Differences in the returns to education for males by disability status and age of disability onset.

Hollenbeck, Kevin ; Kimmel, Jean

1. Introduction

In 1990, the U.S. Congress passed the Americans with Disabilities Act (ADA), legislation perceived as a hallmark in the quest by individuals with disabilities for equal access to labor market opportunities. Numerous studies have been published in recent years assessing the outcomes to date of this legislation, and much debate continues regarding its success or failure in improving labor market outcomes. However, the ADA is only one means of addressing labor market difficulties faced by individuals with disabilities. Because its primary intent is to expand the number of employment opportunities via employer accommodation, the ADA may be characterized as a demand-side intervention. (1) Numerous education, training, and rehabilitation programs operate on the supply side of the labor market. One example of increased attention on supply-side programs may be seen in the Workforce Investment Act of 1998. This sort of disability management model program has the goal of enhancing the skills and knowledge of workers with disabilities to increase their productivity and employability. (2) The returns to education for the population with disabilities is indicative of the potential efficacy of such supply-side interventions. If the rates of return to formal education are large, then it might be argued that there is an underinvestment in such interventions. If the rates are low, then expanding this human capital intervention as a way to improve employment or earnings may not be indicated. Formal educational attainment and its differential effects by disability status and age of onset are the focus of this paper.

For an individual with a work-limiting condition, the age of onset of this condition plays a role in education choices and employment. A priori, one might expect that among the population with disabilities, those with early onset of the disability are the most able to adjust human capital investments and obtain job matches that minimize the potentially negative labor market impacts of the disabling condition (see Loprest and Maag 2003). However, the early onset of the disability might hinder human capital investment or affect its quality.

Many studies and data sets have been used to document the disparities in employment-related outcomes between people with and without disabilities, and many nice summaries exist in the literature. (3) Depending on the data used and the definition of disability, (4) approximately 8-15% of individuals of working age have disabilities. In general, individuals with a disability report lower employment rates, lower levels of educational attainment, and lower wages. Additionally, many researchers have reported a post-ADA decline in employment rates, although the cause of this trend is not clear and might be related to a worsening in recent years of disability severity amongst the population reporting disabilities, or the trend might be confounded by changes in the labor force participation rates among the disabled. (5)

The disability literature is lacking in detailed information regarding the role that age of onset of disability may play in education and employment outcomes. Using data from the Health and Retirement Survey, Burkhauser, Butler, and Weathers (2002) find that the average age of disability onset amongst those with working-age onset is 47.9 years for men and 47.4 years for women. (6) Burkhauser and Daly (1996) report that approximately 70% of individuals with disabilities in their study (using the Panel Survey of Income Dynamics) became disabled during their working lives. The authors show that employment rates decline after disability declaration, but incomes fall more, with male labor earnings falling approximately 25-30% and female labor earnings falling nearly 50% (p. 72). Additional evidence regarding age of disability onset can be gleaned from published reports from the 2000 Census (see, for example, Waldrop and Stern 2003), which show clearly, as expected, that disability rates rise with age. While only 5.8% of the population aged 5 to 15 years report any disability, that percentage increases to 18.5% for the working-age population and to 41.9% for those aged 65 years or older. Related to factors associated with age of disability onset is variation in disability status (i.e., disability status is not always permanent and a large percentage of those reporting disabilities in any given year report themselves as newly disabled) (Kruse and Schur 2003b, p. 290).

Loprest and Maag (2003) provide a detailed portrait of educational attainment and age of onset of disability. Using data from the National Health Interview Survey, they indicate that only about 16.3% of their sample reporting disabilities reported onset prior to age 22 years. Fully one third of this early-onset group failed to achieve a high school education, while only 7.8% reported college completion. While the late-onset group (those who experience disability onset after age 22 years) also reported surprisingly similar high rates of failure to complete high school, they reported twice the rate of college completion. Loprest and Maag find that the two groups have comparable employment rates but that the late-onset group earns higher average hourly wages. (7)

2. Wage Regression Evidence for Workers with and without Disabilities

Looking broadly at the literature that provides estimates of wage equations for people with disabilities, we classify the relevant literature into three groups: papers that focus on measurement and endogeneity concerns, papers that focus on particular forms of disability, and papers that focus on discrimination. The first component of the disability literature recognizes significant heterogeneity by type of disability and how disability is measured. (8) Kruse (1997) uses the 1992 and 1993 Panels of the Survey of Income and Program Participation, otherwise known as the SIPP, to focus on those with disabilities using a broad definition of disability. He finds that while individuals with severe disabilities have much lower employment rates than the non-disabled, rates for those with moderate disabilities are not so different. He suggests that unobserved (or unmeasured) characteristics such as ability, job access, or discrimination partially explain the lower employment rates, although he also allows that diminished returns to human capital (the focus of this paper) might also play a role.

Kruse and Schur (2003a) focus further on different measures of disability and the role that these different measures play in explaining employment-related disparities and ADA effects. They explain that, in general, the negative association between employment and disability is "due in part to high reservation wages associated with many disabilities, particularly resulting from disability income and extra demands on time and energy" (p. 3). (9)

Also relating to measurement issues is the literature that focuses on the endogeneity of self-reported disability (often tied with the ability to work). Stern (1989) finds that the endogeneity bias may be minimal, however. He finds that two self-reported health status measures (a work-limiting condition measure and an overall health status measure) serve as good proxies for estimating the effect of disability on labor force participation.

Moving beyond broader measurement issues is the literature that focuses on employment outcomes experienced by individuals with particular forms of disability. As examples, Ettner, Frank, and Kessler (1997) focus on psychiatric disorders, and Mitchell and Burkhauser (1990) focus on arthritis. Both find important wage and employment differences for those suffering from these particular forms of disability and offer as explanations differences in productivity and differential outcomes by sex. (10)

The final category of disability research that we highlight here includes studies that attempt to identify the presence of employer discrimination or prejudice. It is this strand of the disability literature that is most relevant to our paper because of its estimates of returns to education by disability status as a by-product of the focus on the role that disability plays in wage determination. This literature presents mixed results about the returns to education for men with disabilities versus those without. Johnson and Lambrinos (1985) estimate the returns to a year of education for non-handicapped men to be 0.054, whereas they find men with handicaps generate a 0.040 return. Baldwin and Johnson (1994) estimate returns to education of 0.059 for men without disabilities, 0.055 for men with disabilities (i.e., having impairments that are subject to little or no prejudice--LP), and 0.023 for men with handicaps (i.e., impairments subject to greater prejudice--MP). Baldwin and Johnson (2000) find returns to education for the non-disabled, the LP impairments, and the MP impairments to equal 0.045, 0.044, and 0.053, finding a difference only for the latter, disabled group. These three papers generally find that workers with disabilities have somewhat lower estimated wage returns to education. Note, however, that a study using more recent data (Hotchkiss 2003, pp. 65-6), finds that those with disabilities experience higher wage returns to education. (11) Additionally, DeLeire (2001), using 1993 data, finds that workers with disabilities experience a 9.3% return to education. (12) These two studies indicate that education may serve as a wage buffer for workers with disabilities.

3. Our Contribution to This Literature

Our research contributes to this empirical debate in three important ways. First and most importantly, we address disability heterogeneity as influenced by age of disability onset. We provide estimates of the returns to education for those with early-onset disability separate from those with late disability onset and thus are able to address two issues: (i) potential endogeneity of educational investment arising as a result of early disability onset and (ii) possible differential abilities to achieve job matches with accommodating employment positions. According to Loprest and Maag (2003, p. 26), workers with late-onset disability may benefit more from return-to-work and work incentive approaches, while those with early-onset disability may be most helped by assistance in formal education completion.

Second, we exploit the availability of self-reported productivity-relevant functional limitations to control for heterogeneity that exists for all workers, not just the "officially" disabled. Analyses of the labor market outcomes for individuals with poor health or disability must acknowledge the heterogeneity of this population with respect to health status. Considerable variation exists in type of impairment, severity of condition, age at onset, prognosis for recovery, and access to health care resources.

Finally, we improve on the econometrics used in previous estimates of returns to education for the disabled. We use an endogenous switching model with a trinomial switching mechanism, thereby producing more refined estimates of the differential returns to education by disability status and age of onset. The next section of this paper summarizes the empirical model we implement, including the "full," or most appropriate, econometric specification. Then we describe our data and discuss the results from our estimation, and, finally, we conclude with a summary of our findings.

4. Empirical Specification of the Wage Equation in the Presence of Multiple Types of Selection

The empirical work is based on the standard human capital wage equation, as developed by Mincer (1974). We develop our estimating equation by modifying the Mincer equation step by step. The wage equation, written out below, relates the calculated wage (constructed by dividing earnings on the primary job by hours on the primary job) to years of formal education, demographic control variables, controls for the three broad categories of activity limitations, job characteristic variables, and regional controls. The wage is included in its natural logarithmic form so that the resulting estimated coefficients are more easily interpretable as percentage returns to education.

Equations 1-3 present a linear estimable specification of this basic model estimated separately by disability status:

ln [W.sub.Ni] = [[alpha].sub.1] + [[beta].sub.11] [X.sub.Ni] + [[beta].sub.11] [S.sub.Ni] + [[epsilon].sub.1i] (1)

ln [W.sub.Ei] = [[alpha].sub.2] + [[beta].sub.21] [X.sub.Ei] + [[beta].sub.22] [S.sub.Ei] + [[epsilon].sub.2i] (2)

ln [W.sub.Li] = [[alpha].sub.3] + [[beta].sub.31] [X.sub.Li] + [[beta].sub.32] [S.sub.Li] + [[epsilon].sub.3i], (3)

where [W.sub.i] = hourly wage of individual i; [X.sub.i] = vector of characteristics describing individual i that are thought to be related to wages; [N.sub.i] indicates that individual i has no disability; [E.sub.i] indicates that individual i has early-onset disability; [L.sub.i] indicates that individual i has a later-onset disability; [S.sub.i] = years of education that have been completed by individual i, or three education spline measures; (13) and [[epsilon].sub.1i], [[epsilon].sub.2i], and [[epsilon].sub.3i] = standard error terms.

The parameters of most interest are [[beta].sub.12], [[beta].sub.22], and [[beta].sub.32]. Comparing these three estimated returns to education will tell us whether or not returns to education vary by disability status. Unfortunately, there are several econometric complexities that influence these coefficients arising from unobserved heterogeneity by disability status. Equations 1-3 assume orthogonality between the regressors and the error term, but unobserved heterogeneity bias may confound the differential estimated returns. If individuals with disabilities face greater barriers to access to education or employment, for example, then those individuals with disabilities who are in the labor force, employed, or have high educational attainment may be the most motivated and thus have higher wages.

A relatively high opportunity cost of labor force participation because of attractive non-earned income alternatives may also bias the return to education for individuals with disabilities. The presence of non-earned income will dampen labor force participation and, consequently, employment. Individuals with disabilities have transfer income opportunities from Supplemental Security Income (SSI) and, arguably, worker's compensation. SSI has experienced accelerated growth in beneficiaries over the last couple of decades, indicating that take-up rates are growing. The presence of such an alternative increases the likelihood that eligible individuals who are in the labor force and working have relatively lucrative wage opportunities because those with less lucrative wage opportunities choose to not work and receive public benefits. (14)

Finally, our estimation relies on self-reported health status and disability measures. It is likely that the age of onset of disability plays a role in educational attainment creating heterogeneity even within the disabled population. For example, those with childhood disability onset may be the most able to create a fully accommodating lifestyle and achieve more productive job matches. However, for these individuals, educational choices (including quality of that education) may be constrained by disability status. Thus, in addition to a focus on differential returns to education by disability status, we also focus on age of disability onset in our regressions using the individual's self-reported age of onset.

The declaration of such disability indicators may be a source of measurement error and heterogeneity. If an individual's health condition is not observable by the data collector, then the respondent has a choice to make about whether to report the condition. (15) Generally speaking, there are no incentives or disincentives in a national survey like SIPP to influence reporting behavior systematically, but the individual may still misreport.

To isolate unbiased education returns requires some attention to these sources of unobserved heterogeneity. We address potential employment heterogeneity through the standard technique of estimating a preliminary employment equation in order to construct an Inverse Mills Ratio (IMR) term that will serve as a statistical correction when estimating wage equations only for those individuals with observed wages (i.e., for those currently working) (Heckman 1974). This correction is particularly important for our analyses because of the possible differential importance of education in employment selection for those with and without disabilities (see Neal 2002). Additionally, we estimate regressions that distinguish between early versus late onset of disability, another potential source of heterogeneity. (16)

Our empirical work relies on the following specification of the employment probit model, in which the continuous latent variable [p.sup.*.sub.i] (reflecting preferences for paid work) is expressed as the observed discrete employment outcome:

[EMP.sub.i] = 1 if [p.sup.*.sub.1] > 0 = 0 otherwise, (4)

where [p.sup.*.sub.i] = [a.sub.2] = + [C.sub.1]' [Y.sub.i] + [C.sub.2][S.sub.i] + [u.sub.Ei]; [EMP.sub.i] = 1 if individual i participates in the labor force and has positive earnings/wages, and equals 0 otherwise; [Y.sub.i] = vector of characteristics describing i that are thought to be related to labor force participation; [a.sub.2], [C.sub.1], [C.sub.2] = parameters to be estimated by probit; and [u.sub.Ei] = standard error term. We then use the predicted IMR, which we denote as [[??].sub.i] for each observation in the sample of workers. (17)

The second important source of endogeneity arises from stratifying the full sample into three groups: those without disabilities, those with early-onset disability, and those with late-onset disability. Stratifying the samples in this way is a reasonable approach to test for structural differences in employment and wage behavior. However, as noted earlier, reporting a disability may be endogenous. A robust estimation strategy is to explicitly model the presence of a (self-reported) disability. We implement this strategy by using the early-onset or late-onset disability status indicators as endogenous switching mechanisms. We treat them as endogenous because the probability of self-reporting disability is itself an outcome determined by both personal characteristics and state characteristics. The explanatory variables that we use to explain and identify disability status are state-level variables reflecting variations in the take-up rate of SSI disability benefits, industrial accident rates, physicians per capita, and smoking rates.

The precise model is a two-step endogenous switching regression model analogous to that described by Maddala (1983). The underlying switching mechanism is determined by a trivariate disability reporting model. For empirical convenience, we will estimate the disability reporting model by conditional logit. (18) In Equation 5, [d.sup.*.sub.ij] will be a latent variable describing the propensity of individual i to report status j, where j = early disability onset, late disability onset, or non-disability and [D.sub.ij] is the observed status.

[D.sub.ij] = j if [d.sup.*.sub.ij] > [d.sup.*.sub.ik], for all k [not equal to] = j, (5)

where [d.sup.*.sub.ij] = [[gamma].sub.j][Z.sub.ij] + [u.sub.ij], [Z.sub.ij] = vector of characteristics describing i that are thought to be related to the likelihood of reporting status j; [[gamma].sub.j] = vector of parameters to be estimated by logit; [u.sub.ij] = error term; [Z.sub.ij] = vector of characteristics describing i that are thought to be related to the likelihood of reporting status j; [[gamma].sub.j] = vector of parameters to be estimated by logit; and [u.sub.ij] = error term. Assuming that the [u.sub.ij] are independent and identically distributed with a Weibull distribution, then

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (6)

and the gamma parameters can be estimated by a maximum likelihood logit model. (19)

Because the errors in the disability reporting equations are correlated with the errors in the wage Equations 1-3, following Maddala's terminology, we refer to the model as a switching regression model with endogenous switching. Generalizing on Maddala's specification, Equations 1'-3' present the switching wage regressions above that include a "correction" term in each to adjust for the endogenous stratification on reported disability status as well as the employment selection correction term.

ln [W.sub.Ni] = [[alpha].sub.1] + [[beta].sub.11] [X.sub.Ni] + [[beta].sub.12] [S.sub.Ni] + [b.sub.13][[??].sub.i] + [b.sub.14][[??].sub.Ni] + [e.sub.Ni] (1')

ln [W.sub.Ei] = [[alpha].sub.2] + [[beta].sub.21] [X.sub.Ei] + [[beta].sub.22] [S.sub.Ei] + [b.sub.23][[??].sub.i] + [b.sub.24][[??].sub.Ei] + [e.sub.Ei] (2')

ln [W.sub.Li] = [[alpha].sub.3] + [[beta].sub.31] [X.sub.Li] + [[beta].sub.32] [S.sub.Li] + [b.sub.33][[??].sub.i] + [b.sub.34][[??].sub.Li] + [e.sub.Li], (3')

where [[??].sub.ki] = [[??].sup.*.sub.ki]ln ([[??].sub.ki])/(1 - [[??].sub.ki]), k = N, E, L, and [[??].sub.ki] = predicted probability [D.sub.ij] = k from Equation 6. Estimation of Equations 1'-3' yields consistent parameter estimates. The disability selection term should resolve the potential negative correlation between education and disability status, particularly for those with late-onset disability. (20)

5. Summary of Estimation Strategy

We approach the problem of differentiating the returns to education between individuals with and without disabilities using the econometric model shown in Equations 1'-3'. Note that this model utilizes two generated regressors produced via estimation of Equations 4 and 6.

In addition, the above estimation is implemented with two different measures of education: a continuous years of education measure and then a piecewise representation of years of education: min(yrseduc,12); max[0, min(yrseduc - 12, 4)]; max(0, yrseduc - 16). We label those three variables as years of elementary and secondary education, years of college, and years beyond the baccalaureate, respectively. This alternative set of education regressors produces estimates of marginal effects differentiated by three levels of education. This sort of education spline has the advantage of producing marginal effects that incorporate the degree effect noted by Hungerford and Solon (1987). (21)

Our empirical specification does not incorporate three modeling extensions that have appeared in the recent returns to education literature. First, we ignore any systematic unobserved heterogeneity in education across the disabled and non-disabled population. Our data do not include reasonable correlates for education that are not related to wages for use in a corrective instrumental variable (IV) procedure, and potential state-level data instruments performed poorly in preliminary runs. While it is convenient to assume away the education selectivity, we recognize the sizeable returns to education literature that attempts to compensate for the fact that investment in human capital is an endogenously determined outcome that is correlated with unobserved ability, so its direct inclusion as a right-hand side regressor can be problematic. Somewhat reassuringly, Card (2000) finds that IV estimates of education returns are relatively imprecise, but still at least as large (and sometimes much larger) than non-IV estimates. This alleviates concern that the non-IV approaches overestimate the returns to education. (22) Note also that one significant factor contributing to potential endogeneity of education is the age of disability onset. We do incorporate this factor in our empirical analyses.

Card (2000) also discusses the issue of modeling unobserved heterogeneity (across all individuals) in wage equations; in other words, the use of panel data techniques. He states that, "Since people with a higher return to education have an incentive to acquire more schooling, a cross-sectional regression of earnings on schooling is likely to yield an upward-biased estimate of the average marginal return to schooling ..." (p. 9). Our estimation relies on cross-sectional data, and so our results will likely be overestimates of the returns to education. While the SIPP data are available in a panel, we rely on a single wave in our estimation because of the difficulties inherent in incorporating panel data methods with a switching mechanism (disability status) that does not vary much over the course of the panel. Additionally, the SIPP information regarding age of onset and health condition details is not available longitudinally. As our focus is on relative estimates of wage returns to education, with proper modeling of heterogeneity by disability status and age of onset, our findings do not suffer substantively from problems due to the lack of panel data. (23)

6. Data Description

We use data from the 1993 Panel of the Survey of Income and Program Participation. The SIPP data come from the third interview of the 1993 Panel (called wave 3) that covers the time period from September through December 1993. These data contain detailed disability status measures that were designed to be consistent with the ADA 1990 definition of disability. To retain comparability with previous research, we define disability status using the self-reported response to the question concerning work-limiting conditions. (24) However, as is done in Baldwin and Johnson (2000), we include information regarding a broad list of limitations in daily activities and other functional disabilities as control variables in the wage regressions because of their potential productivity-related effects.

To obtain information on the age of disability onset, we merge our 1993 Panel wave 3 data with relevant survey responses from wave 2 data of the same panel. Specifically, the age of onset of disability information is obtained from part A of the wave 2 topical module on the topic of work disability history. For those individuals reporting a work-limiting condition, the following question was asked: "When did ... become limited in the kind or amount of work that ... could do at a job?" Three responses were possible: the exact month/year of onset; an indicator that "person was limited before person became of working age"; and an indicator that "person became limited after retiring." Using this information, we constructed two measures of age of onset: early onset and late onset. Early onset includes all those with onset at or before age 25 years. Using this age cutoff means that individuals in our late-onset sample likely completed their human capital investment prior to disability onset, alleviating concerns about educational investments being influenced by disability status. (25)

We restrict our SIPP data to include only individuals aged 25 to 62 years in order to minimize problems associated with modeling formal human capital investment decisions or retirement decisions, both of which are decisions made jointly with the employment choice. We present variable means stratified by employment status and disability status (broken down by age of onset) in Table 1. Our data include 9048 non-disabled employed men; 1663 non-disabled, not-employed men; 601 men with disabilities who work; and 760 men with disabilities who are not currently working. (26) Of the full estimation sample, approximately 12% report having a health condition/disability that limits work. (27) In the sample of 1361 individuals reporting disabilities, 21.6% report early onset and 72.4% report late onset of disabilities. (28) This results in 294 individuals in the sample reporting early-onset disability and 1067 reporting late onset. Note that the sample size for workers with early-onset disability is small, a concern that will be reflected later in the strength of the empirical results. For individuals who report a specific age of disability onset, the average age of onset is 45.3 years, very similar to that reported by Burkhauser, Butler, and Weathers (2002).

Considering first the demographic characteristics, Table 1 shows that individuals with early-onset disability are somewhat younger than the non-disabled, while those with late onset of disability are considerably older, reflecting the fact that disability probabilities rise with age. Individuals with disabilities achieve lower average educational attainment (even those with late onset). Twenty-six percent of the sample of employed individuals with early-onset disability report less than high school education (vs. just 17% for the late-onset group), while twice as many of the employed late-onset group report greater than a college education (15% vs. 7%). As might be expected, individuals with early-onset disability are considerably less likely to be married. The role of race is somewhat complex. Fourteen percent of the sample of non-disabled workers is nonwhite, but the percentage of nonwhite that are not employed rises to 21%, consistent with national non-employment rates being higher for nonwhites. This pattern is repeated for the early- and late-onset samples, but it is more extreme for the former: while only 4% of the working sample is nonwhite, 30% of the nonworking sample is nonwhite.

For the non-disabled, non-workers report higher non-labor income than workers (likely resulting from higher wages from spouses), while those with early-onset disability report nearly equal incomes by employment status, and for those with late-onset disability, the workers report higher non-labor income. Finally, individuals with disabilities (regardless of employment status or age of onset) tend to have fewer children, on average.

Considering the aggregate figures for each of the six subsamples, 84.5% of the non-disabled males are working for pay, with the employment rates considerably lower for the two groups reporting disabilities: 51% for early onset and 42% for late onset. The fact that fewer than half of those reporting late-onset disability are employed probably reflects the greater incentives to report a disability for those nearing retirement age. Average hourly wages vary substantially by disability status, with the non-disabled earning $14.32 an hour, workers with early-onset disability earning $9.18 an hour, and those with late-onset earning $12.20 an hour. These wage differences represent a 36% wage disadvantage for individuals with early-onset disability and a 15% wage disadvantage for those with late onset. Reported work hours also vary by disability status, although not by as much, in percentage terms, as the wage differences. Non-disabled workers work on average 45.1 hours per week, while the early- and late-onset groups report 40.5 and 41.7 hours per week, respectively. (29)

Following Baldwin and Johnson (1994, 2000), we exploit the availability in the SIPP of detailed health-related work productivity measures. In addition to chronic conditions and self-reported work-limiting disabilities, the topical module available with our wave of the 1993 SIPP contains detailed information that encompasses three categories of health and functional limitations: limitations in activities of daily living (ADLs), such as bathing, dressing, or eating; limitations in instrumental activities of daily living (IADLs), such as going outside the home, preparing meals, or doing housework; and other functional limitations, such as difficulties in seeing, hearing, or climbing stairs. (30) While Baldwin and Johnson (2000) rely on a complicated method for constructing these three measures, we rely on a more straightforward approach of simply adding up the occurrences to construct three count measures as additional regressors in the wage equations. This simplification should not have much impact on our estimates of the returns to education, particularly since Baldwin and Johnson's estimates find little statistical significance for these measures' estimated coefficients. (31)

As would be expected, the mean numbers of limitations on ADLs, limitations on IADLs, and other functional limitations for men with disabilities exceed considerably the mean numbers for non-disabled males. Furthermore, among all men with disabilities, the means for non-workers are considerably higher than those for workers. Interestingly, those in the group with early onset of disability who are employed have a higher incidence of ADLs and functional limitations than do the corresponding employed group for those with late-onset disability.

7. Estimation Results

Estimates of the returns to education by disability status from the switching wage regressions are presented in Table 2. First-stage estimation results of employment selection (Eqn. 4) and of disability reporting regime (Eqn. 5) can be found in Appendix Tables A1 and A2, respectively. As discussed previously, our definition of disability relies on an affirmative response to whether the survey respondent has a health condition that limits her/his ability to work. Entries in the tables are coefficients from regressions on constructed hourly wages in logarithmic form, so they may be interpreted as percentage effects. The other independent variables in the wage regressions control for other demographic and human capital characteristics, as well as the selection terms. (32)

As seen in Table 2, workers with early-onset disability appear to have no statistically significant return to education. This may be the result of the failure of special education programs to prepare these individuals for employment, or possibly it may simply reflect the small sample sizes. (33) The former explanation is consistent with Kruse (1997), who suggests that lower returns to education partially explains lower employment rates for those with disabilities. The story is very different for individuals who experience late disability onset. These workers enjoy very high returns to education, much higher than those of non-disabled workers (21.1% vs. 10.2%). Focusing on the education spline measures reveals that workers with no disabilities experience an approximate 9-11% return to education, with no evidence that more advanced degrees carry greater degree effects. For those with late-onset disability, the "kick" for an additional year of pre-college or post-baccalaureate education is the greatest, at approximately 24-25%. Overall, our estimates of the returns to education for those with late-onset disability imply that earlier findings of high returns for workers with disabilities (e.g., DeLeire 2001) may have been driven by the high percentage of the samples comprising workers with late-onset disability.

The full set of coefficients for the wage equations estimated using the endogenous switching model are also seen in Table 2. Overall, the coefficient results vary considerably across disability status, indicating the importance of permitting the full set of coefficients to vary. Many of the wage equation regressors impact wages in the expected way: marital status, age, and living in a metropolitan area are related positively to wages. For men with late-onset disability, the male marriage wage premium is substantially larger than that for the non-disabled (approximately 80-83% vs. 40-42%), indicating some combination of the following two explanations: (i) the ability to specialize in market work might be particularly advantageous for men with disabilities; or (ii) only the most productive men with disabilities marry. Being an enrolled student depresses wages, possibly as a result of its likely correlation with part-time employment. Nonwhite race/ ethnicity and residing in the South also reduce wage rates.

Despite the fairly high incidence of functional limitations for those with disabilities, not many of these measures are statistically significant. Note, however, that even with the small sample, workers with early-onset disability do suffer reduced wages for two of the three measures, although the statistical power is weak. Recall that Baldwin and Johnson (2000) found little explanatory power for these measures.

8. Conclusions

Our findings provide evidence concerning the role of educational investments for individuals with disabilities. The most notable finding is in the importance of examining returns to education by age of disability onset. Estimates focusing on workers with early-onset disability show that education may not have a wage advantage for these individuals. On the other hand, our estimates indicate very large returns to education for individuals whose disability onset was past the age of 25 years. An interpretation of these results is that the quality and quantity of education received by individuals who are disabled at the time of their educational decisions (i.e., those with early onset) are not serving them well in terms of finding productive job matches. This could be caused by increased severity of disabilities; by programmatic approaches that are emphasizing employment over skill development; or by low-quality special educational services. Of course, the small sample size for this group may be responsible for the lack of precision in the education returns estimates.

The handsome returns to education for disabled men whose disability had a later onset highlight the insurance value of formal education. These individuals gained general skills and knowledge during their education that allowed them to continue to earn higher wages once becoming disabled. The policy implication is that education has an important insurance value, in addition to its labor market returns, and this insurance value extends beyond just the risk of job loss, for example, to the risk of becoming disabled. The relative magnitude of our estimates for the late-onset population compared to those of the non-disabled yields a rough estimate of the size of this insurance value.

At a minimum, our findings indicate that age of onset is an important factor to consider when identifying which services are best suited to particular individuals with disabilities. Additionally, given the possibility that the quality of early education is an issue for those with early-onset disability, we modify the policy recommendation of Loprest and Maag (2003) to include not just a focus on education completion but also increased attention to the type of that education, particularly its skill-enhancement component. Of course, this suggestion carries with it the caveat that the small sample size may have played a substantial role in the lack of statistical significance in most of the wage equation regressors, including the education measures.

Individuals with late-onset disabilities do appear to benefit from formal education, perhaps because higher education improves one's ability to change jobs to achieve a better job match with an accommodating employer. To the extent that education may serve to lessen the negative employment effects of disability, the education buffer appears to be more effective for those with late-onset disabilities (see, for example, Daly and Bound 1996). Additionally, it may be the case that vocational rehabilitation services (unobserved in our data), typically targeted toward this group, are most beneficial for those with higher levels of education. Interpretation of the high returns to education for those with late-onset disability must be considered in the context of a substantial wage gap between the populations with and without disabilities. The average wages of individuals with poor health or disability, regardless of age of disability onset, are below those of the non-disabled. Thus, while the positive returns to education for workers with late-onset disability indicate that education may serve as a buffer to protect against potential negative wage effects of becoming disabled, this education "insurance" is not sufficient to close the gap.

As a final note, the reader will recall our earlier discussion of three potential limitations to our empirical approach: no IV estimation to address endogeneity of schooling, no panel data methods, and no explicit modeling of the simultaneity between health and education. Our research focus is on the estimate of the returns to education for those with disabilities relative to those of the non-disabled. Because we have incorporated econometric extensions critical to producing reliable relative estimates, we feel our conclusions do not suffer from these omissions. However, to the extent that concerns associated with the potential endogeneity of schooling may be more serious for workers with early-onset disability, we suggest that future research address this endogeneity.

Appendix A1
Coefficients from Employment Probit Model
(Standard Errors in Parentheses)

 Employment Probit
Variable Coefficients

Intercept 1.714 (0.280)
Age 0.092 * (0.013)
[Age.sup.2] (scaled in 00s) 0.139 * (0.015)
Non-white 0.310 * (0.046)
Marital status 0.507 * (0.039)
South 0.090 * (0.034)
Metro 0.132 * (0.037)
Education 0.090 * (0.006)
Enrollment 0.671 * (0.071)
Adjusted family income (scaled in 0000s) 0.280 * (0.067)
Number of children in the household aged 0-17 years 0.019 (0.015)

* Significant at the 0.01 level.

Appendix A2
Coefficients from Disability Reporting Logit
Model (Standard Errors in Parentheses)

 Early-Onset Late-Onset
 Disability Logit Disability Logit
Variable Coefficient Coefficient

Intercept 0.604 (1.630) -4.854 *** (1.042)
Age 0.068 * (0.056) 0.141 *** (0.035)
[Age.sup.2] (scaled in 00s) 0.094 (0.068) -0.073 * (0.038)
Non-white 0.155 (0.199) 0.144 (0.109)
Marital status -1.466 *** (0.156) -0.659 *** (0.089)
South 0.088 (0.179) -0.225 ** (0.103)
Metro 0.149 (0.152) -0.163 ** (0.085)
Education 0.241 *** (0.023) -0.177 *** (0.012)
Enrollment 0.661 ** (0.0275) 0.529 *** (0.203)
Adjusted family income
 (in 0000s) 0.556 (0.275) 0.252 (0.156)
Number of children in the 0.121 * (0.070) 0.063 * (0.035)
 household aged 0-17 years
Smoking 0.011 (0.040) 0.033 (0.023)
Per capita physicians 0.004 (0.007) -0.008 ** (0.004)
Accident 0.010 (0.012) 0.009 (0.007)
SSI 7.337 (55.179) 37.730 (30.861)

* Significant at the 0.10 level.

** Significant at the 0.05 level.

*** Significant at the 0.01 level.

Received June 2003; accepted January 2007.

References

Acemoglu, Daron, and Joshua D. Angrist. 2001. Consequences of employment protection? The case of the Americans with Disabilities Act. Journal of Political Economy 109:915-57.

Baldwin, Marjorie L., and William G. Johnson. 1994. Labor market discrimination against men with disabilities. Journal of Human Resources 29:l-19.

Baldwin, Marjorie L., and William G. Johnson. 2000. Labor market discrimination against men with disabilities in the year of the ADA. Southern Economic Journal 66:548-66.

Barnow, Burr S. 1996. Policies for people with disabilities in U.S. employment and training programs. In Disability, work and cash benefits, edited by Jerry L. Mashaw, Virginia Reno, Richard V. Burkhauser, and Monroe Berkowitz. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research, pp. 297-330.

Berger, Mark C., and J. Paul Leigh. 1989. Schooling, self-selection, and health. The Journal of Human Resources 24:433-55.

Bound, John, and Richard V. Burkhauser. 1999. Economic analysis of transfer programs targeted on people with disabilities. In Handbook of labor economics, volume 3, edited by Orley Ashenfelter and David Card. New York: Elsevier Press, pp. 3417-529.

Bound, John, Michael Schoenbaum, and Timothy Waidmann. 1995. Race and education differences in disability status and labor force attachment in the health and retirement survey. The Journal of Human Resources 30(Supplement):S227-67.

Burkhauser, Richard V., J. S. Butler, and Robert R. Weathers, II. 2002. How policy variables influence the timing of applications for social security disability insurance. Social Security Bulletin 64:37-52.

Burkhauser, Richard V., and Mary C. Daly. 1996. Employment and economic well-being following the onset of a disability: The role for public policy. In Disability, work and cash benefits, edited by Jerry L. Mashaw, Virginia Reno, Richard V. Burkhauser, and Monroe Berkowitz. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research, pp. 59-102.

Card, David. 2000. Estimating the return to schooling: Progress on some persistent econometric problems. NBER Working Paper No. 7769.

Daly, Mary C., and John Bound. 1996. Worker adaptation and employer accommodation following the onset of a health impairment. Journal of Gerontology, Social Sciences 51b:S53-60.

DeLeire, Thomas. 2000. The wage and employment effects of the Americans with Disabilities Act. The Journal of Human Resources 35:693-715.

DeLeire, Thomas. 2001. Changes in wage discrimination against people with disabilities: 1984-1993. Journal of Human Resources 36:144-58.

Dubin, Jeffrey A., and Daniel L. McFadden. 1984. An econometric analysis of residential electric appliance holdings and consumption. Econometrica 52:345-62.

Ettner, Susan L., Richard G. Frank, and Ronald C. Kessler. 1997. The impact of psychiatric disorders on labor market outcomes. Industrial and Labor Relations Review 51:64-81.

Griliches, Zvi. 1977. Estimating the returns to schooling: Some econometric problems. Eeonometrica 45:1-22.

Haveman, Robert, and Barbara Wolfe. 2000. The economics of disability and disability policy. In Handbook of health economics, edited by A. J. Culyer and J. P. Newhouse. Amsterdam: North-Holland, pp. 995-1051.

Heckman, James J. 1974. Shadow prices, market wages and labor supply. Econometrica 42:679-94.

Hendricks, Wallace, Chrisann Schiro-Geist, and Emir Broadbent. 1997. Long-term disabilities and college education. Industrial Relations 36:46-60.

Hotchkiss, Julie L. 2003. The labor market experience of workers with disabilities: The ADA and beyond. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research.

Hotchkiss, Julie L. 2004. A closer look at the employment impact of the Americans with Disabilities Act. The Journal of Human Resources 39:887-911.

Hungerford, Thomas, and Gary Solon. 1987. Sheepskin effects in the returns to education. The Review of Economics and Statistics 69:175-7.

Johnson, William G., and James Lambrinos. 1985. Wage discrimination against handicapped men and women. The Journal of Human Resources 20:264-77.

Kruse, Douglas. 1997. Employment and disability: Characteristics of employed and non-employed people with disabilities, Report to the U.S. Department of Labor, Final Report (September).

Kruse, Douglas, and Lisa Schur. 2003a. Employment of people with disabilities following the ADA. Industrial Relations 42:31-66.

Kruse, Douglas, and Lisa Schur. 2003b. Does the definition affect the outcome? In The decline in employment of people with disabilities: A policy puzzle, edited by David C. Stapleton and Richard V. Burkhauser. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research, pp. 279-300.

Lee, L. F. 1983. Generalized econometric models with selectivity. Econometrica 51:507-12.

Loprest, Pamela, and Elaine Maag. 2003. The relationship between early disability onset and education and employment. Unpublished manuscript, The Urban Institute: Washington, DC.

Loprest, Pamela, Kalman Rupp, and Steven H. Sandell. 1995. Gender, disabilities, and employment in the health and retirement study. The Journal of Human Resources 30:S293-S318.

Maddala, G. S. 1983. Limited-dependent and qualitative variables in econometrics. New York: Cambridge University Press.

Mincer, J. 1974. Schooling, experience, and earnings. New York: National Bureau of Economic Research.

Mitchell, Jean M., and Richard V. Burkhauser. 1990. Disentangling the effects of arthritis on earnings: A simultaneous estimate of wage rates and hours worked. Applied Economics 22:1291-309.

Neal, Derek. 2002. The measured black-white wage gap among women is too small. Unpublished manuscript, University of Chicago.

Oi, Walter Y. 1996. Employment and benefits for people with diverse disabilities. In Disability, work and cash benefits, edited by Jerry L. Mashaw, Virginia Reno, Richard V. Burkhauser, and Monroe Berkowitz. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research, pp. 103-28.

Pezzin, Liliana E., and Barbara Steinberg Schone. 1999. Intergenerational household formation, female labor supply and informal caregiving: A bargaining approach. The Journal of Human Resources 34:475-503.

Schwochau, Susan, and Peter Blanck. 2003. Does the ADA disable the disabled? More comments. Industrial Relations 42:67-77.

Stapleton, David C., and Richard V. Burkhauser, eds. 2003. The decline in employment of people with disabilities: A policy puzzle. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research.

Stern, Steven. 1989. Measuring the effect of disability on labor force participation. The Journal of Human Resources 24:361-95.

Waldrop, Judith, and Sharon M. Stern. 2003. Disability status: 2000. Census 2000 brief Washington, DC: U.S. Census Bureau.

Yelin, Edward, and Miriam Cisternas. 1996. The contemporary labor market and the employment prospects of persons with disabilities. In Disability, work and cash benefits, edited by Jerry L. Mashaw, Virginia Reno, Richard V. Burkhauser, and Monroe Berkowitz. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research, pp. 33-58.

(1) For recent analyses of the complex issues involved in determining the impact of ADA, see DeLeire (2000), Kruse and Schur (2003a), Schwochau and Blanck (2003), and Hotchkiss (2004).

(2) See Barnow (1996) for a detailed description of public policies focusing on employment and training for individuals with disabilities.

(3) See, for example, Burkhauser and Daly (1996), Yelin and Cisternas (1996), and Stapleton and Burkhauser (2003).

(4) See Kruse and Schur (2003b) for a discussion of disability definitions.

(5) For a comprehensive discussion of this issue, see Acemoglu and Angrist (2001), Hotchkiss (2003), and the collected papers included in Stapleton and Burkhauser (2003).

(6) These Health and Retirement Survey (HRS) data are for the cohort born between 1931 and 1941 and interviewed in 1992.

(7) Loprest and Maag (2003) define disability based on the presence of activity and functional limitations (p. 6). Additionally, their sample is near retirement age, thereby producing a sample with a relatively high percentage of late-onset disabled.

(8) See Oi (1996) for further details concerning work and benefits for the diverse disabled population.

(9) With regard to monitoring the effects of ADA, Kruse and Schur (2003b) explain that there is no single data set that plots labor market outcomes over time for the population affected by ADA, in part because "eligibility" is evolving over time and partially because of the difficulties inherent in determining the relevant population. Additionally, self-identification as "disabled" is not stable over time.

(10) Further literature important to this area includes Bound, Schoenbaum, and Waidmann (1995), who study race and education differences for men in disability status and labor force attachment, as well as time of disability onset, and Loprest, Rupp, and Sandell (1995), who focus on sex differences. Both papers also show that access to and choice of occupation are important determinants of outcome differences.

(11) Hendricks, Schiro-Geist, and Broadbent (1997) study the importance of a broader array of human capital investments (including university education and rehabilitation services) and find that both serve to mitigate employment-outcome disparities.

(12) DeLeire's (2001) definition of disabled includes those who self-report a work-limiting condition as well as those with functional limitations. The estimate cited in text is not statistically significant.

(13) In addition to this continuous measure of years of education, the estimation will be repeated Using instead three distinct piecewise education measures explained later in text. Reliance on the continuous education measure relies on the assumption of a constant returns to education for each year of additional education, an assumption not always supported by empirical evidence (see, for example, Hungerford and Solon 1987).

(14) See Bound and Burkhauser (1999) and Haveman and Wolfe (2000) for in-depth discussions of the relationship between disability policies and labor market behavior for the disabled.

(15) This will be the case for most surveys. For example, if the data are collected by telephone or if a single respondent provides information for a family or household, the individuals who are collecting the information will be unaware of poor health or disabilities. Furthermore, a health condition or disability may be internal and unseen, even when the data collector is interacting personally with the individual.

(16) The selection on employment may be particularly important for early onsetters, who are more likely to receive disability benefits. According to Loprest and Maag (2003), early onsetters' lower employment rates are due primarily to this benefit receipt.

(17) Variables that are included in the estimation of employment but not wages, thus helping to identify it, include adjusted family income and number of children in the household.

(18) This procedure replicates the empirical work in Dubin and McFadden (1984). See Lee (1983) for the theoretical justification for using a logit model in a heckit-type selection correction model. that distinguish between early versus late onset of disability, another potential source of heterogeneity. (16)

(19) Estimation of the model normalizes one of the gamma vectors to zero.

(20) It is likely that much of late-onset disability affects lesser-skilled workers (as a result of a higher risk of on-the-job injury), so that the age of onset and education are potentially negatively correlated. Were this correlation not addressed via our selection mechanism, our estimates might overstate the returns to education for the late-onset group.

(21) Estimated coefficients produced with these education splines are simple to interpret. As an example, if the estimated coefficient associated with the second spline is equal to 0.10, this implies that each extra year of education beyond high school generates a 10% wage return.

(22) Also see Griliches (1977) for further details on the econometric difficulties inherent in estimating returns to education.

(23) There is also a strand of the health/education literature that examines the positive correlation of education and health outcomes. Berger and Leigh (1989) conclude that the direct effect of schooling on health is more important than the effect of unobservables jointly on both outcomes.

(24) We also experimented with an alternative, more broadly defined, disability measure that combines the self-reported assessment measure with other more-specific types of disability or poor health. Our overall empirical results were not altered by the use of this alternative measure and so are not reported in text or tables. See Kruse and Schur (2003b) for a detailed discussion of disability definitions. Also, we attempted to replicate the disability definitions used by Baldwin and Johnson (2000) that rely on self-reports of specific health conditions to construct two measures of disability based on the relative visibility of the disability (and thus the likely existence of discrimination), but we were unable to do so because of an insufficient number of disabled defined this way. According to O'Hara at the Census Bureau (during a 2001 phone conversation with one of the authors), this sample size problem occurs because of fundamental changes in sample design implemented following the 1990 SIPP panel.

(25) Loprest and Maag's (2003) analyses of age of onset uses age 22 as the cutoff'. We prefer the later age for the reason stated in the text.

(26) The rightmost columns of the table do not include 137 observations in which the individual self-reported a disability but for which age of onset or functional limitations could not be determined because of missing data or no match between wave 2 and wave 3 of the SIPP. Of these, 85 were employed and 52 were not employed.

(27) Weighted means are comparable. These percentages are quite close to those reported by Kruse (1997), who also used later SIPP panels, but are somewhat larger than what is reported by other researchers using other data or earlier panels of the SIPP. We thank Douglas Kruse for his assistance in assuring the accuracy of our data.

(28) This compares nicely to Loprest and Maag's (2003) finding of 16.3% reporting onset prior to age 22 years and to the findings of Burkhauser and Daly (1996), who report that 70% of individuals with disabilities report onset during the working years.

(29) These mean hours are consistent with the mean weekly hours calculated from the outgoing rotation group of the 1993 Current Population Survey (CPS). The CPS mean for male workers is 40.98, and this sample does not include salaried workers who are likely to work longer hours. Additionally, the CPS mean is for the primary job only, while the SIPP mean reflects usual weekly hours, which includes hours from multiple jobs.

(30) See Pezzin and Schone (1999) for an example of an application of the IADL and ADL disability measures. Included in the ADL list are activities associated with bathing, dressing, eating, toileting, getting in and out of bed, and preparing meals. Included in the IADL list are activities related to housecleaning, using the telephone, keeping track of money and bills, climbing a flight of stairs, lifting, walking, climbing, and getting around inside and outside the house. Other functional limitations include seeing, hearing, speaking, and using an aid to get around. We thank Anthony LoSasso for his expertise in dividing the broad list of functional limitations into the three subcategories.

(31) All estimation was repeated using a second, more broadly defined disability measure that includes the self-report on disability status plus many additional medical conditions. The empirical findings were not substantively affected by this alternative definition. Additionally, we include the three measures of activities limitations. However, as a result of potential concerns about endogeneity or appropriateness, we repeated the estimation without these three measures and found no substantive difference in results. These results are available from the authors on request.

(32) The first stage regressions on employment include sex, age, age-squared, minority status, marital status, presence of children, region, years of education, current student status, and monthly family income excluding own earnings.

(33) Alternative (simpler) specifications did produce some statistically significant education effects for individuals with early-onset disabilities. The first specification was a single wage equation that included dummy variables for early onset and late onset, plus these dummies interacted with the education measures. The second specification had wage equations estimated separately by disability status (early and then late onset), but this endogenous stratification was not corrected. Estimation results from these specifications are available from the authors.

Kevin Hollenbeck * and Jean Kimmel ([dagger])

* W. E. Upjohn Institute, 300 South Westnedge Avenue, Kalamazoo, MI 49007, USA; E-mail Hollenbeck@ upjohninstitute.org.

([dagger]) Department of Economics, Western Michigan University, Kalamazoo, MI 49008-4023, USA; E-mail Jean.Kimmel@wmich.edu; corresponding author.

This paper is a substantially revised version of a paper that was first presented at the January 2001 ASSA meetings held in New Orleans, Louisiana, and then circulated as Upjohn Institute Working Paper No. 01-72. The authors thank our discussant, Robert Haveman, for his excellent suggestions; Andrew Houtenville for generous assistance with the disability literature; Doug Kruse for input and data assistance above and beyond the call of duty; and Noyna DebBurman, Jason Preuss, and Yuanlei Zhu for their skillful research assistance. Additionally, we are very appreciative to Claire Black and Sue Berkebile for their superb assistance in preparation of the manuscript. Finally, we thank our journal editor plus two anonymous journal referees.

Table 1. Variable Means (Standard Deviation in Parentheses)

 Non-Disabled

 Employed Not Employed

Number of observations 9048 1663
Demographic characteristics
 Age 39.84 (9.6) 42.11 (11.7)
 Education 13.48 (2.8) 12.69 (3.0)
 Non-white 0.14 (0.3) 0.21 (0.4)
 Number of children aged 0-6 years 0.39 (0.7) 0.30 (0.7)
 Number of children aged 7-17 years 0.58 (0.9) 0.46 (0.9)
 Marital status 0.70 (0.5) 0.56 (0.5)
 Number of children 0.96 (1.2) 0.76 (1.2)
 Monthly family income, excluding $1897.57 $2090.75
 own earnings (2390.7) (2443.05)
Activity limitations
 Activities of daily living
 (ADL) limitations 0.00 (0.1) 0.01 (0.2)
 Instrumental ADL limitations 0.02 (0.2) 0.05 (0.4)
 Functional limitations
Regional characteristics 0.06 (0.3) 0.08 (0.3)
 Live in metro area 0.77 (0.4) 0.76 (0.4)
 South 0.34 (0.5) 0.30 (0.5)
Job characteristics
 Hourly wage $14.32 (9.1) --
 Weekly hours 45.1 (11.0) --
Educational categories
 Less than 12 years 0.12 (0.3) 0.24 (0.4)
 High school (12 years) 0.36 (0.5) 0.34 (0.5)
 Some college (13-15 years) 0.22 (0.4) 0.21 (0.4)
 College or higher (16+ years) 0.30 (0.5) 0.21 (0.4)

 Early-Onset Disabled

 Employed Not Employed

Number of observations 149 145
Demographic characteristics
 Age 37.17 (9.0) 39.31 (10.5)
 Education 12.31 (2.2) 10.66 (3.1)
 Non-white 0.04 (0.2) 0.30 (0.5)
 Number of children aged 0-6 years 0.22 (0.5) 0.23 (0.7)
 Number of children aged 7-17 years 0.34 (0.7) 0.36 (0.9)
 Marital status 0.38 (0.5) 0.23 (0.4)
 Number of children 0.56 (0.9) 0.59 (1.2)
 Monthly family income, excluding $1734.68 $1798.48
 own earnings (1973.3) (1602.3)
Activity limitations
 Activities of daily living
 (ADL) limitations 0.23 (0.8) 0.69 (1.4)
 Instrumental ADL limitations 0.60 (1.3) 1.89 (2.2)
 Functional limitations
Regional characteristics 0.39 (0.6) 0.69 (0.9)
 Live in metro area 0.75 (1.3) 0.72 (0.4)
 South 0.29 (0.6) 0.43 (0.5)
Job characteristics
 Hourly wage $9.18 (7.3) --
 Weekly hours 40.5 (12.7) --
Educational categories
 Less than 12 years 0.26 (0.4) 0.45 (0.5)
 High school (12 years) 0.42 (0.5) 0.42 (0.5)
 Some college (13-15 years) 0.25 (0.4) 0.09 (0.3)
 College or higher (16+ years) 0.07 (0.3) 0.04 (0.2)

 Late-Onset Disabled

 Employed Not Employed

Number of observations 452 615
Demographic characteristics
 Age 44.92 (9.7) 48.79 (9.7)
 Education 12.63 (2.5) 11.03 (3.0)
 Non-white 0.11 (0.3) 0.21 (0.4)
 Number of children aged 0-6 years 0.30 (0.6) 0.20 (0.6)
 Number of children aged 7-17 years 0.58 (0.9) 0.43 (0.9)
 Marital status 0.71 (0.5) 0.58 (0.5)
 Number of children 0.88 (1.1) 0.62 (1.2)
 Monthly family income, excluding $2094.27 $1811.35
 own earnings (2383.1) (1675.2)
Activity limitations
 Activities of daily living
 (ADL) limitations 0.18 (0.7) 0.69 (1.4)
 Instrumental ADL limitations 0.67 (1.2) 2.21 (2.1)
 Functional limitations
Regional characteristics 0.28 (0.6) 0.66 (0.9)
 Live in metro area 0.73 (0.4) 0.68 (0.5)
 South 0.34 (0.5) 0.37 (0.5)
Job characteristics
 Hourly wage $12.20 (9.1) --
 Weekly hours 41.7 (12.0) --
Educational categories
 Less than 12 years 0.17 (0.4) 0.45 (0.5)
 High school (12 years) 0.44 (0.5) 0.35 (0.5)
 Some college (13-15 years) 0.24 (0.4) 0.13 (0.3)
 College or higher (16+ years) 0.15 (0.4) 0.07 (0.3)

Table 2. Coefficients from Wage Regressions
(Standard Errors in Parentheses)

 Late-Onset
Variable Non-Disability Disability

Years of schooling 0.102*** (0.005) 0.211*** (0.041)
Years of elementary/secondary
 schooling -- --
Years of college -- --
Years of post-baccalaureate -- --
Age 0.161 *** (0.010) 0.097 ** (0.048)
[Age.sup.2] (scaled in 00s) 0.202 *** (0.013) 0.163 *** (0.053)
Non-white 0.420 *** (0.029) 0.346 ** (0.134)
Married 0.415 *** (0.028) 0.803 *** (0.181)
South 0.024 ** (0.014) 0.139 ** (0.065)
Metro 0.240 *** (0.015) 0.358 *** (0.082)
Number of children in the
 household aged 0-17 years 0.016 *** (0.006) 0.062 ** (0.031)
Enrollment 0.677 *** (0.052) 0.729 *** (0.240)
Activities of daily living
 (ADL) limitations 0.020 (0.066) 0.020 (0.052)
Instrumental ADL limitations 0.016 (0.029) 0.024 (0.031)
Functional limitations 0.035 (0.22) 0.028 (0.052)
Intercept 2.967 *** (0.253) 1.618 (1.552)
Disability reporting mills 0.478 *** (0.067) 0.202 * (0.106)
Employment mills 2.876 *** (0.206) 2.062 *** (0.608)
Adjusted [R.sup.2] (n) 0.2768 (8169) 0.2940 (402)

 Early-Onset
Variable Disability Non-Disability

Years of schooling 0.039 (0.099) --
Years of elementary/secondary
 schooling -- 0.094 *** (0.009)
Years of college -- 0.100 *** (0.006)
Years of post-baccalaureate -- 0.110 *** (0.011)
Age 0.165 * (0.089) 0.161 *** (0.010)
[Age.sup.2] (scaled in 00s) -0.212 (0.130) 0.202 *** (0.013)
Non-white -0.603 (0.441) 0.418 *** (0.029)
Married -0.155 (0.612) 0.404 *** (0.029)
South -0.174 (0.133) 0.025 ** (0.014)
Metro 0.002 (0.143) 0.236 *** (0.015)
Number of children in the
 household aged 0-17 years 0.029 (0.098) 0.017 *** (0.006)
Enrollment -0.815 (0.553) 0.670 *** (0.052)
Activities of daily living
 (ADL) limitations -0.213 (0.106) 0.021 (0.066)
Instrumental ADL limitations 0.093 (0.071) 0.018 (0.029)
Functional limitations -0.095 (0.090) 0.034 (0.022)
Intercept -3.675 (2.707) 2.861 *** (0.267)
Disability reporting mills -0.203 (0.223) 0.508 *** (0.072)
Employment mills 1.629 (1.542) 2.874 *** (0.206)
Adjusted [R.sup.2] (n) 0.3530 (136) 0.2767 (8169)

 Late-Onset Early-Onset
Variable Disability Disability

Years of schooling -- --
Years of elementary/secondary
 schooling 0.234 *** (0.046) 0.009 (0.119)
Years of college 0.184 *** (0.046) 0.055 (0.112)
Years of post-baccalaureate 0.246 ** (0.080) 0.032 (0.198)
Age 0.111 ** (0.050) 0.149 (0.096)
[Age.sup.2] (scaled in 00s) 0.179 *** (0.055) -0.190 (0.139)
Non-white 0.365 *** (0.136) -0.555 (0.458)
Married 0.827 *** (0.182) -0.176 (0.621)
South 0.124 * (0.067) -0.191 (0.140)
Metro 0.362 *** (0.082) -0.005 (0.146)
Number of children in the
 household aged 0-17 years 0.063 ** (0.031) 0.037 (0.100)
Enrollment 0.747 *** (0.244) -0.749 (0.575)
Activities of daily living
 (ADL) limitations 0.017 (0.052) -0.213 (0.108)
Instrumental ADL limitations 0.022 (0.031) 0.094 (0.072)
Functional limitations 0.026 (0.052) -0.098 (0.092)
Intercept 2.406 (1.706) -2.934 (3.148)
Disability reporting mills 0.180 ** (0.107) -0.193 (0.227)
Employment mills 2.276 *** (0.643) 1.365 (1652)
Adjusted [R.sup.2] (n) 0.2936 (402) 0.3421 (136)

Samples restricted to individuals aged 25-62 years with wages.
Entries are coefficients from OLS regressions of log wages.

* Significant at 0. 10 level.

** Significant at 0.05 level.

*** Significant at 0.01 level.