Health and Work Capacity of Older Adults: Estimates and Implications for Social Security Policy

The simultaneous growth in longevity and mounting budget deficits in the U.S. have increased interest in raising the age of eligibility for public health and retirement benefits. The consequences of this policy depend on the health of the near elderly, and on the distribution of health by demographic group. We simulate the work capacity and likely disability experience of near elderly individuals (62-64 year-olds) based on the work, disability, and retirement status of slightly younger people. Our estimates, from two distinct data sets, indicate that work capacity is substantial at this age. Because health deteriorates very slowly from ages 60-65, labor force participation could rise by 15-20 percent for all demographic groups, while overall disability rates would change very little. However, less advantaged groups would face challenges in the labor market. The expected earnings of current non-workers without any college education are lower than similar workers, by 15 to 25 percent, indicating the uneven burden of changes in the age of eligibility.

The long-range deficit in Social Security requires reform to bring future benefits and revenues into alignment. Similar concerns surround Medicare, a program for which budget forecasts are even more dire. One common suggestion for reform of both programs is to increase the age of early and normal retirement (Munnell et al. 2004), a suggestion described in detail by the bipartisan National Commission on Fiscal Responsibility and Reform (National Commission on Fiscal Responsibility and Reform 2010). Analysts argue that this is good policy because people are living longer and are healthier at any given age than in the past, thus making employment at older ages more feasible (Munnell et al. 2004;Steuerle, Spiro and Johnson 1999).
The equity implications of raising the age of eligibility may be troubling, however.
People with poor health will find the return to increased work substantially more difficult than people with better health, both because they cannot work as much, and because they earn less when they do work. Since health varies systematically by socioeconomic group, these differences could be important. In this paper, we examine the potential impact of policies that would raise the age of eligibility for early and normal retirement benefits and for Medicare eligibility on work capacity and disability of different groups. Our basic question is: If eligibility ages for public programs increase, would more people go to work, or would more people instead be out of work and classified as disabled? If the latter, how does this vary by demographic group?
To answer this question, we simulate the labor force and disability participation of people aged 62-64, using their own health status and the relationship between health and work status of people just younger than them. We assume that the relationship between health and an individual's capacity to work is the same among people aged 62 to 64 (the ages at which individuals can currently claim early entitlement for Social Security benefits) as it is for people aged 57 to 61. Thus, by estimating the impact of health on labor force participation among people aged 57 to 61, we can forecast work capacity for people aged 62 to 64. If private pension and health benefits, as well as employment opportunities, became the same for ages just above and below 62, this would predict actual labor market impacts of these changes. However we interpret our results as a measure of capacity rather than actual behavioral changes, so our analysis does not rely on this assumption. We do the same exercise for self-reported disability.
Our primary empirical finding is that health deteriorates very slowly in the 60s. In particular, the health of people aged 62 to 64 is only marginally worse than the health of people aged 57 to 61, and virtually indistinguishable from the health of individuals aged 59 to 61. This is true overall and for most demographic groups. It is also true through age 69. As a result, when we simulate the work capacity of people aged 62 and older and assume that there were similar job opportunities and no differential financial consequences to working, we find that labor force participation in these ages would increase by about 15 or more percentage points for most demographic groups, while self-reported disability rates would increase relatively little.
However, among less educated adults aged 62 to 64, the expected earnings of current non-workers predicted to work if the early eligibility age (EEA) were raised to 65 are 15 to 25 percent lower than the earnings of current workers aged 62 to 64. Finally, disability rates would rise more for the less educated than among older adults with some college education. For some groups, the increase in disability could be substantial, as much as 6 percentage points (a 60 percent increase). And since life expectancy is shorter for these groups, the percentage reduction in healthy, non-working years at the end of life is higher -a reduction of 9 percent among less educated men and women compared with 6.5 percent for the better educated. Thus, the equity implications of such a change would be unfavorable.
Electronic copy available at: https://ssrn.com/abstract=2577858 The paper is structured as follows. Section 1 discusses past research on health and work capacity. Section 2 describes the data we analyze. Section 3 presents descriptive calculations of health for different race, gender, and education groups. Section 4 estimates models of work capacity and simulates labor force status in the wake of alternative Social Security and Medicare changes in age eligibility. Section 5 discusses the implications of these results.

Section 1. Background
There is a large literature on determinants of retirement decisions and the consequences of retirement policy for that decision. Much of this work has been summarized in Gustman and Steinmeier (2005) and Blau and Goodstein (2010). We focus our summary on the potential impact of proposed policies to raise the early and normal retirement age for Social Security benefits and age-eligibility for Medicare benefits. Although there is controversy regarding the exact role Social Security benefits play in retirement decisions, a few points of consensus emerge. First, most of the literature finds that health shocks, although important determinants of work, do not explain retirement trends over time, partly because few individuals have truly poor health (Burkhauser, Couch and Phillips 1996;McGeary 2009;Mitchell and Phillips 2000;Smith 1999). Second, Social Security policies cannot explain large declines in labor force participation through 1990, but changes in rules that make work relatively more favorable may have contributed to rising labor force participation in recent years (Blau and Goodstein 2010;Gustman and Steinmeier 1986;Gustman and Steinmeier 2005;Lumsdaine, James and David 1996;Lumsdaine, Stock and Wise 1994;Rust and Phelan 1997). Recent evidence on proposed (rather than actual) policy change comes from simulations of the impact of several policy changes among singles and married couples without private pension benefits. Van der Klaauw Electronic copy available at: https://ssrn.com/abstract=2577858 and Wolpin predict large increases in employment rates among married males aged 62-69 (10 percentage points for raising the EEA to 70 and 15 points for the elimination of all benefits before 70), but only modest changes among married females and single males or females. Gustman and Steinmeier (2005) simulate that if the EEA were raised to 64, the share of 62-63 year olds who would be retired would fall by 5 percentage points. Finally, health insurance benefits provided by employers also influence labor force participation, as does Medicare eligibility, but modestly (Blau and Gilleskie 2008;Gustman and Steinmeier 1994;Lumsdaine, James and David 1996).
Because retirement rates spike at ages 62 and, more moderately at 65, much of the retirement literature attempts to model these peaks at ages 62 and 65 as a function of pension and/or health benefits. Since delaying retirement from 62 to 65 is approximately actuarially fair, the literature cannot attribute the spike in retirement at age 62 to the incentive effects of Social Security. Thus, researchers need other explanations to understand why people retire when they do (Blau and Goodstein 2010;Borsch-Supan 1999;Gustman and Steinmeier 1986;Gustman and Steinmeier 2005;Rust and Phelan 1997). Gustman and Steinmeier (1986; posit that discounting is the issue. Their key insight is that individuals with high discount rates (i.e. they discount the future heavily) will retire as soon as early retirement benefits are permitted at age 62. For individuals with low discount rates, retirement is optimal when the actuarial adjustment to delayed SSA benefits become less favorable, at age 65. An alternative model suggests that individuals would like to retire before 62, but due to market imperfections that leave them unable to borrow against future income, they "wait" until age 62 and retire at rapid rates (Rust and Phelan 1997). Rust and Phelan argue similarly that retirement spikes at age 65 reflect a lack of access to health insurance Electronic copy available at: https://ssrn.com/abstract=2577858 in private markets that is relieved with age eligibility for Medicare.
Most relevant to our work is the literature focusing on health status and/or health insurance and retirement. A study of older male workers in the Health and Retirement Study (HRS) describes characteristics of Social Security beneficiaries claiming early retirement (Burkhauser, Couch and Phillips 1996). In the HRS, early retirees were slightly more likely to be in poor health than non-early retirees, but only 3% of the early retirees reported poor health.
Reports of poor health were much more predictive of adverse economic outcomes (such as low employment rates, low levels of income and wealth) than early retirement. In closely related work, McClellan (1998) demonstrates that functional status, rather than a health condition, is an important determinant of retirement decisions. This finding is not surprising given the heterogeneity in outcomes among individuals with a particular health condition. Mitchell and Phillips (2000) also explore potential changes in age eligibility for Social Security, carefully assessing streams of income, the availability of disability benefits, pension benefits, and health insurance individuals can expect over their lifetimes. On the basis of behavioral responses to the possible change in financial incentives, they estimate that retirement would fall by much more than disability would rise. McGarry (2004) uses the Health and Retirement Study to show that one's perception of health is an important determinant of retirement.
The literature described above aims to understand what people might do in response to policy changes, and in each case authors are forced to make strong assumptions regarding an individual's ability to borrow against future Social Security Benefits, or to restrict empirical populations to narrow groups such as working single males from a single firm or people facing a similar ability to access pension and health benefits, to overcome some of the complexity introduced by joint decisions and by variation in private retiree benefits. However, the literature in this area fails to answer an important question regarding the potential burden of reforms that raise the age of eligibility for retirement and health benefits: what are the young retired able to do?
Our work focuses on this question and thus complements existing studies of health and retirement. Further, our analysis does not force us to restrict our study population in order to avoid complicated private pension rules or make assumptions about liquidity constraints.
Because we model a rise in the Early Eligibility Age to 65, our simulation does not rely on the resolution of the debate regarding whether Medicare benefits induce the spike in retirement at age 65, or whether the financial implications of postponing retirement after age 65 can explain these spikes. Rather, we focus on capacity, not simulating expected behavior. Finally, we focus more attention on the distributional impact of such policy changes by describing how these rules may impact sex, race, and education groups differentially.

Data to Estimate Health and Work Capacity
To match health limitations and conditions with labor force status, we use data from two Both surveys also include a rich set of demographics (exact age, gender, race, ethnicity, education, geographic information) and information on self-reported health and functional limitations. Still, there are some differences that make each one worthwhile.

MEPS Data
Our Data describing whether individuals are "in the labor force," "retired," or "disabled" come from two questions. First, individuals were asked, "Do you currently have a job for pay or own a business?" If the answer is no, respondents were asked, "What is the main reason you did not work since (START DATE)?" Possible responses include: could not find work; retired; unable to work because ill/disabled; on temporary layoff; maternity/paternity leave; going to school; take care of home or family; wanted some time off; or waiting to start new job. We classify individuals as disabled if they self-report "unable to work because ill/disabled" and term them retired if they self-report being "retired, " if they never worked in their life, say they are "taking care of home or family," or give an unclassified reason for not being employed. 3 All other individuals are coded as "in the labor force."

HRS Data
We complement the MEPS with data from the HRS. The HRS is representative of the entire community-dwelling population aged 50 and older. We focus on adults aged 55 to 74, those most likely to work, including 14,679 females and 10,985 males. Our models of work capacity (from which we later predict labor force status) focus on 5,449 males and 7,403 females aged 59-61, whom we observe in 1994-2008.
The HRS has a larger sample of older adults than the MEPS, with richer information on economic and household characteristics and the characteristics of one's spouse. Thus, we can simulate work capacity for 62 to 64 year olds based on the health and work status of 59 to 61 year olds, arguably a better comparison group than all adults aged 57 to 61. The HRS asks about a common measure of depression symptoms, the 8 item Community Epidemiological Scale of Depression,4 To measure labor force status in the HRS, we use the question: "Are you working now, temporarily laid off, unemployed and looking for work, disabled and unable to work, retired, a homemaker, or what?" Individuals who were working, laid off, unemployed, or "partially retired" are coded "in the labor force". Individuals who described themselves as "retired" are coded as "retired". Individuals self-reporting that they were disabled are included as "disabled", in addition to those who reported applying for or receiving Social Security Disability Insurance.
Note that defining disability in this way is, if anything, likely yields an upper bound on rates of disability by age since some individuals may rationalize their decisions to exit the labor force by reporting they are disabled.
and unlike the MEPS, the HRS asks about former smoking status as well as current status. Thus, we can include risk factors as a measure of health. The disadvantage of the HRS is that self-reported health is only characterized on a five-point scale: excellent, very good, good, fair, and poor.
Data on labor force participation in the two surveys are presented in the first rows of Demographic averages for each survey are presented in Appendix Table A1. Both surveys are representative of the US population when weighted.

The Self-Reported Health Status of Older Adults
We begin our empirical analysis by describing the age profile of self-reported health status. The remaining rows of As people enter their 70s, health starts to decline. The thermometer rating of health in the MEPS falls by 6 points from the late 60s to the 70s, and the share of the population in excellent or very good health in the HRS falls more rapidly after age 70 as well.
For our analysis, we care about health both overall and for different demographic groups.
Accordingly, we move to a demographic decomposition. We describe groups separately by education (those attending college versus those with a high school degree or less education), which is determined early in life, in addition to gender and race/ethnicity (white non-Hispanic, black non-Hispanic, and Hispanic). Thus, the membership of the groups is determined well before the minimum age for our analyses. Figures 1 and 2 show the averages of self-rated health over ages 50 to 85. Figure 1 shows health for men, and figure 2 shows health for women. Within each cell, we plot the average thermometer score at each age from the MEPS and the share of people in excellent, very good, or good health from the HRS.
There are clear differences in the level of health by gender, education and race/ethnicity.
Within any gender and education group, average health for whites is higher than for blacks and for Hispanics. Similarly, the better educated report better health than the less educated, holding constant gender and race/ethnicity. The magnitude of the differences varies by group, with the largest gaps showing for white college attendees compared with blacks or Hispanics with less education. Scores for men and women reveal modestly better health for women, despite the significant longevity advantage for women.
Looking at the age profiles, we see again that self-reported health changes very little over ages 50-69. Among white, college educated women and men, for example, the health scale is virtually constant at 80 throughout the age range. Although the average health scale for less educated white women and men is slightly lower, between 70 and 74, the trend is also flat over this age range. The data for blacks and Hispanics are noisier, but show similar patterns.

Estimating Work Capacity
In this section, we consider additional dimensions of health and examine how the multiple dimensions of health taken together are related to labor force participation. We then use Before we present multinomial logit models of labor force status, we describe how labor force participation and self-reported disability vary by self-reported health status in figures 3 and 4. The relationship between health status and disability rates is nonlinear, but generally speaking, a 10 point decline in the 100 point health scale accompanies a rise in disability of 2 to 5 percentage points near the top of the health scale, and closer to a 10 percentage point increase in disability as one moves below 60 on the health scale. Among 59-61 year olds in the HRS as well, there is a non-linear relationship between self-reported health and labor force participation.
Participation is low for people with fair or poor health, and flattens out above that. The same is true, in the opposite direction, for reported disability. A notable finding in both data sets is that labor force participation and disability change very little over the top half of the health scales.
This suggests that work capacity remains very high for the vast majority of older adults, given relatively high ratings of health in the MEPS and the HRS samples.
Appendix table A2 reports relative risk ratios from multinomial logit models of labor force status in the MEPS. The 100-point health scale has a strong association with self-reported disability, with a risk ratio of .51 for males and .56 for females for each 10 points on the linear health scale. Given the higher order terms, the relationship between the thermometer scale of health and labor force status is non-linear. We plot the relationship between labor force participation, disability, and health scores in Figure 3.
Comparable multinomial results for the HRS population aged 59-61 are reported in Appendix table A3. Several indicators for health limitations, impairments, or conditions are strongly associated with self-reported disability. Being in poor self-rated health has a very large impact on reported disability, as do functional limitations and a variety of specific conditions. Figure 4 shows the relationship between self-reported health and retirement and disability for the Electronic copy available at: https://ssrn.com/abstract=2577858 population aged 59-61.
The MEPS and HRS data both show that health variables are much less predictive for retirement than they are for disability. Functional limitations, ADLs or IADLs lead to a greater probability of retirement in the MEPS, but the effects are generally smaller than for disability.
None of the non-health variables are significantly associated with retirement for males. For females, college education is associated with lower retirement, as is being unmarried. Hispanics are more likely to be retired than non-Hispanics. Second, the pattern for disability is quite variable by group. In the MEPS, the fraction of less educated white males who are disabled increases by 5 percentage points in the absence of early eligibility for Social Security benefits -a 60 percent relative increase. Less educated black males would also witness a 7 percentage point rise in disability rates in the MEPS.
The HRS results show little or no rise in disability rates among most groups, with the exception of black males and females in lower education groups, among whom disability rises by 3 to 5 percentage points. For several groups in the HRS, the models predict, if anything, lower disability among 62 to 64 year olds. This is predominantly due to low rates of reported ADLs/IADLs at these ages. Overall, despite some large relative increases in self-reported disability, the share of individuals likely to report they are disabled would remain very low, less than 10 percent among college attendees, and 20 percent or lower among the remaining individuals.

Earnings among individuals simulated to be in the labor force
An important question for individuals who would delay retirement is what they would likely earn were they to work. To examine this question, we use the detailed information on earnings in the HRS to simulate the earnings of non-working individuals aged 62 to 64 based on the earnings of workers 59-61. Our model is similar to equation (1), except that the dependent variable is the earnings of current workers. We estimated these models using generalized linear models with a log link function to account for the log normal distribution of earnings while permitting us to use model parameters to predict earnings, rather than the natural log of earnings.
We then predicted earnings for each individual aged 62-64 in the older population. 7 To compare predicted earnings of non-workers likely to be affected by any increase in the EEA to actual earnings of currently employed workers age 62-64, we take a weighted average of the predicted earnings for non-workers, where the weight for each non-worker is his/her predicted probability of entering the labor force from our earlier simulation. We focus these results on less educated workers in each sex, race, and ethnicity group, since these potential workers tend to have fewer labor market prospects. All earnings are expressed in constant 2002 dollars based on the Consumer Price Index. Effectively, we ask the question, "Would workers who are induced to stay in the labor force because of an increase in the EEA likely earn much less than individuals who currently work at these ages?" On the basis of health and demographic information, the answer, shown in Table 2 7 To compare simulated earnings with actual earnings of people aged 62-64, we need to adjust for average earnings differences as individuals age. We make such an adjustment using the ratio of average earnings of working 62-64 years olds to the predicted earnings based on the model from 59-61 year olds. For estimates reflecting predicted earnings and the difference between predicted earnings of non-workers and actual earnings of workers, we report bootstrapped standard errors. appears to be "Yes." The table shows earnings comparisons for the less educated groups, for whom these differences would be the greatest concern. 8 For women, the situation is similar. Average earnings are lower for women than for men, but the difference in predicted earnings of those not in the labor force is roughly the same range: 15-20 percent. Thus, the earnings consequences of this change would be significant. The lower earnings for non-participants may offer some insight into why these workers have chosen early retirement instead of labor market participation.
Average earnings for current 62-64 year-old low education workers (row 1) range from $17,000 to $33,000, depending on race and gender. The predicted earnings of those who may be induced to work if Social Security EEA was raised to age 65 are generally lower (the second row of the table). As the third row shows, these predicted earnings are 15 to 25 percent lower than the actual earnings of those currently in the labor force.

Expected retirement years and changes in lifetime benefits
These findings have implications for the progressivity of the retirement system. Since less educated people and racial/ethnic minorities live less long than their counterparts, reductions in retirement years for these groups are a greater reduction in lifetime welfare. To show these implications quantitatively, we consider how raising the Social Security and Medicare eligibility ages would affect the expected amount of time spent in retirement. To do this, we calculate the expected retirement years (ERY) for each demographic group currently, and compare this to the ERY using the predicted labor market status in our simulations.
We form ERY as the number of years a person alive at a given age can expect to spend retired. Defining survival probabilities as π g,a|x for group g at age a, conditional on being alive at age x, and the fraction of the group that is retired at each age as r g,a , ERY is defined as: (2) ERY g,x = .
If monthly benefits remain approximately the same for each group, changes in lifetime benefits amounts will be closely related to changes in the length of retirement.
To obtain survival probabilities, we use the nearly universal sample of death certificates compiled in the Mortality, Multiple-Cause-of-Death (MCD) file for 2000 (see Meara, Richards and Cutler 2008). This gives us death totals by age, gender, race/ethnicity, and education. We match these with population estimates from the Public-Use Micro Sample of the 2000 Census to calculate mortality rates for each group. Although there are known inconsistencies with the education levels reported on death certificates relative to self-reports while alive (Sorlie and Johnson, 1996), these data allow the use of two broad educational categories with reasonable accuracy: high school or less vs. any college. 9 Table 3 shows the loss of retirement years from this policy. For brevity, we display only results from the MEPS data; the patterns are similar within the HRS. Because the expected increase in labor force participation is relatively similar across groups, the expected reduction in retirement years is similar as well. Among whites, the simulated reduction in expected retirement years is 0.6 for men with a high school degree or less, and 0.5 for men with any college education. 10 As a percentage of current expected years in retirement, however, the changes are very different. Because the better educated live so much longer than the less educated, the proportional loss in retirement years is greater for the less educated (5 percent) than for the better 9 For further details on these data sources and education measurement, see Meara, Richards, and Cutler (2008). 10 To understand why ERY could decrease by only 0.5 when the EEA increases by three years, recall that in our simulations the share retired at ages 62-64 drops by about 0.2. Thus 0.2 times three years is approximately the change in ERY.
Electronic copy available at: https://ssrn.com/abstract=2577858 educated (3 percent). A similar differential is apparent for women, for the same reason.
If we only count years in retirement from age 65, when Social Security benefits would be received under the new policy, the proportional loss is 9 percent for less-educated white males and females and 6.5 percent for the high-education groups. 11 All else equal, this change would reduce the progressivity of the Social Security system.

Conclusion and Discussion
In this paper, we simulated the work capacity of individuals in age groups targeted by policies that raise the age of eligibility for Social Security. Specifically, we examined a circumstance equivalent to increasing the age of early eligibility for Social Security to 65 (instead of 62) and no change in the age of full eligibility. Our primary output is work capacitywhat the people in this age group would do differently from their current status using only the fact that their health is different from those of younger aged individuals.
We find that until age 70, health appears to decline very slowly, and thus work capacity is large. Our results suggest that, based upon the health of today's young retirees, 15 to 20 percent more individuals could work than currently do. However, rates of new disability may increase.
We estimate that about 2 to 6 percent of the population would report disability in this circumstance.
To put our findings in context, consider the drop in labor force participation between 1970 and 1994 (Burkhauser, Couch and Phillips, 1996). Between 1970 and 1994, the percentage of 63-year old men in the labor force fell by 24 percentage points. At age 68, the drop was 17 11 Our calculation assumes that current retirees aged 62-64 all receive Social Security retirement benefits, which is supported by evidence that only a small number of eligible retirees delay claiming benefits. Coile et al. (2002) show that only 10% of men who retire before age 62 delay claiming benefits by more than one year after their 62 nd birthday, and fewer than 5% of those who retire after age 62 delay claiming by more than one year. percentage points. This is roughly the same as our estimate of additional work capacity at those ages assuming that only health status changed across ages. Thus, our analysis suggests that work capacity is just as high now as actual labor force participation was several decades ago.
On the downside, however, there are significant disparities in these findings across groups, and the earnings capacity of less educated older adults who may be induced to work with a raise in the EEA is significantly lower than for current workers. The simulated rise in disability is large for blacks and people who never attended college. Indeed, as much as 6 percent of this group does not have the capacity to work at ages 62-64 due to their poor health. For other groups, the simulated rise in disability is smaller at 3 percentage points, but this is still a considerable amount relative to their baseline rates. In addition, raising the ages of eligibility for partial Social Security benefits disproportionately reduces the expected retirement years of the less educated and racial minorities. The potential rise in disability suggests the need for alternative sources of income among older workers, especially older workers facing higher risk of disability or barriers to work.
Despite these concerns, our results offer reasons to be cautiously optimistic about the ability for many older Americans to continue working beyond current retirement ages. We find that the good health enjoyed by individuals aging throughout their 60s implies a tremendous potential for labor force participation. Even among those groups that would observe the highest disability rates under a policy that delayed early retirement until age 65, the majority of individuals aged 62 to 64 could work, based on their reported health status. If workers translate this good health, and a new norm of providing public retirement benefits at older ages into longer working lives, such a shift has the potential to both reduce the costs of the Social Security program and to increase revenue raised through payroll and income taxes that would not be realized in the absence of such a policy change.

Figure 1: Average Health Scale in the MEPS and HRS, Males
The chart shows the average thermometer score by age and demographic group in the MEPS (age 55-84), and the share of people reporting good, very good, or excellent health in the HRS (age 55-74). All data are weighted to national totals.

Figure 2: Average Health Scale in the MEPS and HRS, Females
The chart shows the average thermometer score by age and demographic group in the MEPS (age 55-84), and the share of people reporting good, very good, or excellent health in the HRS (age 55-74). All data are weighted to national totals.

Figure 3: Probability of Labor Force Participation and Disability by Health Status in the 2000-2003 MEPs, ages 57-61
The chart shows rates of disability and labor force participation by health status.

Figure 4: Probability of Labor Force Participation and Disability by Self-Reported Health Status in the HRS, ages 59-61
The chart shows rates of disability and labor force participation by health status.
Electronic copy available at: https://ssrn.com/abstract=2577858       , 2000-2003and Health and Retirement Study, 1994 Data are weighted to national totals. * Asthma in the MEPS; lung disease in the HRS. Standard errors of predicted earnings for non-workers and the difference between earnings of workers and non-workers were obtained using bootstrapping techniques, based on parameter estimates from generalized linear models of earnings for workers aged 62-64. The predicted wage for current non-workers is weighted by the predicted probability of being in the labor force (based on the work capacity models). The difference in median earnings observed -predicted was 1,335 (4.5%), 7,119 (26%), and 1,658 (7.4%) for male whites, blacks, and Hispanics, respectively. For women the same figures are: 5,227 (28%), 4,063 (30%) and 4,211 (31%).  Risk ratios reflect multinomial logit models of reporting disability or retirement, relative to being in the labor force. Models include dummies for region. p-values: *p<.1; ** p<.05; *** p<.01 7,403 Risk ratios reflect multinomial logit models of reporting disability or retirement, relative to being in the labor force. Models include dummies for census region and survey wave. p-values: *p<.1; ** p<.05; *** p<.01.