62% of women would prefer to work fewer hours that they currently work (Bond, Galinsky et al. 2005). The expectation of longer working lives but preference for fewer work hours later in life may make ‘gradual’ or ‘phased’ retirement (in which workers reduce their hours of work over time but remain in the workforce longer) attractive to workers: researchers found that 57% of workers in the HRS want to reduce hours of work as they age if they can3 (Zissimopoulos and Karoly 2007). However, workers may vary in their ability to reduce their hours of work as they age due to an unwillingness of employers to accommodate this. In the most comprehensive examination of labor market rigidities facing older workers, Hurd provides a series of indirect evidence for employer‐imposed constraints on hours of work: 1) part‐time work is more common among self‐employed workers than among waged workers, 2) if more flexible hours are possible for the self‐employed, we expect the self‐
employed should remain in the workforce at older ages, and some waged workers should transition to self‐employment to obtain flexibility, 3) as expected, in cross‐section the proportion of self‐employed workers increases with age, 4) in panel data from the Retirement History Survey (RHS), more self‐
employed workers transition to part‐time work than waged employees (Hurd 1996). More recent work using the HRS and the English Longitudinal Study of Ageing (ELSA) finds evidence consistent with Hurd’s line of argument, showing that the increased prevalence of self‐employment with age in cross‐section is driven primarily by differential retirement: older workers who became self‐employed before they were 50 years old are much less likely to withdraw fully from the labor force than wage and salary workers at any given age in both the United States and England (Zissimopoulos, Maestas et al. 2007). However, as noted by these authors, there are other differences in incentives between wage‐and‐salary workers and self‐employed workers. For example, the differences in types of pensions used by the two groups are shown to explain a considerable portion of the difference in age‐specific exit rates of the different worker types within each country (Zissimopoulos, Maestas et al. 2007). A multinomial logit analysis of the HRS suggests that self‐employed workers may be more likely to transition into partial retirement and less likely to transition to full retirement, even after controlling for differences in assets (Kim and DeVaney 2005). 3
Author calculation based on Table 3.1 of the cited work 2 Research on ‘bridge jobs’ provides further indirect evidence that wage‐and‐salary workers are often not free to adjust their hours of work while retaining their hourly wage. Two influential studies in this field note that older workers who switch jobs and reduce their hours of work have a much lower hourly wage than the jobs that they leave (Gustman and Steinmeier 1985; Ruhm 1990). Theoretically, there may be characteristics of these new jobs that are so attractive that workers are making a tradeoff that compensates for the wage reduction (e.g. a worker may be moving into a lower‐paid occupation later in life that they enjoy more, or that is more psychologically rewarding; or a worker may be so bored with their previous workplace that any change is an improvement). However, the simpler conclusion that the authors draw is that older workers who take these jobs wanted to reduce their hours of work and increase their leisure time, could not do so in their career job, and so had to settle for a bundle of increased leisure and disproportionately decreased earnings.4 There are a variety of reasons why wage‐and‐salary workers may face rigidities in the labor market that prevent them from reducing hours in the same job, and also make it difficult to garner a high wage with fewer hours of work with a different employer. Fixed costs of employment (such as health benefits or training costs) and team production requirements mean lower wages must be offered to people who are working part‐time: a greater percentage of the part‐time worker's product must go towards paying the fixed employment costs, and team members operating on different schedules may reduce team productivity (with variation across industry and occupation). These aspects make reduction of hours in a career job unlikely, and make it unlikely that a worker can move to a job with the same industry and occupation at a different firm with reduced hours (Hurd 1996). Some jobs, such as those with management functions, may not easily be filled by a part‐time worker or shared among several part‐
time workers, and some industries may require more teamwork than others. Conversely, businesses that already utilize large numbers of part‐time workers in their normal plan have been shown to be more likely to be able to accommodate phased retirement within their company (Hutchens 2007); and some organizations with unpredictably fluctuating labor requirements5 may benefit from maintaining a pool of on‐call part‐time employees (Rappaport 2009). An interesting recent paper shows that businesses in the United States that employ a relatively high proportion of female workers under the age of 30 also have lower retirement hazard rates for their older workers: the argument is that some firms have more flexible ‘technology’ and can accommodate workers who may wish to work non‐
standard work weeks – such as older workers who wish to reduce their work hours, or women of child‐
bearing age who may need to take a period of maternity leave and/or work part‐time at hours that can be scheduled around child‐rearing duties (Blau and Shvydko 2011). The importance of flexibility in work schedules is cited elsewhere as a reason for women and older workers to transition into self‐
employment, implicitly because this flexibility is not readily available in the wage sector (Lombard 2001; Zissimopoulos and Karoly 2007). 4
Those with DB pensions may face an additional incentive to switch jobs rather than reducing hours in the same job, if their final pension benefits are dependent on final annual earnings 5
The cited article suggests that, e.g., major storms may lead utilities to require additional labor to restore power, insurance companies and banks to require additional labor to process claims and finance repairs, etc. 3 In addition to gradually reducing hours in a career job, changing jobs at older ages to reduce hours is also problematic for business reasons: job‐specific human capital is no longer valuable, and employers are unlikely to invest in training for workers with short expected service (Hurd 1996), although this may be becoming a less pertinent issue than previously due to the declining average tenure of workers in younger cohorts (Friedberg and Owyang 2004). In general, older workers may find their chances of finding a new job hampered by the relatively high health insurance and pension costs associated with older workers relative to younger workers (Scott, Berger et al. 1995). Older job‐seekers may also have to grapple with age discrimination: while it can be difficult to measure discrimination directly, an audit study found that submitting resumes of older women was more than 40% less likely to result in an interview request than submitting otherwise identical resumes with more recent high school graduation dates (Lahey 2008). The significant importance of employer‐based health insurance in the United States provides a unique challenge to achieving labor supply flexibility for older workers: the costs of providing health insurance make employers less likely to offer full work flexibility to employees6, and reliance on employer‐based health insurance can make it infeasible for a worker to reduce their labor supply below the level required to receive full health benefits. For an employer, health insurance (if offered) is a fixed per‐
employee cost: faced with rising health insurance costs, employers have an incentive to maintain a workforce with a smaller number of employees each working a larger number of hours, in order to minimize the fixed costs (Cutler and Madrian 1998). As health insurance is typically a lumpy benefit – either fully offered or not offered at all – an employer faces a disincentive to allowing workers to retain health benefits and reduce their hours of work. On the employee side, a significant amount of literature explores the “job‐lock” phenomenon, whereby workers are less likely to change jobs due to the fear of losing health insurance benefits, through not having continuous coverage or through insurance company underwriting rules concerning pre‐existing conditions (Gruber and Madrian 2002). Recent work extends this idea to another aspect of labor supply: workers are discouraged from entrepreneurship due to fear of losing health benefits, and are therefore more likely to enter into self‐
employment if they have low demand for health insurance or have coverage through their spouse (Fairlie, Kapur et al. 2009). The same concept has been shown to apply to retirement decisions, with retiree health insurance and Medicare eligibility7 increasing retirement hazard, and working spouses also more likely to retire once their dependent spouse reaches Medicare eligibility (Gruber and Madrian 2002). While most literature in this field refers to staying in or leaving a particular job (hence “job‐lock”), a logical extension is that workers in a particular job may face “labor‐lock”, whereby they are compelled to supply some baseline number of hours in order to retain full health benefits for themselves (and potentially for their dependents), making it infeasible to begin gradual retirement until some other 6
Employers can choose to offer different benefits to full‐time and part‐time workers, but cannot provide benefits to young (and inexpensive) part‐time workers while systematically excluding older (and more expensive) part‐time workers. Employers are also unlikely to wish to manage fine‐grained sliding scales of health benefits to accommodate fully flexible working arrangements. The most likely scenarios appear to be that employers will offer little‐to‐no flexibility for employees with benefits, and/or allow flexibility for part‐time workers but provide them no health benefits. Employers who provide no health benefits to any of their employees likely face fewer difficulties in offering flexible working options. 7
and continuation coverage from COBRA through to Medicare eligibility age 4 source of health insurance (such as Medicare) is available. A recent working paper finds that married women who depend on their own employment for their health insurance and are diagnosed with breast cancer reduce their labor supply by 5.5‐7% less than those who have access to health insurance through their spouse’s insurance (Bradley, Neumark et al. 2012). It is not clear a priori what the various effects would be of increased labor supply flexibility and increased prevalence of gradual retirement; importantly, the effect on an individual’s total labor supply is ambiguous. In an atypical labor setting (the University of North Carolina system), one study found that workers choosing a newly offered phased retirement scheme would otherwise likely have continued in full‐time work rather than fully retiring, so the net effect was to decrease labor supply; however, the least productive professors (those who were previously receiving low or no annual pay rises) were much more likely to enter phased retirement, such that the phased retirement scheme may have accelerated the departure of less productive workers (Allen, Clark et al. 2004). In contrast to this narrow finding, a much more general projection exercise suggested that allowing all workers the option of working half time at half pay would lead to a modest increase in full time equivalent (FTE) employment, with the increase in labor supply from those who would be persuaded not to fully retire more than offsetting the decrease in labor supply from those who would take the opportunity to reduce from full‐time work to half‐time work (Gustman and Steinmeier 2007). Little work has directly addressed the effect of stated labor supply constraints on labor transitions at older ages. In the paper closest to our approach, Charles and DeCicca compared the retirement hazards of workers with different levels of flexibility over their working hours, finding that workers (aged 55‐64) who reported no ability to reduce their working hours in 1992 were more likely to retire by 1996 than workers who reported some flexibility (Charles and DeCicca 2007). However, their analysis does not examine relevant labor outcomes other than retirement, ignores the possibility of effects of flexibility that differ by age, and does not utilize the full information available to them regarding individual labor flexibility. In particular, Charles and DeCicca’s theoretical model contrasts workers who are fully constrained with those who face no constraints whatsoever regarding their personal labor supply, but their empirical implementation includes in the ‘unconstrained’ category a significant number of workers who are able to reduce their hours of work only at a lower wage, or with a reduction (or elimination) of their health benefits or pension eligibility. In the present paper, I distinguish between those who cannot reduce hours, those who can reduce hours but face some penalty for doing so, and those who are free to reduce hours without any penalty. Additionally, I examine not only labor force participation, but also hours worked, and examine the effects of labor supply constraints on hours worked at different ages. 5 Data
I use data from the first nine waves of the Health and Retirement Study (HRS), a panel study of Americans aged 50 and over, funded by the National Institute of Aging and administered by the Institute for Social Research at the University of Michigan. I use a public‐use version of the data that has been processed and augmented by researchers at the RAND Center for the Study of Aging.8 In my main analyses, and descriptive statistics, I restrict the sample to include only those respondents from the 1931‐41 birth cohort who were present in the survey for all nine waves.9 Ability to Reduce Hours The main variable of interest in this study is the ability of workers to reduce their hours of work at older ages. For all workers, HRS has a question asking whether or not it is possible for the worker to reduce their hours. For people working 30 hours a week or more, additional questions are asked about whether the worker could reduce their work to half‐time and, if so, how this would affect their hourly wage, health benefits and pension eligibility. Using the answer to these questions, I separate individuals into three groups: 1) No Reduce: the individual cannot reduce their hours to half‐time (or at all) 2) Constrained Reduce: the individual can reduce their hours to half‐time, but reducing their hours to half‐time would result in a drop in hourly wage or a negative consequence for their health benefits or pension plan eligibility10 3) Free Reduce: the individual can reduce to half‐time at the same hourly wage rate without any negative consequences to their health benefits or pension plan eligibility Outcomes: hours change, retirement The outcomes of interest pertain to labor force participation and labor supply. I measure changes between waves in the hours worked by an individual, with one version of this outcome measuring changes in hours worked conditional on remaining in the labor force, and an unconditional version measuring changes in hours worked and counting people who have dropped out of the labor force between waves as experiencing a reduction to zero hours of work between waves.11 I also investigate 8
See http://www.rand.org/labor/aging/dataprod.html for more details At baseline, individuals present in all nine waves did not differ from those who were not present in all waves in regards their net wealth, income or age, but were more likely be female, married, better educated and in better health 10
Note that an individual must currently have health benefits or a pension plan in order to suffer negative consequences in relation to them 11
Given the very high prevalence of the 40 hour work week, I censor the hour changes at ‐40 and +40 to decrease the statistical influence of a small number of outliers 9
6 the transition to retirement, where individuals are considered retired if they both self‐report being retired and are not working for pay.12 Income and wealth variables A priori, income and wealth are likely to affect labor supply. There are some data issues and definitional issues with these variables; my standard approach was to choose the definitions that provided the best combination of practicality and conceptual applicability, and then rerun my analyses to ensure that my results were not being driven by these definitional decisions. Given the focus on changes in weekly hours worked and constraints on those changes, the most relevant income measure is the marginal value (in terms of annual income) of a one hour change in weekly hours worked. This measure conceptually provides the annual benefit of working an extra hour per week or the annual cost of taking an extra hour of leisure per week, taking into account the fact that not all workers work the same number of weeks. It is conceptually superior to a simple annual earnings measure because workers who have flexibility over their working hours and have chosen to reduce hours are likely to have systematically lower annual earnings than workers constrained to work more hours, even if their hourly wage and weeks worked per year are identical. I construct this variable from the hourly wage reported in the RAND HRS and the number of weeks worked per year.13 Turning to wealth, I define net wealth as the sum of primary residence wealth (net of mortgage owed), financial wealth, retirement account wealth, business wealth, and other real estate wealth. This measure does not include wealth associated with a secondary residence due to a data problem in wave 3 of the original HRS data that prevents consistent calculation of second home equity across waves. 14 12
The focus in this paper is on the constraints a worker faces in managing the transition to full retirement, and the consequences of these constraints; individuals who report being retired but are nonetheless working for pay are not ‘fully’ retired in the sense most relevant to this paper 13
It is also possible to calculate this by dividing Annual Earnings by Weekly Hours, but there are data differences that lead me to prefer the formulation above. The Annual Earnings variable is retrospective, asking about the previous year’s earnings, whereas the Weekly Hours, Hourly Wage and Annual Weeks variables refer to the current job. In dual‐earner households, each spouse/partner answers questions from which the Hourly Wage and Weekly Hours variables are created, but the Annual Earnings for each spouse/partner are reported by whichever is considered the ‘financial respondent’ at the household level. The Annual Earnings / Weekly Hours calculation produces fewer missing values than the Hourly Wage * Annual Weeks calculation. Looking within individuals over time, all variables appear to have some suspicious values, but there does not appear to be an advantage of one over another. Overall, because the Annual Earnings variable does not necessarily reflect the respondent’s current job, and other issues appear not to favor either measure strongly, I choose the Hourly Wage * Annual Weeks option. For all the analyses presented later in this paper, I verified that the choice of income variable does not materially influence the main results for the key variables of interest 14
An alternative would be to use more full wealth measures but omit individuals in wave 3; this results in estimates that are less precise but not qualitatively different. I also rerun the analysis on the smaller sample, splitting out the 2nd home wealth / mortgage variables from net wealth to see if the omission of these 2nd home variables is likely to have an effect on the results; the coefficients on these variables were small and insignificant (implying they are conditionally unrelated to the main outcomes of interest), and the omission of these variables have virtually no effect on the coefficients on the other variables. 7 Other explanatory variables I use various categorical variables for demographics and job characteristics. Health is self‐reported on a 5‐point Likert scale from Excellent to Fair; Education is constructed from various survey responses into 5 categories: Less than High School, GED holder, High School graduate, Some College, and College and above. Race is defined as White/Caucasian, Black/African American, and Other. I code people as ‘married’ if they self‐report being married or partnered; I code people as ‘not married’ if they self‐report being separated, divorced, widowed or never married. Current jobs are coded into 13 industries and 17 occupations based on standard categorizations from the 1980 census. Individuals report how often their job requires ‘lots of physical effort’ on a 4‐point scale from Always/Almost Always to None/Almost None of the Time. I code people into four different pension categories based on their description of their main pension plan on their current job: No Pension, Defined Benefit pension, Defined Contribution pension, or Defined Benefit + Defined Contribution pension. DescriptiveStatistics
Table 1, Table 2 and Table 3 show descriptive statistics for the sample in Wave 1. The sample here is restricted to age‐eligible members of the original HRS cohort, who work at least 30 hours per week in Wave 1 and are present in each of the first nine waves of the survey. The statistics are broken out by the worker’s ability to reduce hours in their current job; p‐values for adjusted Wald tests (in the case of the continuous variables) and Chi‐squared tests (for discrete variables) are presented, testing the hypothesis that these characteristics are the same across flexibility groups. While the flexibility groups appear to be similar in many characteristics – a hypothesis of no difference across groups cannot be rejected for age, race, education, marital status, health or wealth – they are significantly different in several respects. In Table 1, we see that the ‘no reduce’ and ‘constrained reduce’ group have higher hours worked and more income per weekly hours worked than the ‘free reduce’ group. In Table 2, we see that there is a significant gender difference, with females much more likely to be in the ‘constrained reduce’ or ‘free reduce’ group than males; but other demographics appear similar. Turning to job characteristics, there is a very dramatic difference in pension coverage across groups, with the ‘free reduce’ group much less likely to have a pension than the other groups. There is also a difference in types of plan: people in the ‘no reduce’ group are very likely to have a Defined Benefit pension, while Defined Contribution plans are relatively more prevalent in the ‘constrained reduce’ group. Differences in the physical effort required on the job are not significant. Given the possible occupation‐ and industry‐specific reasons for employer constraints on hours flexibility mentioned in the background section, it is not surprising that people in different jobs have different levels of flexibility. In Table 3 we see that people with occupations in the managerial specialty operations category are particularly likely to be in the ‘no reduce’ group, and that people working in sales are particularly likely to be in the ‘free reduce’ group. Various service occupations tend towards the ‘free reduce’ group, while a broad range of mechanical/production/operator occupations tend 8 towards the ‘no reduce’ group. The ‘constrained reduce’ group has a notable spike in the health services occupation. These occupational differences are mirrored in the industry distribution, where manufacturing, transportation and public administration appear to be low‐flexibility industries, while retail, business & repair services, and personal services appear to be high‐flexibility industries. People in the ‘constrained reduce’ group appear relatively more likely to be in the professional & related services industry. To provide a crude summary: it appears that people in the ‘free reduce’ category are more likely to be female, with jobs that are lower paying with very low pension coverage, often in sales or services. Members of the ‘no reduce’ group tend to be more concentrated in production industries and operational occupations, and are likely to have Defined Benefit pensions. Those in the ‘constrained reduce’ group tend to work more in technical service roles, often with a Defined Contribution pension. The groups do not differ on health, wealth or the physical requirements of their jobs, nor on age, race, education and marital status. The differences in benefits across groups are potentially problematic and require careful examination of controls: with a larger sample, it would be ideal to compare outcomes based on flexibility among those with good health and pension benefits; with only a small percentage of the small ‘free reduce’ group having good benefits, including controls for health and pension benefits and seeing how these affect the models is the more feasible approach. 9 Theoretical model In a typicaal model of th
he individual’s labor supplyy decision, in dividual workkers receive u
utility from Consumpttion (C) and LLeisure (L), an
nd are endow
wed with Timee (T) to allocatte across paid
d work (h) – aat exogenou
us wage rate W
W and leisure
e time (l). The single
e‐period labor supply utilitty maximization is then Max U(C(h
hW), L(l)) subject to tthe constraintt T = hh + l UC and UL may be assumed to be po
ositive, and U
UCC and ULL maay be assumeed to be negattive. Standard manipulation
n of the utilityy function and
d budget con straint yield tthe familiar result that thee interior so
olution for the optimal cho
oice of hours occurs wheree the marginaal rate of substitution of leeisure for consumption equals the wage raate, or, in graaphical terms,, where the in
ndifference curve representing the utilityy function is taangential to tthe budget co
onstraint. Figuure 1 below rrepresents the solution (h**, l*) of a consttrained optim
mization of a ssimple Cobb‐D
Douglas utilityy function of the form U(C
C,L) =  ln C + (1‐) ln L. h*(t) l*(t) Figure 1: Styylized constrained optimization
n 10 e, individual p
preferences re
egarding leisu
ure and consuumption mayy change: at o
older ages, peeople Over time
may place
e greater valu
ue on leisure ttime and lesss value on connsumption go
oods. This is cconsistent witth the observed reduction in the hours people work as they age, an d stated prefferences for reducing hourrs gradually over time as they age. If u
utility function
ns change ov er time and rreal wages remain consisteent at older agess, individuals should rebalance their tim
me budget tow
wards more lleisure and leess work, as shown in Figure 2
2 below, whicch maintains the same bud
dget constrai nt as the prevvious graph b
but changes the paramete
er ‘‘ in the uttility function
n. h*(t+1) l*(t+1) Figure 2: Co
onstrained optim
mization with ad
djusted utility paarameter However, when worke
ers are not offfered a smooth budget co nstraint, the situation beccomes more complex. Employers raarely allow a ffree choice off hours at a g iven wage ratte; some offeer no flexibilitty in hours at aall, others allo
ow workers to
o reduce hours only with aa penalty in in
ncome or ben
nefits. Facing discontinu
uous constraiints on their leisure/consu
umption decission, a slight cchange in an individual’s preferencces may not le
ead to a moderate change
e in hours of laabor supplied
d: depending on the options available with their em
mployer, a wo
orker may nott adjust their labor supply at all, may make a uous change, or may drop out of the labor force altoogether. The graphical dep
pictions below
w discontinu
show how
w different disscontinuous cchoice sets avvailable to woorkers interacct with their p
preferences to produce d
different resu
ults; in all case
es, the worke
er is worse offf (on a lower indifference curve) than they 11 oth budget co
onstraint, and
d slight differeences in the cconstraints/p
preferences w
would would be given a smoo
lead to drramatically different labor outcomes. Figure 3: Un
nderemploymen
nt: offered only one option, wo
orker works less than if offered a smooth budgeet set 12 Figure 4: Ovveremploymentt: offered only o
one option, workker works more than if offered a smooth budget set 13 Figure 5: Un
nstable equilibriium – offered tw
wo choices, sligh
ht changes in prreferences could
d lead to a disco
ontinuous increaase or drop in labo
or supply odels is that w
workers who face discontinuous choicee sets may bee The impliccation of thesse stylized mo
more likely to stay worrking full‐time
e than workers with continnuous budgett constraints,, but may also
o be more likely to drop outt of the workforce altogether. A further implication – though it iss difficult to observe this directly – is that discon
ntinuous choice sets reducce utility. In th
he next subseection, I show
w prima facie evidence th
hat discontinuous choice ssets of workeers do affect lifetime labor supply trajectorie
es. In theory,, discontinuitiies in the cho
oice sets offerred by individdual employerrs need not m
matter if: a) collectively employers offer a wide range of choices, b) collecctively the cho
oices offered
d match the preferencces of individu
uals, and c) it is costless fo
or employees to switch em
mployers at older ages. Wh
hile it is true thaat different em
mployers may offer differe
ent choices, itt is difficult to
o assess whetther they match collectively match the individual pre
eferences of w
workers. How
wever, as disccussed in the literature secction, switching costs can be very high forr older worke
ers, whose jobb‐specific hum
man capital m
may be less vaalued in a new job, who mayy have difficulties finding a new employyer due to higgh costs of frin
nge benefits, and generally may face age
e discrimination in the labo
or market. If sswitching cossts are high and worker preferencces tend to sh
hift at older agges towards w
working feweer hours, the cconstraints posed by discontinu
uous choice ssets offered b
by their own e
employers wi ll be at least partially bind
ding. 14 LaborSupplyTrajectories
Before providing a quantitative empirical analysis of the effect of hours flexibility on individual labor supply, it is helpful to show graphically how the labor supply decisions of the three groups vary over time. In the graphical analysis, I limit the sample to those who were age 50‐57 in Wave 1 of HRS, but it is otherwise the same sample described in the Data section above. I code individuals into groups based on their Wave 1 ability to reduce hours, and then follow these groups as they age. I do not allow people to switch between groups for two reasons: first, when people drop under 30 hours of work, they are no longer asked the detailed questions that allow categorization into the three groups; second, while in the earliest waves there is no statistically significant difference across the groups in the desire to reduce hours gradually over time15, this does not hold true in later waves, suggesting non‐random switching between groups at older ages and potentially conflating the effects of constraints on labor supply trajectory with differences in desired labor supply trajectory. In Figure 6, we see how average hours worked changed in the three groups, averaging across those in each group who were still in the labor force in later waves and collaping individuals into 2‐year age bands. As suggested by Table 1, the groups start off with different average hours worked, with the ‘no reduce’ group somewhat higher than the other two groups. Each group has a mostly similar pattern, with average hours decline moderately in the 50s and more steeply in the 60s. The difference in average hours between the ‘free reduce’ and ‘no reduce’ groups widens in the early‐mid 50s, but narrows again by age 62. The constrained group fluctuates more, with the smallest decrease in hours between 50 and 60, but the largest decrease between 60 and 68. In Figure 7, the same graph is plotted for median hours; unsurprisingly, 40 hours remains the focal point for each group up to age 58; the ‘free reduce’ group median drops below 40 from age 60 onwards, the ‘constrained reduce’ group median drops below 40 at age 62, and the ‘no reduce’ group median does not drop below 40 until age 66. 15
See the Empirical Model section for further discussion of this issue and Table 4 for the statistics 15 Average Hours worked per WORKER
25
30
35
40
45
20
50
52
54
56
58
60
Age
62
Reduce Status:
Constrained reduce
64
66
68
70
No reduce
Free reduce
Figure 6: Change in Average Hours of Work with Age, conditional on remaining in the labor force 16 Median Hours worked per WORKER
25
30
35
40
50
52
54
56
58
60
Age
62
Reduce Status:
Constrained reduce
64
66
68
70
No reduce
Free reduce
Figure 7: Change in Median Hours of Work with Age, conditional on remaining in the labor force Turning to labor force participation in Figure 8, there is a steady decline in the proportion of people working for pay in each group between age 50 and 58, with each group dropping 17‐18% in that time. However, the groups begin to diverge at age 60 as the ‘no reduce’ and ‘constrained reduce’ groups drop out of the much faster than the ‘free reduce’ group. The gap in participation is widest at age 62 and mostly sustained through age 66 before narrowing significantly at age 68. Combining these two effects – at the extensive and intensive margins – the mean hours of work averaged across all members of each group are plotted in Figure 9. The ‘free reduce’ and ‘constrained reduce’ group begin working fewer hours than the ‘no reduce’ group, and continue to work fewer hours through age 58 as all three groups see a modest decline. However, while the ‘free reduce’ group continues to follow a similar trend of moderate decline through their 60s, the ‘no reduce’ group and ‘constrained reduce’ groups experience accelerated decline in their early 60s, so that by age 62 the ‘free reduce’ group is actually working more hours (25.0) on average than the ‘constrained reduce’ (19.4) and ‘no reduce’ (20.0) groups. This substantial gap persists until age 66, before narrowing at age 68. Overall, the pattern seems to suggest that while those with the ability to reduce hours do take advantage of this in their 50s and work less than those without the ability to reduce hours, this is offset by their reduced retirement hazard in their 60s, particularly around key Social Security eligibility ages. Or, stated another way, those with less flexibility over their working hours tend to work more in their 17 50s , but then drop out of the labor force in large numbers as soon as they become eligible for Social Security. The relatively abrupt drops in labor supply from the ‘no reduce’ and ‘constrained reduce’ groups, compared with a fairly smooth smooth labor supply path for the ‘free reduce’ group, provide prima facie evidence of some workers being “over‐employed” in their 50s due to employer‐imposed constraints, and then choosing “under‐employment” when they have access to pension income in their 60s. Looking over the whole period, it is possible that total labor supply for a given individual might work out to be the same whether he faces constraints or has a free choice of hours16, but the distribution of these hours when constrained – too many in his 50s and too few in his 60s – could be associated with significant loss of utility.17 .2
Proportion working for pay
.4
.6
.8
1
50
52
54
56
58
60
Age
62
Reduce Status:
Constrained reduce
64
66
68
70
No reduce
Free reduce
Figure 8: Change in Labor Force Participation with Age 16
Over the age range studied, from 50‐51 to 68‐69, the ‘no reduce’ and ‘free reduce’ groups average 21.5 and 21.3 hours per person respectively. However, I do not capture work before age 50 (where it is possible the ‘free reduce’ group may already have been working less) or after 70 (where we might expect the ‘free reduce’ to continue the later trend of working more than the ‘no reduce’ group) 17
Nevertheless, this distribution of hours may have positive aspects for employers, who may prefer to employ a smaller number of people for a larger number of hours when there are fixed costs per employee rather than per employee‐hour 18 Average Hours worked per PERSON
10
20
30
40
50
50
52
54
56
58
60
Age
62
Reduce Status:
Constrained reduce
64
66
68
70
No reduce
Free reduce
Figure 9: Change in Average Hours of Work with Age, averaged across all individuals EmpiricalModel
Unconditional on any other factors, it appears that demand‐side constraints on hours of work have a noticeable effect on individual labor supply. However, it is possible that the difference in labor supply trajectories between the groups are due to the influence of observed differences in characteristics between the groups. In order to isolate the effect of demand‐side constraints, I estimate a series of multivariate regressions, controlling for a wide range of observable characteristics. As my base models I use pooled Ordinary Least Squares estimation18, and then examine the sensitivity of my results to alternative estimation assumptions. The basic models I estimate take the form: 18
With pooled OLS, it is important to consider possible autocorrelation in the error terms within individuals over time. In non‐survey‐weighted data, it is possible to create robust standard errors in Stata using clustering estimation subcommands. Using this method, it appears that there is negative autocorrelation in the within‐
person error terms: the cluster standard errors are smaller than the normal standard errors. While Stata does not offer options to deal with the negative autocorrelation in pooled OLS with complex survey data, the standard errors I estimate may be considered to be conservatively large. 19 Where is the individual’s ability to reduce hours of work (no reduce, constrained reduce, free reduce) in wave 1, is a vector of demographic variables including age, race, gender, education and marital status at time t, is a self‐reported assessment of health at time t, represents net worth at time t, and is a vector of job characteristics at time t including industry, occupation, physical demands and a standardized measure of income. A few variables warrant further description. The labor change outcomes I estimate are change in weekly working hours conditional on remaining in the labor force, change in retirement status, and a composite measure of change in weekly working hours where retirees are treated as having working hours of zero. Insofar as policy‐makers may be most interested in aggregate labor supply, the last outcome provides the net effect of changes in labor participation and changes of working hours for those who remain working. All changes are 1‐wave (2‐year) changes. The standardized measure of income I use is the hourly wage multiplied by weeks worked per year: this variable is the marginal change in annual income accompanying a one hour change in weekly working hours, and thus is more closely related to the outcome variables than a simple measure of annual income or hourly wage. The ability to reduce hours of work variable is assigned to individuals in Wave 1 of the survey, and I then keep this fixed for individuals throughout the nine waves over which I observe them. As mentioned briefly in the Labor Supply Trajectory section, there are two reasons for making this choice, compared with allowing this variable to change over time. First, the questions that allow me to distinguish between my preferred categorization are only asked to people working more than 30 hours per week, so people dropping below 30 hours can no longer be categorized contemporaneously. Second, the identification of the model relies on the premise that – conditional on the rich set of covariates – there are no relevant underlying differences between those who face constraints on their labor and those who do not. However, if people select into these categories, that assumption may not hold, and this is likely to be increasingly problematic as people approach retirement. In Wave 2, people across the three groups express similar preferences when asked whether they would prefer to gradually reduce their hours of work over time as they age19; but by Wave 3, there are statistically significant preference differences, suggesting that some selection has taken place through job switching or (potentially) negotiation of job characteristics with the employer. The ability of workers to reduce hours in Wave 1 thus allows me to follow labor trajectories when people dip below 30 hours per week, and also provide the most plausible conditionally‐random assignment of demand‐side constraints on working hours. 19
See Table 4 20 Results
Baselineresults
Given the labor supply trajectories seen in Figure 9, I separate individuals by age, to examine the effects of the ability‐to‐reduce variables at different stages in the transition to retirement. The results from OLS regressions are presented in Table 8, for change in hours worked (conditional on remaining in the labor force), likelihood of retirement, and unconditional change in hours worked. The ‘ability to reduce’ variables do not have much effect on hours changed for workers who continue to work. The coefficients for both the ‘constrained reduce’ and ‘free reduce’ group are insignificant in all three age categories, though some of these coefficients are substantively large. Consistent with the labor supply trajectories seen in Figure 8, the OLS regression results for retirement in Table 8 highlight that the main effects of the ‘ability to reduce’ variables are felt by those aged 60‐64, who reach age 62‐66 by the following wave. At those ages, the ‘free reduce’ group were more than 7 percentage points less likely to retire than the ‘no reduce’ group. With an average 2‐year retirement hazard of 0.25, this represents an effect size of 30%. There were no significant differences across groups at either the younger (50‐59) or older (65 and over) ages. The retirement results are mirrored in the unconditional hours‐change results – no difference across groups at the younger or older ages, but for the people aged 60‐64 the hours change was more positive (or , less negative) for the ‘free reduce’ group than for the no reduce baseline group. When the changes in labor force participation and changes in average hours per worker are combined, the OLS estimates suggest that the ‘no reduce’ group has a decline in hours per person that is 2.6 hours larger than the ‘free reduce’ group. With an average decline of 11.9 hours, 2.6 hours represents an effect size of 22%. Turning to other variables: in Table 8, I do not report coefficients for model control dummy variables for age, gender, race, marital status, education, industry, occupation, and physical effort required on the job.20 The other variable coefficients reported suggest the expected health gradient, with less healthy people more likely to retire, particularly at older ages. Wealth does not appear to affect hours of work but may affect participation at older ages; the marginal income for an hour worked per week does not affect the participation decision, but does seem to increase hours worked, particularly at older ages. Having a Defined Benefit pension appeared to strongly increase the likelihood of retiring in the next 2 years, with a particularly strong effect at ages 60‐64; those with Defined Benefit pensions did not decrease average hours per person significantly more than the no‐pension group prior to age 60, but did reduce hours significantly more from 60 onwards. Those with Defined Contribution pensions were neither more nor less likely to continue participating in the labor market than those without a pension on their current job, at any age. However, at older ages, people with Defined Contribution pension plans did reduce their average hours of work significantly more than the no‐pension group. 20
Gender, race, marital status, education and physical effort required on the job are insignificant for the most part. Unsurprisingly, older people were significantly more likely to retire across the whole age distribution, and tended to reduce their hours more than younger people. Industry/occupation variables appear to have more effect on hours changed and less effect on retirement. 21 Alternativespecifications
The OLS baseline models described in the section above provide a simple model with intuitive coefficients. In this section, I report the results of some alternative specifications suggested by the nature of the data21, and examine alternative ways of capturing the ‘ability to reduce hours’. Given that retirement is a binary outcome by my definition, I provide a comparison of OLS and probit specifications in Table 9. The results are quite consistent between the two models, with nearly all coefficients of interest maintaining the same sign and significance level. The main result of interest – the statistically significant negative coefficient on ‘free reduce’ for 60‐64 year olds – is consistent across models, with the Probit estimate of ‐0.278 implying a marginal effect of ‐0.078, compared with ‐0.075 in the OLS model. Turning to the hours‐change model, the outcome in this case is a continuous variable, but subject to censoring. In particular, regardless of one’s disutility for work, it is impossible to drop below zero hours of work, which places a bound on the maximum size of hours change for an individual depending on their hours in the previous wave. Overall, left‐censored observations make up 12% of the 50‐59 year old group, 28% of the 60‐64 year old group, and 27% of the 65+ year old group. I fit an interval regression model taking into account individual‐level censoring, and present the results in Table 10. In general, we see that point estimates are generally larger in absolute value but are less precisely estimated. This is understandable, as the latent values of preferred changes in hours of work take on a wider range than the observed changes, but we do not have data in that range. Nevertheless, it is worth noting that a coefficient of particular interest – the effect on hours‐change of the ‘free reduce’ category for 60‐64 year olds – is quite substantially larger for the interval regression (3.975) than for OLS (2.557). Given that the outcome is censored on an individual basis depending on the initial hours of work, it is unlikely that the assumption of normally distributed error terms on the latent variable is correct, and therefore is not clearly an improvement on the original OLS model; but the interval regression results make us more confident that the significant OLS coefficients are not an artifact of the way the data are constructed. Finally, Table 11 and Table 12 provide alternative versions of the key explanatory variable, ability to reduce hours. For each age group, I compare the base model (3 categories set in Wave 1) with two versions of an alternative binary model, combining the ‘constrained reduce’ and ‘free reduce’ groups to form one ‘any reduce’ group and treating this either as fixed in Wave 1, or as changing over time. For the most part, the ‘any reduce’ results fixed in Wave 1 are similar to the base models, with point estimates for the ‘any reduce’ group falling between the point estimates for the ‘constrained reduce’ and ‘free reduce’ groups in the base models. The most notable difference between the base model and the third model, where individuals are categorized as ‘no reduce’ or ‘any reduce’ based on their contemporaneous employment, comes in the 65+ age group. Here we see the contemporaneous ‘any 21
As the ‘hours change [workers]’ and ‘hours change [all people]’ outcomes are similar in nature, I omit alternative specifications for the former 22 reduce’ variable seems to have a marginally negative effect on retirement (p‐value <0.1), and has a statistically significant (and substantively large) positive effect on hours worked. This difference in results is consistent with the concerns expressed earlier in the paper, that workers with more flexible jobs at older ages are more likely to have selected into those jobs precisely because of the flexibility to adjust hours of work. An alternative interpretation might be that workers intending to work for many years to come are more willing to incur switching costs to a job that better suits their preferences, while workers planning on complete retirement in the near future are more likely to just put up with their (inflexible) job for a short period and then quit the workforce altogether. Thus, the difference we see between the Wave‐1‐fixed and contemporaneous results, attributable to people switching groups after Wave 1, may reflect differences in unobservable preferences or differences in work horizon between those who switch groups and those who keep the same classification from Wave 1 onwards. Either way, these results reinforce the advantage of retaining the Wave 1 classifications as the least likely to be subject to non‐random selection. 23 Conclusion
In this paper, I have demonstrated that people without constraints on their flexibility to reduce working hours are likely to remain in the labor force longer, and consequently tend to provide greater aggregate labor supply at older ages. The effect of flexibility is felt most strongly around key ages connected to Social Security benefit eligibility: when workers reach the Social Security early retirement age of 62, those with the least flexibility over their hours begin to drop out of the labor force more precipitously than those who are free to adjust their hours. However, the implications for total labor supply over the lifecycle are unclear: workers who had the flexibility to reduce their hours were, on average, working fewer hours in their 50s than the workers who did not have flexibility. We do not have information on hours of work prior to people entering the HRS dataset, so we do not know if the initial difference in hours is the result of flexible workers beginning to taper down their work output at some point earlier in life than the less flexible workers, or reflects a persistently lower output over the lifecycle. The results suggest that policies aimed at improving the part‐time work options for older workers over the age of 60 might allow workers to stay in the labor force for longer. This may be of interest for both employers and government (and government employers). Employers who wish to retain human capital, in the form of experienced workers who can share knowledge with (and perhaps provide an example to) younger workers, may find that they get more out of their older workers by allowing them to reduce hours prior to normal retirement age. Government may be interested in increasing total labor supply for fiscal reasons and reducing the reliance of older people on public programs, and may be generally interested in enabling people to make their preferred ‘optimal’ choices rather than choosing between suboptimal alternatives. While more research is required on the underlying causes of constraints on hours of work, some areas of public policy seem particularly relevant: for example, employers may be reluctant to allow workers to reduce their hours and maintain their health benefits, due to the fixed per‐
person costs involved; providing subsidies for health insurance for older part‐time workers, or allowing Medicare to be the primary insurance for individuals over the age of 6522, could alleviate this barrier to flexibility. Similarly, rules surrounding public and private pension systems should be examined to make sure that people planning to slowly reduce their hours of work at older ages are not disadvantaged compared with their colleagues who might work full‐time to a certain age and then retire completely. Finally, while I have demonstrated a quantifiable effect of hours flexibility on labor output, I have not estimated the utility lost by individuals when they are forced to choose between a limited set of suboptimal alternatives, nor have I examined in this paper any other individual‐level consequences of labor supply flexibility. In theory, the ability to adjust individual labor supply could affect a range of other factors, such as: physical, mental or psychosocial health; savings and investment behavior; 22
In some cases, individuals ‘retire’ from their job on reaching Medicare eligibility, and then are re‐hired by the same employer into a position with a slightly different job description that doesn’t provide health benefits, because current law requires employer insurance to be the primary insurance plan, and employers are not allowed to discriminate in their benefit packages on the basis of age (i.e. for 2 people in identical jobs, it is not permitted to provide the 50‐year‐old with health insurance and deny health insurance to someone who happens to have reached Medicare eligibility age) 24 caregiving to a spouse or parent; or even the propensity to start a business. One might expect that increased control of one’s labor options could only be beneficial, but testing this hypothesis, and quantifying the size of any effects, fall beyond the scope of this paper and are left for future research. 25 Bibliography
Allen, S. G., R. L. Clark, et al. (2004). "Phasing into Retirement." Industrial & Labor Relations Review 58(1): 112‐127. Blau, D. and T. Shvydko (2011). "Labor Market Rigidities and the Employment Behavior of Older Workers." Industrial & Labor Relations Review 64(3). Bond, J. T., E. M. Galinsky, et al. (2005). "The Diverse Employment Experiences of Older Men and Women in the Workforce." Boston: The Center on Aging & Work/Workplace Flexibility. New York: Families and Work Institute. Bradley, C. J., D. Neumark, et al. (2012). Does Employer‐Provided Health Insurance Constrain Labor Supply Adjustments to Health Shocks? New Evidence on Women Diagnosed with Breast Cancer, National Bureau of Economic Research. Charles, K. K. and P. DeCicca (2007). "Hours flexibility and retirement." Economic Inquiry 45(2): 251‐267. Cutler, D. M. and B. C. Madrian (1998). "Labor market responses to rising health insurance costs: evidence on hours worked." RAND Journal of Economics 29(3): 509‐530. Fairlie, R. W., K. Kapur, et al. (2009). "Is employer‐based health insurance a barrier to entrepreneurship?" UCD Centre for Economic Research Working Paper WP09/03. Friedberg, L. and M. Owyang (2004). "Explaining the evolution of pension structure and job tenure." NBER Working Paper. Goldin, C. and L. F. Katz (2007). "Long‐run changes in the wage structure: Narrowing, widening, polarizing." Brookings Papers on Economic Activity 2: 135. Gruber, J. and B. C. Madrian (2002). Health Insurance, Labor Supply, and Job Mobility: A critical review of the literature. NBER Working Paper w8817. Gustman, A. L. and T. L. Steinmeier (1985). "The Effect of Partial Retirement on the Wage Profiles of Older Workers." Industrial Relations 24(2): 257‐265. Gustman, A. L. and T. L. Steinmeier (2007). "Projecting Behavioral Responses to the Next Generation of Retirement Policies." NBER Working Paper 12958. Hurd, M. D. (1996). "The Effect of Labor Market Rigidities on the Labor Force Behavior of Older Workers." Advances in the Economics of Aging. Hutchens, R. (2007). "Phased Retirement: Problems and Prospects." Work Opportunities for Older Americans Series. 26 Kim, H. and S. A. DeVaney (2005). "The selection of partial or full retirement by older workers." Journal of Family and Economic Issues 26(3): 371‐394. Lahey, J. N. (2008). "Age, Women, and Hiring: An Experimental Study." J. Human Resources 43(1): 30‐56. Lombard, K. V. (2001). "Female self‐employment and demand for flexible, nonstandard work schedules." Economic Inquiry 39(2): 214‐237. Mermin, G. B. T., R. W. Johnson, et al. (2006). "Why Do Boomers Plan to Work So Long?" The Retirement Project: Discussion Paper 6(04). Rappaport, A. (2009). "Phased Retirement ‐ An Important Part of the Evolving Retirement Scene." Benefits Quarterly 25(2): 38‐50. Ruhm, C. J. (1990). "Bridge Jobs and Partial Retirement." Journal of Labor Economics 8(4): 482‐501. Scott, F. A., M. C. Berger, et al. (1995). "Do Health Insurance and Pension Costs Reduce the Job Opportunities of Older Workers?" Industrial and Labor Relations Review 48(4): 775‐791. Zissimopoulos, J. and L. A. Karoly (2007). "Work and well‐being among the self‐employed at older ages." Research Study. AARP Public Policy Institute. Zissimopoulos, J., N. Maestas, et al. (2007). The Effect of Retirement Incentives on Retirement Behavior: Evidence from the Self‐Employed In the United States and England. RAND Working Paper Series, RAND Corporation. Zissimopoulos, J. M. and L. A. Karoly (2007). "Transitions to self‐employment at older ages: The role of wealth, health, health insurance and other factors." Labour Economics 14(2): 269‐295. 27 Tables
Ability to reduce hours on Wave 1 job Constrained reduce No reduce Wave 1 Observations Free reduce p‐value of difference test 2353
133
153 Age 55.08
55.03
55.42 0.57 Hours worked 43.09
42.42
39.83 0.00 Income 745.5
675.5
560.8 0.00 191850
200490
202410 0.91 Continuous variable means Net Wealth Table 1 Note: Income is calculated as hourly wage * weeks worked per year, to give annual income per hour in a normal work week (or, the opportunity cost in annual income of reducing a normal work week by one hour) 28 Ability to reduce hours on Wave 1 job Constrained reduce No reduce Free reduce p‐value of difference test Discrete variable proportion distribution Individual Characteristics Gender Male 54.72
31.27
35.01 Female 45.28
68.73
64.99 0.00 88.49
89.58
87.77 Black/African American 8.30
7.73
7.88 Other 3.22
2.69
4.34 0.85 15.22
16.33
15.93 5.78
3.31
5.40 High school graduate 32.36
33.37
39.20 Some college 20.66
18.94
20.52 College plus 25.98
28.05
18.96 0.54 Married/Partnered 77.32
76.02
71.84 Single 22.68
23.98
28.16 0.33 Health Excellent 30.59
27.39
33.90 Very good 35.68
34.64
32.30 Good 25.70
29.49
22.66 6.50
7.50
8.65 Race White/Caucasian Education Less than high school GED Marital Status Fair 29 Poor 1.52
0.98
2.49 0.75 Job Characteristics Pension plan on current job No pension 22.86
26.98
72.35 DB pension 48.75
39.48
14.97 DC pension 25.40
31.21
11.70 2.98
2.32
0.99 0.00 6.12
5.71
17.20 Employer 80.00
77.17
48.25 Spouse 10.77
13.18
24.84 Government 1.28
1.21
1.60 Other 1.82
2.73
8.11 0.00 All/almost all the time 18.56
27.22
16.68 Most of the time 16.43
18.98
16.25 Some of the time 31.06
23.38
31.54 None/almost none of time 33.94
30.41
35.54 0.16 DB+DC pension Source of Health Insurance None Job requires physical effort Table 2 30 Ability to reduce hours on Wave 1 job Constrained reduce No reduce Occupation Free reduce p‐value of difference test Managerial specialty oper 19.00
7.91
12.40 Prof specialty opr/tech sup 19.85
22.98
15.57 5.65
14.26
18.77 17.84
16.16
17.54 Svc prv hhld/clean/bldg svc 0.35
0.29
2.25 Svc protection 1.91
1.08
1.93 Svc food prep 1.94
7.50
4.31 Health svc 1.06
10.96
5.31 Personal svc 4.24
5.63
6.85 Farming/forestry/fishing 1.20
0.00
1.77 Mechanics/repair 4.49
2.22
1.50 Constr trade/extractors 2.52
2.31
1.25 Precision production 4.20
2.06
0.71 Operators machine 8.31
1.91
2.29 Operators transport etc 4.84
1.93
4.48 Operators handlers etc 2.37
2.81
3.06 Member of armed forces 0.22
0.00
0.00 0.00
Agric/forest/fish 1.28
0.00
2.50 Mining and constr 5.37
3.77
3.47 Mnfg non‐durable 10.23
2.92
1.90 Mnfg durable 14.68
3.88
4.85 Sales Clerical/admin supp Industry 31 Transportation 8.33
6.66
1.71 Wholesale 3.83
4.44
4.96 Retail 7.25
19.23
20.64 Finan/ins/realest 5.90
7.99
9.61 Busns/repair svcs 2.82
5.32
10.07 Personal services 1.48
2.11
6.63 Entertn/recreatn 1.13
0.00
2.24 Prof/related svcs 30.06
43.24
28.06 7.64
0.45
3.37 0.00
Public administration Table 3 Ability to reduce hours on Wave 1 job Constrained reduce No reduce Prefer to reduce hours gradually Free reduce p‐value of difference test Strongly Agree 12.27
18.11
13.72 Agree 40.88
43.92
51.55 Disagree 41.48
35.98
31.88 4.75
1.99
1.93 0.10 Strongly Disagree Table 4 32 age 51‐59 at t age 60‐64 at t
age 65+ at t
Outcome: Hours Change [workers] Mean: ‐1.14 Outcome: Hours Change [workers] Mean: ‐2.91Outcome: Hours Change [workers] Mean: ‐2.14
Reduce 1 Constrain ‐0.405 2 ‐0.320 3 ‐0.328 4
5
6
7
8
9 10
11
12
13
14
15
‐0.969
0.427
‐0.593
‐0.720
‐0.550
‐0.658
(1.111)
(1.426)
(1.410)
(1.420)
(1.438)
‐0.692
‐1.542
‐1.646
‐1.610
‐1.603
(0.724)
(0.961)
(0.949)
(0.994)
(1.178)
2.161
2.016
2.034
2.189
‐0.360
‐0.199
‐0.954
‐0.788
‐0.848
‐0.759 (0.432) (0.452) (0.450) (0.456)
(0.499)
(0.889)
(0.800)
(0.809)
(0.814) (0.921)
Free ‐0.0993 0.0910 0.226
0.134
0.235
0.00241 ‐0.150 (0.269) (0.371) (0.381) (0.388)
(0.425)
(0.680)
(0.719)
(0.710)
Health 0.149 0.161
Very Good 0.0788 0.0913 0.111
Good ‐0.322
(0.694) (0.777)
0.0782
‐0.684
‐0.663
‐0.582 (0.329) (0.337) (0.343)
(0.351)
(0.561)
(0.574)
(0.567) (0.555)
(1.136)
(1.098)
(1.084)
(1.146)
0.482 0.404
‐1.526* ‐1.546* ‐1.543* ‐1.484*
2.484*
2.347
2.468*
2.639*
(0.290) (0.291) (0.295)
(0.311)
(0.584)
(0.578)
(0.581) (0.573)
(1.227)
(1.212)
(1.207)
(1.158)
Fair ‐0.890 ‐0.931
‐1.040
0.0683
0.178
0.0616 1.840
1.652
1.742
2.222
(0.534) (0.561) (0.554)
(0.537)
(0.923)
(0.923)
(0.908) (0.949)
(1.388)
(1.398)
(1.424)
(1.502)
Poor ‐0.393 ‐0.482
‐0.502
2.455
2.434
2.185 2.973
3.546
3.839
3.482
(1.460) (1.577) (1.555)
(1.615)
(2.889)
(2.856)
(2.882) (2.901)
(1.919)
(2.288)
(2.130)
(2.060)
‐3.564 ‐3.476 ‐3.455
2.871
2.618
2.364
3.750
Net 0.500 ‐0.914 ‐0.415 0.516
‐0.471
0.189
2.466
‐3.316
‐2.405
‐2.383
‐2.353 ‐2.308
33 Wealth (2.056) (1.992) (1.961)
Income (1.859)
(2.810)
(2.757)
(2.731) (3.105)
(4.340)
0.468*** 0.474*** 0.461*** 0.518***
0.171
0.211
0.217 1.127** 1.316** 1.281** 1.085**
(0.130) (0.131) (0.132)
(0.200)
(0.202)
(0.198) (0.311)
Pension DB Pen 0.264 0.278
DC Pen (0.143)
0.559
(0.405)
(4.385)
(0.392)
(4.498)
(0.374)
(4.854)
(0.333)
0.275
‐0.986* ‐0.573 (0.308) (0.348)
(0.359)
(0.474)
(0.482) (0.524)
‐0.372 ‐0.352 ‐0.293 ‐0.604 ‐0.198 (0.344) (0.355)
(0.377)
(0.500)
(0.531) (0.585)
DB+DC Pen ‐0.609 ‐0.585 ‐0.571 ‐2.572 ‐2.197 (0.743) (0.758)
(0.767)
(1.636)
(1.637) (1.604)
Health Ins. Employer ‐0.849
‐1.014
(0.720)
Spouse Gov.plan ‐0.825
‐0.410 ‐2.905** ‐2.252* ‐2.443*
(0.896)
(0.981)
(1.130)
‐
2.915*** ‐2.351** ‐2.289**
(0.720)
(0.812)
(0.821)
‐2.193 ‐1.376 ‐1.472 (3.063)
(3.181)
(2.863)
‐2.482** ‐2.586**
11.42
12.82
(0.741)
(0.889) (0.903)
(8.623)
(7.922)
‐1.301
‐1.478
‐2.434* ‐2.310
12.26
13.77
(0.777)
(0.776)
(1.188) (1.179)
(8.454)
(7.850)
‐0.601
‐0.275
‐0.619 12.74
13.74
‐1.567 ‐1.279
34 (0.896)
(1.017)
(1.408) (1.453)
(8.568)
(7.840)
Other ‐1.635
‐1.668
‐1.400 16.07
16.91*
(1.163)
(1.171)
(1.358) (1.376)
(8.740)
(8.245)
Obs. 5855 5461 5391 5391
5350
2992
2712
2663
2663 2620
1695
1451
1432
1432
1318
R‐squared 0.005 0.011 0.013 0.014
0.020
0.010
0.014
0.016
0.019 0.031
0.013
0.036
0.049
0.061
0.084
Notes: ‐1.519
all models include sex, education, race and marital status; models 5, 10, 15 include additional job dummies for industry/occupation/physical effort required omitted categories: Reduce ‐ None ; Health ‐ Excellent ; Pension ‐ None ; Health Insurance ‐ None. Income is $000s p.a. / hrs per wk. Net Wealth is in $million Table 5 35 age 51‐59 at t
age 60‐64 at t age 65+ at t
Outcome: Retire by t+2 Mean: 0.12
Outcome: Retire by t+2 Mean: 0.42
Outcome: Retire by t+2 Mean: 0.64 Reduce 1 2 3 4
5
6
7
8 9
10
11
12
13
14
Constrain ‐0.00277 ‐0.0119 ‐0.000478 0.00004960.000612 ‐0.0832
‐0.0884*
‐0.0307 ‐0.0297
‐0.0382
‐0.107*
‐0.0766
0.0379
0.0380
0.0515 (0.029) (0.031) (0.019) (0.043)
(0.031) (0.031)
(0.033)
(0.052)
(0.050)
(0.052)
(0.051)
(0.050) Free ‐
0.0604** ‐0.0665** ‐0.0279* ‐0.0314* ‐0.0274 ‐0.160***‐0.139*** ‐0.0829** ‐0.0835** ‐0.0737* ‐0.108* ‐0.0820* ‐0.0274 ‐0.0245 0.00314 (0.019) (0.020) (0.012) (0.031)
(0.039)
(0.032)
(0.035) Health (0.019)
(0.012)
(0.020)
(0.014)
(0.044)
(0.031)
(0.027) (0.027)
(0.030)
(0.041)
(0.031)
Very Good ‐0.00530 ‐0.0171 ‐0.0177* ‐0.0185
0.0400*
0.0426* 0.0428*
0.0462*
0.0129
0.0435
0.0470
0.0567 (0.014) (0.009) (0.009)
(0.009)
(0.018)
(0.018) (0.018)
(0.019)
(0.024)
(0.029)
(0.030)
(0.030) Good 0.0128 0.00287
0.00152
0.000155
0.0638*** 0.0513* 0.0511*
0.0479*
0.0485*
0.0746*
0.0791*
0.0832* (0.015) (0.010) (0.010)
(0.010)
(0.017)
(0.021)
(0.022)
(0.024)
(0.034)
(0.035)
(0.034) Fair 0.0922*** 0.0360*
0.0357*
0.0332
0.115*** 0.0876** 0.0877** 0.0876**
0.0937*** 0.0519
0.0565
0.0624 (0.020) (0.018)
(0.019)
(0.023)
(0.026) (0.026)
(0.026)
(0.025)
(0.041)
(0.042)
(0.049) Poor 0.191*** 0.0775*
0.0753*
0.0750*
0.178*** 0.0791 0.0737
0.0486
0.151*** 0.0611
0.0641
0.0879 (0.038) (0.036) (0.036)
(0.037)
(0.039)
(0.072)
(0.076)
(0.037)
(0.095)
(0.102) 0.300 0.106 Net Wealth (0.017) (0.021) (0.071) (0.095)
0.112
0.113
0.173
‐0.0341 ‐0.0323
‐0.0398
0.433*** 0.0919*
0.0902*
36 15 0.114** (0.261) (0.125) (0.128)
Income ‐0.0340*** ‐0.000761 ‐0.000583 ‐0.000884
‐0.0957***‐0.00329 ‐0.00366 ‐0.00116
‐0.188*** ‐0.0121* ‐0.0122* ‐0.0116 (0.010) (0.003) (0.019)
(0.042)
Pension DB Pen 0.0519*** 0.0618*** 0.0576***
0.128*** 0.130*** 0.127***
0.0463
0.0389
0.0459 (0.009) (0.009)
(0.017) (0.019)
(0.020)
(0.039)
(0.037)
(0.038) DC Pen ‐0.00707 0.00190
0.00305
0.00627 0.00790
0.00402
‐0.0134
‐0.0212
‐0.00268 (0.009) (0.010)
(0.010)
(0.017) (0.017)
(0.019)
(0.027)
(0.026)
(0.025) DB+DC Pen 0.0417 0.0517*
0.0458*
0.0437 0.0442
0.0245
0.00351
‐0.00349 0.00758 (0.021) (0.021)
(0.020)
(0.038) (0.039)
(0.039)
(0.071)
(0.075)
Health Ins. Employer ‐0.0162
‐0.0209
‐0.00245 ‐0.00153
0.152
0.0912 (0.015)
(0.016)
(0.027)
(0.029)
(0.078)
(0.081) Spouse 0.0241
0.0200
‐0.000341 0.0115
0.0858
0.00874 (0.018)
(0.019)
(0.034)
(0.036)
(0.087)
(0.093) Gov.plan 0.0498
0.0247
0.0697
0.0897
0.141
0.0918 (0.052)
(0.053)
(0.068)
(0.072)
(0.077)
(0.078) Other ‐0.00286 ‐0.00612
‐0.0102
0.00843
0.00808
‐0.0575 (0.020)
(0.045)
(0.045)
(0.135)
(0.137) (0.003)
(0.121)
(0.003)
(0.179)
(0.038) (0.007) (0.038)
(0.007)
(0.038)
(0.009)
(0.101)
(0.042)
(0.006)
(0.042)
(0.006)
(0.009)
(0.080) 37 (0.007) (0.020)
(0.041) Obs. 7400 6895 6103 6103
6051
6632
6161
3749
3749
3683
7068
6604
2020
2020
1816 R‐squared 0.049 0.076 0.036 0.039
0.046
0.039
0.093
0.034
0.035
0.053
0.032
0.146
0.028
0.031
0.052 Notes: all models include sex, education, race and marital status; models 5, 10, 15 include additional job dummies for industry/occupation/physical effort required omitted categories: Reduce ‐ None ; Health ‐ Excellent ; Pension ‐ None ; Health Insurance ‐ None. Income is hrly wage *wks per year. Net Wealth is in $million Table 6 age 51‐59 at t
age 60‐64 at t age 65+ at t
Outcome: Hours Change [all people] Mean: ‐3.88
Outcome: Hours Change [all people] Mean: ‐6.24
Outcome: Hours Change [all people] Mean: ‐2.20 Reduce 1 2 3 4
5
6
7
8 9
11
12
Constrain 0.267 0.318 ‐0.218 ‐0.202
‐0.151
0.0266
0.526
0.571
0.651
0.363
‐0.366
‐0.403
‐0.0156
0.119
‐0.194 (0.579) (0.710) (0.902) (0.874)
(0.898)
(0.736)
(0.765)
(1.015) (1.021)
(1.145)
(0.608)
(0.606)
(1.856)
(1.840)
(1.737) Free 1.162* 1.315* 1.224 1.396
1.066
1.725*
2.322*
3.159** 2.967**
2.454*
‐1.030*
‐1.211*
‐0.416
‐0.454
‐0.914 (0.479) (0.532) (0.758) (0.773)
(0.782)
(0.725)
(0.872)
(1.075) (1.178)
(0.400)
(0.466)
(1.054)
(1.075)
(1.232) Health Very Good 0.643 0.745 0.769
0.791
‐1.589*
‐1.968** ‐1.905** ‐1.978**
‐0.265
0.399
0.354
0.206 (0.479) (0.539) (0.540)
(0.575)
(0.698)
(0.706) (0.553)
(1.208)
(1.238)
(1.235) Good ‐0.0387 ‐0.0826 ‐0.0188
0.0148
‐1.951** ‐3.062*** ‐3.076*** ‐3.092***
0.0240
‐0.238
‐0.214
‐0.143 (0.534) (0.579) (0.581)
(0.590)
(0.660)
(0.826) (0.537)
(1.061)
(1.050)
(1.110) Fair ‐1.997* ‐3.246** ‐3.211** ‐3.282**
‐0.978
‐3.297** ‐3.393*** ‐3.381**
‐0.160
‐1.277
‐1.371
‐0.654 (1.095)
10
13
14
(0.706)
(0.823)
(0.702)
(0.864)
38 15 (0.838) (1.062) (1.073)
(1.092)
(0.844)
(0.977) (0.971)
(1.035)
(0.558)
(1.537)
(1.575)
(1.714) Poor 0.751 ‐5.190* ‐5.092*
‐5.153*
2.185
‐1.561 ‐1.754
‐0.737
0.524
0.563
0.478
‐1.006 (1.213) (1.956) (1.998)
(2.060)
(1.322)
(3.606) (3.613)
(3.635)
(0.577)
(3.255)
(3.361)
(3.277) Net Wealth ‐2.612 ‐5.739 ‐6.078
‐6.160
4.710
0.304
0.384
0.763
5.139*
‐1.363
‐1.343
‐1.147 (2.945) (5.476) (5.639)
(5.141)
(2.412)
(2.096) (2.075)
(2.323)
(2.454)
(1.759)
(1.787)
(1.924) Income ‐0.485 0.439* 0.433*
0.492**
‐2.361*** 0.422* 0.436*
0.740*
‐1.788*
1.396*** 1.361*** 1.301*** (0.271) (0.166) (0.167)
(0.173)
(0.557)
(0.194)
(0.280)
(0.855)
(0.341)
Pension DB Pen ‐1.079* ‐1.618** ‐1.660**
‐7.549*** ‐6.958*** ‐6.874***
‐5.208*** ‐3.893** ‐4.108* (0.483) (0.494)
(0.530)
(0.675) (1.370)
DC Pen 0.530 0.0322
‐0.0598
‐2.775*** ‐2.217** ‐2.096**
‐4.055*** ‐2.948** ‐3.150** (0.508) (0.500)
(0.534)
(0.712) (0.739)
(0.729)
(1.030)
(1.080)
(1.045) DB+DC Pen ‐1.186 ‐1.732
‐1.552
‐5.465** ‐4.962*
‐3.577
‐4.099
‐2.821
‐2.620 (1.038) (1.036)
(0.992)
(1.937) (1.871)
(3.311)
(3.293)
(3.364) Health Ins. Employer 1.515
1.556
‐2.636*
‐2.895*
2.519
5.152 (1.077)
(1.101)
(1.068)
(1.244)
(6.368)
(5.796) Spouse ‐0.414
‐0.519
‐1.721
‐2.196
5.794
9.322 (0.197) (0.325)
(0.766)
(1.967)
(0.701)
(1.415)
(1.557) 39 (0.351) (1.191)
(1.203)
(1.301)
(1.490)
(6.538)
(6.101) Gov.plan ‐1.655
‐0.972
‐1.033
‐2.302
5.215
7.529 (2.198)
(2.290)
(2.354)
(2.531)
(6.366)
(5.775) Other 1.447
1.666
‐1.057
‐1.720
10.96
12.65 (1.452)
(1.512)
(1.615)
(1.655)
(7.639)
(7.276) Obs. 7328 6858 6075 6075
6023
6512
6099
3707
3707
3642
6896
6535
1976
1976
1786 R‐squared 0.007 0.011 0.025 0.027
0.033
0.014
0.040
0.046
0.047
0.065
0.004
0.020
0.037
0.043
0.067 Notes: all models include sex, education, race and marital status; models 5, 10, 15 include additional job dummies for industry/occupation/physical effort required omitted categories: Reduce ‐ None ; Health ‐ Excellent ; Pension ‐ None ; Health Insurance ‐ None. Income is hrly wage *wks per year. Net Wealth is in $million Table 7 Hours Change [workers] Retire by t+2 Hours Change [all people] Age groups
Age groups
Age groups
51‐59 at t
60‐64 at t
65+ at t
51‐59 at t
60‐64 at t 65+ at t
51‐59 at t
60‐64 at t
65+ at t
Ability to reduce (omitted: None) Constrained Reduce ‐0.794
‐1.202
‐0.738
‐0.00174
‐0.0404 0.0536
‐0.612
0.0882
‐0.372
(0.586)
(0.992)
(1.489)
(0.021)
(0.031) (0.052)
(0.997)
(1.078)
(1.780)
Free Reduce ‐0.235
‐0.282
‐1.329
‐0.0194
‐0.0749* 0.00834
0.278
2.557*
‐0.943
(0.581)
(0.831)
(1.105)
(0.015)
(0.033) (0.036)
(0.901)
(1.250)
(1.273)
40 Health (omitted: Excellent) Very Good ‐0.314
‐0.668
2.099
‐0.0193*
0.0532** 0.0487
0.348
‐1.841*
0.385
(0.438)
(0.671)
(1.156)
(0.009)
(0.019) (0.029)
(0.593)
(0.750)
(1.209)
Good 0.00960
‐1.693*
2.540*
0.0122
0.0760** 0.0743*
‐0.692
‐3.701***
0.0487
(0.362)
(0.805)
(1.174)
(0.011)
(0.023) (0.034)
(0.675)
(0.983)
(1.151)
Fair ‐0.916
0.195
1.952
0.0406*
0.0869** 0.0567
‐3.685***
‐2.798*
‐0.548
(0.554)
(1.177)
(1.475)
(0.018)
(0.029) (0.047)
(0.996)
(1.280)
(1.649)
Poor ‐0.0133
4.273
3.378
0.103*
0.127 0.0870
‐5.815*
‐2.300
‐0.896
(1.871)
(3.717)
(2.273)
(0.041)
(0.084) (0.102)
(2.341)
(4.274)
(3.242)
Wealth/Income Net Wealth [$million] ‐2.525
‐1.914
4.037
0.0768
‐0.0405 0.114**
‐4.001
0.973
‐1.046
(1.305)
(3.079)
(4.398)
(0.100)
(0.033) (0.042)
(3.766)
(2.304)
(1.935)
Income [$000s p.a. / hrs per wk] 0.476 0.606
1.132**
0.00276
‐0.00526 ‐0.0117
0.256
0.743*
1.348***
(0.247)
(0.330)
(0.332)
(0.005)
(0.007) (0.007)
(0.253)
(0.308)
(0.357)
Pension on job (omitted: None) DB Pension 0.405 ‐1.491*
‐2.876**
0.0383***
0.111*** 0.0505
‐0.767
‐7.100***
‐5.214**
(0.383)
(0.560)
(1.028)
(0.010)
(0.019) (0.040)
(0.584)
(0.720)
(1.548)
DC Pension ‐0.137
‐1.097
‐2.491**
‐0.0127
‐0.000875 0.00164
0.740
‐2.681**
‐4.013***
(0.439)
(0.617)
(0.723)
(0.011)
(0.021) (0.026)
(0.608)
(0.815)
(1.019)
41 DB+DC Pension 0.729 ‐2.763
‐1.818
0.0242
‐0.0210 ‐0.00740
0.460
‐3.022
‐2.854
(0.533)
(2.004)
(2.696)
(0.026)
(0.041) (0.072)
(1.094)
(2.131)
(3.099)
Observations 4295 2196
1287
4864
3072 1776
4841
3045
1746
R‐squared 0.021 0.038
0.072
0.050
0.056 0.048
0.035
0.067
0.059
Table 8 42 Outcome: Retire by t+2 OLS Probit 51‐59 at t 60‐64 at t 65+ at t 51‐59 at t 60‐64 at t 65+ at t Ability to reduce (omitted: None) Constrained Reduce ‐0.00174 ‐0.0404 0.0536 0.0025 ‐0.0374 0.0514 (0.021) (0.031) (0.052) (0.0178) (0.0337) (0.053) Free Reduce ‐0.0194 ‐0.0749* 0.00834 ‐0.0260 ‐0.0760* 0.0032 (0.015) (0.033) (0.036) (0.0122) (0.0315) (0.0378) Health (omitted: Excellent) Very Good ‐0.0193* 0.0532** 0.0487 ‐0.0148 0.0506* 0.0597 (0.009) (0.019) (0.029) (0.0085) (0.0202) (0.0343) Good 0.0122 0.0760** 0.0743* 0.0003 0.0517* 0.0907* (0.011) (0.023) (0.034) (0.0086) (0.0239) (0.0403) Fair 0.0406* 0.0869** 0.0567 0.0311 0.0964** 0.0690 (0.018) (0.029) (0.047) (0.0188) (0.0305) (0.0587) Poor 0.103* 0.127 0.0870 0.0771* 0.0603 0.0955 (0.041) (0.084) (0.102) (0.0387) (0.0841) (0.1153) Wealth/Income Net Wealth [$million] 0.0768 ‐0.0405 0.114** 0.0732 ‐0.0698 0.1273** 43 (0.100) (0.033) (0.042) (0.0558) (0.083) (0.049) Income [$000s p.a. / hrs per wk] 0.00276 ‐0.00526 ‐0.0117 ‐0.0008 ‐0.0019 ‐0.0205 (0.005) (0.007) (0.007) (0.0029) (0.0109) (0.0151) Pension on job (omitted: No pension) DB Pension 0.0383*** 0.111*** 0.0505 0.0550*** 0.1303*** 0.0462 (0.010) (0.019) (0.040) (0.01) (0.0216) (0.0404) DC Pension ‐0.0127 ‐0.000875 0.00164 0.0027 0.0063 ‐0.0047 (0.011) (0.021) (0.026) (0.0111) (0.0215) (0.0273) DB+DC Pension 0.0242 ‐0.0210 ‐0.00740 0.0520 0.0260 0.0055 (0.026) (0.041) (0.072) (0.0244) (0.042) (0.084) Table 9 44 Hours change, all people: [t+2] ‐ t OLS Int Reg 51‐59 at t 60‐64 at t 65+ at t 51‐59 at t 60‐64 at t 65+ at t Ability to reduce (omitted: None) Constrained Reduce ‐0.612 0.0882 ‐0.372 ‐0.508 0.664 ‐1.344 (0.997) (1.078) (1.780) (1.134) (1.496) (2.440) Free Reduce 0.278 2.557* ‐0.943 0.570 3.975* ‐0.845 (0.901) (1.250) (1.273) (1.021) (1.703) (1.674) Health (omitted: Excellent) Very Good 0.348 ‐1.841* 0.385 0.446 ‐2.508* 0.000359 (0.593) (0.750) (1.209) (0.658) (0.986) (1.672) Good ‐0.692 ‐3.701*** 0.0487 ‐0.864 ‐5.007*** ‐0.920 (0.675) (0.983) (1.151) (0.763) (1.329) (1.579) Fair ‐3.685*** ‐2.798* ‐0.548 ‐4.349*** ‐3.995* ‐1.063 (0.996) (1.280) (1.649) (1.174) (1.737) (2.236) Poor ‐5.815* ‐2.300 ‐0.896 ‐7.143* ‐5.141 ‐1.887 (2.341) (4.274) (3.242) (2.777) (6.138) (4.339) Wealth/Income Net Wealth [$million] ‐4.001 0.973 ‐1.046 ‐4.573 1.341 ‐2.888 45 (3.766) (2.304) (1.935) (4.310) (2.766) (3.432) Income [$000s p.a. / hrs per wk] 0.256 0.743* 1.348*** 0.277 0.889* 1.567*** (0.253) (0.308) (0.357) (0.278) (0.377) (0.427) Pension on job (omitted: No pension) DB Pension ‐0.767 ‐7.100*** ‐5.214** ‐0.720 ‐8.685*** ‐5.480* (0.584) (0.720) (1.548) (0.667) (1.077) (2.089) DC Pension 0.740 ‐2.681** ‐4.013*** 1.119 ‐2.252* ‐3.817** (0.608) (0.815) (1.019) (0.685) (1.110) (1.309) DB+DC Pension 0.460 ‐3.022 ‐2.854 0.632 ‐2.449 ‐2.913 (1.094) (2.131) (3.099) (1.268) (2.883) (3.954) Table 10 46 Outcome: Retire by t+2 51‐59 at t
60‐64 at t
65+ at t
Ability to reduce (omitted: None) Constrained Reduce [Wave 1] ‐0.00174
‐0.0404
0.0536
(0.021)
(0.031)
(0.052)
Free Reduce [Wave1] ‐0.0194
‐0.0749*
0.00834
(0.015)
(0.033)
(0.036)
Any Reduce [Wave 1] ‐0.00635
‐0.0773***
‐0.000607
(0.012)
(0.018) (0.025)
Any Reduce [Contemp.] 0.00170
‐0.0249
‐0.0448
(0.010)
(0.015)
(0.026)
Health (omitted: Excellent) Very Good ‐0.0193*
‐0.0196*
‐0.0193*
0.0532**
0.0505** 0.0560***
0.0487
0.0538
0.0730*
(0.009)
(0.009)
(0.009)
(0.019)
(0.018) (0.016)
(0.029)
(0.029)
(0.030)
Good 0.0122
0.0125
0.0101
0.0760**
0.0762** 0.0719**
0.0743*
0.0829*
0.0847*
(0.011)
(0.011)
(0.010)
(0.023)
(0.022) (0.021)
(0.034)
(0.034)
(0.039)
Fair 0.0406*
0.0395*
0.0352*
0.0869**
0.0866** 0.0829**
0.0567
0.0606
0.0995*
(0.018)
(0.017)
(0.015)
(0.029)
(0.029) (0.026)
(0.047)
(0.046)
(0.041)
Poor 0.103*
0.102*
0.105*
0.127
0.117 0.145
0.0870
0.0931
0.100
(0.041)
(0.041)
(0.040)
(0.084)
(0.077) (0.091)
(0.102)
(0.101)
(0.100)
47 Wealth/Income Net Wealth [$million] 0.0768
0.0698
0.0539
‐0.0405
‐0.0266 ‐0.0177
0.114**
0.105*
0.151
(0.100)
(0.092)
(0.086)
(0.033)
(0.037) (0.041)
(0.042)
(0.044)
(0.175)
Income [$000s p.a. / hrs per wk] 0.00276
0.00290
0.000825
‐0.00526
‐0.00454 ‐0.0130
‐0.0117
‐0.0112
‐0.00787
(0.005)
(0.005)
(0.001)
(0.007)
(0.007) (0.009)
(0.007)
(0.007)
(0.005)
Pension on job (omitted: No pension) DB Pension 0.0383***
0.0404***
0.0439***
0.111***
0.108*** 0.123***
0.0505
0.0489
0.0428
(0.010)
(0.010)
(0.010)
(0.019)
(0.018) (0.019)
(0.040)
(0.041)
(0.041)
DC Pension ‐0.0127
‐0.0116
‐0.00673
‐0.000875
0.00482 0.0125
0.00164
‐0.000164
‐0.0254
(0.011)
(0.011)
(0.010)
(0.021)
(0.020) (0.019)
(0.026)
(0.026)
(0.023)
DB+DC Pension 0.0242
0.0249
0.0299
‐0.0210
‐0.00613 ‐0.00713
‐0.00740
‐0.00280
‐0.0377
(0.026)
(0.026)
(0.024)
(0.041)
(0.040) (0.040)
(0.072)
(0.071)
(0.070)
Table 11 48 Hours change, all people: [t+2] ‐ t 51‐59 at t
60‐64 at t
65+ at t
Ability to reduce (omitted: None) Constrained Reduce [Wave 1] ‐0.612
0.0882
‐0.372
(0.997)
(1.078)
(1.780)
Free Reduce [Wave1] 0.278
2.557*
‐0.943
(0.901)
(1.250)
(1.273)
Any Reduce [Wave 1] ‐0.187
1.931** ‐0.420
(0.512)
(0.675) (0.817)
Any Reduce [Contemp.] ‐0.196
2.157**
1.645*
(0.450)
(0.660)
(0.795)
Health (omitted: Excellent) Very Good 0.348
0.386
0.225
‐1.841*
‐1.792* ‐1.860*
0.385
0.196
‐0.480
(0.593)
(0.594)
(0.570)
(0.750)
(0.752) (0.704)
(1.209)
(1.178)
(1.216)
Good ‐0.692
‐0.656
‐0.690
‐3.701***
‐3.582***
‐2.789**
0.0487
‐0.332
‐0.703
(0.675)
(0.677)
(0.588)
(0.983)
(0.966) (0.871)
(1.151)
(1.170)
(1.191)
Fair ‐3.685***
‐3.635***
‐3.349***
‐2.798*
‐2.759* ‐2.408
‐0.548
‐1.024
‐1.008
(0.996)
(0.973)
(0.941)
(1.280)
(1.288) (1.248)
(1.649)
(1.675)
(1.668)
Poor ‐5.815*
‐5.820*
‐5.788**
‐2.300
‐1.433 ‐3.546
‐0.896
‐0.997
‐2.241
(2.341)
(2.332)
(2.147)
(4.274)
(4.021) (4.646)
(3.242)
(3.239)
(3.728)
49 Wealth/Income Net Wealth [$million] ‐4.001
‐3.857
‐3.444
0.973
0.888 0.315
‐1.046
‐0.765
‐7.625
(3.766)
(3.615)
(3.717)
(2.304)
(2.077) (2.872)
(1.935)
(1.998)
(7.533)
Income [$000s p.a. / hrs per wk] 0.256
0.263
0.0750
0.743*
0.727* 1.087*
1.348***
1.325***
1.103*
(0.253)
(0.253)
(0.060)
(0.308)
(0.310) (0.521)
(0.357)
(0.370)
(0.460)
Pension on job (omitted: No pension) DB Pension ‐0.767
‐0.906
‐1.088
‐7.100***
‐7.034***
‐6.495***
‐5.214**
‐5.251**
‐5.406***
(0.584)
(0.562)
(0.573)
(0.720)
(0.680) (0.719)
(1.548)
(1.579)
(1.545)
DC Pension 0.740
0.634
0.378
‐2.681**
‐2.875***
‐2.383**
‐4.013***
‐3.858***
‐3.206**
(0.608)
(0.588)
(0.579)
(0.815)
(0.807) (0.860)
(1.019)
(0.963)
(1.001)
DB+DC Pension 0.460
0.327
0.0272
‐3.022
‐3.545 ‐2.563
‐2.854
‐2.816
‐0.450
(1.094)
(1.090)
(1.085)
(2.131)
(2.127) (2.043)
(3.099)
(3.086)
(2.989)
Table 12 50 Paper2:Laborsupplyflexibilityandportfoliochoice
Introduction23
Life expectancy in the United States continues to increase (Miniño, Murphy et al. 2011), Defined Benefit pensions are declining in prevalence (Poterba, Venti et al. 2009), and Social Security provides a lower income replacement rate now than in previous generations (Butrica, Iams et al. 2004) . In order to sustain their high standard of living at older ages, Americans need to accumulate significant wealth over their lifetime by working, saving and investing. If opportunities exist, they may also wish to delay retirement or supplement their retirement income with post‐retirement part‐time employment. This paper provides new insights into three aspects of this lifetime work/save/invest decision process. At older ages, workers may find it difficult to sustain the same workload as at younger ages; this paper provides evidence on job stressors over time and the flexibility workers have to adjust their workload. At older ages, a negative shock to an individual’s savings through unexpected costs or poor investment performance can be very damaging; this paper provides new evidence on the extent to which workers expect to be able to compensate for such shocks with increased work. Finally, most individuals make a series of investment decisions over their lifetime, and these decisions may be related to their labor behavior and expectations; this paper provides an empirical exploration of the relatively under‐
researched link between labor behavior and portfolio choice. More specifically, our empirical work investigates how an individual’s ability to adjust their lifetime labor supply affects their stock holding behavior. Building on a limited existing literature, we examine the impact of labor supply flexibility at the intensive margin (the ability to adjust hours of work within a career) on an individual’s investment in the stock market. Going beyond previous research, we also examine how factors that affect labor supply at the extensive margin (job characteristics and demands that make it easier or more difficult to continue a career at older ages) are related to stock investment behavior. In contrast with previous work, we find little evidence of any impact from intensive margin flexibility. However, we do find suggestive evidence supporting a link between extensive margin flexibility and investment behavior. BackgroundandPreviousLiterature
Theexistenceof‘rigidities’inthelabormarket,andtheireffectsonlaboroutcomes
While the simplest lifecycle models of labor supply assume that workers are free to adjust their personal labor supply over time, many workers face labor market rigidities: their employer offers a limited choice 23
Thanks are due to the Roybal Center for Financial Decision Making, which provided funding for our survey through a pilot project, and to Joanne Yoong and Angela Hung for providing access to data from their asset allocation survey 51 of work options, and switching employers can be prohibitively costly (particularly at older ages)24. Several papers find indirect evidence of constraints on labor supply by examining work transitions at older ages. Comparisons between the labor supply of employees and the self‐employed show that employees do not transition from full‐time work to part‐time work at the same rates as the self‐
employed, implying that there are constraints preventing them from doing so (Hurd 1996). Similarly, employees have been shown to be more likely to withdraw fully from the labor market than the self‐
employed, perhaps due in part to a lack of suitable part‐time work options25 (Zissimopoulos, Maestas et al. 2007). Research on ‘bridge jobs’ – part‐time employment taken as a bridge to retirement after leaving a ‘career job’ – finds that workers who switch employers at older ages receive significantly lower hourly wages than at their previous employer, and argues that these workers must therefore not have had the option available to continue to work for their previous employer on a part‐time basis (Gustman and Steinmeier 1985; Ruhm 1990). Switching jobs at older ages may be difficult for several reasons: job‐specific human capital is no longer valuable (Hurd 1996), relatively high health insurance and pension costs may dissuade employers from hiring older workers (Scott, Berger et al. 1995), and older job‐seekers may face discrimination (Lahey 2008). Thus, older workers may face real constraints on their labor supply decision: their employer may not allow them to adjust their labor in their current job, and it may not be feasible to effect an adjustment by switching to a new job. The effect of labor supply constraints on overall labor supply is theoretically ambiguous: some individuals might prefer to work part‐time, but due to constraints choose to work full time; other individuals might most prefer to work part‐time, but prefer retirement to full‐time work; the net result of these two effects is theoretically ambiguous. A study of the University of North Carolina system found that allowing the flexibility of a phased retirement plan had the net effect of reducing an individual’s labor supply (Allen, Clark et al. 2004). In contrast, a more general projection exercise suggested that allowing workers to work half‐time at half‐pay would increase net labor supply (Gustman and Steinmeier 2007). Two papers examine longitudinal effects of labor supply constraints using the Health and Retirement Study, in which workers are asked directly whether or not they are able to reduce their hours of work in their main job. Charles and DeCicca found that older workers who faced a strong constraint on labor supply in 1992 had increased an likelihood of retiring by 1996 (Charles and DeCicca 2007). Clift confirms that older workers with greater flexibility tend to have lower retirement hazard (particularly around Social Security’s early retirement age), but finds that by reducing their hours of work in their 50s, more flexible workers may simply be smoothing labor supply over time; thus, labor supply constraints may affect when an individual supplies labor without having much effect (negative or positive) on how much labor an individual supplies (Clift 2012). 24
For more detailed discussion of literature on labor supply flexibility and its effect on labor outcomes, see discussion by Clift (Clift 2012) 25
Although (as the authors note) a large part of the observed difference in labor supply behavior can be explained by reference to pension differences and other factors 52 Portfoliochoiceandtheroleoflaborsupply
Over the life cycle, individuals decide not only how much to work and how much to save, but also how to allocate those savings across a portfolio of assets that vary in their risk and return. Simple economic models of asset allocation between a ‘safe’ and ‘risky’ asset – assuming stochastic asset returns and constant relative risk aversion (CRRA) – predict that individuals should keep the same balance between safe and risky assets throughout the lifecycle; but this conflicts with the typical advice of financial professionals, that individuals should reduce their exposure to risk as they age (Samuelson 1969; Gollier 2005). However, various extensions of the standard CRRA lifecycle model can produce this type of time‐varying investment strategy. For example, an intuitive explanation for having higher risk tolerance in investment early in the life‐cycle is that if pension contributions can be adjusted over time, then an early wealth shock can be smoothed through adjustments to savings and consumption over a long period, while later wealth shocks require abrupt adjustments and steep declines in utility (Gollier 2005). A potentially more interesting and nuanced explanation for time‐varying investment strategies is examined in a number of papers examining the role of future labor income in determining optimal portfolio allocation strategy. Future income represents a large portion of a young person’s wealth, but it is not clear how it should be regarded in terms of expectations and risks. If future income is considered to be a ‘safe’ asset with finite horizon, and this wealth diminishes over time as the individual approaches retirement, then CRRA models suggest that the reduction in the riskless labor income asset over time should be met with a rebalancing of financial wealth away from risky assets towards riskless assets, in order to maintain the same balance of risk in the total wealth portfolio (Jagannathan and Kocherlakota 1996). However, treating future labor income in this way may be too simplistic – a number of different assumptions may be made to bring the modeling closer to reality. One modeling exercise shows that if future income has a great deal of prospective uncertainty early in a career, but becomes revealed over the course of a career, this might combine with the long‐term smoothing justification above to produce an investment trajectory that is conservative when future income is uncertain, rises as future income becomes more certain, and falls again as the ability to smooth future consumption and savings decreases (Gollier 2005). In addition to uncertainty, there may also be uninsurable risk involved in future labor income, and this risk may be heterogeneously distributed across individuals. One study uses the 1979‐1990 Panel of Individual Tax Returns to document significant heterogeneity in wage income and proprietary income streams, and combines this with prior research on the holding of employer stocks by employees in order to show differences in ‘background risk’ that might explain differences in portfolio behavior across different people (Heaton and Lucas 2000). A further issue with future labor income risk is the extent to which it is correlated with stock market risk, such that investments in stocks do not provide a diversified risk portfolio: if future labor income is highly correlated with the performance of the stock market, this will tend to reduce the optimal portfolio share invested in stocks. On the one hand, research to date suggests that labor income is not strongly tied to stock market performance for most people 53 (Jagannathan and Kocherlakota 1996; Cocco, Gomes et al. 2005); on the other hand, other research has found that modeling labor income and asset returns as jointly related to the business cycle leads to an improved match of simulated asset profiles over the life‐cycle with those empirically observed (Lynch and Tan 2011). Theeffectoflaborsupplyconstraintsonportfoliochoice
One other aspect of lifecycle labor income may affect portfolio allocation behavior: the degree to which people have an ability to adjust their labor supply (and thus labor income) over time in response to shocks. In the earliest full examination of the topic from a theoretical standpoint, researchers showed that in a continuous‐time consumption and portfolio allocation model, optimal portfolio allocation rules vary when labor supply is flexible, rather than predetermined over the lifecycle (Bodie, Merton et al. 1992). Specifically, a worker with flexibility over future labor supply should be able to bear more risk in their investments than people with fixed labor supply, because in the event of a shock a flexible worker can adjust not only their consumption/savings tradeoff, but also their leisure/consumption tradeoff – a worker can work more hours to make up for poor performance in their investments, rather than just cutting back on consumption.26 Additionally, the investment effect of labor flexibility may vary by age: earlier in life, a worker has a long period over which work hours can be increased very slightly to make up for a negative wealth shock; later in life, workers have a shorter period over which to adjust their labor supply and may also have less flexibility to do so. A subsequent theoretical paper confirms that workers with labor flexibility should be able to invest more riskily at all ages (until retirement) than workers without labor flexibility. Furthermore, an additional relevant source of labor supply flexibility is the ability to choose when to retire: flexibility to choose timing of retirement or partial retirement allows more risky assets to be held at older ages than otherwise would be advisable (Gollier 2005). From the literature discussed thus far, simple CRRA models of savings and investment behavior can be extended in various ways to justify the time‐varying portfolio allocation strategies typically proposed by financial advisors.27 The most important extensions focus on the role that future labor income plays as an asset, and the varying ability of workers to smooth past and present consumption and leisure in response shocks. The ability to adjust labor supply is a potentially important smoothing mechanism, reducing the lifetime utility losses associated with negative investment shocks (while also increasing the lifetime utility gains associated with positive investment shocks). To date, very little empirical work has addressed the relationship between labor supply flexibility and portfolio allocation. In an unpublished working paper, Benitez‐Silva conducts an empirical analysis of the 26
Recent work on the effects of the financial crisis find that households near retirement reacted by reducing consumption, working longer, and reducing bequests (Hurd and Rohwedder 2010) 27
Although it is beyond the scope of this paper, risk tolerance in investment behavior is also likely to be affected by other sources of background risk, such as medical expenditures. International differences in the medical expenditure risks over the life cycle have been cited as a reason for international differences in life cycle investments in risky assets (Kapteyn and Panis 2005); heterogeneity in the supplemental coverage held by Medicare recipients has been shown to be related to differences in risky asset holdings among older Americans (Goldman and Maestas 2012). 54 relationship between labor supply flexibility and portfolio decisions using panel data from HRS, alternately using self‐employment status and self‐reported ability to increase28 working hours as measures of labor supply flexibility for individuals aged 51 and over (Benitez‐Silva 2006). He chooses as his outcome of interest the logarithm of an individual’s wealth held in stocks, while attempting to control carefully for the individual’s level of wealth held in non‐risky assets. He finds that labor supply flexibility increases wealth held in stocks by 12‐14%. In contrast to Benitez‐Silva, we do not have the benefit of repeated observations over time for individuals in our dataset. However, our dataset is unique in containing information on both labor supply constraints and portfolio allocation for people aged 25‐67; as previously discussed, the effects of labor supply flexibility on portfolio choices are predicted to be larger earlier in the lifecycle29. We also attempt to control for heterogeneity in intrinsic risk‐aversion (potentially important for explaining variation in portfolio risk exposure). Finally, we use broader concepts of labor supply flexibility than Benitez‐Silva, examining flexibility on both the intensive and extensive margin. While the discussion to this point has focused on the ability of individuals to make small adjustments to the number of hours worked in a given period, an alternative paradigm involves workers with relatively fixed weekly work hours over their working life but with flexibility on the length of their career. While the results are open to interpretation30, we find that the variables pertaining to labor supply flexibility on the extensive margin have a stronger relationship to portfolio choice than the intensive margin flexibility variables. We now go on to present a description of our data sources and sample characteristics. There follows an array of descriptive statistics of interest; a section discussing the logic behind our choices in the empirical modeling; a section detailing our empirical results; and a section discussing our findings. We conclude with a conclusion. 28
He argues that the ability to reduce hours is less important for the interaction between labor supply and portfolio decisions, citing as supporting evidence that individuals did not react to the 2001 economic upturn with a reduction in hours worked 29
However, other factors (such as uncertainty about future income) may also be stronger earlier in the lifecycle 30
We do not have plausibly exogenous variation in job characteristics, reducing the level of causal inference we can draw; and the pathways through which our independent variables may affect investment behavior are also subject to alternative interpretations 55 DataDescriptionandSampleCharacteristics
TheAmericanLifePanel
For this research, we fielded a new survey31 in RAND’s American Life Panel (ALP)32, and combined this data with previous ALP surveys answered by the same group of respondents. The ALP is a large‐scale internet panel survey administered by RAND, fielded to approximately 5,000 participants aged 18 and above. The initial cohort of survey participants was recruited from the University of Michigan Survey Research Center’s random digit dial Monthly Survey, and has been supplemented with members of the Stanford / Abt SRBI National Survey Panel. In order to avoid a common sample bias issue for internet surveys, RAND provides internet access and an internet browsing device33 to the relatively small number of households that do not have their own access. Respondents are compensated approximately $40 per hour of survey‐taking. The ALP asks a detailed set of demographic questions each quarter, and these are used to construct sampling weights to align the ALP sample with the larger benchmark Current Population Survey (CPS). Efficient sampling weights require a parsimonious set of matching variables/categories. Given our particular interest in job characteristics and their relationship to portfolio choice, we include among our matching variables a parsimonious occupation variable, in addition to gender, age group, educational level and income level. After weighting, the basic descriptive statistics for our dataset are generally as expected, although our sample appears to be slightly more ALP CPS educated than the CPS (the difference in the Male 0.51 0.53 Married/Partnered variable is likely to be definitional in nature, with unmarried cohabiting partners included in the White 0.79 0.80 definition in ALP). Married/Partnered 0.70 0.64 Age 25‐40 0.41 0.41 41‐50 51‐60 61‐67 0.26 0.25 0.08 0.27 0.24 0.08 Education High School or Less 0.34 Some College but no BA 0.31 Bachelor’s Degree 0.22 Higher Degree 0.14 N 1341 31
0.40 0.27 0.21 0.11 83,402 The full survey protocol for the new survey we fielded is reproduced in Appendices AppendixA For full information on the American Life Panel, see https://mmicdata.rand.org/alp/ 33
Laptop or Microsoft TV2 32
56 Occupation Management Business and Financial Operations Computer and Mathematical Architecture and Engineering Life, Physical, and Social Science Community and Social Services Legal Education, Training, and Library Arts, Design, Entertainment, Sports, and Media Healthcare Practitioner and Technical Healthcare Support Protective Service Food Preparation and Serving Related Building and Grounds Cleaning and Maintenance Personal Care and Service Sales and Related Office and Administrative Support Farming, Fishing, and Forestry Construction and Extraction Installation, Maintenance, and Repair Production Transportation and Material Moving ALP 0.12 0.06 0.03 0.02 0.01 0.02 0.01 0.07 0.02 0.06 0.04 0.02 0.03 0.02 0.02 0.10 0.13 0.01 0.03 0.04 0.08 0.05 CPS 0.11 0.05 0.03 0.02 0.01 0.02 0.01 0.07 0.02 0.06 0.02 0.02 0.04 0.04 0.03 0.10 0.13 0.01 0.06 0.04 0.06 0.06 N 1341 83,402
Table 14: Occupations
Table 13: Demographics
57 Oursurvey
In August 2011, we fielded a new survey in the ALP focusing on labor supply flexibility, and combined the data with data from a previously fielded survey that focused on detailed asset holdings. The survey was fielded to 1527 individuals who had responded to the previous asset survey and were aged between 25 and 67 inclusively. Our survey received 1498 responses. Of these, 157 were dropped because they indicated that they were not currently working. This gives a base sample of 1341 for examining questions about labor supply flexibility and other job characteristics. Of this 1341, 102 had not given complete information in the asset holding survey to determine their complete wealth measure. Of the remaining 1239, a further 43 did not provide sufficient information to determine whether they held stocks, so cannot be used for determining stock holdings. An additional 126 do not provide enough information to calculate their full financial holdings, with one respondent providing enough information for gross wealth but not enough for net wealth. With the most restrictive sample selection, the sample characteristics were very similar to those in tables 1 and 2; however, the selected sample included a larger proportion of White respondents. Our new survey included several variables that also appear for the age 51+ population in the Health and Retirement Study. Of these, the key variables we will explore are: whether the individual can increase hours of work in their current job; whether the individual can reduce hours of work in their current job; how often the individual’s job requires lots of physical effort; how often the individual’s job requires intense concentration. We also included in the survey several new variables relating to labor supply flexibility, and the ability to adjust labor supply in response to negative wealth shocks: whether the individual can “buy” additional vacation days from their employer; whether the individual can “sell” surplus vacation days back to their employer; how far the individual agrees that, in their current job, it is difficult to keep up with changes in technology; how far the individual agrees with the statement “If I lost some of my savings, I would just have to make do with less when I retired”; and how far the individual agrees with the statement “If I lost some of my savings, I could always work longer to make up for it”.34 The first two are an extension of being able to increase or reduce hours of paid work, looking at being able to increase or reduce days of paid work. The third variable explores one of the ways in which workers may not be able to continue to work indefinitely: obsolescence of technology‐specific skills. The last two variables are novel, and attempt to understand whether people expect to be willing and able to work more intensely (or retire later) if they had a negative wealth shock, or whether they would mostly have to cut back on consumption. The logic expressed in the literature previously discussed is that individuals may consciously choose to invest in more risky assets precisely because they believe that they can mitigate potential negative consequences; these new variables attempt to capture that sentiment. 34ExactquestionphrasingisavailableinAppendices
Appendix A 58 Outcomemeasures
There are several potential measures that could be used to summarize portfolio choice. The most basic measure we will use is a simple binary outcome for whether or not an individual owns any stocks. We can also examine the intensity of stock holdings: in their seminal theoretical paper, Bodie et al. mention the proportion of financial wealth held in risky assets as one standard measure, and the proportion of ‘total’ wealth held in risky assets as the more relevant measure35 (where total wealth includes not only financial and real assets but also an estimation of the asset equivalent of future labor income) (Bodie, Merton et al. 1992). While seemingly straightforward, the measures above pose some questions for implementation. In order to focus on ‘active’ investment choices in liquid assets, we do not include assets held within Defined Contribution (DC) plans in the calculation of financial wealth held in risky assets; this prevents us from conflating employer‐driven pension effects with more active portfolio risk‐taking. Additionally, the measures of DC wealth and allocation across asset classes in our dataset contain a significant number of clear reporting errors; using measures of DC allocation and excluding the known errors would reduce our sample without producing measures in which we have confidence. Nonetheless, in sensitivity analysis we rerun our models incorporating DC assets on the reduced sample. For our main analyses, we define financial wealth as the sum of stocks, bonds, certificates of deposit, checking and savings account balances. Creating a variable describing the proportion of ‘total’ wealth held in risky assets poses additional challenges. Bodie et al.’s measure is imagined to be bounded by 0 and 1, but given debts and liabilities (particularly after the housing crash left some individuals with mortgage debt exceeding their home values), it is possible for individuals to have net asset wealth that is less than their stock holdings, or even to have negative net wealth; this would cause the ratio of stocks to net wealth to exceed 1 or be less than 0. This particular challenge could be overcome by focusing on gross wealth, rather than net wealth, ignoring debts and mortgages, but this would misrepresent an individual’s total wealth. Furthermore, we lack sufficient data to judge the asset‐equivalent value of future labor income, and therefore our ‘total’ wealth measure options fall far short of those suggested by Bodie et al.. We do not believe that the potential ‘total’ wealth measures are a sufficient theoretical improvement to justify the increased complexity described above. 35
Their concept of ‘total’ wealth includes a ‘human capital’ value placed on the potential future labor income in their theoretical model, divided into a ‘risk‐free’ and ‘risky’ portion to reflect the equivalent asset characteristics. Any attempt for us to do this in an empirical setting would be highly speculative. Nonetheless, our results relating portfolio choice to job characteristics can be interpreted in this spirit. 59 DescriptiveStatistics:ByGender,AgeandEducationLevel
Hoursandvacationflexibility
Between the ages of 25 and 60, there seems to be little change in the hours flexibility for men, hovering around 20% who can reduce weekly work hours and around 30% who can increase work hours. Women tend to have more flexibility to reduce hours than men at all ages, but women in the 51‐60 age group have less flexibility than those in the 25‐40 range, for both increasing or reducing hours. This may reflect women in the younger age group working more part‐time jobs in combination with raising children. An alternative explanation is that there are differences between cohorts, such that older cohorts have been subject to different opportunities throughout their lives and been placed in different types of work (or different labor force statuses) from younger cohorts.36 Women in the oldest age category (61‐67) seem to have the most flexibility, but it is not clear whether this reflects women switching to more flexible jobs at older ages, or that differences in retirement hazard by flexibility status (as seen in Clift(2012)) lead to a biased sample in this age category. Buying and selling of vacation days is not a common option for men or women, though it appears that the option to sell vacation days may be more prevalent than the option to buy. Can reduce weekly work hours Can increase weekly work hours Can 'buy' extra vacation days Can 'sell' back vacation days Male Female 25‐40 41‐50 51‐60 61‐67 25‐40 41‐50 51‐60 61‐67 0.15 0.14 0.23 0.25 0.34 0.30 0.25 0.38 0.26 0.28 0.31 0.19 0.34 0.37 0.25 0.37 0.06 0.03 0.03 0.04 0.04 0.07 0.02 0.02 0.11 0.15 0.06 0.12 0.05 0.10 0.07 0.10 Table 15: Hours and vacation flexibility by sex and age. All figures are survey weighted proportions. [N=1072] Male Female 25‐40 41‐50 51‐60 61‐67 25‐40 41‐50 51‐60 61‐67 Chances of having a flexible job when approaching retirement 0: 0‐49% 0.45 0.60 0.55 0.57 0.41 0.41 0.58 0.52 1: 50% 0.26 0.17 0.14 0.18 0.34 0.27 0.19 0.10 2: 51‐100% 0.29 0.23 0.31 0.24 0.26 0.32 0.23 0.38 Table 16: Expected future hours flexibility by sex and age. All figures are survey weighted proportions. [N=1072] In general, it appears that people with higher levels of education may work in less flexible jobs than those with lower education ‐ men with higher education are less likely to be allowed to increase their work hours, while women with higher education are less likely to be allowed to reduce their work hours. 36
Although the descriptive tables we present segmented by age groups generally seem to reflect plausible differences in work conditions, expectations and sorting over the lifecycle, all ‘age differences’ in our cross‐
sectional data admit of cohort‐difference explanations. This may be particularly relevant for women, given the significant rise in female labor force participation: in 1970, when the oldest of our sample were in their early 20s, the female labor force participation rate was 43.3%; in 2006, when the youngest of our sample were in their early 20s, the rate was 59.4%. (US Bureau of Labor Statistics 2011) 60 This may reflect a preponderance of hourly sales and service jobs among the lower educated, with more highly educated workers in more formal salaried positions. Can reduce weekly work hours Can increase weekly work hours Can 'buy' extra vacation days Can 'sell' back vacation days HS 0.21 0.33 0.02 0.13 Male Coll BA 0.12 0.21 0.30 0.21 0.04 0.06 0.07 0.10 MA+ 0.15 0.17 0.06 0.14 HS 0.34 0.26 0.04 0.07 Female Coll BA 0.35 0.27 0.38 0.31 0.04 0.04 0.09 0.07 MA+ 0.17 0.30 0.03 0.04 Table 17: Hours and vacation flexibility by sex and education. All figures are survey weighted proportions. [N=1072]
Male HS Coll BA MA+ Chances of having a flexible job when approaching retirement 0: 0‐49% 0.56 0.52 0.44 0.58 1: 50% 0.15 0.25 0.24 0.17 2: 51‐100% 0.29 0.23 0.32 0.24 HS 0.64 0.21 0.16 Female Coll BA MA+ 0.37 0.32 0.32 0.47 0.23 0.30 0.39 0.26 0.34 Table 18: Expected future hours flexibility by sex and education. All figures are survey weighted proportions. [N=1072]
61 Jobstressors
Of the job stressors, it appears that older men may have jobs that require intense concentration less often than younger men, while older women may have less physically demanding jobs than younger women. We see the expected negative education / physical effort gradient across both sexes. Men with high school education only may face more retirement pressure than those in other groups. Male Female 25‐40 41‐50 51‐60 61‐67 25‐40 41‐50 51‐60 61‐67 Coworkers make people feel they should retire by 65 0.16 0.13 0.13 0.23 0.12 0.16 0.13 0.05 Agree 0.69 0.78 0.82 0.67 0.72 0.76 0.78 0.88 Disagree 37
0.15 0.09 0.05 0.10 0.16 0.08 0.09 0.07 DNA My job requires lots of physical effort 0.32 0.28 All or Most time 0.25 0.31 Some time 0.42 0.41 No time or DNA My job requires intense concentration or attention 0.91 0.79 All or Most time 0.06 0.21 Some time 0.03 0.00 No time or DNA 0.34 0.27 0.39 0.76 0.21 0.03 0.14 0.28 0.58 0.28 0.23 0.49 0.18 0.35 0.46 0.21 0.23 0.56 0.17 0.27 0.56 0.79 0.16 0.05 0.79 0.16 0.04 0.81 0.19 0.00 0.79 0.19 0.01 0.82 0.15 0.04 Table 19: Job stressors by sex and age. All figures are survey weighted proportions. [N=1072] Male HS Coll BA Coworkers make people feel they should retire by 65 0.24 0.16 0.11 Agree 0.65 0.71 0.78 Disagree 0.11 0.14 0.11 DNA My job requires lots of physical effort 0.53 0.37 0.13 All or Most time 0.29 0.31 0.27 Some time 0.18 0.31 0.60 No time or DNA My job requires intense concentration or attention 0.83 0.83 0.81 All or Most time 0.11 0.16 0.16 Some time 0.06 0.01 0.03 No time or DNA MA+ 0.16 0.73 0.12 HS 0.12 0.73 0.15 Female Coll BA 0.15 0.76 0.09 0.15 0.73 0.12 0.14 0.81 0.05 0.08 0.21 0.70 0.37 0.26 0.37 0.29 0.31 0.40 0.16 0.25 0.59 0.11 0.22 0.67 0.86 0.13 0.01 0.77 0.20 0.04 0.79 0.17 0.05 0.83 0.15 0.02 0.87 0.13 0.00 MA+ Table 20: Job stressors by sex and education. All figures are survey weighted proportions. [N=1072] 37
Unless otherwise noted, all Agree/Disagree answers are recoded from a five point scale (Strong Agree, Agree, Disagree, Strong Disagree, Does Not Apply), and all ‘time’ answers are also recoded from a five point scale (All of the Time, Most of the Time, Some of the Time, None of the Time, Does Not Apply). We believe that for ‘time’ questions ‘Does Not Apply’ should be considered conceptually similar to a very strong negative (equivalent to ‘No time’, but the same cannot be said for Agree/Disagree questions 62 Jobrequirementbeliefs
With the job requirement beliefs, the physical fitness variable mirrors the physical demand variable from the previous section: higher importance for men, not much age gradient, and a negative relationship between education level and the importance of physical fitness. The importance of learning new skills tends to decline with age and increase with education, across both sexes. Men appear to find it increasingly difficult to keep up with changes in technology at older ages, and, interestingly, find it more difficult to keep up with changes in technology if they are more highly educated (perhaps because they are in jobs that involve more quickly changing technology). Male Female 25‐40 41‐50 51‐60 61‐67 25‐40 41‐50 51‐60 61‐67 Have to keep physically fit if you want to keep up 0.43 0.49 0.48 0.40 0.24 0.27 0.31 0.38 Agree 0.52 0.40 0.46 0.45 0.59 0.56 0.51 0.35 Disagree 0.05 0.11 0.06 0.15 0.18 0.16 0.18 0.26 DNA Have to keep learning new skills if you want to keep up 0.82 0.85 0.74 0.80 0.78 0.74 0.72 0.86 Agree 0.16 0.14 0.24 0.17 0.15 0.25 0.22 0.05 Disagree 0.02 0.02 0.02 0.02 0.06 0.02 0.06 0.09 DNA Difficult to keep up with changes in technology 0.17 0.22 0.32 0.38 0.20 0.24 0.21 0.29 Agree 0.78 0.76 0.64 0.58 0.72 0.74 0.74 0.67 Disagree 0.04 0.02 0.04 0.04 0.09 0.02 0.06 0.04 DNA Table 21: Job requirement beliefs by sex and age. All figures are survey weighted proportions. [N=1072] Male HS Coll BA Have to keep physically fit if you want to keep up 0.63 0.55 0.26 Agree 0.31 0.35 0.63 Disagree 0.06 0.10 0.11 DNA Have to keep learning new skills if you want to keep up 0.73 0.79 0.89 Agree 0.22 0.17 0.10 Disagree 0.05 0.04 0.02 DNA Difficult to keep up with changes in technology 0.21 0.22 0.31 Agree 0.74 0.74 0.67 Disagree 0.05 0.04 0.02 DNA MA+ 0.22 0.66 0.11 0.92 0.07 0.00 0.40 0.59 0.01 HS 0.36 0.49 0.15 0.71 0.22 0.07 0.21 0.70 0.09 Female Coll BA MA+ 0.33 0.53 0.14 0.75 0.22 0.03 0.20 0.74 0.06 0.22 0.63 0.15 0.87 0.10 0.04 0.24 0.74 0.02 0.21 0.54 0.24 0.76 0.18 0.07 0.26 0.70 0.04 Table 22: Job requirement beliefs by sex and education. All figures are survey weighted proportions. [N=1072]
63 Likelyresponsetosavingsshock
Unsurprisingly, older workers are more likely to agree that they would have to ‘make do with less’ in the event of a negative wealth shock, and less likely to agree that they could ‘work more to make up for it’. Education patterns are less clear‐cut, and could be presented different ways by different groupings of education or agreement categories: males with high education tend to be more likely to disagree that they would have to ‘make do’, and more likely to agree that they could ‘work more’; highly educated females, somewhat paradoxically, agree more with both statements more than less educated females, agreeing that they would have to work more and that they would have to make do with less. For both statements, the ‘5’ category is a major focal point, representing 1 out of 11 possible categories but being selected roughly 20% of the time. Male Female 25‐40 41‐50 51‐60 61‐67 25‐40 41‐50 51‐60 61‐67 If I lost some savings, would just have to make do with less 0.26 0.16 0.14 0.11 0.30 0.26 0.18 0.10 0‐4: Disagree 0.25 0.22 0.23 0.22 0.25 0.23 0.18 0.15 5 0.49 0.62 0.63 0.67 0.45 0.51 0.64 0.75 6‐10: Agree If I lost some savings, could always work more to make up for it 0.15 0.19 0.13 0.26 0.13 0.17 0.17 0.25 0‐4: Disagree 0.19 0.11 0.25 0.27 0.16 0.22 0.17 0.24 5 0.66 0.71 0.62 0.47 0.72 0.61 0.66 0.51 6‐10: Agree Table 23: Expected response to savings shock by sex and age. 10 point scale recoded to 0, 1‐9, 10. All figures are survey weighted proportions. [N=1072]
Male HS Coll BA MA+ If I lost some savings, would just have to make do with less 0.18 0.17 0.21 0.27 0‐4: Disagree 0.29 0.20 0.19 0.23 5 0.54 0.64 0.60 0.50 6‐10: Agree If I lost some savings, could always work more to make up for it 0.18 0.18 0.11 0.16 0‐4: Disagree 0.27 0.18 0.12 0.14 5 0.55 0.64 0.77 0.70 6‐10: Agree HS 0.30 0.22 0.48 0.16 0.22 0.62 Female Coll BA MA+ 0.20 0.26 0.54 0.21 0.20 0.58 0.16 0.20 0.64 0.09 0.10 0.80 0.27 0.15 0.57 0.10 0.14 0.76 Table 24: Expected response to savings shock by sex and education. 10 point scale recoded to 0, 1‐9, 10. All figures are survey weighted proportions. [N=1072] 64 Distributionofportfolioallocationvariables
In this section we report the distribution of several portfolio allocation variables: our main outcomes (owns stocks, and the proportion of financial assets held in stocks); and three additional outcomes (owns stocks as personal assets or within Defined Contribution (DC) employer‐based pension plans; and the proportion of gross or net wealth held in stocks). A consistent story appears across variables and sexes: stocks are more widely held and more important the older and more educated a person is. This almost certainly reflects the positive relationship between age/education and wealth, and the positive relationship between wealth and stock ownership, rather than a rising preference for risk at older ages. 25‐40 Owns Stocks 0.29 Owns Stocks or DC Stocks 0.49 Stocks / Financial Assets 0.17 Stocks / Gross Wealth 0.12 Stocks / Net Wealth 0.23 Male 41‐50 51‐60 0.43 0.45 0.63 0.63 0.31 0.29 0.15 0.17 0.26 0.24 61‐67 0.56 0.77 0.35 0.20 0.28 25‐40 0.20 0.31 0.15 0.08 0.19 Female N 41‐50 51‐60 61‐67 0.31 0.35 0.46 1196 0.41 0.51 0.61 1196 0.20 0.21 0.29 1072 0.10 0.11 0.18 1070 0.21 0.18 0.21 1069 HS 0.20 0.22 0.13 0.05 0.08 Female Coll BA 0.23 0.43 0.39 0.59 0.19 0.24 0.11 0.13 0.22 0.25 Table 25: Distribution of portfolio allocation by sex and age Owns Stocks Owns Stocks or DC Stocks Stocks / Financial Assets Stocks / Gross Wealth Stocks / Net Wealth HS 0.29 0.47 0.23 0.10 0.16 Male Coll BA 0.33 0.52 0.54 0.73 0.22 0.29 0.14 0.17 0.25 0.32 MA+ 0.52 0.73 0.32 0.21 0.34 N MA+ 0.47 0.68 0.27 0.15 0.32 1196 1196 1072 1070 1069 Table 26: Distribution of portfolio allocation by sex and education Finally, we present some differences by level of risk aversion. Using wealth‐gamble hypothetical questions fielded in a previous survey, we can divide our sample into people with more or less risk‐
averse preferences.38 As we would hope to find if our measures are meaningful, the more risk‐averse People were asked a series of three hypothetical questions designed to elicit risk preferences. Each question was in the same format, presenting a hypothetical ‘risky investment’ that succeeds with probability p % of the time, increasing the total value to X, and fails with probability (1‐p), decreasing the value to Y. An example is presented in 38
Appendix B. In the series of questions, different values of p, X and Y, representing riskier or less risky investments, were presented to individuals based on their initial response, and allowing us to distinguish several categories of people, from those who were willing to accept very risky investments to those who were unwilling to accept any of the risky investments presented to them. The measure we use here is the simplest available, dividing the sample into two categories of roughly equal size, with the ‘less risk‐averse’ being those whose answers indicated willingness to accept more risky investments than those in the ‘more risk‐averse’ group. 65 individuals are less likely to own stocks, and stocks make up a relatively smaller proportion of their wealth. Male Owns Stocks Owns Stocks or DC Stocks Stocks / Financial Assets Stocks / Gross Wealth Stocks / Net Wealth Less RA 0.43 0.62 0.26 0.16 0.27 Female More RA 0.32 0.52 0.24 0.12 0.21 Less RA 0.35 0.51 0.24 0.13 0.24 More RA 0.23 0.32 0.15 0.07 0.15 N 1192
1192
1068
1067
1066
Table 27: Distribution of portfolio allocation variables by sex and risk aversion Methodology:Empiricallinksbetween‘flexibility’andportfoliochoice
Conceptionsofflexibilityandtheirrelationshiptoeachother
As discussed previously, Bodie et al. consider a model with a fixed working lifetime, within which individuals can choose to work more or fewer hours, and can make small adjustments for all future periods in response to shocks in the present (Bodie, Merton et al. 1992). The most relevant empirical variables we have to address this logic are what we might call ‘instantaneous hour flexibility’ variables – whether people can increase or reduce hours in their current job. This is the type of variable Benitez‐
Silva uses to examine the relationship between labor supply flexibility and portfolio choice (Benitez‐Silva 2006). A more full version of this would include whether an individual has that flexibility at all relevant points in the lifecycle; while this is impossible to determine fully, we do have a “future hour flexibility” variable, which may also be the more relevant concept if we expect that the main response individuals will have to savings shocks will be towards the tail end of their working life. On the other hand, Gollier recognizes that the ability to prolong work life (i.e. delay retirement) may be an important form of flexibility (Gollier 2005). Empirically speaking, most workers do not rebalance their 401(k) portfolios regularly39, so it is possible that the flexibility to make small adjustments to work over the course of the working life is not as relevant to investment decisions as the ability to prolong the working life if necessary. As mandatory retirement ages are now prohibited in most walks of life, and age discrimination law offers further protections to older workers, workers now have more opportunities to continue to work than previous generations; but the physical, mental and psychological demands of work may prevent people from continuing to work indefinitely. To explore these issues, we have ‘job sustainability’ variables, regarding the demands of work and the difficulty of meeting those demands. In the analysis that follows, we focus on (a) the amount of time that a job involves lots of 39
Using a dataset featuring 1.2 million workers, Mitchell et al find that 80% of 401(k) participants did not make a single trade over a two‐year period (Mitchell, Mottola et al. 2008) 66 physical effort, (b) how often a job involves intense concentration and (c) how difficult it is to keep up with technology on a job.40 Related to these variables we have two variables that directly ask individuals about their likely reaction to negative savings shock: one asks the individual how far they agree with the statement “If I lost some of my savings, I could always work more to make up for it”, the other asks how far they agree with the statement “If I lost some of my savings, I would just have to make do with less”. While we might expect those two variables to be mirror images of each other and thus provide no additional information, in fact some people agreed (or disagreed) strongly with both statements, apparently suggesting that they would engage in both (neither) courses of action in the event of a negative wealth shock. As they are apparently distinct, we explore both variables. ExploratoryRegressionModel
While we do not have adequate multi‐period data to estimate a full dynamic model linking lifecycle labor and portfolio allocation decisions, the theoretical models discussed in the background literature sections of this paper suggest that greater flexibility over labor supply should enable individuals to take on more risk in their investments. The intuition for this result is that in the event of a negative wealth shock, individuals with fixed labor supply can only adjust their (current and future) consumption, while individuals with flexible labor can adjust both their consumption and their leisure, working slightly more in order to face a less drastic reduction in consumption. In order to test whether this intuition is borne out in a statistically significant association, we fit an exploratory regression model relating portfolio allocation to labor flexibility. This takes the basic form: where Di is a vector of demographic characteristics including age, age2, education, race, partnership status; Ri is a binary measure of risk aversion taken from a previous ALP survey; Wi is an indicator for net wealth decile (including all financial wealth, business wealth, vehicles and housing); Ji is a vector of job characteristics including detailed occupation information and income quintile; 0 is a common constant term and  is a stochastic error term. As discussed previously, the ‘allocation’ outcome is either an indicator for stock ownership or a measure of the percentage of financial assets held in stocks. For the ‘flexibility’ variables, we report results for separate models with these variables entered in conceptual groups and individually. For this exploratory work, we report results for simple OLS models (with Huber‐White robust standard errors) for ease of interpretation. However, the direction and statistical significance of our results are broadly robust to different model specifications (e.g. probit and logit for binary choice, tobit with two censoring points for a continuous variable bounded by 0 and 1). 40
These variables were chosen because they address the different types of demand placed on a worker by a job, are not strongly correlated with each other, and (as demonstrated in the descriptive statistics) differ in their gender/age/education trends 67 Results
In this section, we report simple OLS results relating various labor supply flexibility variables to asset allocation outcome variables. All regressions control for age, age2, race, partnership status (single or in a marriage / long‐term partnership), education (four categories), risk aversion (a binary measure), wealth decile, income quintile, and occupation (22 detailed categories). First, we focus on labor flexibility at the intensive margin (flexibility over hours of labor supply). Second, we examine variables pertaining to labor flexibility at the extensive margin (factors affecting the sustainability of a career at older ages). Finally, we investigate the impact of intentions at the extensive margin, and whether the intent to compensate for any negative savings shocks in the future by working longer affects investment behavior in the present. Currentandfuturehoursflexibility:intensivemarginoflaboradjustment
Examining the relationship between hours flexibility and portfolio outcomes, we do not find any statistically significant relationships. In contrast to Benitez‐Silva (2006), who found a strong positive relationship between the ability to increase hours and wealth held in stocks, we find a mixed set of results for this variable, with none close to being statistically distinguishable from zero.41 The current ability to reduce hours, and the prospective future ability to have hours flexibility, are similarly mixed and insignificant. Results do not change much whether these flexibility variables are entered jointly or separately into the models. Outcome: Hold any stocks? [mean: 0.35] Female Male Can increase hours ‐0.00670 0.00262 ‐0.0594 ‐0.0380 (0.0356) (0.0349) (0.0564) (0.0545) Can reduce hours ‐0.00206 0.00587 ‐0.0341 ‐0.0229 (0.0358) (0.0363) (0.0652) (0.0617) Chances (0‐100) of having flexible job in future 0 ‐0.0574 ‐0.0567 0.00750 (0.0488) (0.0492) (0.0664) 1‐49 ‐ ‐ ‐ 50 ‐0.0145 ‐0.0152 0.0406 (0.0451) (0.0448) (0.0654) 51‐99 ‐0.0180 ‐0.0196 0.0890 (0.0508) (0.0499) (0.0717) 100 0.0871 0.0845 0.138 (0.0783) (0.0780) (0.102) 0.0166 (0.0662) ‐ 0.0367 (0.0652) 0.0720 (0.0710) 0.104 (0.0968) Table 28: Effect of hours flexibility on holding stock. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation 41
However, there are various differences between our data and model from that of Benitez‐Silva that may explain differences in outcome: see Discussion section. 68 Outcome: % of Fin. Wealth held in stocks [mean: 0.22, median: 0] Female Male Can increase hours 0.0171 0.0270 ‐0.0272 ‐0.0196 (0.0312) (0.0312) (0.0438) (0.0455) Can reduce hours 0.000317 0.0162 ‐0.00608 ‐0.00431 (0.0310) (0.0324) (0.0511) (0.0491) Chances (0‐100) of having flexible job in future 0 ‐0.0503 ‐0.0531 0.0348 (0.0422) (0.0432) (0.0568) 1‐49 ‐ ‐ ‐ 50 ‐0.0494 ‐0.0481 0.0250 (0.0394) (0.0392) (0.0509) 51‐99 ‐0.0272 ‐0.0241 0.0519 (0.0456) (0.0438) (0.0529) 100 0.113 0.118 0.0598 (0.0781) (0.0779) (0.0849) 0.0383 (0.0579) ‐ 0.0246 (0.0506) 0.0454 (0.0512) 0.0474 (0.0807) Table 29: Effect of hours flexibility on percentage of financial wealth held in stock. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation 69 Jobsustainability:extensivemarginoflaboradjustment
In contrast to the hours flexibility variables which deal with flexibility on the intensive margin, we see statistically stronger effects of several variables relating to job characteristics and demands which affect an individual’s ability to have flexibility over their career length. In particular, working in a job where it is difficult to keep up with technology is associated with lower propensity to own stocks for both men and women, and is associated with lower percentages of financial wealth held in stocks for women. The effect sizes for women are particularly large. In addition, working in the most physically demanding jobs is associated with lower propensity to hold stock for men, and a lower percentage of financial wealth held in stocks for women. However, differences in the intensity of concentration required are not associated with portfolio differences. Outcome: Hold any stocks? [mean: 0.35] Female Male Difficult to keep up with technology in job Strong Agree ‐0.236*** ‐0.242*** ‐0.0780 ‐0.0876 (0.0543) (0.0502) (0.0955) (0.0902) Agree 0.0226 0.0205 ‐0.122* ‐0.123* (0.0542) (0.0539) (0.0578) (0.0568) Disagree ‐ ‐ ‐ ‐ Strong Disagree ‐0.0711 ‐0.0677 ‐0.107 ‐0.112 (0.0412) (0.0414) (0.0590) (0.0610) DNA ‐0.0397 ‐0.0638 ‐0.165 ‐0.214* (0.0497) (0.0460) (0.0936) (0.0894) Job requires intense concentration All Time ‐ ‐ ‐ ‐ Most Time ‐0.0348 ‐0.00267 ‐0.0233 0.00657 (0.0414) (0.0406) (0.0543) (0.0538) Some/No Time 0.00349 0.0340 0.0261 0.0611 (0.0495) (0.0475) (0.0710) (0.0698) DNA ‐0.0711 ‐0.0702 ‐0.187 ‐0.260* (0.113) (0.0904) (0.153) (0.109) Job requires lots of physical effort All Time ‐0.0762 ‐0.0843 ‐0.190* (0.0602) (0.0573) (0.0951) Most Time 0.0427 0.0406 ‐0.129 (0.0593) (0.0610) (0.0861) Some Time ‐0.0378 ‐0.0353 ‐0.106 (0.0462) (0.0456) (0.0674) No Time ‐ ‐ ‐ DNA ‐0.0436 ‐0.0644 ‐0.138 (0.0680) (0.0650) (0.0866) ‐0.188* (0.0949) ‐0.114 (0.0871) ‐0.0899 (0.0672) ‐ ‐0.185* (0.0826) Table 30: Effect of job sustainability on holding stock. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation 70 Outcome: % of Fin. Wealth held in stocks [mean: 0.22, median: 0] Female Male Difficult to keep up with technology in job Strong Agree ‐0.137** ‐0.168*** ‐0.0403 ‐0.0599 (0.0485) (0.0477) (0.0711) (0.0687) Agree 0.0420 0.0369 ‐0.0819 ‐0.0883 (0.0472) (0.0482) (0.0467) (0.0475) Disagree ‐ ‐ ‐ ‐ Strong Disagree ‐0.0427 ‐0.0396 ‐0.0651 ‐0.0711 (0.0333) (0.0339) (0.0500) (0.0505) DNA ‐0.00641 ‐0.0520 ‐0.0714 ‐0.121 (0.0499) (0.0461) (0.0912) (0.0903) Job requires intense concentration All Time ‐ ‐ ‐ ‐ Most Time 0.00764 0.0247 0.0249 0.0480 (0.0351) (0.0353) (0.0451) (0.0431) Some/No Time 0.0302 0.0482 0.0448 0.0747 (0.0437) (0.0415) (0.0596) (0.0595) DNA ‐0.0360 ‐0.0772 ‐0.221 ‐0.220** (0.102) (0.0914) (0.129) (0.0818) Job requires lots of physical effort All Time ‐0.0950 ‐0.106* ‐0.127 (0.0512) (0.0477) (0.0717) Most Time 0.00729 0.00534 ‐0.00280 (0.0540) (0.0549) (0.0719) Some Time ‐0.0282 ‐0.0303 ‐0.00684 (0.0387) (0.0383) (0.0525) No Time ‐ ‐ ‐ DNA ‐0.0852 ‐0.101* ‐0.0804 (0.0478) (0.0444) (0.0645) ‐0.129 (0.0702) 0.0000815 (0.0728) 0.00344 (0.0524) ‐ ‐0.108 (0.0624) Table 31: Effect of job sustainability on percentage of financial wealth held in stock. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation We present the most detailed coding of the relevant variables here in order to provide more information rather than less. However, some of the statistically significant results might be thought to be driven by relatively small categories or the choice of reference category. The results are not fully robust to collapsing the categories into their simplest versions; alternative versions of these tables using categories that have been further collapsed are available in the appendices. 71 Expectedreactiontosavingsshock:intenttoadjustlabor,consumption,orboth
In our final set of regressions, we examine the relationship between how individuals believe they would react to negative wealth shocks in the future, and their current portfolio choices. While the previous questions provide evidence on how having the ability to adjust labor affects investment behavior, these questions directly address how the intent to adjust labor/consumption in the event of a negative savings shock relates to investment behavior. Here, we do not find any statistically significant results for women over any outcome or variable of interest, but there are several interesting results for men. When we examine agreement with the statement “If I lost some of my savings, I would just have to make do with less when I retired”, controlling for the normal controls, men who disagree that they would have to ‘make do with less’ in the event of negative wealth shocks are more likely to hold stocks, with statistical significance at the 5% level when this variable is included without the ‘work longer’ variable, or at the 10% level when both variables are included. Outcome: Hold any stocks? [mean: 0.35] Female Male If I lost some of my savings, I could always work longer to make up for it ‐0.0334 ‐0.0187 0.0684 0.0984 0‐4: Disagree (0.0600) (0.0560) (0.0795) (0.0755) 5 6‐10: Agree ‐ ‐ ‐ ‐0.0000214 (0.0464) 0.0152 (0.0426) 0.0920 0.126 (0.0678) (0.0658) If I lost some of my savings, I would just have to make do with less when I retired 0.0625 0.0575 0‐4: Disagree (0.0553) (0.0524) 5 6‐10: Agree ‐ 0.122 (0.0704) 0.156* (0.0695) ‐ ‐ ‐ ‐ 0.00781 (0.0453) 0.00256 (0.0415) 0.0621 (0.0616) 0.0875 (0.0590) Table 32: Effect of expected reaction to savings shock on stock holding. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation Turning to our other outcome of interest, for men, expecting to work longer in the event of a future negative wealth shock is associated with a greater percentage of financial wealth being held in stocks. These results are slightly weaker but still statistically significant at the 5% level when controlling for expectations about ‘making do with less’. 72 Outcome: % of Fin. Wealth held in stocks [mean: 0.22, median: 0] Female Male If I lost some of my savings, I could always work longer to make up for it 0.0505 0.0557 0.0356 0.0642 0‐4: Disagree (0.0581) (0.0561) (0.0559) (0.0536) 5 6‐10: Agree 0.0128 (0.0431) 0.0227 (0.0413) 0.0986* 0.114* (0.0461) (0.0443) If I lost some of my savings, I would just have to make do with less when I retired 0.0442 0.0512 0‐4: Disagree (0.0482) (0.0473) 5 6‐10: Agree 0.0282 (0.0588) 0.0649 (0.0563) ‐ ‐ ‐ ‐ 0.00411 (0.0387) 0.0141 (0.0364) 0.0652 (0.0497) 0.0841 (0.0484) Table 33: Effect of expected reaction to savings shock on percentage of financial wealth held in stocks. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation As with the analysis in the preceding section, we present our preferred coding of variables in the specifications above, but have run additional analyses. With the ‘5’ category being a focal point, we prefer not to combine it with responses above it or below it. Nevertheless, while the main focus is on statistically significant results, it may be noted that the statistically insignificant results do not consistently have the sign that is expected under the hypotheses. For robustness, we provide alternative analyses in the appendices folding the ‘5’ category into the 0‐4 category; the statistically significant results are consistent with those above, fewer results are significant, but a greater proportion of non‐
significant results demonstrate the expected sign. AdditionalAnalyses
As noted in the text, some additional analyses were conducted after the main analyses. One set of analyses collapsed the ‘sustainability’ categorical independent variables into their simplest forms (typically 3 categories including ‘does not apply’) in order to increase power. These analyses showed generally slight improvements in precision, but led generally to smaller, statistically insignificant results. We choose to present the more detailed results in the main text, but the less detailed results are available in Appendix C. Also in Appendix C, we present the simplified version of the ‘expected reaction to savings shock’ variables which, as noted above, also reduces the number of statistically significant findings. On a restricted sample, we also ran analyses incorporating information concerning DC pension plans. One set of analyses incorporated the stocks held in DC plans into the outcome measures; another set of analyses reran the original models but with a dummy controlling for an individual holding some stocks in a DC plan. The results from these analyses were broadly consistent with the results presented in the main analysis, with most differences largely attributable to sample selection effects when excluding 73 individuals with low quality data on their DC plan wealth and asset allocations. An additional finding from these analyses is that holding stock in a DC plan is highly correlated with holding stocks outside DC plans; insofar as both types of stock holding might be influenced by work factors, including DC stock holdings as a control may have the effect of controlling for the outcome.42 These results are available on request. Discussion
In our empirical work, we examined factors affecting individual labor supply flexibility and their relationship to portfolio choice. We distinguished between two broad types of flexibility that might be used to adjust to wealth shocks: the flexibility to increase or reduce hours within a career (intensive margin), and the flexibility to extend a career (extensive margin). In theory, both of these types of flexibility could affect the degree of risk an individual might be willing to accept in their investment portfolio. While our results are open to wide interpretation (and are not fully robust to alternative specification), and are best thought of as descriptive rather than causal, they provide more support for the importance of extensive margin flexibility than intensive margin flexibility. Men who believe that they would be able to work longer in order to make up for negative wealth shocks are more likely to hold stocks and likely to hold a larger proportion of their financial wealth in the stock market. Men and women who struggle to keep up with technological change in their jobs ‐ and who may therefore be less able to sustain their career as they age – are shown to be less likely to hold stocks. Individuals with very physically demanding jobs may expect that the natural processes of aging may prevent them from continuing this type of work indefinitely; we find that men with the most physically demanding jobs are less likely to own stock than other men, and women with the most physically demanding jobs are likely to have a lower percentage of their financial wealth held in stocks than other women. In contrast, we do not find any evidence that differences at the intensive margin ‐ in the current or future ability to adjust hours of work within a career – have any relationship whatsoever with portfolio choice. Generally speaking, our results for those variables are mixed and fall far short of standard thresholds for statistical significance. In some ways, our intensive margin results are unsurprising: empirically speaking, work hours tend to be lumpy rather than continuous, with extreme clumping at focal values such as 40 hours per week, with little evidence that people are making adjustments; and making small marginal adjustments to labor over the lifecycle in response to changing portfolio performance seems an unlikely behavioral model given how little attention individuals tend to give to their portfolio on a monthly (or even yearly) basis. 42
It is beyond the scope of this paper to investigate the relationship between holdings inside and outside DC plans. The limited evidence in this paper might suggest that individuals do not look to balance very risky DC plans by taking less risky investment positions outside their DC plan, or vice versa (as might be the case if individuals were constrained in the choices available to them within their DC plans); rather, more risk‐tolerant individuals tend to tolerate more risk in all forms of their investments. 74 However, our results are surprising insofar as they run counter to the findings of previous empirical work by Benitez‐Silva, who did not examine the extensive margin but found large and statistically significant effects of intensive margin flexibility on the amount of stock wealth an individual had (Benitez‐Silva 2006). There are several potential explanations for the discrepancy. The first possible explanation is that we use different outcome measures and (additional / more detailed) control variables from Benitez‐Silva, but our (non‐)findings here are robust to using the same outcome measure and similar control variables as he did in his paper. A second possibility is that his dataset focuses on workers over the age of 50, while ours includes workers as young as 25, and effects may be concentrated in older workers. This explanation runs counter to the logic expressed by Bodie et al., who anticipate smaller effects among older workers who have fewer remaining work years over which to adjust their labor supply43 (Bodie, Merton et al. 1992), but might be correct if the uncertainty about future labor income faced by younger workers dominates any effects of labor supply flexibility. Moreover, when we run the intensive margin models on a sample restricted to older ages, we do not see any evidence in support of this contention.44 A third possibility is that individuals in the Health and Retirement Study used by Benitez‐Silva differ in other ways from the individuals in the American Life Panel; the waves of HRS used in his paper cover people from 1992 to 2000, a period of rapidly rising stock values due to the dot‐com bubble; our ALP data were collected in 2011, in the middle of the recession brought about by the sub‐prime mortgage crisis. Finally, it is possible that the panel data techniques employed by Benitez‐Silva provide a better control for unobserved heterogeneity than the cross‐sectional approach used in this paper; however, the cross‐sectional results he reports are consistent with his panel results and inconsistent with our results. Before concluding, it is also worth taking a moment to discuss the interpretation of the effects seen for women vs. men. Given that wealth is a household variable, and male labor supply has traditionally played the larger role in household income, it is worth considering whether the effects seen for women might be driven by their spouses. In sensitivity analysis of the significant effects seen for women, we reran the models separately for single and partnered women.45 While the smaller sample sizes reduced precision significantly, it appears that the variable concerning difficulty keeping up with technology has a roughly similar effect for single and partnered women, while the variable concerning physical demands of the job may have had a larger effect for single women. In neither case does it appear that the result is driven by spousal effects. 43
Indeed, Benitez‐Silva uses this logic to assert that his own findings should be considered a lower bound estimate Results available on request; the reduction in sample leads to imprecise estimates opposite in sign to those of Benitez‐Silva 45
Results available on request 44
75 Conclusion
In this research, we described a novel data set and a range of variables pertaining to labor supply flexibility, broadly interpreted. Some of these questions are common to the Health and Retirement Study, but are being applied to a broader age spectrum of respondents; others are completely novel. We have documented how these various types of labor supply flexibility differ across sex, age and education groups. In our empirical work, we explored the relationship between labor supply flexibility and portfolio allocation decisions, and found that the ability to prolong a career at older ages (extensive margin) may be more important than the ability to adjust the intensity of work through a career (intensive margin). These results contribute to a sparse empirical literature which has investigated the intensive margin flexibility effects for older workers, but has not addressed the intensive margin effects for younger workers or the extensive margin effects at all. In general, the policy environment has changed over time – and continues to change ‐ in ways that can increase labor supply flexibility at both the intensive and extensive margins. Legislation making mandatory retirement ages illegal and prohibiting age‐based discrimination have helped to increase employment opportunities at older ages and flexible working practices are more prevalent now than in previous decades. In the future, the Patient Protection and Affordable Care Act may lead to decreased reliance on employer‐based healthcare, which would in turn remove one of the barriers to flexible working practices for pre‐Medicare‐age workers. It is unclear whether broader changes in the economy, such as the move towards more highly educated workers and technology intensive industries, will allow people to prolong their careers: although physically demanding work may be difficult to sustain at older ages, in a knowledge‐based economy the mental demands and the pace of technological change may also prove challenging at older ages. Our work suggests that increases in the ability to prolong a career may result in individuals making increased investment in risky assets, which on average in the long‐run is likely to lead to increased wealth at retirement. Individuals with the flexibility to prolong their career thus have a double benefit: they are likely to be financially better off at retirement age than people with fixed labor supply, and in the event of other shocks (such as medical bills for a family member) would also be able to continue to work. Workers without the ability to prolong their working life are faced with a challenge: they must accumulate significant wealth over their fixed working life, but can ill‐afford to take on high‐risk, high‐
reward investments because they have limited ability to cope if the risks do not work out. However, it is important to note two major limitations to the work. The first is the instability of our results: given that several results are not robust to relatively small differences in specification, our work may provide more value in highlighting under‐researched areas into which future research can delve further, than in providing definitive answers to the questions we pose. The second major limitation is common to nearly all work dealing with labor supply flexibility: despite using a rich set of covariates to control for observed differences, we cannot rule out possible selection effects; as individuals sort into jobs, there may be unobserved heterogeneity that leads to a spurious correlation between job characteristics and investment behavior. Given the importance of ‘extensive margin’ flexibility suggested 76 by our work, it is possible that future research may be able to exploit quasi‐exogenous variation in career extendibility due to legal changes (e.g. the abolition of mandatory retirement ages in most professions, or the removal of disincentives to work after normal retirement age) in order to shed further light on the impact of labor supply flexibility on investment behavior. 77 Bibliography
Allen,S.G.,R.L.Clark,etal.(2004)."PhasingintoRetirement."Industrial&Labor
RelationsReview58(1):112‐127.
Benitez‐Silva(2006).LaborSupplyFlexibilityandPortfolioDecisions:AnEmpirical
Analysis,(ReviseandResubmitatJournalofAppliedEconometrics,
downloadedon8/13/2001from
http://ms.cc.sunysb.edu/~hbenitezsilv/portfhbs.pdf).
Bodie,Z.,R.C.Merton,etal.(1992)."Laborsupplyflexibilityandportfoliochoicein
alifecyclemodel."JournalofEconomicDynamicsandControl16(3‐4):427‐
449.
Butrica,B.A.,H.M.Iams,etal.(2004)."ThechangingimpactofSocialSecurityon
retirementincomeintheUnitedStates."SOCIALSECURITYBULLETIN‐
WASHINGTON‐65(3):1.
Charles,K.K.andP.DeCicca(2007)."Hoursflexibilityandretirement."Economic
Inquiry45(2):251‐267.
Clift,J.(2012).LaborMarketRigiditiesandtheLaborSupplyofOlderWorkers,
PardeeRANDGraduateSchool.
Cocco,J.F.,J.G.Gomes,etal.(2005)."ConsumptionandPortfolioChoiceoverthe
LifeCycle."TheReviewofFinancialStudies18(2).
Goldman,D.andN.Maestas(2012)."MEDICALEXPENDITURERISKAND
HOUSEHOLDPORTFOLIOCHOICE."JournalofAppliedEconometrics.
Gollier,C.(2005).Optimalportfoliomanagementforindividualpensionplans,
Citeseer.
Gustman,A.L.andT.L.Steinmeier(1985)."TheEffectofPartialRetirementonthe
WageProfilesofOlderWorkers."IndustrialRelations24(2):257‐265.
Gustman,A.L.andT.L.Steinmeier(2007)."ProjectingBehavioralResponsestothe
NextGenerationofRetirementPolicies."NBERWorkingPaper12958.
Heaton,J.andD.Lucas(2000)."Portfoliochoiceinthepresenceofbackgroundrisk."
TheEconomicJournal110(460):1‐26.
Hurd,M.D.(1996)."TheEffectofLaborMarketRigiditiesontheLaborForce
BehaviorofOlderWorkers."AdvancesintheEconomicsofAging.
Hurd,M.D.andS.Rohwedder(2010)."TheEffectsoftheEconomicCrisisonthe
OlderPopulation."MichiganRetirementResearchCenterWorkingPaper231.
Jagannathan,R.andN.Kocherlakota(1996)."Whyshouldolderpeopleinvestlessin
stocksthanyoungerpeople?"FederalReserveBankofMinneapolisQuarterly
Review20(3).
Kapteyn,A.andC.Panis(2005)."InstitutionsandSavingforRetirement."Analyses
intheeconomicsofaging:281.
Lahey,J.N.(2008)."Age,Women,andHiring:AnExperimentalStudy."J.Human
Resources43(1):30‐56.
78 Lynch,A.W.andS.Tan(2011)."Laborincomedynamicsatbusiness‐cycle
frequencies:Implicationsforportfoliochoice."JournalofFinancial
Economics.
Miniño,A.M.,S.L.Murphy,etal.(2011)."Deaths:finaldatafor2008."Nationalvital
statisticsreports:fromtheCentersforDiseaseControlandPrevention,
NationalCenterforHealthStatistics,NationalVitalStatisticsSystem59(10):
1.
Mitchell,O.,G.Mottola,etal.(2008)."Theinattentiveparticipant:Portfoliotrading
behaviorin401(k)plans."
Poterba,J.M.,S.F.Venti,etal.(2009).Thedeclineofdefinedbenefitretirement
plansandassetflows.SocialSecurityPolicyinaChangingEnvironment.J.
Brown,J.LiebmanandD.A.Wise,UniversityofChicagoPress.
Ruhm,C.J.(1990)."BridgeJobsandPartialRetirement."JournalofLaborEconomics
8(4):482‐501.
Samuelson,P.A.(1969)."LifetimePortfolioSelectionByDynamicStochastic
Programming."TheReviewofEconomicsandStatistics51(3):239‐246.
Scott,F.A.,M.C.Berger,etal.(1995)."DoHealthInsuranceandPensionCosts
ReducetheJobOpportunitiesofOlderWorkers?"IndustrialandLabor
RelationsReview48(4):775‐791.
USBureauofLaborStatistics(2011).Womeninthelaborforce:Adatabook.US
BureauofLaborStatistics.1034:1‐98.
Zissimopoulos,J.,N.Maestas,etal.(2007).TheEffectofRetirementIncentiveson
RetirementBehavior:EvidencefromtheSelf‐EmployedIntheUnitedStates
andEngland.RANDWorkingPaperSeries,RANDCorporation.
79 Appendices
AppendixA:SurveyProtocol
intro intro
We would like to ask you a few questions about your current work, the flexibility you have in your
current job, and your plans and expectations for the future. In previous surveys, you may already have
answered some similar questions. Though you may have seen some of these questions before, we are still
interested in your answers today, particularly as some circumstances may have changed.
Q1 reduce paid hours in regular work schedule
First we would like to ask some questions about job flexibility: Can you reduce the number of paid hours
in your regular work schedule?
1 Yes
2 No
Q2 increase paid hours in regular work schedule
Can you increase the number of paid hours in your regular work schedule?
1 Yes
2 No
Q3 employer allow buy or sell vacation days
Some employers allow their employees to "buy" extra vacation days (work fewer days in the year but
with a lower annual salary) or "sell" vacation days back to the employer (take less vacation but get a
higher annual salary). Does your employer allow you to "buy" or "sell" vacation days?
1 Can neither "buy" nor "sell"
2 Can "buy" but cannot "sell"
3 Can "sell" but cannot "buy"
4 Can both "buy" and "sell"
Q4 chances approach retirement job able to increase or decrease paid work hours
Please answer the next two questions on a scale from 0 to 100, where 0 equals absolutely no chance and
100 equals absolutely certain: On a scale from 0 to 100, when you are approaching retirement, what are
the chances you will have a job where you will be able to increase or decrease the number of paid hours
in your regular work schedule? (please use whole numbers only and no %)
Range: 0..100
Q5 chances lose job during next 3 years
Sometimes people are permanently laid off from jobs that they want to keep. On a scale from 0 to 100,
what are the chances that you will lose your job during the next three years? (please use whole numbers
only and no %)
Range: 0..100
[The following questions are displayed as a table]
Q6_intro intro
Now we would like to ask about the demands and characteristics of your current job: How often are these
statements true?
Q6 My job requires lots of physical effort
80 My job requires lots of physical effort
1 All or almost all of the time
2 Most of the time
3 Some of the time
4 None or almost none of the time
5 Does not apply
Q7 My job requires intense concentration or attention
My job requires intense concentration or attention
1 All or almost all of the time
2 Most of the time
3 Some of the time
4 None or almost none of the time
5 Does not apply
[End of table display]
[The following questions are displayed as a table]
Q8_intro intro
How strongly do you agree or disagree with the following statements?
Q8 In my job, you have to keep learning new skills if you want to keep up with other workers
In my job, you have to keep learning new skills if you want to keep up with other workers
1 Strongly agree
2 Agree
3 Disagree
4 Strongly disagree
5 Does not apply
Q9 In my job, you have to keep physically fit if you want to keep up with other workers
In my job, you have to keep physically fit if you want to keep up with other workers
1 Strongly agree
2 Agree
3 Disagree
4 Strongly disagree
5 Does not apply
Q10 In my job, it is difficult to keep up with changes in technology
In my job, it is difficult to keep up with changes in technology
1 Strongly agree
2 Agree
3 Disagree
4 Strongly disagree
5 Does not apply
Q11 My co-workers make older workers feel that they ought to retire before age 65
My co-workers make older workers feel that they ought to retire before age 65
1 Strongly agree
2 Agree
3 Disagree
4 Strongly disagree
5 Does not apply
81 [End of table display]
Q12 impairment or health problem
Now we would like to know how your health affects your work: Do you have any impairment or health
problem that limits the kind or amount of paid work you can do?
1 Yes
2 No
Q13 chances health limit work next 10 years
On a scale from 0 to 100 where 0 equals absolutely no chance and 100 equals absolutely certain, what are
the chances that your health will limit your work activity during the next 10 years? (please use whole
numbers only and no %)
Range: 0..100
[The following questions are displayed as a table]
Q14_intro intro
As people grow older, they sometimes have to pay for unexpected costs out of their savings, and
sometimes their investments or retirement pension plans lose value. For example, people sometimes face
unexpected medical bills at older ages, or lose money when there is a stock market crash. On a scale from
0 to 10 where 0 equals very strongly disagree and
10 equals very strongly agree, to what extent do you agree with the following statements?
Q14 If I lost some of my savings, I would just have to make do with less when I retired
If I lost some of my savings, I would just have to make do with less when I retired
Range: 0..10
Q15 If I lost some of my savings, I could always work longer to make up for it
If I lost some of my savings, I could always work longer to make up for it
Range: 0..10
[End of table display]
CS_001 HOW PLEASANT INTERVIEW
Could you tell us how interesting or uninteresting you found the questions in this interview?
1 Very interesting
2 Interesting
3 Neither interesting nor uninteresting
4 Uninteresting
5 Very uninteresting 82 AppendixB:RiskAversionWealthGamble
Wealth Gamble hypothetical questions: initial question Suppose that you unexpectedly inherited 1 million dollars. You have the chance to take a risky but possibly rewarding investment option that has a 50‐50 chance of doubling the money to 2 million dollars in a month, and a 50‐50 chance of reducing the money by one third, to 667 thousand dollars in a month. In other words, you could keep $1,000,000 or invest in a risky asset that has an equal chance of doubling your money to $2,000,000 or reducing it to $667,000. Would you choose to invest in the risky asset? 83 AppendixC:AlternativeAnalyses
Outcome: Hold any stocks? [mean: 0.35] Female Male Difficult to keep up with technology in job Agree 0.00324 ‐0.00281 ‐0.0745 ‐0.0722 (0.0455) (0.0450) (0.0484) (0.0471) Disagree ‐ ‐ ‐ ‐ [DNA] ‐0.00962 ‐0.0263 ‐0.142 ‐0.175* (0.0442) (0.0404) (0.0892) (0.0760) Job requires intense concentration All Time ‐ ‐ ‐ ‐ Not All Time ‐0.0188 ‐0.0226 ‐0.00748 ‐0.0192 (0.0352) (0.0349) (0.0484) (0.0488) [DNA] ‐0.0283 ‐0.0543 ‐0.0807 ‐0.203 (0.0916) (0.0744) (0.164) (0.105) Job requires lots of physical effort No Time 0.0371 0.0394 0.132* (0.0434) (0.0422) (0.0624) Not No Time ‐ ‐ ‐ [DNA] ‐0.00989 ‐0.0134 0.00336 (0.0660) (0.0621) (0.0890) 0.126* (0.0625) ‐ ‐0.0403 (0.0781) Table 34: Effect of job sustainability on holding stock. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation Outcome: % of Fin. Wealth held in stocks [mean: 0.22, median: 0] Female Male Difficult to keep up with technology in job Agree 0.0394 0.0285 ‐0.0404 ‐0.0440 (0.0397) (0.0404) (0.0398) (0.0394) Disagree ‐ ‐ ‐ ‐ [DNA] 0.00611 ‐0.0220 ‐0.0878 ‐0.125 (0.0454) (0.0406) (0.0913) (0.0831) Job requires intense concentration All Time ‐ ‐ ‐ ‐ Not All Time ‐0.0367 ‐0.0389 ‐0.0498 ‐0.0567 (0.0301) (0.0298) (0.0399) (0.0395) [DNA] ‐0.0265 ‐0.0789 ‐0.188 ‐0.255** (0.0713) (0.0564) (0.123) (0.0806) 84 Job requires lots of physical effort No Time 0.0434 (0.0366) Not No Time ‐ [DNA] ‐0.0446 (0.0448) 0.0451 (0.0355) ‐ ‐0.0484 (0.0420) 0.0338 (0.0464) ‐ ‐0.0360 (0.0650) 0.0302 (0.0465) ‐ ‐0.0680 (0.0627) Table 35: Effect of job sustainability on percentage of financial wealth held in stock. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation Outcome: Hold any stocks? [mean: 0.35] Female Male If I lost some of my savings, I could always work longer to make up for it 0‐5: Disagree/Neutral ‐ ‐ ‐ ‐ 6‐10: Agree 0.0255 0.0238 0.0817 0.0813 (0.0369) (0.0364) (0.0527) (0.0526) If I lost some of my savings, I would just have to make do with less when I retired 0‐5: Disagree/Neutral ‐ ‐ ‐ 6‐10: Agree ‐0.0288 ‐0.0272 0.0162 (0.0360) (0.0355) (0.0492) ‐ 0.0155 (0.0494) Table 36: Effect of expected reaction to savings shock on stock holding. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation Outcome: % of Fin. Wealth held in stocks [mean: 0.22, median: 0] Female Male If I lost some of my savings, I could always work longer to make up for it 0‐5: Disagree/Neutral ‐ ‐ ‐ ‐ 6‐10: Agree ‐0.00238 ‐0.00248 0.0871* 0.0842* (0.0358) (0.0357) (0.0360) (0.0357) If I lost some of my savings, I would just have to make do with less when I retired 0‐5: Disagree/Neutral ‐ ‐ ‐ ‐ 6‐10: Agree ‐0.0114 ‐0.0114 0.0573 0.0533 (0.0311) (0.0310) (0.0386) (0.0390) Table 37: Effect of expected reaction to savings shock on percentage of financial wealth held in stocks. OLS with controls: age, age2, race, partnership status, education, risk aversion, wealth decile, income quintile, and occupation 85 Introduction46
Public pension systems around the world play a vital part in providing retirement income security for the elderly, and in most Western countries comprise a significant portion of government spending. As populations age, many countries will have (or have had) to reform unsustainable pension systems, expecting individuals to take on greater responsibility for their income in retirement by saving and investing during their working life. To understand the consequences of current and future pension reforms, it is important to understand how individuals will change their behavior in response to any changes in incentives; but this is generally difficult to predict. In this paper, we lay the foundations for significant extensions to research in this field to date, constructing a multi‐country quasi‐panel of individual‐level lifetime earnings histories, along with detailed modeling of the different incentives created by institutional differences in pension systems between countries. This dataset features a combination of individual‐level detail and national‐level structures that provides unique opportunities to exploit cross‐national variation in order to identify models of individual behavior, which can provide crucial insights into the likely effects of pension reforms. While standardized cross‐national measures of pension system generosity and progressivity have been produced by other researchers, these have typically been based on stylized hypothetical workers with simple (often, flat) earnings trajectories over the lifecycle (OECD 2011); in contrast, we are able to generate pensions for real individuals and their reported earnings, and, importantly can capture for each individual at every age the marginal effect on pension income of working for an additional year, including both accrual and deferral effects. The interaction of real earnings trajectories with pension systems produces effects that might not be seen in in stylized models: two individuals with the same total lifetime earnings may be entitled to very different pensions if they have different trajectories, if (for example) the pension system emphasizes earnings at older ages, or has a binding cap on pensionable earnings. In our use of data on real individuals across several countries, our work is closest in spirit to the series of parallel analyses described in Gruber and Wise’s seminal volume (Gruber and Wise 1999). In that work, national research teams produced standardized analyses using the most appropriate data available to them. In our work, we are able to construct a single dataset with data that are harmonized across countries and earnings information for specific individuals over their whole life, which is better suited for exploiting cross‐national variation in institutional structures and lifecycle behavior. After documenting the construction and describing the characteristics of this unique dataset, we demonstrate the clear relationships between measures of realized replacement rates from pensions and the financial wealth individuals accumulate, and between pension accrual incentives and the age at which individuals choose to retire, which confirm the pivotal role that the institutional environment plays in determining individual work, savings and retirement decisions. 46
This work was supported by a grant from the National Institutes of Aging (NIA P01 AG022481). Thanks are also due to Alessandro Malchiodi and Jose Castillo for their assistance with national websites. 86 Model
To help motivate the usefulness of this dataset, and foreshadow the exploratory analysis we conduct in the closing sections of this paper, we here present the simple lifecycle model that underpins this paper. The model is similar to that discussed by Hurd et al., to be applied to observed lifetime earnings histories rather than the synthetic earnings histories used in that paper (Hurd, Michaud et al. 2012); we draw heavily on their discussion in our presentation here. We assume an individual optimizing labor and consumption decisions over a lifecycle from time t=0 to their death at t=T. Individuals are assumed to be rational and to have perfect information over their future potential earnings and potential pension income, and do not face any uncertainty about mortality, asset returns or any other potential shock. The individual works from time t=0 until retiring at time t=R, earning income yt when working and receiving pension benefit bt in retirement, where bt is a function of prior income and retirement age R. The individual makes a decision about when to retire, and how much to consume in each period, in order to maximize lifetime utility. Assume that the individual derives utility from consumption prior to retirement, and from consumption and leisure after retirement, such that 1
is utility while working and is utility once retired. The individual discounts future utility with discount rate ρ . Preference parameters in this model are therefore γ which determines the marginal utility of consumption; Γ , which determines the marginal utility of leisure from retiring (or, the marginal disutility of work from delaying retirement), and the time preference ρ which determines consumption growth over the life‐cycle. Assume also that the individual has no bequest motive but must die with non‐negative assets a, and starts with no assets. Assets (and liabilities) grow with at a certain real interest rate of r; in this set‐up, the individual does not face liquidity constraints. The individual then solves the following optimization, selecting the retirement time R that maximizes utility given optimal consumption over the lifecycle, subject to the inter‐temporal constraint on assets: max ,
s.t. ,
0,
0 87 The solution for optimal consumption given retirement age R is then: 1
,
where Y(R) and B(R) are the discounted present value of lifetime income from earnings and retirement 1 and . benefits respectively; /
Thus the initial consumption given any retirement age is a fraction of total lifetime income, determined by the asset growth rate r and the preference parameters ρ and γ. Consumption changes by a fixed factor from period to period, rising over time if the real interest rate is greater than the discount rate, or remaining flat if r= ρ. The rate of change in consumption in inversely related to the parameter γ that defines the curvature of the utility function. When γ approaches zero, the utility function approaches a linear function of consumption, leading to extreme distributions of optimal consumption (either consuming a very high proportion of lifetime wealth early if ρ> r, or consuming very little and allowing assets to build for very high later consumption if ρ<r). Conversely, when γ is large, there are sharply diminishing marginal returns to consumption and consumption will tend to be spread more evenly across the life cycle. Turning to the decision variable R, the first order condition can be written as where B’(R) represents the marginal change in pension wealth for a change in R. This yields the familiar intuition equating marginal utilities: at the optimal value of R, the consumption utility to be gained by additional work (due to both an additional year of income, and a larger pension in retirement) is equal to the leisure utility lost. This leads to some additional insights: when the marginal benefit of additional consumption is relatively low, the individual is likely to retire earlier; if the individual will earn a large amount in the marginal year relative to his lifetime earnings, they are more likely to delay retirement; and if the increase in pension wealth for an additional year of work is relatively large, retirement is more likely to be delayed. This last point is relevant for comparisons across countries, comparisons within countries over time, and between individuals with different earnings profiles. In countries with stiff penalties for early retirement, for example, the increase in pension for an additional year of work may be larger than for countries with lesser penalties. In countries with pensions calculated based on the basis of some number of last or best years of earnings, a high earning year at the end of a career may significantly increase pension benefits in retirement; the magnitude of this effect, in turn, will be strongly affected by whether a worker has had a steep earning profile over time or has had flat earnings for a large number of years. Under certain circumstances, a low earning year may have a minimal or even negative effect on pension benefits. Capturing these differences between individuals within and across countries at different ages is one of the key motivations and contributions of this paper. 88 Data
IntroductiontoSHAREandSHARELIFE
The Survey of Health, Ageing and Retirement in Europe (SHARE) is a large panel survey of the 50+ population in Europe, modeled after the US Health and Retirement Study (HRS) and the English Longitudinal Study of Ageing (ELSA). The first wave was fielded in 2004; the second wave was fielded in 2006 with a broadly similar survey. The survey covers a wide range of important topics for understanding the behavior of the older population across Europe: demographics, health, economic status, family and social networks, and other topics. Full details of the survey construction and implementation, including imputation of missing values, are available in the SHARE methodology book (Borsch‐Supan and Jurges 2005). Particularly valuable for our work are the detailed income, wealth, and standardized cross‐national educational attainment variables, which we employ directly or indirectly in constructing our analytic dataset. In 2008, the third wave of the SHARE project, dubbed ‘SHARELIFE’, was fielded. In contrast with the first two waves, which dealt primarily with the contemporaneous condition of older people at the time the survey was fielded, SHARELIFE focused on retrospective life histories, eliciting details on key life events from childhood to the present day. The survey was designed to allow analysis of the European welfare state by providing information on a range of life decisions that could then be placed in a national institutional context. For our purpose, the most useful variables are those dealing with employment status and earnings at different times in the individual’s life: these allow us to construct full earnings histories for retirees, and partial earnings histories for those still in the work force. There are many practical benefits to the retrospective life history method47, and it provides access to a great deal of information that cannot be obtained in the present by other means48. However, there are some clear difficulties with using retrospective life history as a substitute for a panel survey conducted at regular intervals over an individual’s life. The most problematic is data quality: when asking older individuals to recall events from 50 years ago, an element of error is expected, some of which may be systematic rather than random. To minimize the potential problems with dating specific events, SHARELIFE implemented a state‐of‐the‐
art event history calendar structure based on an earlier ELSA survey. This survey technique begins with the interviewer asking questions about highly salient life events (the birth (and, if applicable, death) of children, the beginning and ending of cohabiting relationships, places of living), and placing these events on a computerized calendar grid that both the interviewer and respondent can see. The interviewer goes on to other survey modules, where events might be subject to greater error, using the calendar of life events to help enhance recall by relating these other life events (such as the start of a particular job) to the salient life events (e.g. prompting recall by asking “Did you start that job before or after your 47
The SHARELIFE team cite speed and cost as important benefits In some cases, it may be possible to link survey data to national administrative data (such as tax returns) to obtain information on certain aspects of life, but this is often difficult and the scope of administrative data is often narrowly focused around the purpose for which the data is being collected 48
89 wedding?”). The SHARELIFE interface also allowed interviewers to relate the date of an individual’s life events to notable country or world events, providing the interviewer with a searchable database of relevant events that the interviewer could use to anchor an event date (e.g. a respondent might indicate that an event took place a few months after the first moon landing; the SHARELIFE survey interface would allow the interviewer to search for that date. Alternatively, the respondent might indicate with uncertainty that a life event took place in 1969 or 1970, and the interviewer can then find salient public events from those years and ask (e.g.) whether the life event took place before or after the moon landing). A full discussion of the challenges of retrospective lifecycle data collection, and the means by which the event history calendar seeks to mitigate these challenges, is available in the SHARELIFE methodology book (Schröder 2011). While the SHARELIFE data undergoes a meticulous data cleaning process before being released, it is important to note that the cleaning team takes a conservative approach to altering original responses and does not make edits to the data based purely on grounds of subjective implausibility. In the next sub‐section, we describe the variables we use and the additional processing that we conduct. CleaningandProcessingtheSHARELIFEEarningsData
In this section we describe our processing of the SHARELIFE earnings data to produce lifetime earnings profiles for each individual in our final dataset, the validation of the data, and sample trimming to remove low quality data. To provide structure for the following section, the basic building blocks of our dataset are questions that cover for each individual: a)
b)
c)
d)
the start year of each job the starting monthly salary of each job, net of taxes (for those still working) the net monthly salary on the current job (for those who are retired) the end net monthly salary on the ‘main’ job of the career Our basic strategy is to take these known salary points (every starting salary and one current/end salary), and impute the salary for the intervening years. Cleaningthedata:initialsampleselection;ambiguous,misrememberedandforgottendata
The initial selection criteria restricted the sample to individuals who were male; still alive; and aged 50‐
75 at the time of the SHARELIFE survey, and for whom it was possible to obtain education level data from SHARE Wave 1 or 2. We then conducted a number of internal quality checks on the data, and removed individuals from the data who exhibited clear errors of coding or recall. We removed individuals who claim to finish a job before starting it (Nth job end year < Nth job start year) or who claim to start one job before starting a preceding job ((N+1)th job start year < Nth job start year). To simplify the eventual imputation, we also drop individuals who report starting a new job before ending a previous job ((N+1)th job start year < Nth job end year). Note that in the questionnaire, 90 individuals are asked to combine series of small part‐time jobs into a single reported ‘job’, so this last restriction only excludes a small number of people who report substantial overlapping jobs. For each salary point, the individual was asked in which currency the salary was denominated. In some cases this currency was not codable, and the individual was removed. In other cases, the currency was ambiguous (e.g. the generic “Francs” could theoretically refer to French, Belgian, Swiss, or some other currency); when choosing among possible currencies, we assumed that the currency was the national currency of the individual wherever possible, and otherwise excluded the individual. We also took the simplifying step of excluding the relatively small number of individuals who reported a currency that was not the currency of their home country. We believe these currency reports are subject to above average degrees of error, but the exclusion also dramatically simplifies pension modeling by not requiring that we calculate partial pensions based on work across different countries. Finally on currency, we assumed that individuals who reported an amount in a currency that does not fit the timing of the introduction of the Euro (i.e. reporting in Euros prior to its launch, or in national currencies after they had been replaced by the Euro) were nonetheless representing a correct ‘value’. By inspection of countries with large fixed exchange rates against the Euro, it appeared clear that the majority of these temporal currency discrepancies represented people reporting a national‐currency‐
equivalent amount in national currency, rather than reporting a Euro amount and misreporting the name of the currency (and, vice versa for the corollary discrepancy of reporting values in Euros prior to the introduction). Moreover, survey respondents were permitted to give answers in whichever currency they felt they could give the most accurate answer, and so we would expect these apparent discrepancies. Based on this decision, we convert all these amounts into their temporally correct currency via the fixed exchange rates. In some cases individuals reported that they did not remember the starting salary on one or more jobs. We excluded individuals who started their career with 2 or more jobs for which they forgot the salary, who began their career with more than 5 years of jobs with forgotten wages, or have more than 10 years of forgotten wages. ProcessingtheData:RealGrossValues,PlausibilityChecksandInterpolation
We converted all known income values to real PPP‐adjusted 2009 German Euros, using standardized national CPI values and PPP conversions from the OECD where possible, and extending the date range using national websites where possible.49 Where values were not available for the earliest years of the dataset, we imputed values backwards using an average of the 10 earliest years available. After converting all values to real income, we made some additional trims to the data before imputing values. First, we remove anyone who claims (after conversion to real currency) to earn a net income of more than 40,000 Euros per month. These people most likely represent data errors, and even if correct are extreme earnings outliers. We also remove individuals whose real income values show a drop of 49
For a starting point, several useful suggestions and sources are provided in a SHARE working paper (Trevisan, Pasini et al. 2011) 91 more than 10% on their main or current job (the only jobs for which we have both a reported start and a reported end/current value). While it is possible for wages to decrease in real terms, a large number of these cases appear likely to be memory recall errors, most likely underestimating how currency has changed over time with inflation: these cases typically report modest nominal increases over time that are overwhelmed by compounded inflation. SHARELIFE income values are reported as net income rather than gross income, but gross income is used for calculating pensions. We therefore used OECD taxation programs to create conversion factors for net income to gross income. For years prior to the collection of OECD taxation program data, we used the same tax brackets (in real terms) and tax levels as the earliest year of data. After converting to gross income, we interpolated between known income points. Several combinations of known income points were possible: 1) Known start and end salary of main or last job: we interpolate between the two points log‐
linearly 2) Known start salary of one job and start of a following higher salary job: as above, we interpolate log‐linearly 3) Known start salary of one job and start of a following lower salary job: as we generally believe wages are sticky, and it is more likely to be a step down to a lower salary than a gradual process over the course of the previous job, we keep real salary flat on the original job 4) Known start salary of one job followed by a known period of zero earnings: as with point (3), we hold real earnings flat After this imputation, we convert these calendar‐year income histories into age‐year income histories using the individual’s birth year and compared the resulting data with income values from SHARE Waves 1 and 2. In general our constructed dataset seemed to be of high quality. One possible source of error in the SHARELIFE responses was giving annual income responses to questions about monthly income: when an individual in our dataset had income values that differed from observed values in SHARE Waves 1 and 2 by a factor of 10, we flagged these observations as suspicious and excluded them from our final analyses. Finally, because we need an estimate of a person’s potential earnings at a given age even if they have retired, we create a version of our earnings variable which imputes salaries for individuals up to age 70 on a flat real trajectory, assuming that wage increases at this age will tend to be in line with increases in the cost of living. Countryselection
SHARELIFE includes 13 countries. From this 13, we dropped the Czech Republic and Poland at the outset, as their earnings histories would include periods of communist rule during which real earnings would be difficult to calculate and non‐comparable to Western European countries. We also dropped Denmark due to difficulties modeling the overall retirement structure: despite a normal retirement age of 67, many Danes effectively retire as early as 60 with unemployment benefits that may be larger than future pension benefits, but vary with income and contributions to unemployment insurance funds over their lifetime. Other countries were dropped from the analysis for this paper after the cleaning process due to 92 small country‐education cell sizes, excessively high prevalence of suspicious values after cross‐validation with other sources, or both. 93 Descrip
ptionofEa
arningsTrrajectoriessbyCounttryandEd
ducation
After significant processsing, our bassic earnings d
data are depiccted in the Figgures 1 and 2
2. First, we see that there
e is a fairly common patterrn across countries, with m
median earnin
ngs rising shaarply in a individualls in their 20ss and then gro
owing more slowly throug h the rest of tthe working llife before levveling off. In ourr dataset, Gerrmans consisttently have th
he highest ea rnings, follow
wed by the Beelgians; Italian
ns tend to le
evel off earlierr and have lower earnings than the othher countries from their 50
0s onwards.
Figure 10: M
Median Annual E
Earnings by Cou
untry 94 of countries in the previou
us graph, in thhe set of grap
phs below thee main compaarison After the comparison o
is within ccountries. Witthin countries, different groups have d ifferent lifetime earnings and earning trajectorie
es, which in turn affect pensions and saavings. We chhoose to disagggregate by eeducation gro
oup, with ISCED definitions correspondin
ng to “less thaan secondaryy education”, “secondary eeducation” an
nd “tertiary e
education”. A
As is to be exp
pected, in all ccountries thee tertiary educated group h
have zero earrnings at the me
edian at age 2
20 (as most will not have co
ompleted eduucation), but rise sharply sso that by agee 30 they are tthe highest eaarners in all countries. However, the differencces between education groups vary acrross countriess: in Germanyy, France and
d the Netherlan
nds, the two lower educatiion groups haave fairly simiilar earnings, while in Italyy and Belgium
m the three edu
ucation group
ps seem more
e distinct. In m
most countriees, the gap beetween high aand low educaated does not ggrow much after age 30, b
but in France the gap contiinues to wideen until the laate 50s, resultting in the larggest income d
disparity by ed
ducational status of any o f the countriees. Figure 11: M
Median Annual E
Earnings by Education within co
ountries These earrnings trajecto
ories are important, becau
use different pension systeems interact with real earnings trajectorie
es in different ways. For exxample, in pe
ension system
ms based on the average o
of some numb
ber of last or besst earning yeaars, a person with a steeply increasing earnings trajeectory will receive a higheer pension than a person with the iden
ntical lifetime
e earnings sprread more evvenly across ttheir working life. On the other hand, pension systems with caps o
on pensionablle earnings teend to give higgher pensions to 95 those who earn at the capped amount every year than those who earn the same total amount with some years below the cap and some above the cap. InstitutionalDetailsonPensions
In this section we describe institutional differences across countries regarding the calculation of pensions. In gathering this information we found the OECD’s ‘Pensions at a Glance’ series particularly helpful, along with chapters from the NBER volume on Social Security programs and retirement around the world and the European edition of the International Social Security Association’s review of Social Security programs around the world (Gruber and Wise 2007; OECD 2011; International Social Security Association 2012). Where relevant, we also consulted websites of the relevant national governmental departments, particularly in order to understand how reforms enacted during the lifetime of our sample are applied. Germany
In Germany, the main pension system is a points‐based scheme, in which workers earn ‘points’ based on their earnings each year, and the pension received on retirement is calculated by multiplying the sum of accumulated points by the redemption value of the points. The ‘cost’ of a point varies year to year, and is equal to the average earnings of all the contributors in a given calendar year. Pensionable earnings are also subject to a cap, slightly more than 200% of average earnings. In 2008, the cost of a pension point was 30,625 Euros, with a redemption value of 317 Euros. This implies that an individual working at the average level for 44 years from 21 to 65 would receive a pension that replaced approximately 46% of earnings. Given the accrual structure, there is no progressivity in the pension system at lower levels: low earners and average earners receive the same replacement rate. High earners are subject to the earnings cap on pensionable earnings, and so receive a lower replacement rate than low and average earners. Two workers with the same average earnings may earn different pension amounts if one has a steeper earnings trajectory, as they will be more affected by the earnings cap. Normal retirement age for members of our sample50 is 65. Early retirement age is 63; there are no actuarial penalties for retiring early if the individual has an earnings history with at least 35 years of contributions, otherwise the adjustment is a 3.6% reduction for each year early. Late retirement is rewarded with a 6% increase in pension for each year the pension is deferred. Overall, the pension rules increase the pension by approximately 1% for each year worked at average earnings, and then by 6% for each year of deferral. France
In contrast with Germany, France has a complex pension system, made more complicated to model by the changes in policy that have been enacted during the working lives of our analytic sample members. We model two overarching tiers: the public earnings‐based contributory pension, and the mandatory occupational pension. 50
At the time of writing, pension reform in Germany is set to increase the retirement age incrementally up to 67 for the cohort born in 1964 96 The public earnings‐based pension is a defined benefit pension, calculated as 50% of the average earnings over some number of ‘best years’, subject to reduction factors. The number of ‘best years’ was 25 in 2008, having incrementally increased from a base of only 10 years in 1993. For given individuals with similar average earnings over their entire careers, this pension may therefore be very different depending on the earnings trajectory and cohort: ceteris paribus, steeper earnings trajectories lead to higher pensions, particularly for those retiring in 1993 or earlier. The pensionable earnings are subject to a cap, which is approximately equal to national average earnings, which reduces the effective replacement rate received by higher earners from this pension. Reduction factors depend on both the age at which the individual retires and the number of quarters of contribution required for a full career; this latter number incrementally increased from the 150 (37.5 years) to 160 (40 years) between 1993 and 2003. Normal retirement age for this tier is 60. Deferral is rewarded with a 5% increase per year. If an individual works a full career but still generates a very low pension, the pension value is raised to the ‘minimum contributif’ level, but this mostly applies to those with very low wages and/or significant amounts of part‐time work. France has several occupational pension systems. We model the ARRCO system which covers the majority of private sector workers. This scheme is a points‐based system, with contributions made each year divided by the cost of a pension point, and the final pension depending on the redemption value of the total number of points accrued. Benefits are earned on 6% of earnings under the public sector earnings cap, and on 16% of earnings between the public sector earnings cap and three times that cap. Earnings higher than three times the public sector earnings cap earn no additional points. However, this upper threshold affects only very high earners (approximately: those earning more than 100,000 Euros per year). As with the public pension, any reductions for early retirement depend on age and number of years of contribution. Late retirement does not lead to actuarial adjustment, but the worker continues to accrue ARRCO points. Italy
While the Italian pension system has undergone several reforms in the past 20 years, nearly all the individuals in our sample are grandfathered under the old pre‐reform system, so it is this that we describe here.
The Italian pension system (as modeled) has one main component: a public defined benefit pension. The pension aims to provide a replacement rate of 80% of the last 5 years of earnings, accruing at a rate of 2% per year for a maximum of 40 years of work, with no cap on earnings. A relatively low minimum pension (around 6300 Euros) provides a floor on pensions for those with extremely low pension entitlements. The system is only slightly progressive (few full‐career workers fall under the minimum pension), and the focus on just five years of earnings for calculating the base amount means that workers with fairly high average earnings can receive a lower pension than workers who had lower earnings on average over their career but received a high wage in the last five years of their career. Normal retirement age is 60 (with at least 15 years of work), though it is possible at a younger age after 35 years of work; there are no actuarial adjustments made for deferred pensions, though an individual may wish to prolong their career if this will significantly raise the average of their final five years of earnings. 97 Belgium
The main pension provision in Belgium is a public defined benefit system. This system aims to replace 60% of average lifetime income (or 75% if the individual has a dependent spouse), with a cap on pensionable earnings in a given year of roughly 47,000 Euros (around 120% of average earnings in Belgium); earnings less than roughly 19,000 Euros are credited with this value through the minimum annual credit for low earners. A minimum pension of roughly 12,000 Euros is available to those who have worked full careers but have not earned enough to qualify for a higher pension. The system is therefore progressive for low earners and capped for high earners. The full pension requires 45 years of work, and is reduced proportionally for shorter careers. Normal retirement age is 65, but retirement is possible from 60 after 35 years of work. No actuarial adjustments are made for earlier retirement, but most early retirees will face reduced pension due to not having contributed a full 45 years. No actuarial adjustment is made for deferring retirement, but delaying may allow an individual to increase their pension by filling in gaps in their contributions to reach 45 years, or allow the base amount to increase (the pension is based on the last 45 years of earnings, so an additional year of high earnings at the end of the career may displace a year of low earnings at the start of the career in the pension calculation); but for most individuals, there is little to gain from additional years of work at older ages. Netherlands
The pension system in the Netherlands consists of two main components: a basic fixed pension that does not depend on earnings, and an array of occupational pensions that cover the vast majority of Dutch workers. For a single person, the basic pension amount is roughly 13,000 Euros (around 30% of average earnings), accrues at 2% for each year of living or working in the Netherlands, and is payable from age 65 with no early receipt or deferral possible. The occupational schemes have significant heterogeneity, with hundreds of significant pension funds and the option for employers to offer their own schemes. A large majority of these plans are defined benefit plans, and a large majority of defined benefit plans are based on lifetime average earnings. A typical plan provides an accrual rate of 2% of pensionable earnings per year of work; pensionable earnings are not subject to any maximum cap, but a ‘franchise’ amount (typically: 10/7 of the public pension amount) is subtracted from the raw earnings. Thus, an individual who earns exactly this franchise amount every year over the course of their working life would receive no occupational pension, but the public pension would be sufficient to replace 70% of the individual’s average earnings. An individual with lower earnings would receive a greater replacement rate; the overall system is very progressive for low earners, but individuals with high earnings receive similar replacement rates to those with average earnings. Normal age of pension receipt is 65. Different occupational pension plans vary in their rules on pension deferral; based on a sampling of large firm plans, we assume that deferral of an occupational pension leads to an 8.5% increase per year of deferral. 98 PENSIONCALCULATION
In this section, we provide more detail on the assumptions and modeling we used for each country, and demonstrate graphically the generosity, progressivity, and accrual properties of the modeled pension systems. Germany
Public pension: The pension calculation in Germany is a simple multiplication of total ‘points’ accrued by the redemption value of a point: ∗
/
,
where ycap is the maximum pensionable earnings in a given year, pcost is the cost of a pension point in a given year, and pvalue is the value of a pension point in the year it is drawn. Normal retirement age is 65. Early retirement results in a reduction of 3.6% for each year. Delayed retirement results in an increase of 6% for each year. France
Public pension: The French public pension is a defined benefit based on a number of years of salary, with a replacement factor that varies with age and years of contribution, calculated as: Z ∗ min 1,
∗
,
/ where q is the number of years of work, qreq is the number of years required for a full pension (which ranges from 40 to 42 for people for our sample based on cohort), X is the number of ‘best years’ over which the pension is calculated (which ranges from 10 to 25 for people in our sample), and Z is a replacement factor, calculated as 0.50
0.05 ∗
0,
65
,
Deferred retirement provides a gain of 5% per year of deferral. The value is raised to the ‘minimum contributif’ pension value if an individual has a full career but still generates a pension below this floor. Mandatory occupational pension: The mandatory occupational pension is a points‐based system, with the total points accrued multiplied by the redemption value, calculated as: 99 ∗
/
pvalue is the value of a pension point, and pcost is the cost of a pension point in a given year. ypens is the pensionable earnings in a given year, equal to the sum of: 6% of earnings below threshold k ; and 16% of earnings between threshold k and the earnings cap of 3k. Early retirement from occupational plans is possible, subject to a similarly structured penalty as in the reduction of the replacement factor in the public pension: the occupational pension is reduced by approximately 4% for each year of age/work the individual is short from the normal retirement age or the required number of years of work, whichever results in the smaller deduction. Belgium
Public pension: The Belgian public pension is a defined benefit, based on 45 years of pensionable earnings (subject to a cap and floor), calculated as: ∗ min
45
,1 ∗
⟨
,
,
⟩ / where Targ is the target rate of 0.6 for independent individuals or 0.75 for individuals with a dependent spouse, ycap is the limit on pensionable earnings, yfloor is the minimum entitlement to which very low earnings are raised, and q is the number of years worked. If this pension calculated does not meet the minimum pension level, it is replaced by the minimum pension at public retirement age. Italy
Public pension: The Italian public pension modeled is a simple defined benefit based on the last 5 years of salary, calculated as: 40,
∗ 0.02 ∗
/5 where q is the number of years worked. 100 Netherlands
Public pension: The basic public pension in the Netherlands is assumed to be flat amount for all people who have lived in the Netherlands for at least 50 years; it is received at age 65, with no early receipt or deferral possible. Quasi‐mandatory occupational pension: The typical occupational pension in the Netherlands is a defined benefit, based on average lifetime pensionable earnings (earnings over a threshold tied to the public pension), calculated as ∗ 0.02 ∗
max 0,
10
7
/ where q is the number of years worked, and frch is the ‘franchise’ deduction equal in size to the public pension. The actuarial adjustment for deferral is assumed to be 8.5% per year. 101 aracteristicssAcrossCou
untries
PensionSystemCha
de across countries is a verry basic meassure of pensio
on generosityy. In The simplest comparison to be mad
e below, we summarize the
e median rep
placement ratte from pensions of averagge lifetime the figure
earnings ffor each coun
ntry in the dattaset, as accrued by age 600. This is calcculated as thee annual penssion amount an individual w
would receive
e on reachingg retirement aage if they ceaased workingg at age 60, divided byy their averagge lifetime eaarnings up to age 60. Whilee this measurre is not perfeect and conflaates various pe
ension characcteristics, it does demonsttrate a basic leevel of generrosity across ccountries. This is highest in
n Italy, where under the old
d system it w
was typically ppossible to recceive 80% of their averagee salary of tthe last 5 years of their carreer, which m
might easily trranslate to 1000% of averagge salary overr the lifetime given the earn
nings trajectories shown prreviously. It iss lowest in Geermany, where the relative cost of pe
ension points and redeemaable value of pension poinnts keep replaacement ratess low.51 Figure 12: G
Generosity of Pe
ensions However, these countrry‐level medians obscure sseveral imporrtant featuress of pensionss. Next we turrn to the progressivity of pensions. Figure
e 4 is not scaled to emphaasize the diffeerent pension
n levels acrosss 51
The rankk order is the same as that caalculated for sttylized individuuals by OECD (O
OECD 2011), exxcept that theiir modeling o
of the new Italian pension system – rather than the pensiion system into
o which our saample was grandfathe
ered – places Ittaly as the seco
ond most gene
erous rather thhan the most ggenerous. 102 onstrate within countries h
how the replaacement ratees differ for in
ndividuals bassed countries,, but to demo
on the ave
erage lifetime
e earnings. Here we ssee sharply prrogressive pe
ension system
ms in Belgium and the Neth
herlands for vvery low earners, due to the
e relatively hiigh minimum pensions offfered in thosee countries to
o guarantee a minimum standard of living, but becoming lesss progressive
e for mid‐to‐hhigh earners. France and G
Germany both
h demonstrrate steadily d
downwards sloping relatio
onships in theese graphs. Itaaly has very liittle progresssivity built into the pension ssystem, and tthe replacement rates – ass measured compared with average lifeetime earnings –
– vary widely,, due to the ssmall numberr of years at thhe end of thee career that aare used to calculate pensions. Italians with higgh earning job
bs early in theeir career who
o take lower earning jobs later will receivve a very low pension relattive to lifetim
me earnings. Figure 13: P
Pension Replacement Rates by C
Country and Me
ean Lifetime Earrnings The progrressive pensio
on systems in most countrries suggest thhat individuals with very lo
ow earnings m
may need veryy little financial wealth at rretirement wiith which to ssustain their sstandard of living, as pension systems m
may replace m
more than 100
0% for these people; and tthe precipitous drops in reeplacement rate for low‐m
mid earners in Belgium and the Netherlaands might le ad to particular differencees in wealth accumulation as individuals move from low earn
nings to slighttly higher earnings. 103 portraying the same inform
mation is giveen below, witth pension en
ntitlements An alternaative way of p
(accrued b
by age 60) on
n the y‐axis. H
Here we get a better sensee of the impacct of the earn
nings caps on pensionab
ble earnings: Germany hass the most bin
nding earninggs cap in our dataset, refleected in the seevere flatteningg of the pensio
on when meaan lifetime eaarnings approoach 40,000 Euros. There iss also reasonably wide dispersion in Germany, reflectting the fact tthat individuaals with differring distributiion of incomee eir lifetime will be more orr less affected
d by the earn ings cap. Belggium also hass an earnings cap across the
that come
es into play att higher earnings, and again there is disspersion caussed by the intteraction betw
ween earnings ttrajectories and the cap. France has different earninngs caps for different pillarrs in the penssion system, and the overall effect of the
ese is not obvvious in the ggraph. As in th
he previous seet of graphs, the relatively generous minimum pensions in Belgium and the Neetherlands afffect the patteern of the scatterplo
ots at low earrnings, and Itaaly displays litttle progressiivity and a lott of dispersion
n. Figure 14: P
Pension Entitlem
ments by Countrry and Mean Life
etime Earnings
104 dividual workkers is The third key characteristic of pension systems tthat affect thee marginal deecisions of ind
the effecttive rate of acccrual and adjjustment in p
pensions: how
w much the peension increaases for each additional year of workk. Prior to normal retireme
ent age, this ttypically invo
olves increasin
ng a number of years of contributions (up to a limitt), along with country‐speccific ways in w
which the prim
mary pension
n amount iss increased – for example, in Germany, it is simply a n accrual of p
points; in Fran
nce and Italy,, depending on the earn
nings trajectory, it may invvolve increasinng the averagge of the salary years used
d for calculation of the base
e amount. After normal rettirement age,, pension systtems in somee countries (particularly Germany and the Neth
herlands) makke significant actuarial adjustments to p
pension amounts if pension
ns are deferre
ed, while othe
ers do not. Th
hese effects leead to the patterns seen b
below. Figure 15: A
Accrual of Annua
al Pension by Co
ountry and Educcation Group There are
e clear differences across ccountries in th
he way pensioons increase over time. Geermany and tthe Netherlan
nds have big kkinks in the m
marginal pension increase ffor an additio
onal year of w
work when no
ormal retiremen
nt age is reach
hed, significantly changingg the incentiv e to stay in th
he labor forcee. Also, the ch
hange in pension
n values repre
esented by th
he slopes of these lines difffers greatly aacross countries: working ffrom 60 years o
old to 70 yearrs old leads to
o extremely laarge increasees in pension in Germany, France and th
he Netherlan
nds, but only a moderate increase in Be
elgium and virrtually no increase in Italy. This leads to
o dramatic difference in incentives accross countrie
es, with workkers under thee Belgian and
d Italian pension systems faacing relative
ely little incen
ntive to contin
nue, workers in France faccing a relativeely high incentive to continu
ue, and worke
ers in German
ny and the Ne
etherlands fa cing moderatte incentives for a time 105 ncentives at o
older ages. W
We expect thatt these differrences should
d have a signifficant followed by stronger in
impact on
n retirement d
decisions. Recastingg the same data in a differe
ent way, the ggraph below sshows how m
much the annual pension payable changes from year to year, assuming a d
discount rate of 2% and co
ountry‐specifiic survival ratte to the next yyear. Figure 16: P
Percentage Chan
nge in Pension A
Accrued by Coun
ntry and Educatiion. Pension acccrued by followiing year discoun
nted by mortalityy risk and 2% discount factor. e than other ppensions in th
he years up to
o age 60, duee to Here we ssee that Frencch pensions increase more
the multip
ple ways earlyy retirement and receipt o
of pensions arre penalized aand the largee increase in b
base pension amount possib
ble by workin
ng a few more
e high earningg years in a pension system
m that focusees on some finaal years of earrnings. The syystems for Ge
ermany and thhe Netherlands generally produce small decreasess prior to norm
mal retirement age but prrovide an incrrease after no
ormal retirem
ment age. Italyy and Belgium aare fairly conssistently negaative once discounting andd survival probabilities are taken into account. TThese effects influence the
e retirement date that maaximizes penssion wealth, w
which in turn we expect to influence rettirement behaavior.
106 EffectofPensionsonRetirementandSavingsDecisions
Having documented the significant differences across countries in the functioning of their pension systems, we now relate these differences to variation in observed behavior across and within countries. We restrict our sample to individuals who had retired by the time of SHARE Wave 2, to ensure our wealth measures do not reflect people who are still in the wealth acquisition phase of life and to allow us to use a consistent sample for both our wealth analysis and retirement age analysis. In the graph below, we relate pension replacement rates for individuals in our dataset (as calculated using SHARELIFE earnings histories and our own country‐specific pension calculators) to a standardized measure of financial wealth derived from SHARE Wave 2. Each data point represents a country‐
education cell, as previously defined. 52 Our wealth measure is the financial wealth (checking accounts, stocks, bonds and balances held in Individual Retirement Accounts) held by the individual in SHARE Wave 2, divided by the mean lifetime earnings of the individual between age 20 and age 60. On the x‐
axis, we have the average income replacement rate accrued by age 60, calculated as the pension entitlement accrued by age 60 divided by mean lifetime earnings of the individual. This replacement rate reflects a combination of the generosity and progressivity of the pension systems, as applied to the real life earnings histories in our dataset. 52
With the exception of Germany, for whom we omit the less‐than‐secondary education cell due to insufficient sample size after restricting to those who had retired by SHARE Wave 2 107 Figure 17: R
Replacement Rates and Accumu
ulated Financial Wealth. Cells coollapsed at the ccountry‐education level. Educaation level indicatted by parenthe
eses: (1) less thaan secondary; (2
2) secondary; (3)) tertiary It appearss that there iss a clear negative relationsship betweenn pension replacement ratees and accumulated wealth in
n the graph ab
bove. For cou
untry‐educatioon cells with high median replacementt umulation of financial weaalth. While th e result admiits of alternattive explanatiions, rates, there is low accu
it is very p
plausible to believe that hiigh replaceme
ent rates from
m pensions discourage acccumulation of financial w
wealth: individuals who caan rely on the
eir pensions m
maintaining th
heir income aaround their average lifetime level ccan maintain their average
e lifetime stanndard of livin
ng without rellying on accumulated savings and investmen
nts. It is also iinteresting too note that th
his relationshiip is borne ou
ut with variaation both bettween and within countrie
es, decreasingg the likelihoo
od that this iss driven by so
ome other cou
untry‐specific factors that w
we are not modeling. Our prefe
erred measure
e of wealth above is a meaasure of finanncial wealth, tthe type of w
wealth that is typically m
most liquid an
nd most easilyy used to finaance consumpption. We can
n also look att a measure o
of total wealth, including housing and other assets,, as an alternative; althouggh there are likely major sources of heterogeneity in housingg markets acrross countriess53 and the wealth may bee less liquid, itt is undeniable that some people may ffinance consu
umption in reetirement parrtly by downsizing their 53
And, indeed, differentiial access to th
he housing marrket for differeent socio‐econo
omic groups accross countriess 108 entering into reverse morttgages, or using some otheer method for realizing thee value of housing housing, e
wealth. In
n general, the
e countries in
n our sample aare not counttries who havve embraced reverse morttgage / equity re
elease schem
mes, as compaared to the En
nglish‐speakinng countries o
of the OECD w
where these schemes aare more prevalent (Clerc‐‐Renaud, Pére
ez‐Carillo et aal. 2010). Thiss reduces ourr concern that omitting h
housing wealtth dramatically misstates retirement reesources, butt nonetheless we perform a similar an
nalysis using aa net wealth m
measure (see Appendix). TThe result is aa marginally significant neggative relationsh
hip driven by cross‐countryy variation. Next we tturn to the im
mpact of accru
ual incentivess on retiremennt age. Again, each cell in the graph beelow representts a country‐e
education gro
oup. The y‐axiis is the meann retirement aage within the cell. The x‐aaxis measure iis the marginal increase in
n pension for w
working an a dditional yeaar at age 65, ccalculated forr individualls and collapssed at the median. Figure 18: M
Marginal Pension Accrual and M
Mean Retiremen
nt Age. Cells coll apsed at the co untry‐education
n level. Educatio
on level indicatted by parenthe
eses: (1) less thaan secondary; (2
2) secondary; (3)) tertiary. Here, as w
we might expect, we see a strong positiive relationshhip between tthe benefit off working an additional year and the
e age at which people cho
oose to retire.. Countries where pension
ns receive a sttrong actuarial aadjustment te
end to have h
higher retirem
ment ages; coountries with little to no acctuarial adjustment tend to have lower retirement ages. This again is a very plausiible result; ho
owever, in thiis 109 persuasive thaan the previo
ous graph duee to the relian
nce on differeences betweeen case the rresult is less p
countries to drive the rrelationship, without such
h a clear corolllary between
n education ggroups within countries. or graph deal t with the capacity to incrrease pension
ns in Next, we look at longer term incenttives. The prio
one year aat a specific aage; but the kkinked accrual curves mea n that individ
duals making a decision at a given age must take into account no
ot only the m
marginal beneffit of a single year, but also the future e graph below
w, we examinne how retirement ages arre related to tthe schedule of benefit inccreases. In the
pension in
ncentives invo
olved with wo
orking a decaade from age 60 to age 70 (a period thaat averages ou
ut the accruaal kinks in Germany and th
he Netherland
ds). Figure 19: D
Decade Pension Deferral and Re
etirement Age. C
Cells collapsed aat the country‐eeducation level. Education levell indicated byy parentheses: ((1) less than seccondary; (2) seco
ondary; (3) tertiiary. In Figure 1
10 we see a ssimilar relatio
onship betwee
en gains from
m pension defferral and retiirement age, with increased benefits from
m working to older ages asssociated witth increased rretirement agges. Again, thiis appears to be mostly d
driven by diffe
erences betw
ween countriees; but here w
we perhaps seee a little morre support frrom the within‐country vaariation – in th
hree countriees (Belgium, Ittaly, Germany), the group with the lowesst median gain
n also has the
e lowest mean retirement age, and it iss only the Nettherlands which appears to have the op
pposite relationship betwe
een pension ggain and retirrement age. 110 we adapt a co
oncept from tthe “peak val ue” literaturee (Coile and G
Gruber 2007). Finally in tthis section, w
Here, we estimate the value of pension wealth aat each possibble retiremen
nt age, discou
unting future onal discount rate of 2% an
nd country‐sp
pecific mortality income sttreams into present value using a perso
rates. The
e resulting net present valu
ues of pensio
on wealth alloow us to estim
mate the age of retirementt that maximizes pension we
ealth. This doe
es not take in
nto account thhe disutility o
of work / utilitty of leisure, ely causes peo
ople to retire earlier than tthe date at w
which pension
n wealth would be maximized, which like
but does provide an alternative measure of the p
purely financcial incentivess built into thee pension systems. e again plot m
mean retireme
ent age on thhe y‐axis, and on the x‐axiss take the meean of In the graph below, we
the year aat which the p
present value
e of pension b
benefits are m
maximized. W
We see a signifficant positivee relationsh
hip between tthe pension b
benefit peak yyear and retirrement age, aagain mostly d
driven by cross‐
country differences, w
which reflect aa combination
n of the accruual differencees seen in prevvious graphs and the earliest date at wh
hich people caan begin receiving a pensioon. Figure 20: P
Pension Benefit Peak Year and M
Mean Retiremen
nt Age. Cells colllapsed at the co
ountry‐educatio
on level. Educatiion level indicatted by parenthe
eses: (1) less thaan secondary; (2
2) secondary; (3)) tertiary. Clearly, pe
eople do not choose retire
ement purelyy to maximize discounted p
present valuee of retiremen
nt wealth, ass the best fit line does not have a slope
e of exactly 1. This is unsurrprising, given
n that working additional years involvves a loss of le
eisure utility ((and may alsoo displace oth
her productive activity). A 111 further step to take would be to use an “option value” framework to capture the value of consumption and leisure at each age more explicitly (Stock and Wise 1990). This approach requires explicit specification of utility functions and parameters (as in the model that motivates this paper), is sensitive to parameter values that must be chosen, and would benefit from work that has yet to be done estimating these parameters. Therefore, this is left for future research. Overall, these analyses of pension institutions taken together provide prima facie evidence that pension incentives influence work, saving and retirement decisions. Using a unique dataset to capture the real pension environment unique to each real individual, we were able to show that groups with high replacement rates tend to accumulate less financial wealth and groups with low accrual/adjustment rates tend to retire earlier. Helpfully, our set of countries include at least one in each quadrant of the set [High Replacement Rates, Low Replacement Rates],[Strong Accrual Incentive, Weak Accrual Incentive], which means these two effects are unlikely to be conflating each other. Nonetheless, these basic patterns at the aggregate level could certainly be explained by other factors at the country level, with national pension systems designed to reflect national preferences in the balance between private and public retirement provision. The analyses presented here are best seen as a first step in describing broad correlations in our dataset between wealth accumulation and labor supply outcomes at the individual level and the incentives created by institutional pension structures. The next section describes future directions for research. 112 Futureresearch:fromaggregateanalysistotheestimationoflife‐cyclemodels
Finally, we turn to future directions for research using this unique dataset with dynamic programming models, which can more completely model individual behavior and provide a better basis for policy simulation than the aggregate analyses we have performed in this paper. The dataset we have constructed includes for the first time the lifetime earnings histories of a large number of people across several countries, together with a detailed modeling of the pension incentives faced by each individual at each stage of their working lives, the age at which the individual retired, and a measure of their accumulated wealth. This provides unique opportunities for the type of lifecycle model discussed at the beginning of this paper, which in turn can provide useful insights into the likely reaction of individuals to different types of pension reform. More sophisticated models can be applied to this data in the future, utilizing a range of additional information to help capture individual circumstances more precisely, and relaxing assumptions that allowed for exploratory analysis but which may be unrealistic. On the informational side: as an adjunct to the main thrust of our work, we imputed spousal income for each person in our dataset at each age of life, using regression coefficients derived from the larger European Community Household Panel (ECHP) dataset. Household income measures can therefore be constructed and analyzed. Similarly, we have processed the SHARELIFE family variables to ascertain household size at each age: variations in family size may have a significant impact on consumption over the lifecycle, and likely are related to country and education level. Real asset returns have been treated as common to all individuals, but data on historic asset returns across countries could be easily incorporated. The model can also be extended in a number of ways. As a consequence of the decision rule driving the program, consumption is a given fraction of a known discounted lifetime income. We found in exploratory modeling that this decision rule often led those in countries with high pension replacement rates to get into very large amounts of debt in young and middle age, which might not be feasible in the real world, as it is not easy to borrow against future income and pension entitlements. So, liquidity constraints may be a sensible extension, particularly for modeling the generous replacement rate countries. As mentioned, individuals are treated in the model as having perfect information on lifetime income, and also on asset returns. Relaxing these assumptions to allow for uncertainty and concern about possible shocks may bring the model closer to reality. In addition, our modeling of the utility of retirement leisure / disutility of work as an age‐independent parameter could be adjusted simply to allow tastes for leisure to depend on age. Finally, the model as constructed focuses exclusively on individual decision‐making, but other research has developed concepts of joint decision‐making for couples and noted its importance for understanding labor force behavior (Hurd 1990; Blau 1998; Gustman and Steinmeier 2000). While data limitations do not allow us to construct lifetime labor histories for a full set of matched husband‐wife pairs, a range of spousal characteristics are available through SHARE and could be incorporated into further extensions of this work. 113 Conclusion
In this paper, we describe the construction, characteristics and value of a novel dataset that combines lifetime earnings histories with a detailed modeling of national pension institutions, uniquely capturing the interaction between lifetime earnings levels and trajectories with the rules of pension systems in a way that allows us to understand the pension incentives and decisions faced by individuals in each year of their working lives. With a large quasi‐panel of respondents with comparable data across multiple countries, this dataset provides new opportunities to exploit cross‐national institutional variation for research purposes without resorting to stylized assumptions about individual income over the lifecycle. In addition, we have provided additional support to two hypotheses: pension systems that provide more generous replacement rates are associated with lower accumulation of wealth to finance retirement; and sharper incentives to continue working at older ages are associated with later retirement. The former effect receives stronger support from our analysis, with both within‐ and between‐country variation showing a negative relationship between replacement rates and wealth, while the retirement age effect is mostly driven by between‐country variation. The main findings in this paper are based on an aggregation of individual pension situations up to the country‐education cell level, but the greatest value in our dataset lies in future work that can calibrate lifecycle models to explain individual behavior as a function of the incentives facing them. In turn, this will allow more accurate policy simulations to address the likely consequences of (necessary) future pension reforms both at an aggregate and an individual level. These important modeling tasks are far from straightforward, but our work has provided a solid foundation on which they can be built. 114 Bibliography
Blau, D. M. (1998). "Labor force dynamics of older married couples." Journal of Labor Economics 16(3): 595. Borsch‐Supan, A. and H. Jurges, Eds. (2005). The Survey of Health, Aging, and Retirement in Europe – Methodology, Mannheim Research Institute for the Economics of Aging (MEA). Clerc‐Renaud, S., E. Pérez‐Carillo, et al. (2010). Equity Release Schemes in the European Union, European Commission. Coile, C. and J. Gruber (2007). "Future social security entitlements and the retirement decision." The Review of Economics and Statistics 89(2): 234‐246. Gruber, J. and D. A. Wise, Eds. (1999). Social Security and Retirement around the World, University of Chicago Press. Gruber, J. and D. A. Wise, Eds. (2007). Social security programs and retirement around the world: Fiscal implications of reform, University of Chicago Press. Gustman, A. L. and T. L. Steinmeier (2000). "Retirement in dual‐career families: a structural model." Journal of Labor Economics 18(3): 503‐545. Hurd, M., P. C. Michaud, et al. (2012). "The Displacement Effect of Public Pensions on the Accumulation of Financial Assets." Fiscal Studies 33(1): 107‐128. Hurd, M. D. (1990). The joint retirement decision of husbands and wives. Issues in the Economics of Aging. D. A. Wise, University of Chicago Press, 1990. International Social Security Association (2012). Social Security Programs Throughout the World: Europe, 2012, United States Social Security Adminstration. OECD (2011). Pensions at a Glance 2011: Retirement‐income Systems in OECD and G20 Countries, OECD Publishing. Schröder, M., Ed. (2011). Retrospective Data Collection in the Survey of Health, Ageing and Retirement in Europe. SHARELIFE Methodology. Mannheim Research Institute for the Economics of Aging (MEA). Stock, J. H. and D. A. Wise (1990). "Pensions, the Option Value of Work, and Retirement." Econometrica 58(5): 1151‐1180. Trevisan, E., G. Pasini, et al. (2011). Cross‐country comparison of monetary values from SHARELIFE, SHARE Working Paper Series 02‐2011, Survey of Health, Ageing and Retirement in Europe. 115 Append
dix
Figure 21: R
Replacement Rates and Accumu
ulated Net Weallth. Cells collapssed at the counttry‐education leevel. Education llevel indicated byy parentheses: ((1) less than seccondary; (2) seco
ondary; (3) tertiiary 116