#9910 September 1999

Transcription

1 #9910 September 1999 The Impact of Pay Inequality, Occupational Segregation, and Lifetime Work Experience on the Retirement Income of Women and Minorities by Olivia S. Mitchell University of Pennsylvania-Wharton School Phillip B. Levine Wellesley College-Department of Economics John W. Phillips University of Pennsylvania-Population Studies Center

2 #9910 September 1999 The Impact of Pay Inequality, Occupational Segregation, and Lifetime Work Experience on the Retirement Income of Women and Minorities by Olivia S. Mitchell University of Pennsylvania-Wharton School Phillip B. Levine Wellesley College-Department of Economics John W. Phillips University of Pennsylvania-Population Studies Center The Public Policy Institute, formed in 1985, is part of the Research Group of the AARP. One of the missions of the Institute is to foster research and analysis on public policy issues of interest to older Americans. This paper represents part of that effort. The views expressed herein are for information, debate, and discussion, and do not necessarily represent formal policies of the Association. 1999, AARP. Reprinting with permission only. AARP, 601 E Street, N.W., Washington, DC 20049

3 Acknowledgments We acknowledge research support from the AARP and the Pension Research Council at the University of Pennsylvania, and computer support from the University of Pennsylvania s Aging Research Center. The views expressed herein are those of the authors and do not necessarily represent the views of the institutions with which they are affiliated. Olivia S. Mitchell Phillip B. Levine John W. Phillips i

4 Foreword In response to the persistent problem of poverty among older women and minorities, analysts and policymakers have sought to determine why these vulnerable groups have such a high likelihood of being poor in old age. Women and minorities bring different labor market experiences to retirement than do white men. For women, lower earnings and fewer years in the work force mean less retirement income because pensions and Social Security typically reward those with higher pay and more years of paid work. Likewise, for blacks and Hispanics, lower levels of education and lower earnings have a negative impact on their anticipated retirement income. The causes of the wage gap that affects women and minorities have been the subject of considerable research. Some of this work has examined the impact of occupational segregation and other labor market experience on these wage differentials. How wage inequality, occupational segregation, and lifetime work experience influence the future retirement incomes of women and minorities, however, has not been as thoroughly studied. Even less information is available about the impact of these factors on the three major components of retirement income Social Security benefits, pension income, and savings. This vital information the impact of pay inequality, occupational segregation, and lifetime work experience on the retirement income of women and minorities is the focus of this report by researchers Olivia Mitchell and John Phillips of the University of Pennsylvania and Phillip Levine of Wellesley College. By merging the first wave of the Health and Retirement Study with two additional files the Pension Provider File and the Earnings and Benefits File (Social Security data) these researchers examine the anticipated wealth available to groups, by sex and minority status, from Social Security, employer-provided pensions, and other financial resources, including housing. The authors find that occupational segregation and pay differences account for most of the difference in the retirement income of women and men. Even controlling for factors such as number of years worked and level of education, two-thirds of the annual difference in the retirement income of nonmarried women and men reflects two key measures: the proportion of women in an occupation and the pay gap during prime earning years. The losses women incur as a result of these pay differences show up primarily in their pension and other income, not in their Social Security benefits. This is because Social Security s progressive calculation formula and the availability of spousal benefits play an important role in helping women counteract the retirement impact of working in female-dominated jobs and having lower earnings. The anticipated retirement incomes of nonwhite men and women do not differ markedly; both have low levels of wealth to draw on in retirement. These findings raise important policy questions about how to improve the retirement income prospects of women and minorities, particularly how to enhance their opportunities to obtain employer-provided pensions and to increase their ability to save. In addition, women and ii

5 minorities could benefit from better information about the long-term consequences of leaving school early, taking low-paying jobs with no career ladders, and working in female-dominated occupations. The findings also raise questions about Social Security reform proposals that might reduce the program s ability to provide a foundation of retirement income security for women and minorities. Jules H. Lichtenstein, Ph.D. Senior Policy Advisor AARP Public Policy Institute iii

6 Table of Contents Acknowledgments...i Foreword...ii List of Tables...v Executive Summary... vi Introduction...1 Prior Studies...2 Data and Statistical Analysis...4 Explanatory Factors...5 Subsample Issues...6 Household Resources...7 Methodological Approach...7 Empirical Findings...8 Regression Results...10 Decomposition Results...12 The Role of Marital History...13 Discussion and Conclusions...14 Data Appendix...22 References...38 iv

7 List of Tables Table 1. Median Projected Retirement Wealth by Race/Ethnicity, Sex, and Current Marital Status 17 Table 2. Median Projected Annual Retirement Income by Race/Ethnicity, Sex, and Current Marital Status Table 3. Predicted Changes in Total Projected Annual Retirement Income Associated with Key Explanatory Variables.. 19 Table 4. Decomposing Differences in Projected Annual Retirement Income by Type: Dollar Gap Attributable to Differences in Respondent Characteristics Appendix Table 1. Unweighted Sample Sizes by Race/Ethnicity, Sex, and Current Marital Status 25 Appendix Table 2. Mean Projected Retirement Wealth by Race/Ethnicity, Sex, and Current Marital Status Appendix Table 3. Mean Projected Annual Retirement Income by Race/Ethnicity, Sex, and Current Marital Status.. 27 Appendix Table 4. Regression Results for Projected Annual Retirement Income by Type 28 Appendix Table 5. Mean Values of Explanatory Variables by Sex and Current Marital Status 32 Appendix Table 6. Unweighted Sample Sizes by Sex, Race, and Marital History. 33 Appendix Table 7. Median Projected Annual Retirement Income by Source and Marital History. 34 Appendix Table 8. Regression Results for Total Projected Annual Retirement Income by Marital History. 35 Appendix Table 9. Percentage of Men and Women in HRS Occupation Codes 37 v

8 Executive Summary Introduction Women and minorities are especially vulnerable in retirement: both groups face a disproportionately high risk of poverty in their old age. Women over the age of 65, for example, are about twice as likely to live in poverty as their male counterparts. Although more women are working and their pension coverage is rising, the labor market experience of women still differs substantially from that of men. By the time women are in their fifties, they have spent fewer years working, have earned less over their lifetimes, and have held lower quality jobs than similar-aged men. Lower earnings and fewer years in the work force translate into less retirement income because pensions and Social Security typically reward those with higher pay and more years of paid work. Minorities, who tend to have lower levels of education and lower earnings, are at a similar disadvantage in retirement. Purpose This report reviews and synthesizes existing research in the area of pay inequality differences in earnings between working men and women and between minority and white workers. It then uses the Health and Retirement Study (HRS) to evaluate the relative importance of personal characteristics and labor market experiences in determining the retirement income prospects of older women and minorities. The report examines the anticipated wealth available, by sex and minority status, from three sources: Social Security, employer-provided pensions, and other financial assets, including housing. This research helps evaluate how the next generation of retirees may fare compared with prior cohorts. Few studies have followed workers into retirement to determine whether labor market differences continue to have an impact at older ages. Methodology The first section of the study reviews the labor economics literature on pay inequality. It discusses the difference between the concept of pay inequality and pay inequity. It examines the sources of the wage gap between men and women, including the effects of occupational segregation, lifetime labor market attachment, education, and experience. The second section of the study estimates anticipated retirement income using the first (1992) wave of the HRS, a nationally representative sample of U.S. households on the verge of retirement. The HRS focuses on respondents aged 51 to 61 in 1992, along with their spouses, and contains information on earnings and labor market work patterns as well as the pension plans of people anticipating or in the early stages of retirement. Two additional information sources were merged with the HRS data to provide information about Social Security benefits and pension income. The Pension Provider File (PPF) contains data from pension plan descriptions and provides estimates of anticipated pension benefits of respondents; the Earnings and Benefits File (EBF) contains information about expected retirement income from the Social Security vi

9 Administration s administrative records. Together, the HRS, PPF and EBF represent one of the richest data resources available to analyze retirement income: no other current data source has detailed, linked administrative records for this group of households. Using these three sources, the anticipated retirement wealth of each household in the HRS was computed from the annuitization of wealth from three sources: Social Security, employer-provided pensions, and other financial assets, including housing. The analysis has two phases. The first explores how much of the difference in projected retirement income between groups is represented by pay inequality, occupational segregation, and differences in labor market attachment. The second phase assesses how women might be expected to fare in retirement if their labor market experiences were similar to those of men, with race/ethnicity and other important factors held constant. Principal Findings Occupational segregation and pay differences account for most of the difference in the retirement income of women and men. Even controlling for factors such as the number of years worked and the level of education, the proportion of women in an occupation and the pay gap during prime earning years together account for two-thirds of the more than $8,000 per year difference, on average, between the retirement incomes of nonmarried women and men. In both cases, the losses women incur as a result of pay differences show up primarily in their pension and other income, not in their Social Security benefits. This is because Social Security s progressive calculation formula and availability of spousal benefits play an important role in helping women counteract the retirement impact of working in female-dominated jobs, and having lower earnings. For instance, for each 10 percent rise in the proportion of women in a job, nonmarried women lose about $157/year in pension income and about $389/year in other income, but only $20/year in Social Security income. Married women lose $375/year in pension income, $173/year in other income, and just $9/year in Social Security income. The study estimates projected retirement income based on anticipated retirement wealth for each household. Married couples have three times the wealth and, when wealth is translated into annual retirement income flows, much more retirement income than their nonmarried counterparts. The median wealth of a married couple on the verge of retirement exceeds half a million dollars. By contrast, more than half of nonmarried people have less than $200,000 in all forms of wealth. The projected annual retirement incomes of nonmarried people are only onethird the level for married men and women. Moreover, nonmarried women can anticipate less retirement income than nonmarried men. The retirement income gap among nonmarried men and women is greatest for whites primarily because minorities have little retirement wealth. The anticipated retirement income of nonwhite men and women do not differ markedly from each other. Minorities are particularly vulnerable at older ages because they have very low levels of wealth to draw on in retirement. Hypothetically, it would be possible to close 75 percent of the overall gap between men and women s anticipated retirement incomes. This would entail raising women s educational level vii

10 to that of men and closing the gap in years of work, average pay, and occupational attainment. Closing these gaps would also help equalize retirement incomes of whites and minorities the largest gains would be for blacks and Hispanics, who have the lowest levels of wealth. Conclusions There has been considerable research on the factors affecting the wage gap facing women and minorities. Some has involved examining the impact of occupational segregation and other labor market experience on these wage differentials. Less understood is the impact of wage inequality, occupational segregation, and lifetime work experience on future retirement incomes of women and minorities. Even less information is available about how these impact three major components of retirement income Social Security benefits, pension income, and savings. The results of this study suggest several lessons and policy implications. First, there are close links between earnings during a lifetime and income during retirement, especially income from pension and other financial accumulations. These links imply that women and minorities could benefit from better information about the long-term consequences of leaving school early, taking lower-paying jobs, and working in female-dominated occupations. They could also profit from improved information about the implications of dropping out of the labor force for extended periods of time, although the effects of this behavior have fewer negative consequences for them. Expanding the information available about the benefits of obtaining higher education on their economic status while working and during old age (in retirement) could also be beneficial to them. There are trends which point to improved retirement income prospects. If women s pay levels continue to climb and women are less segregated occupationally than they have been in the last decade or so, women approaching retirement will have earned more over their lifetimes, enhancing their well-being absolutely and relative to men. As women s labor market histories become more congruent with men s over time, the gap in retirement income between men and women should diminish. viii

11 Introduction Persons age 65 and over were once among the poorest segment of the population in the United States, but today the rate of poverty for this group is at least as low as that of younger segments of the population. Nevertheless, pockets of poverty remain within the older population. Specifically, women over the age of 65 are about twice as likely to live in poverty in the U.S. as are similarly aged men. 1 One factor believed to contribute to women s disadvantage in old age is their different labor market experiences during their prime working years. For instance, women who hold lower-wage jobs, work for fewer years, or work in more female-dominated jobs than men may end up with less retirement income. This is because pensions and Social Security retirement benefits typically reward those with higher-paying jobs and more years of paid work. In addition, other important differences between men and women in terms of education, health, and other factors may also translate into differences in retirement wealth. Similarly, minorities face a disproportionate risk of old-age poverty relative to whites. In this paper we use the Health and Retirement Study (HRS) to evaluate how labor market experiences (including differences in years worked, levels of pay, and types of occupations) and related factors help explain why older women and minorities face relatively poor retirement income prospects. This research is important to young workers entering the labor market who could profitably be alerted to the lifetime implications of their job-market patterns. Policymakers may find this research of interest to the extent that poverty in old age is influenced by job-market conditions and experiences earlier in life. And because women workers today spend more of their lives in paid employment and earn more relative to men than they have in the past, our research helps evaluate how the next generation of retirees may fare compared to prior cohorts. To preview our conclusions, we find that the median wealth of a married couple on the verge of retirement today exceeds half a million dollars, with substantial accumulations in Social Security, pensions, housing, and other holdings. By contrast, nonmarried people approaching retirement are in far worse shape. More than half of the nonmarried population has less than $200,000, counting all forms of retirement wealth. Few can look forward to future pension benefits, and even Social Security wealth is not large. Within the nonmarried group, women are particularly disadvantaged, having a level of wealth about one-quarter lower than that of nonmarried men. This difference is concentrated in the white population, since blacks and Hispanics have very low and relatively similar levels of retirement wealth regardless of sex or marital status. The anticipated retirement wealth values captured in this study translate into median annual retirement incomes of about $28,000 per year (in 1992 dollars) for married men and women, but only $13,000 for single men and $9,000 for nonmarried women. Our statistical analysis of the determinants of these differences in anticipated retirement income points to several factors that explain why women and minorities face such poor prospects in retirement. One reason is that women currently on the verge of retirement have lower lifetime average earnings, which translate into lower pensions and Social Security benefits. Another 1 For an extensive list of references see Levine, Mitchell, and Moore (forthcoming). 1

12 explanation is that women have spent fewer years in the paid labor market than men, a difference that also penalizes them somewhat in retirement. Further, we find that working in occupations heavily dominated by females reduces women s eventual retirement income; by contrast, this effect is not evident for their male counterparts. We also examine the extent to which other differences between men and women contribute to the retirement income gap. For instance, among older people on the verge of retirement today, married women have somewhat lower educational attainment than do men, particularly at the post-college level. This accounts for part of the measured retirement income differential by sex. We also find that blacks and Hispanics, who rely on Social Security for the bulk of their retirement income, typically can expect far lower levels of retirement income than can whites. But the anticipated retirement incomes of nonwhite men and women do not differ markedly. Before we proceed with our analysis, it is useful to define the economic terminology used in this study. By pay inequality, we mean differences in earnings between working men and women. This stands in contrast to pay inequity, a term often used interchangeably with comparable worth, which refers to a policy that seeks to redress wage differentials caused by occupational segregation. Our analysis evaluates the determinants of anticipated retirement income differences, from which policy implications can be drawn, rather than assuming specific policy implications. We include indicators of labor market occupational segregation because it may be a potential contributor to the differences in well-being among older persons. As will be explained below, we focus on retirement income that could potentially be derived from the annuitization of wealth from three sources: Social Security, employer pensions, and other financial assets. Prior Studies Labor economics research to date has explored differences in labor market wages (earnings for paid market work) between men and women as well as between minorities and whites, generally focusing on prime age persons. 2 Two distinct research strands may be discerned in this literature. One approach considers whether observed wage gaps can be explained or accounted for in a statistical sense by differences in characteristics likely to be related to worker productivity, such as age, education, and labor market experience. To the extent that a wage differential exists after controlling for these factors, it is typically attributed to labor market discrimination. This strand of the literature consistently finds evidence of discrimination by this definition. 3 A second approach used in explaining sources of the male/female wage gap evaluates occupational segregation and its impact on wage differentials. Studies have shown that more than half of all women would have had to move from female-dominated to male-dominated jobs in 2 There is no unique definition of prime age in the labor economics literature. In what follows we use this term to refer to the period between age 20 and 50 since the dataset we use provides summary labor market measures over this age range. 3 See Blau and Kahn (1997); Gunderson (1989); and Blau and Ferber (1987). 2

13 1990 if the occupational distribution of men and women were to be equalized. 4 This estimate is lower than the 70 percent figure that applied in Yet despite the diminution of occupational segregation by sex over time, some have contended that it continues to play a role in the wage gap between men and women because traditional women s jobs may pay less than traditional men s jobs. These findings have led to calls by some to implement a comparable worth, or pay equity, pay policy that would adjust the wages of female-dominated jobs to bring them in line with their value or worth. The extent of the adjustment required would depend upon the wage penalty experienced by occupants of female-dominated jobs. Therefore, studies in this area have sought to determine the impact of occupational segregation on the wage gap between men and women. However, many economists have noted that observed differences in pay between maleand female-dominated jobs do not necessarily amount to discrimination because factors related to productivity, such as education and experience, may differ between those holding femaledominated jobs and those in other jobs. 5 Therefore, the statistical methodology typically employed in this literature includes a control for the percentage of women in a worker s occupation in a wage regression that also controls for other productivity-related characteristics of workers. The estimated coefficient associated with this sex composition variable then is interpreted as an estimate of the penalty faced by occupants of female-dominated jobs, holding constant these other factors. Results from earnings studies tend to find that people employed in an occupation dominated by women still earn somewhat lower wages, even after controlling for measurable productivity differences. 6 While the literature on younger workers is lengthy, few researchers have followed workers into retirement to determine whether labor market differences continue to have an impact at older ages. Our previous work (Levine, Mitchell, and Moore, forthcoming) is an exception to this generalization, since in that study we sought to determine whether differences in lifetime labor market attachment accounted for differences by sex in projected retirement income. We concluded in that study that health and family responsibilities had only tiny measured impacts on projected retirement incomes. One drawback of the previous study was that only self-reported labor market data could be used, rather than more precise data taken from administrative records on labor market experience. In addition, that analysis focused on total retirement income and did not consider its components separately. Our more ambitious goal in the present study is therefore to derive and use better quality data than heretofore available on pension and Social Security wealth, and to explore how labor market and other factors influence anticipated retirement income by source. We begin by describing our data and analytic approach and then document sources of differences in expected retirement income flows by sex and race/ethnic status, both in aggregate terms and using multivariate statistical analysis. 4 See Blau, Ferber, and Winkler (1998) and Blau, Simpson, and Anderson (1998). 5 See Macpherson and Hirsh (1995). 6 For evidence on this point see Sorenson (1989 and 1990); Killingsworth (1990); Filer (1989) and Johnson and Solon (1986). 3

14 Data and Statistical Analysis Our statistical analysis takes place in two phases. First, we explore how pay inequality, occupational segregation, and differences in lifetime labor market attachment account for observed variability in projected retirement income across different population groups. Second, we use the first-stage results to assess how women might be expected to fare in retirement if their labor market experiences were to be more similar to men s, controlling for race/ethnicity and other important factors. This analysis uses the HRS, which is a nationally representative sample of U.S. households drawn from a cohort on the verge of retirement (age 51 to 61 in 1992). 7 The HRS provides extensive and detailed demographic, health, wealth, income, and family structure data for respondents and their spouses. We also use information from two additional files containing invaluable information on respondents pension and Social Security benefits. One file, known as the Earnings and Benefits File (EBF), provides measures of expected retirement income derived from Social Security benefits as well as labor market history data. A second file, the Pension Provider File (PPF), contains estimates of anticipated pension benefits. These merged files have been obtained for a majority of HRS respondents who gave permission to link their survey data with administrative records supplied by the Social Security Administration, and also with pension plan descriptions provided by respondents employers. 8 Together, the HRS, EBF, and PPF data represent one of the richest data resources available to analyze retirement. There is no other current data source with equivalently detailed linked administrative records for this cohort. 9 7 For more detail on the HRS dataset, see the Data Appendix describes variable creation for the present study. 8 Because of the confidential nature of these data, researchers may access them only under restricted conditions; see for details. 9 The availability of the Social Security and pension matched data makes the HRS uniquely valuable among all datasets covering retiring Americans. Though Social Security benefits were calculated for most of the age-eligible HRS respondents in the sample, in a few cases this information could not be computed, and the respondent had to be omitted from the sample analyzed in this paper (more detail on sample sizes is given below). One reason for missing Social Security benefits was that respondents gave permission for the University of Michigan to request their Social Security records, but no match was obtained because their records did not match SSA identification information. Another reason is that the Social Security Administration excluded from the match file any respondents receiving Social Security Disability Insurance benefits. Also, some age-eligible respondents declined to sign the release form permitting their Social Security data to be matched with the HRS. In this study we rely on Social Security wealth estimates as well as earnings histories provided in the EBF, so respondents lacking these data are excluded from our analysis. This selection might bias results if those with an EBF file differ from those without a match; we have no evidence that results are biased, and, indeed, respondents lacking consents for a Social Security match are quite diverse. Thus, some of the very wealthy (having high levels of financial assets) did not sign the special release, while some blacks and Hispanics also did not provide consent. Inasmuch as people at both ends of the wealth distribution are missing EBF matched records, we believe the direction of potential bias is ambiguous. More formally, an econometric solution to this sample issue would require finding an instrumental variable correlated with the probability of having an EBF match but uncorrelated with Social Security wealth. Such a variable does not exist in our sample. 4

15 Using these three files, we compute anticipated retirement wealth for each household. 10 This wealth value is allocated or spread over the household s retirement period using conventional annuity factors. 11 In other words, we take each household s assets and divide them up to reflect the annual payments that a given level of wealth would yield if it were drawn down to zero over the household s remaining life expectancy. The annuity factors used to convert wealth to annual income flows reflect the different life expectancies of men and women at different ages. Thus, annuity factors for older respondents and men are smaller than those for younger respondents and women, since older respondents and men have shorter life expectancies than younger respondents and women. For example, the annuity factor for a nonmarried 56-year-old male in our sample is at age 62 and at age 65. The values for nonmarried women of these same ages are and , respectively. Turning a stock of wealth into an annual income flow makes it easier to interpret and understand exactly what older Americans command by way of retirement resources. 12 Explanatory Factors We develop and use two indicators to capture respondents employment and earnings over their working lives. One is average annual earnings between the respondent s 20 th and 50 th birthday using data on annual pay up to the Social Security earnings ceiling. 13 We call these prime age earnings since they begin at age 20 when most of the respondents would have completed their schooling, and the average ends at age 50 because the level of labor market activity beyond that age may begin to be influenced by early retirement. In general, we anticipate that people with higher prime age earnings will anticipate higher retirement wealth and hence more annual income in retirement. This is a reflection of the way pension formulas work, and also of the way earnings are translated into Social Security benefits. Empirically, of course, it is of interest to estimate the specific way in which higher earnings result in higher retirement income. The second indicator of labor market attachment used here is a count of the number of years of Social Security-covered employment up to the respondent s 50 th birthday. 14 This factor is invaluable in assessing how another year of work is converted into additional retirement income, via pension, Social Security, and saving mechanisms. To examine the impact of occupational segregation, we determined each worker s longest job along with a summary measure indicating the occupation of that job. 15 Following the 10 Dollar figures throughout this paper are in constant 1992 dollars. 11 Levine, Mitchell, and Moore (forthcoming) discuss several ways to model well-being; here we simply focus on levels of retirement income, since these are more readily understood. Burkhauser et al. (1985), Moon (1977), and Hurd (1989) employ similar measures. 12 The regression analysis uses the natural log of all wealth variables because of the skewness of the distribution of these variables. 13 The natural log of prime-age earnings is used because of the skewness of its distribution; the derivation of the prime-age earnings measure is described in Mitchell, Olson, and Steinmeier (forthcoming). 14 Variable creation is described in Mitchell, Olson, and Steinmeier (forthcoming). 15 The Data Appendix describes creation of this occupation variable. 5

16 literature cited above on job segregation, we sought to assess whether having a higher concentration of women in a given occupation also has long-term negative consequences for their retirement income. This was tested in our multivariate model by evaluating whether a higher proportion of women in the worker s occupation on her/his longest job is associated with higher, or lower, retirement income. For purposes of the study, the sex composition of a worker s occupation was assigned on the basis of the occupation in which the respondent was employed the longest. 16 The remaining information we obtained on respondent characteristics was available directly from the HRS. Thus, for instance, survey respondents supplied extensive information on the economic, social, demographic, and other attributes of household members. The survey delved into household members incomes, assets, debt, and health for respondents aged 51 to 61 in 1992 and their spouses (of any age). In addition, we were able to determine each worker s longest job along with a summary measure indicating the occupation of that job. Subsample Issues Starting with the full HRS sample of 12,652 persons interviewed in 1992, we imposed several criteria to generate the sample used for empirical analysis. A few households (95) lacked a financial respondent responsible for providing financial data to the interviewers, and these had to be dropped. We also restricted the respondent sample to include only age-eligibles, namely the 9,714 respondents who were age 51 to 61 in It should be noted that people in this age bracket were interviewed along with their spouses (irrespective of the spouse s age). 17 Next, since we needed anticipated Social Security benefits to conduct our analysis, the sample included only those respondents and spouses who furnished a consent form; for whom the Social Security Administration could locate a matched file; and who were not receiving disability benefits at the time of the 1992 interview. These conditions generated an interim sample of some 5,800 respondents. Finally, we omitted from the analysis persons whose race/ethnic status was not recorded as white, black, or Hispanic, in order to focus attention on the groups of most interest here. The analysis sample then consisted of 5,684 individuals. One important advantage of using the HRS over other possible surveys is that the survey intentionally oversampled blacks and Hispanics. This was done because these groups are relatively rare in the population at large, and their retirement income status is of special interest to those concerned with the well-being of relatively disadvantaged segments of the population. In the descriptive tables below, we therefore draw distinctions by race/ethnicity when comparing retirement incomes by sex. Nevertheless, even with the HRS oversample of minorities, there prove to be too few respondents from these groups to permit the estimation of separate 16 One limitation of the HRS data used here is that we can only identify each worker s occupation at the 2-digit level; this restriction was required in order to obtain the earnings and Social Security benefits data key for retirement well- being computations. We are able, therefore, to identify professional and technical occupations separately from craft occupations but cannot distinguish further within these categories. Future researchers with access to narrower occupational definitions might assess the impact of such occupational aggregation. 17 In any event, a spouse s wealth is included in the analysis irrespective of the spouse s age. 6

17 regressions by sex, marital status, and race/ethnicity (see Appendix Table 1 for sample sizes). Therefore, the multivariate approach taken below includes indicators of respondents race/ethnicity rather than estimating models for each race/ethnicity group separately. Household Resources In order to evaluate access to anticipated retirement income for HRS respondents, we must distinguish between an individual's own resources and those available to his or her household. 18 In the present analysis we assess projected retiree wealth available to a household without seeking to allocate assets within the household to individual members of a married couple. In other words, the models assume that retirement income generated by different assets is equally available to a husband and a wife in a married couple. Hence, the analysis assumes that household resources are consumed jointly as long as both spouses are living. 19 As a result, sex differences in retirement well-being, in theory, can result only from measured differences in the well being of nonmarried men and women. 20 Methodological Approach The analysis that follows first describes HRS respondents wealth levels along with the anticipated annual income flows they represent. Next, we estimate multivariate models of annual retirement income for men and women. 21 In this analysis we focus on the influence of labor market variables in driving retirement income differences by sex: that is, years in the labor market, average prime-age pay, and a measure of how sex-segregated the respondent s longest job was. We also include socioeconomic factors (e.g., education, marital history, and number of children) along with race/ethnic indicators and measures of respondents health and preferences (including indicators of respondents planning horizon and risk aversion). In the case of married respondents, we also include the same measures for the respondent s spouse, since his/her characteristics may also contribute to differences in family resources available in retirement. These regression models are estimated separately by sex and marital status so that results can be compared across groups. Identical model specifications are estimated for the three dependent variables of key interest, namely, income flows from Social Security, pensions, and other financial wealth including housing. Our methodological 18 All values are computed assuming retirement will occur at age 62. See the Data Appendix for more discussion of this point. 19 After one party dies, the surviving spouse is assumed to keep half the pension in a joint-and-survivor arrangement, and Social Security benefits continue for eligible widows (widowers). Housing and other wealth is bequeathable to the surviving spouse in its entirety. 20 Nonmarried persons in the HRS are those who are not currently married; this population includes the never married, the divorced, and the widowed, based on self-reported marital status. Married persons are likewise self reported. Practically speaking, there are slight differences in married men s and women s measured resources in the HRS because the age-eligible women in HRS couples are slightly younger than the women in couples with HRS age-eligible men. 21 The particular empirical models examined are analogous to those employed by Levine, Mitchell, and Moore (forthcoming). There, as here, we recognize that there may be dual causality in the regression models between retirement income, on the one hand, and earnings as well as work years on the other. That is, more work at higher pay would be anticipated to raise retirement income, but conversely, having higher retirement assets might discourage people from working more years or seeking out higher pay. In order to reduce the possibility of endogeneity of these variables, the labor market variables we adopt are strictly retrospective measures. That is, a worker s years of labor market experience are measured up to age 50 but not thereafter; average pay is likewise computed based on the worker s Social Security earnings reported between ages 20 and 50; and the occupation to which the fraction female refers is the respondent s longest job. 7

18 approach is informed by the approaches followed in prior studies that have sought to explain differences in pay for active workers. Having in hand estimates of the effects of each factor on projected retirement income, we next evaluate how much of anticipated retirement income differences by sex could be attributed to differences in the workers characteristics. 22 To conduct this exercise with regard to differences in labor market characteristics, we ask the hypothetical question: how much would the gap in projected retirement income decrease if lifetime labor market characteristics of men and women were identical? In other words, based on the estimated returns to these characteristics, we predict what women s retirement income would be if they had characteristics that were equal in value to those of men, on average. Since men tend to have had stronger labor force attachment during their working years, one would expect the gap between men s and women s projected retirement income to be smaller, or potentially even zero, when it is based upon this prediction. Finally, we estimate and report the dollar reduction in the sex gap in projected retirement income between the predicted and observed levels. A similar analysis can be conducted regarding other factors included in the model, such as socioeconomic factors. This permits us to evaluate how retirement income gaps might narrow as workers traits grow more congruent. Empirical Findings Median retirement wealth expected by HRS respondents appears in Table 1 by sex, marital status, and race/ethnic group. We focus on medians since differences in averages (Appendix Table 2) may be driven by a relatively small number of individuals with very large levels of wealth. A first point to note regarding Table 1 is that married men and women anticipate similar levels of 22 Oaxaca (1973) devised the statistical technique used here to show how differences in outcomes might be allocated to different sources. In the present context, we have adapted this approach to decompose the difference in projected log annual retirement income between older women and men into two parts: the portion due to differences in characteristics that differ by sex, and the portion due to differences in returns to those characteristics between men and women. These analyses are conducted separately by marital status groups. We use women s returns to characteristics to determine how much of the gap in log retirement income would be closed if women's characteristics became like those of men. Specifically, we compute: m f k m m f ( ) ( ) m f k f RY RY = βi X i X i + X i i i i = = β β 1 1 i,where RY represents a particular measure of economic well-being; b represents the vector of regression coefficients estimated using the multivariate model described above; the X values represent a vector of mean characteristics, f and m represent women and men, respectively; and k indexes characteristics. The first expression on the right hand side of this equation is said to represent the "explained" part of the differential in retirement income because it is attributed to the different characteristics of men and women. The second expression is said to represent the "unexplained" part of the differential because it would result in differences in income even if men and women had the same characteristics. Our simulation computes the percentage reduction in the retirement income gap between men and women that would occur if both had identical characteristics. Formally, this involves estimating: % Gap = k f β i i m RY m f ( X X ) i RY f i 100 This expression represents the gap in log retirement income that can be "explained" by differences in characteristics as a percentage of the size of the gap. Below we also compute the dollar contribution to the gap in retirement income by applying the percentages to the dollar gap in projected annual retirement income. 8

19 retirement wealth, a reasonable expectation in view of the model s assumption that retirement wealth is pooled at the household level. 23 Retirement wealth levels are quite substantial for married couples, exceeding half a million dollars if pensions, Social Security, and other financial assets are counted. By contrast, projected retirement wealth for nonmarried people appears much lower, totaling only about one-third as much as for married couples (between $157,000 and $192,000). There are also striking sex differences disfavoring women; that is, nonmarried men have almost 20 percent more retirement wealth than nonmarried women. These overall differences become even sharper when we examine the subcomponents of wealth. For example, married couples Social Security wealth totals about $180,000, a figure not too different from their $160,000 in housing and net financial assets. Their pension wealth amounts to approximately $98,000-$121,000. By contrast, Social Security wealth represents a dominant portion of total wealth for nonmarrieds, housing is less important, and particularly striking pension wealth is very tiny indeed. The median nonmarried woman, for instance, has no pension wealth at all, compared to her nonmarried male counterpart who has $26,000, and her married female counterpart with close to $100,000 in household pension assets. Patterns of retirement wealth by race/ethnic status in Table 1 indicate that the relative disadvantage faced by nonmarried women versus men is most concentrated among the white population. This is because the wealth gap for black and Hispanic nonmarried men versus women is very small or even nonexistent. Thus, nonmarried black women actually have higher levels of total wealth ($106,000) than their nonmarried black male counterparts ($75,000); for Hispanics total wealth is $91,000 for nonmarried men and $72,000 for nonmarried women. Pension wealth is effectively nil for black and Hispanic nonmarried people, and other wealth is similarly minuscule. In sum, differences in retirement wealth between whites and minorities are considerably larger than those between men and women. How these wealth figures translate into annual retirement income flows by sex, marital status, and race/ethnicity is evident from Table The retirement assets shown previously are estimated to produce annual income equivalents for married men and women that are similar to each other, on the order of about $28,000-$29,000 per year. About one-third of the anticipated retirement income is attributable to Social Security benefits totaling about $10,000 per year for the median married household, exceeding the annuitized value of housing and financial wealth that totals about $8,000-$9,000 annually. Median pension income for married couples might seem low, at only about $5,000-$6,000 per year, but it must be recalled that many in the HRS sample anticipate receiving no pension income (zeros are included in Tables 1 and 2). Projected annual retirement income for nonmarried people is expected to be only about one-third the size of married couples income, at $13,000 for nonmarried men and $9,000 for nonmarried women. The relative disadvantage of nonmarried women stems partly from the fact 23 For respondents and spouses who are both age-eligible, wealth levels would be expected to be identical for men and women. Small differences emerge in the HRS dataset, because some age-eligible women respondents have spouses older than 61, while age-eligible men are more likely to have spouses younger than age An analogous table of means appears in Appendix Table 3. 9

20 that they are anticipated to live longer than men on average, which makes the gap in annual retirement income flows larger than the wealth gap. Furthermore, nonmarried people probably require more than half a married couple s income to maintain a comparable living standard. 25 Hence, the finding that nonmarried respondents expect so much less income in retirement than do married couples does not bode well for their prospective retirement income status. Looking further at the components of retirement income flows, it appears that the redistributive nature of Social Security benefits somewhat narrows the retirement gap between nonmarried men and women. However, median expected annual benefit levels are low, on the order of $5,500 for men and $3,600 for women. A problem confronting the median nonmarried woman approaching retirement is that she has no pension wealth at all, whereas nonmarried men have small accumulations, and the median married couple can expect $5,500-$6,200 of pension income annually. Nonmarried men and women have similar levels of net financial and housing wealth, but it is worth pointing out that more nonmarried men have very high levels of other wealth, since the medians are similar but the means are higher for the men. Finally, focusing on the differences in anticipated annual retirement income by race/ethnicity, we find that the median married black couple would anticipate $19,000 annually, and the married Hispanic couple $12,000-$15,000 annually. This compares to much lower levels expected by nonmarried persons, with black and Hispanic women expecting $4,000-$6,000 per year in total income, and black as well as Hispanic men anticipating income in the same range. Table 2 clearly shows minority groups heavy reliance on Social Security, since they can expect relatively little income from sources other than Social Security. These very low income levels do not differ much by sex for minorities. The exception is for married black women, who have pension income almost comparable to that of married white women. Regression Results Moving beyond simple tabulations of the data, we next evaluate how changes in respondent characteristics might improve retirement well-being. Specifically, we are interested in the returns that people anticipate receiving in the form of higher retirement income for a given increase in earnings and work experience; we are also interested in examining how the sex composition of jobs affects anticipated retirement well-being. Results from multivariate statistical analysis controlling for these factors as well as other socioeconomic factors appear in Table Specifically, the table shows how a change in one of the explanatory variables of the model might be expected to influence the average person s annual retirement income. 27 These should be interpreted as marginal relationships, that is, the likely effects of a specified change in the explanatory variable on the outcome of interest. 25 For a discussion of equivalence scales see Levine, Mitchell, and Moore (forthcoming). 26 Appendix Table 4 reports regression results also controlling for respondents health status and risk preferences; the latter coefficient estimates are not reported in detail here. 27 Values are computed at the sample mean unless otherwise noted. Mean values of independent variables appear in Appendix Table 5. 10