Disability Reporting Choices by Married Couples: Evidence from Census Data.

Michael Zimmer [*]

The purpose of this paper is to analyze determinants of the decision to self-report health impairments that limit or prevent work. Its particular focus is on joint reporting behavior by married couples, with a view toward potential interdependence of spouses' reporting decisions. The model is based on a large sample of married couples taken from the 1990 Public Use Microdata Sample of the U.S. Census. Empirical analysis is based on a simultaneous probit model of spouses' reports. Results of the analysis show evidence of an association between spouses reports, after controlling for important background variables. The extent of association appears to be more pronounced among low earners and among older wives.

1. Introduction

During the past two decades, the United States has witnessed a marked increase in persons seeking and receiving public assistance for physical impairments and work disabilities. Income support programs remain a significant component of national policy toward the disabled. The Social Security Disability Insurance program (DI) provides benefits to disabled persons based on payroll tax contributions during their working years, and the Supplemental Security Income (SSI) program is means-tested and not linked to prior payroll contributions. By 1995, annual SSI payments had grown to $19.5 billion for 4.9 million recipients, and DI paid $40.0 billion to 4.2 million recipients (Rupp and Stapleton 1998, p. 2).

These programs, together with changes accompanying the Americans with Disabilities Act (ADA) in 1990, have engendered considerable research regarding their behavioral consequences within the disabled population. A principal item of interest has been consequences of public policy for supply of labor by disabled persons and demand for their services by employers (Burkhauser and Haveman 1982; Leonard 1991; Rosen 1991; Stern 1989). Other studies have investigated the effects of health impairments on workers' earnings (Baldwin, Zeager, and Flacco 1994; Kahn 1998).

The motivation for this paper is that most previous research has left unexplored the question of behavior regarding reporting of disabilities. The importance of this issue arises from the fact that self-reports of disabilities constitute the substance of available data on the impaired population. Agencies such as the Census Bureau and the Social Security Administration rely on individual or household data in which persons describe the existence and degree of their physical limitations. Moreover, claims to payments under DI or SSI are initiated by persons who self-report their impairments and then submit claims for clinical evaluation as the basis of application for benefits.

In view of the importance of self-reporting behavior, there appears to have been relatively little attention paid to its determinants. Exceptions include Chirikos and Nestel (1984) and Aarts and deJong (1992). The purpose of this paper is to analyze determinants of the decision to self-report physical impairments that limit or prevent work. Its particular focus is on reporting behavior of married couples, with a view toward the potential interdependence of spouses' reporting decisions. Interdependence is important for policy purposes, because it could produce a tendency toward clustering of reported disabilities within the married population. The model is based on a large sample of intact marriages taken from the 1990 Public Use Microdata Sample (PUMS) of the U.S. Census. Results of the model indicate a positive association between spouses' reports, after controlling for important background factors.

The remainder of the paper is organized as follows. Sections 2 and 3 provide a background of previous studies that motivate the work described here. Section 4 describes a statistical framework in which the potential for mutual dependence between spouses can be inferred. Section 5 describes the data and model specification while section 6 presents results of estimation. The final section summarizes principal findings.

2. Descriptive Evidence and Policy Implications

A substantial amount of our knowledge about the disabled population comes from information that individuals report about themselves to the Census Bureau and the Social Security Administration. Rosen (1991, pp. 19-20) reports that in both 1970 and 1980, approximately 13% of people aged 15 and older reported themselves as disabled. Census data contain some hints of matching behavior between spouses. The 1990 PUMS records used in this study, described in more detail in section 5, include responses to questions about health status for each individual and his or her spouse. Respondents indicated whether they had a physical condition that limited their work and whether they were prevented from working by those limitations. The responses are cross-tabulated for a large sample of husband-wife pairs in Table 1. Entries in the tables are sample proportions corresponding to each possible combination of spouses' responses.

The proportions are not remarkable at first glance, revealing that most couples report no disabilities and the least frequent occurrence is for both spouses reporting disabilities. However, the distribution is apparently not consistent with what would have been obtained if disability status were an independent phenomenon between spouses; in each case, chi-square statistics based on the absolute frequencies in each cell are sufficiently large to reject the null hypothesis of independence. Some perspective of the extent of interdependence can be gained by assessing conditional probabilities embodied in Table 1. The probability that a randomly selected husband reports a work-limiting disability is .13; if his wife reports a similar disability, however, his probability increases to .41. For wives, the probability of a work-limiting disability is .10, but it increases to .31 if the husband reports likewise. For work-preventing disabilities, the pattern is similar: The husband's probability is .08 but increases to .38 conditional on a disabled wife, and the wife's probability is .06 while increasing to .32 if the husband is disabled.

An understanding of spouses' joint reporting behaviors is important for future policy considerations. Previous studies have established that self-reports of health problems are associated with reductions in labor force participation (Bound and Waidman 1992) and that families' predominant strategies for coping with loss of income due to disability include application for income assistance (Johnson and Murphy 1975). The implication is that the same factors that lead families to report ill health or physical impairments induce a demand for disability benefits. Researchers have noted this connection in past studies of growth in the DI and SSI programs. Stapleton et al. (1988), analyzing the programs' growth from 1988 to 1992, conclude that expansion of the disability population arose from four causes. First, the recession of 1990-1991 expanded the ranks of unemployed, some of whom sought and received disability benefits. Second, reductions at the federal level in general welfare assistance indirectly caused indi viduals to seek alternative forms of income support, namely disability benefits. Third, an increase in the population of children from single-parent families lead to extension of benefits to disabled children. Finally, an increase in the supply of applicants occurred in response to expansion in the scope and generosity of benefits. Bound and Waidman (1992) corroborate the latter point, citing policy changes that more readily accommodated disability applicants.

Another factor in the programs' growth, which has not received attention in previous research, is the potential clustering of reported disability within families. As Yelowitz (1998, p. 126) suggests, once a household enrolls one member in an income support program, its propensity to enroll others might increase. Some researchers have noted that spouses of disabled persons have a higher tendency to report themselves as disabled. Johnson and Murphy (1975, pp. 89-90) observed such a tendency in the 1966 Social Security Survey of the Disabled. Kennedy, Walls, and Owens-Nicholson (1999, pp. 47-49), using the 1992-93 Surveys of Income and Program Participation, find that spouses of disabled persons are the primary source of caregiving assistance, and caregivers themselves tend to self-report higher rates of disability. LaPlante et al. (1996), using data from the 1990 National Health Interview Survey, report a result similar to that in Table 1: The number of households in which both partners report disabilities exc eeds what one would expect by chance if partners' disabilities were not correlated.

These papers present the spousal connection only in a descriptive sense and do not provide a structural explanation. The purpose of this paper is to determine whether an association exists between spouses' self-reports after controlling for important background variables and whether the clustering phenomenon appears to be uniform across broadly defined age and income groups. The descriptive sample proportions in Table 1 are not useful for that purpose, because they do not control for background variables. Consequently, they are not informative about spouses' interdependence in reporting behavior per se. For example, it is well known that individuals tend to marry persons similar in age to themselves, and physical impairments are more likely to occur as individuals (hence, marriages) advance in years. Sociologists have also noted a tendency for positive assortative mating with respect to education (Mare 1991), which in turn appears to be related to the incidence of reported disabilities (Berkowitz and Hill 19 89). Consequently, observed patterns of assortative mating might be an artifact of association in age and education that characterizes most marriages. The critical question is whether spouses' reporting behaviors remain dependent in the context of a model that includes suitable controls.

3. Determinants of Self-Reported Disability

There are behavioral and epidemiological reasons, as well as economic incentives, for couples' reporting behavior. Spouses share many elements of a common lifestyle, including diet, recreation habits, and location of residence, and they are likely to seek medical care from similar providers (Umberson 1987; Smith and Zick 1994). Over the life cycle, they are likely to develop similar health traits. During the past two decades, spouses have nurtured those mutual traits under economic and policy conditions that were conducive to higher rates of disability. Bound and Waidman (1992) point out several contributing factors. There has been a tendency among clinical evaluators toward earlier diagnosis of preexisting health impairments, coupled with higher survival rates from chronic diseases. In addition, there has been a persistent decline in the proportion of the population institutionalized for mental impairments. Together with a general increase in public access to disability benefits, these factors have combined to produce an increase in self-reported disabilities. In addition, Koitz, Kollman, and Neisner (1992) argue that medical advances have produced a decline in the proportion of benefit recipients whose payments cease due to death. Thus, the past two decades were a period in which spouses' joint health behaviors were realized in the context of a general rise in disability rates. Although researchers have used data from that period to study the shared mortality experiences of married couples (Smith and Zick 1994), they have devoted only limited attention to spouses' disability reporting behavior. Chirikos and Nestel (1984) used National Longitudinal Survey data to estimate probit equations for individuals' self-reports of disability. Estimating separate models for the sample partitioned according to race and gender, their principal finding is that people with higher expected wage rates are less likely to report disabilities, after controlling for job requirements and other family income. Aarts and deJong (1992, chapter 6) used data from the Dutch Disability Insurance Program to estimate probit models of individuals' entry into the program. To gain entry, individuals must first claim a right to eligibility, which entails a self-report of disability. Aarts and deJong's results indicate that claimants tend to be older and less educated and have previous records of unemployment and absenteeism.

Economic determinants of spouses' joint reporting decisions can be seen in the context of Benham's (1973) treatment of human capital within marriage. In this formulation, a husband-wife pair constitutes a decision-making unit that employs its collective human capital in the labor market to generate family earnings. Benham's emphasis is on interdependence between spouses' human capital. Each partner's human capital, consisting of formal training, work experience, job skills, and physical health, is complemented by the human capital of his spouse. In this way, spouses act constructively to allocate the family's human capital to maximize household utility. Benham hypothesizes that the contractual nature of marriage facilitates efficient sharing of information between spouses. The advantage of marriage is that it reduces costs of communication and transaction costs of implicit contracting between partners. Spouses possess opportunities and incentives for learning about available income support benefits and rehab ilitation opportunities available under federal and state programs, and they can educate each other regarding legally mandated accommodations. An example of this is found in the eligibility rules for the DI program. The rules state that the disabled survivor of a deceased spouse can remain eligible for continuation of benefits. To qualify, the survivor's disability must have commenced before or within seven years after the death of the first spouse. [1] Though the rules allow for declarations of disability after the spouse's death, the option to do so earlier might induce some couples to report jointly beforehand. Thus, their claims for benefits are predicated on their willingness to declare themselves disabled. Conversely, as they learn that they are not eligible for benefits, they are prone to report themselves as free of impairments.

Finally, it is possible that the sorting process of marriage matches individuals with similar health characteristics. Consequently, their disability reports subsequent to marriage might reflect permanent health factors that existed before marriage. [2] The critical point is that the decision to report a disability is likely to be the outcome of shared lifestyles and health traits, choices regarding labor supply, and informed attempts to claim legal entitlements and benefits. As Leonard (1989) points out,

The crux of the matter is that disability is not simply a medically defined condition, but depends rather on an as-ray of psychological, sociological, and economic factors. A person who perceives himself or herself as disabled may thereby become disabled. And a person who finds greater economic returns to disability than to work may not struggle so hard to work. This need not be a question of fraud or dissembling, but merely of adapting to the given incentives. (p. 64)

The principle implication of these ideas is that spouses have considerable potential to affect their partners' reported disability status. The approach outlined in the next section describes a model with potential interactions between spouses' reporting decisions.

4. Framework of the Model

The sample data used in this study, described in section 5, contain individual records on a large sample of intact marriages. For each couple, define [[d.sup.*].sub.hi] as the latent propensity of the husband to report himself as disabled. [3] If [[d.sup.*].sub.hi] is positive, he reports himself as disabled; otherwise, he reports no health problems. The latent variable is not observed; instead, we observe a dichotomous indicator of the husband's disability status:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

In a similar manner, define as the wife's latent propensity to report herself as disabled. We observe her behavior only as a dichotomous indicator:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

The latent index is a function of background characteristics and the spouse's latent index:

[[d.sup.*].sub.hi] = [Z.sub.hi][[alpha].sub.1] + [[delta].sub.1] [[d.sup.*].sub.wi] - [[epsilon].sub.1i] (3)

[[d.sup.*].sub.wi] = [Z.sub.wi][[alpha].sub.2] + [[delta].sub.2] [[d.sup.*].sub.hi] - [[epsilon].sub.2i], (4)

where [Z.sub.hi] and [Z.sub.wi] are vectors of observed background characteristics and the [epsilon]'s are normally distributed error terms with zero means and unit variances. The parameters [[alpha].sub.1], [[alpha].sub.2], [[delta].sub.1], and [[delta].sub.2] can be estimated from a suitable sample of husband-wife pairs. The parameters [[delta].sub.1] and [[delta].sub.2] are of particular interest in this paper, because they capture the extent of association between spouses' reporting propensities. If [[delta].sub.1][greater than]0, husbands of wives who are strongly prone to report themselves as disabled are similarly inclined, after controlling for relevant background factors. An analogous interpretation holds for [[delta].sub.2]. Such a finding would support Benham's (1973) theory of information sharing within marriage or hypotheses concerning shared lifestyles (Umberson 1987; Smith and Zick 1994). Alternatively, it supports a hypothesis of marital matching on the basis of health. These explanations are not mutually exclusive.

As Equations 3 and 4 are formulated, they are amenable to estimation by means of maximum likelihood probit methods. The complication that arises is the presence of the endogenous variables [[d.sup.*].sub.wi] and [[d.sup.*].sub.hi] the right-hand sides. Consequently, conventional probit estimates of Equations 3 and 4 will produce biased estimates of [[delta].sub.1] and [[delta].sub.2]. A variant of this model is discussed by Mallar (1977) and Maddala (1983). Mallar suggests a simple two-stage procedure for estimating the structural parameters. Substituting for the latent variables in Equations 3 and 4 yields the corresponding reduced forms:

[[d.sup.*].sub.hi] = [X.sub.i][[pi].sub.1] - [[eta].sub.1i] (5)

[[d.sup.*].sub.wi] = [X.sub.i][[pi].sub.2] - [[eta].sub.2i], (6)

where [X.sub.i], denotes a vector of all explanatory variables contained in [Z.sub.hi] or [z.sub.wi], the [pi]'s are conformable vectors of reduced form coefficients, and the [eta]'s are reduced-form error terms. Step one involves probit estimation of the reduced form to obtain estimates of [[pi].sub.1] and [[pi].sub.2]. The estimated coefficients are used to obtain fitted values of the latent indexes for each spouse pair:

[[d.sup.*].sub.hi] = [X.sub.i][[pi].sub.1] (7)

[[d.sup.*].sub.wi] = [X.sub.i][[pi].sub.2]. (8)

The fitted values can be used as instruments for the latent indexes in the structural form. In step two, the structural form is estimated by probit methods:

P([d.sub.hi] = 1) [phi]([Z.sub.hi][[alpha].sub.1] + [[delta].sub.1] [[d.sup.*].sub.wi]) (3')

P([d.sub.wi] = 1) [phi]([Z.sub.wi][[alpha].sub.2] + [[delta].sub.2] [[d.sup.*].sub.hi]), (4')

where [phi] denotes the cumulative distribution function associated with a standard normal distribution. Because the model includes controls for factors that might correlate spouses' reports, the structural coefficients [[delta].sub.1] and [[delta].sub.2] capture a direct cross-effect between spouses' reporting decisions.

The use of fitted values in the second stage complicates the covariance matrix of the structural coefficients. Maddala (1983, pp. 246-7) derives the correct covariance matrix. In section 6, we present estimates of Equations 3' and 4' that incorporate the corrected covariance matrix, thus preserving the consistency and efficiency of the estimated structural coefficients. [4]

5. Data and Model Specification

The sample consists of 30,567 married couples extracted from the 1990 PUMS of the U.S. Census. Of the total, 21,396 are white couples while the remainder are black couples and couples of mixed race. [5] As part of the information provided in each record, spouses respond to separate questions of whether they are impaired by a physical or medical condition that limits their work or prevents them from working altogether. In addition, each record contains detailed data regarding each spouse's age, education, and other background factors.

The 1990 PUMS is well situated for studies related to issues of federal disability policy, because it is centered in a period of rapid growth in assistance programs. Between 1988 and 1992, applications for DI and SSI increased 8.9% and 10.5% annually while the dollar volume of awards increased 10% and 12% annually (Stapleton et al. 1998, p. 31). Another advantage is that the PUMS file permits collection of a large sample of married couples across a substantial range of education, age, and earnings. The data have their limitations, however, because they record only receipts of total Social Security income, which are not disaggregated by program. In addition, the PUMS does not measure individual job tenure or general work experience. Thus, although we can focus clearly on the existence of dependence among spouses' behaviors, it is difficult to make clear distinctions among the competing hypotheses that explain those choices.

Most empirical studies of disability rely on self-reported measures of health or impairment. In the context of this study, that is a suitable approach, because our objective is to model the joint decision to report. It is worth noting, however, that joint reports might be mismeasurements of spouses' true disabilities. Use of self-reports is defended by Daly and Bound (1996) and Mjelde-Massey and Barak (1998) on grounds that they have been shown to be highly correlated with both objective measures of health and records of chronic conditions or functional limitations. With respect to the policy issue of disability benefits, it can be argued that benefits are the result of objective clinical evaluation and not merely self-reported health status. Again, however, self-reports are statistically informative if they adequately distinguish between persons with and without health impairments, as the evidence seems to indicate. Further evidence on this issue, in a study of the Netherlands, is reported by deJong (1983). In the Dutch Disability Insurance Program he found that, controlling for objective disability, self-reported work capacity plays a dominant role in determination of eligibility. Although conclusions about the Dutch system cannot be safely generalized, the critical mass of evidence indicates that self-reports are adequate and useful in their own right for policy analysis.

The model described in section 4 establishes a framework for testing the crucial independence hypothesis. The model is completed by specifying variables for inclusion in Equations 3 and 4. Consistent with Chirikos and Nestel (1984) and Aarts and deJong (1992), we control for variables that influence earnings and tastes for wealth. In addition, age is a critical control, because spouses' health circumstances change as they grow older. Consequently, the wives' equation includes the wife's age and years of education, an interaction term between age and education, a dummy variable equal to one if the wife is white, number of children in the household below the age of 18, a dummy variable equal to one if the couple owns a home free of a mortgage, and a dummy variable equal to one if they own a home encumbered by a mortgage. The omitted category includes couples who do not reside in owner-occupied housing. The owner occupancy indicators are intended to control for family wealth and tastes for ownership of assets. The husbands' equation omits the number of children but includes a dummy variable for military veteran's status.

Table 2 contains descriptive statistics and definitions for variables in the model. Restricting the sample to intact marriages results in average ages in the mid-40s. Spouses tend to have formal schooling slightly in excess of high school completion. Approximately three-fourths of the sample consists of home owners, either outright or through mortgages. With respect to disability status, the sample means indicate that 13% of husbands and 10% of wives report work-limiting disabilities while 8% of husbands and 6% of wives report being prevented from working due to disabilities. Estimates from other studies show some variation around these levels, depending on the population and the nature of the sample. Burkhauser and Haveman (1982) cite 1978 data from the Social Security Administration on disabilities in the U.S. population aged 18 to 64. Those with severe disabilities constitute 8.5% while those with occupational disabilities make up 3.7%. An additional 4.6% report secondary limitations. Oi (1991) reports th at between 1985 and 1987, work-limiting disabilities were reported by 8.1-9.4% of working age adults; the proportion reporting work-preventing disabilities ranged from 4.4% to 5.0%. Berkowitz and Hill (1989), using data from the 1986 Current Population Survey, estimate that the proportion of disabled varies by age and years of education, and to a lesser extent, by gender. Among males, their estimates range from 3.5% for ages 25-34 with 13 or more years education to 26.6% for ages 45-54 with 8 or fewer years of education. Among females aged 25-34 with 13 or more years of schooling, 3.3% are disabled, compared to 24.6% in the 45-54 age-group with 8 or fewer years of schooling.

It is unlikely that the sample in this study will produce disability numbers that mimic those estimates, due to the restriction of marital status. First, as seen in Table 2, using intact marriages tends to produce an older than average sample. In addition, marital status itself might interact with disabilities. Franklin (1977) reported that disabled persons do not show as great a degree of stability in their marriages. Thus, the focus on married couples, necessitated by the purpose of this study, might affect the disability character of the sample.

6. Results of Estimation

This section presents estimates for two specifications of the model. The basic version includes personal background and demographic variables but excludes individual wages. The second version appends an estimate of each spouse's hourly wage to his or her equation.

The Basic Specification

Estimates of Equations 3' and 4' are presented separately for wives and husbands in Tables 3 and 4. Both include fitted values from the spouses' reduced-form equations, which are obtained in step one in accordance with Equations 7 and 8. The tables present separate results based on models for work-limiting and work-preventing disabilities.

The wives' model shown in Table 3 reveals a number of significant coefficients associated with the control variables. Wives' tendencies to report both types of disabilities tend to increase with age and decrease with education, as expected. The model includes an interaction term between age and education. Its positive sign can be interpreted to mean that among women of a given level of education, older women are more prone to report disabilities. The interaction coefficient is significant in the case of work preventing disabilities. The estimates also reveal that disability reporting is less likely among women with more children and among whites, after controlling for the other factors in the model. The coefficient on children might reflect the fact that bearing or raising children is a natural result of stronger health attributes, or it might reflect family responsibilities associated with children that have an indirect effect on reporting behavior. The home ownership effects are negative in both models. Th is might reflect the fact that maintaining a home is associated with favorable health attributes, or it might reflect an underlying taste for wealth and assets that exerts an indirect effect on reporting behavior. The chi-square statistics test the null hypotheses that the explanatory variables, taken as a group, do not contribute significant explanation to the observed pattern of reported disability. Their large values easily reject the null hypothesis in each case.

The most noteworthy estimates are for the fitted spouse disability variables, which are both positive and highly significant. They indicate that after controlling for age, education, and other characteristics, wives whose husbands report disabilities are themselves more prone to self-report the same status. The magnitude of the coefficients is similar for both work-limiting and work-prohibiting disabilities. The estimates measure the direction but not the magnitude of the mutual spouse effects. The partial effect of variable [x.sub.j] is obtained by [d.sub.j][[phi](P)], where [d.sub.j] is the probit estimate of the coefficient of [x.sub.j], P is the fitted value of the probit equation for a wife with average sample characteristics, and [phi] denotes the density function of a standard normal distribution. Carrying out this computation for the coefficients in Table 3 produces a partial effect of husband's status equal to 0.047 in the work-limiting equation and 0.025 in the work-preventing equation. After contr olling for factors that would naturally correlate spouses' disability states, such as age and education, wives of husbands who report work-limiting disabilities are themselves 4.7 percentage points more likely to report the same state. In the case of husbands' work-preventing disabilities, the marginal effect on wives is 2.5 percentage points.

Results for husbands' self-reports, shown in Table 4, show estimates for the control variables that are similar in many respects to the wives' estimates. One difference is that the husbands' model excludes children but includes a dummy variable to capture military veteran status. Its coefficient is significant and positive in the work-limiting equation but not significant in the work-preventing equation. Another difference is that there appears to be no significant race effect in the husbands' equations.

As was the case in the wives' equations, husbands tend to self-report disabilities similar to their spouses; the coefficients of the wife's disability status are positive and significant in both equations. Their magnitudes are smaller than their counterparts in the wives' equations, and that is reflected somewhat in the estimated partial effects. Husbands whose wives report limited disabilities are 3.5 percentage points more likely to report the same, and those whose wives report serious disabilities are 2.2 percentage points more likely to do likewise. Again, these effects are residual in the sense that the model controls for other factors that would tend to reflect both spouses' health status.

It is useful to ask whether the strength of the mutual self-reporting effects is uniform across broad categories of income. To address that question, I restricted the sample to families with incomes of less than $15,000 and reestimated the probit models. The resulting sample contains 3711 couples. Estimates of the structural cross-spouse effects, [[delta].sub.1] and [[delta].sub.2] in Equations 3' and 4', are presented in column one of Table 5. Each set of entries shows the estimated coefficient, the corrected t statistic in parentheses, and the corresponding partial effect in brackets associated with each spouse's disability status. Estimates in the first column reveal that the spousal interactions in the low income sample are substantially more pronounced than their counterparts for the entire sample (Tables 3 and 4). In terms of work-limiting disabilities, husbands whose wives report disabilities are 7.9 percentage points more likely to report themselves as disabled. That compares to a partial effect of 3 .5 percentage points for the entire sample. For wives, the result is similar: Wives of work-limited husbands are 11.2 percentage points more likely to report disabilities, in contrast to an effect of 4.7 percentage points in the entire sample. A similar pattern emerges in the case of work-preventing disabilities. The mutual spouse effects are substantially larger among low wage families.

These disparities across income groups might reflect a labor supply effect, wherein couples with low income tend to view disability benefits as attractive replacements to wage income from jobs. Consequently, when one spouse reports a disability, the other is prone to follow suit. While it might be argued that the spouse would instead respond with greater labor force participation in order to replace the family's lost earnings, previous research has not found that to be the case (Johnson and Murphy 1975). An alternative explanation relies on the effect of shared lifestyles. Low income couples tend to share health disadvantages (poor diet) and less access to quality medical care (lack of health insurance). Although the evidence in Table 5 does not distinguish between the two possibilities, there is clear evidence of an income effect in spouses' reporting behaviors.

The selection hypothesis suggests that the pattern of association derives from conditions that existed before the marriage: Individuals tend to marry persons with similar health traits, which tends as a matter of course to produce similar disability reports after marriage. Although the PUMS data are not sufficiently detailed to permit a direct test, we can restrict the sample to obtain indirect evidence. If the selection hypothesis is valid, we should be able to measure relatively strong partial effects between spouses' reports in the early stages of marriage. To test this, we restricted the sample to couples in which both spouses were younger than 30 years of age, with the intention of capturing relatively young marriages. The resulting sample contains 3077 couples. Estimates of the structural coefficients and partial effects are presented in column two of Table 5. The results indicate substantially weaker interactions between young spouses. The coefficient differs significantly from zero in only one case, the effect of wives' work-limiting disabilities on their husbands' reports. Even there, the estimated partial effect (2.5 percentage points) is smaller than its counterpart in the entire sample (3.5 percentage points).

Another useful partition of the sample is based on couples' proximities to the age of retirement. [6] To determine whether inclusion of older couples alters the principal conclusions, Table 5 presents results for the subsample of working age couples (at least one spouse younger than 65) and those of retirement age (both spouses 65 or older). For the working age sample (n = 26,960), the essential estimates, shown in column three, are consistent with their counterparts for the complete sample (Tables 3 and 4). In particular, the estimated partial effects for the complete sample are not substantially altered by the exclusion of older couples. For the retirement age sample (n = 3607), a different pattern emerges. In the husbands' equation, the lack of significance in the parameter [[delta].sub.1] suggests only weak dependence of husbands' reports on the status of their wives. In contrast, wives tend to emulate their husbands' reports. The estimated partial effects, 0.14 for work-limiting disabilities and 0.11 fo r work-preventing disabilities, are among the largest in any of the sample partitions.

To test these hypotheses in a more direct fashion, we would need data of richer detail than seem to be currently available. Within the constraints of the data available here, the indirect evidence suggests that couples with low earnings prospects exhibit a relatively strong pattern of mutually reinforcing behavior in their disability reports while recently married couples show negligible patterns.

The Specification Including Individual Wages

Estimates in Tables 3 and 4 are based on a specification that ignores wages of each spouse. Although the equations include such variables as age and education, which tend to be correlated with wages, those variables are intended to measure direct impacts of schooling and age on disability status. Inclusion of wage rates serves the purpose of a control that isolates the schooling and age effects. More important, it permits a test of whether the spouse coefficients remain significant when the model directly controls for earnings.

Because spouses' wages are endogenous in a probit model of disability status, I constructed an estimated wage for all spouses based on their background data. The sample data do not contain wage information on individuals who were not employed at the time of the census survey. Consequently, I based the earnings regressions on data for workers only. To correct for the potential problem of selection bias, I used the standard Heckman (1979) two-stage method to derive fitted wages (measured as the logarithm of hourly earnings) for each individual. [7]

Estimates of selected parameters in the simultaneous probability model that includes spouses' estimated wages are presented in columns five and six of Table 5. Because of the inclusion of potential wages, the sample is restricted to marriages in which at least one spouse is younger than 65 years. For the sake of brevity, estimates of the other coefficients are not included, because they are generally consistent with their counterparts in Tables 3 and 4. The estimates indicate that the spouse coefficients, [[delta].sub.1], and [[delta].sub.2], are similar in sign and significance to their counterparts in the basic model (Tables 3 and 4). The essential conclusion of the analysis is robust with respect to the inclusion of estimated wages. After controlling for background factors, including earnings capacity, there remains a significant positive association between spouses' reported states of disability, although the magnitudes are affected in two cases. The coefficient is larger in the wives' work-limiting equa tion and smaller in the husbands' work-preventing equation. Regarding the wage coefficients, the estimates indicate that high-wage husbands are less likely to report work-limiting disabilities; the coefficient in the work-preventing equation is not significant. For wives, the result is somewhat surprising. Holding other factors constant, including husbands' disability status and their own education and age, wives with high-estimated wages are more prone to report themselves as disabled.

Results from the wage-augmented equation serve to establish the robustness of the structural cross-spouse coefficients. However, that model suffers two shortcomings. First, the standard errors are not corrected to account for the inclusion of estimated wages in the structural equation. Second, the model includes the potentially endogenous dummy variable for work-limiting disability in both the probit and log wage equations. This was necessary to specify the equations in a manner consistent with common practice in the literature. [8] For these reasons, the basic results in Tables 3 and 4 rather than the latter model appear to present the soundest evidence to support the principal conclusion of the study.

These findings have potential policy implications. Recent changes in the direction of policy in the United States have the effect of expanding the population of individuals who might report themselves as disabled. Evidence in this paper suggests that there might be a clustering effect of this phenomenon within the married population, particularly for low-wage couples. Research by LaPlante et al. (1996) reveals an implication for the SSI and DI programs. Their evidence indicates a similar clustering within the disabled population with respect to participation in income support programs. Looking at 1990 data on a sample of households with no members older than the age of 65, they found that the general rates of participation in DI and SSI are 2.9% and 2.0%, respectively. Among the subset in which only one partner reports a disability, those rates increase to 9.5% and 2.7%. When both partners (and no other members) report disabilities, participation in the income support programs escalates to 23.6% and 9.5%, re spectively. Thus, self-reporting behavior appears to be strongly associated with participation in Social Security benefit programs, with relatively strong concentrations among dual-disabled couples. To understand future growth in Social Security's disability obligations, one must recognize that a statistically significant clustering of reported disabilities is likely to occur in the married population. Moreover, passage of legislation such as the Family Leave Act creates additional opportunities for spouses to alter their behaviors in response to their partners' physical and health problems. The consequences of these changes for disability policy in the public sector and employment policy in the private sector are substantial, and they invite further research.

7. Conclusion

Married couples show a marked tendency to report similar states of disability. Based on the data used in this study, only 15% of couples respond differently to census inquiries about impairments that limit their work; only 9% respond differently when asked if they are unable to work for reasons of health. In some respects, this is not surprising. Married persons tend to be similar in age, so a cross-section of couples is likely to produce a pattern of similar health outcomes. They tend to be similar in years of schooling as well, which has been shown by other researchers to correlate with disability status. Consequently, simple contingency tables of disability status between spouses will tend to deviate from patterns that are purely random. What is more interesting is the model in which spousal associations remain after controlling for important background factors.

This paper proposes and estimates a model that serves that purpose. Its principal conclusion is that there appears to be a significant propensity among married couples for mutually reinforcing behavior in self-reported disability. The mutual effects appear to be more pronounced among couples with low incomes and among older wives. They are not prevalent among young couples. The model is not capable of directly identifying the source of association. Married couples tend to adopt similar lifestyles and habits, so it is natural to expect their health fortunes to converge. Other explanations have less to do with lifestyle attributes than with market incentives and labor supply behavior. In a public policy environment that includes income support programs (DI, SSI) and mandated workplace accommodation (ADA), couples might make self-reports consistent with their joint preferences for leisure versus labor income. Marriage appears to be an institution conducive to efficient sharing of information regarding lifestyle s and labor markets in general, and disability reporting in particular. As the baby boom cohort's advancing average age pushes the married population toward potential medical impairments, this issue looms as a research item of substantial importance.

(*.) School of Business Administration, University of Evansville, 1800 Lincoln Avenue, Evansville, IN 47722, USA; E-mail [e-mail].

The author wishes to thank Kathy Hayes of this journal and anonymous referees for prompt and careful reviews. An earlier version of this paper was presented at the 1999 conference of the Southern Economic Association. The author thanks Angie Ritzert and Noel Campbell for helpful comments.

(1.) See U.S. Social Security Administration (1999) publication no. 05-10029.

(2.) A related issue is the potential selection of healthy individuals in general into marriage, without reference to partners' matching. Past research, however, appears not to support this proposition. For a brief discussion and related references, see Smith and, Zick (1994).

(3.) In the empirical work, I distinguish between work-limiting and work-prohibiting disabilities. Because both approaches use the same statistical framework, I do not distinguish them here.

(4.) The corrected variance covariance matrix can be readily incorporated in the LIMDEP package. For more details, see Greene (1995, p. 664).

(5.) One of the original purposes of this study was to determine whether disability reporting behavior differs between same-race and mixed-race marriages. That was motivated by interesting findings in Macpherson and Stewart (1992), who report distinctive patterns of labor force participation in mixed-race marriages. To obtain a suitable number of mixed-race couples, I oversampled for mixed races and weighted the resulting estimates to correct for the oversampling. Preliminary estimates of the model showed no distinctive behaviors based on the race combination of spouses. Consequently, that issue is not emphasized in section 6, which presents weighted estimates from the pooled sample.

(6.) I wish to thank an anonymous referee for suggesting this point.

(7.) Specifications for the probit and log wage equations are similar to numerous studies of earnings. The probit employment equation for each spouse includes his age and its square, two dummy variables indicating the presence of work-limiting disability for his spouse and himself, a dummy variable indicating status as a military veteran, dummy variables to indicate home ownership, either outright or through a mortgage, and number of children under 18 years of age. The log earnings equation includes age and its square, years of completed education, a dummy variable for military veteran status, a dummy variable for the presence of work-limiting disability, five dummy variables for occupation, and eleven dummy variables for industry of employment. The dependent variable is the logarithm of hourly wages, in which the latter is measured as the ratio of annual earnings to the product of usual weekly hours and annual weeks worked. Results of the probit and wage regressions are consistent with numerous other studies of earnings based on cross-sectional census data. For the sake of brevity, they are not reported here.

(8.) I estimated a version of the wage model in which disability status was removed from the probit and log wage equations. The results were similar to those in Table 5.

References

Aarts, Leo, and Philip deJong. 1992. Economic aspects of disability behavior. Amsterdam: North Holland Publishing Co.

Baldwin, Marjorie, Lester Zeager, and Paul Flacco. 1994. Gender differences in wage losses from impairments. Journal of Human Resources 29:865-87.

Benham, Lee. 1973. Benefits of women's education within marriage. In Economics of the family: Marriage, children, and human capital, edited by Theodore Schultz. Chicago: The University of Chicago Press, pp. 375-93.

Berkowitz, Monroe, and M. Anne Hill. 1989. Overview. In Disability and the labor market: Economic problems, policies and programs, edited by Monroe Berkowitz and M. Anne Hill. Ithaca, NY: ILR Press, pp. 1-28.

Bound, John, and Timothy Waidman. 1992. Disability transfers, self-reported health, and the labor force attachment of older men: Evidence from the historical record. The Quarterly Journal of Economics 107:1393-419.

Burkhauser, Richard, and Robert Haveman. 1982. Disability and work: The economics of American policy. Baltimore: The Johns Hopkins University Press.

Chirikos, Thomas, and Gilbert Nestel. 1984. Economic determinants and consequences of self reported work disability. Journal of Health Economics 3:117-36.

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

deJong, Philip. 1983. Self-perceived disability and eligibility for disability benefits. Mens en Maatschappij 58:325-59.

Franklin, Paula. 1977. Impact of disability on family structure. Social Security Bulletin 40:3-18.

Greene, William. 1995. LIMDEP Version 7.0 user's manual. Plainview, NY: Econometric Software, Inc.

Heckman, James. 1979. Sample selection as a specification error. Econometrica 47:153-61.

Johnson, William, and Edward Murphy. 1975. The response of low income households to income losses from disability. Industrial and Labor Relations Review 75:85-96.

Kahn, Matthew. 1998. Health and labor market performance: The case of diabetes. Journal of Labor Economics 16: 878-99.

Koitz, David, Geoffrey Kollman, and Jennifer Neisner. 1992. Status of the disability programs of the Social Security Administration. Washington, DC: Congressional Research Service.

Kennedy, Jae, Craig Walls, and Dawn Owens-Nicholson. 1999. A national profile of primary and secondary household caregivers: Estimates from the 1992 and 1993 surveys of income and program participation. Home Health Care Services Quarterly 17:39-58.

LaPlante, Mitchell, Dawn Carlson, H. Stephen Kaye, and Julia Bradsher. 1996. Families with disabilities in the United States. In Disability Statistics Report (8). Washington, DC: U.S. Department of Education, National Institute on Disability and Rehabilitation Research.

Leonard, Jonathan. 1989. Labor supply incentives and disincentives for disabled persons. In Disability and the labor market, economic problems, policies and programs, edited by Monroe Berkowitz and M. Anne Hill. Ithaca, NY: ILR Press, pp. 64-94.

Leonard, Jonathan. 1991. Disability policy and the return to work. In Disability and work, edited by Carolyn Weaver. Washington: The AEI Press, pp. 46-58.

Macpherson, David, and James Stewart. 1992. Racial differences in married female labor force participation behavior: An analysis using interracial marriages. Review of Black Political Economy 21(1):59-68.

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

Mallar, Charles. 1977. The estimation of simultaneous probability models. Econometrica 58:173-88.

Mare, Robert. 1991. Five decades of educational assortative mating. American Sociological Review 56:15-32.

Mjelde-Massey, Lee Ann, and Michal Mor Barak. 1998. The conceptual and empirical link between health behaviors, self reported health, and the use of health care in later life. Home Health Care Services Quarterly 17:71-89.

Oi, Walter. 1991. Disability and a workfare-welfare dilemma. In Disability and work, edited by Carolyn Weaver. Washington: The AEI Press, pp. 31-45.

Rosen, Sherwin. 1991. Disability accommodation and the labor market. In Disability and work, edited by Carolyn Weaver. Washington: The AEI Press, pp. 18-30.

Rupp, Kalman, and David Stapleton. 1998. Introduction. In Growth in disability benefits: Explanations and policy implications, edited by Kalman Rupp and David Stapleton. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research, pp. 1-27.

Smith, Ken, and Cathleen Zick. 1994. Linked lives, dependent demise? Survival analysis of husbands and wives. Demography 31:81-93.

Stapleton, David, Kevin Coleman, Kimberly Dietrich, and Gina Livermore. 1998. Empirical analysis of DI and SSI application and award growth. In Growth in disability benefits: Explanations and policy implications, edited by Kalman Rupp and David Stapleton. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research, pp. 31-83.

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

Umberson, Debra. 1987. Family status and health behaviors: Social control as a dimension of social integration. Journal of Health and Social Behavior 28:306-19.

U.S. Social Security Administration. 1999. Disability insurance, no. 05-10029. Washington, DC.

Yelowitz, Aaron. 1998. The impact of health care costs and medicaid on SSI participation. In Growth in disability benefits: Explanations and policy implications, edited by Kalman Rupp and David Stapleton. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research, pp. 109-33.

                  Sample Proportions of Self-Reported
                     Disability: Husband-Wife Pairs
Husband Disabled     Wife Disabled  Disability Limits Work
Yes                       Yes               0.040
Yes                       No                0.090
No                        Yes               0.057
No                        No                0.813
chi-square (1 d.f.)                      2330.42
Husband Disabled     Disability Prevents Work
Yes                           0.024
Yes                           0.052
No                            0.040
No                            0.884
chi-square (1 d.f.)        2705.99
Source: 1990 Public Use Microdata
Sample (n = 30,567). Chi-square tests
the null hypothesis that husbands'
and wives reported disability states
are distributed independently.
                         Descriptive Statistics
                                                         Sample
Variable                                                  Mean
WIFE DISAB1    = 1 if wife's disability limits work;          0.098
               = 0 otherwise.
WIFE DISAB2    = 1 if wife's disability precludes work;       0.064
               = 0 otherwise.
HUSB DISAB1    = 1 if husband's disability limits work;       0.130
               = 0 otherwise.
HUSB DISAB2    = 1 if husband's disability precludes          0.076
                 work;
               = 0 otherwise.
HUSB AGE       = husband's age: years.                       47.188
WIFE AGE       = wife's age: years.                          44.356
HUSB EDUC      = husband's years of schooling                12.979
                 completed.
WIFE EDUC      = wife's years of schooling completed.        12.967
HOME OWN       = 1 if family owns its home;                   0.237
               = 0 otherwise.
MORTGAGE       = 1 if family owns home with mortgage;         0.514
               = 0 otherwise.
HUSBAND WHITE  = 1 if husband is white;                       0.767
               = 0 otherwise.
WIFE WHITE     = 1 if wife is white;                          0.891
               = 0 otherwise.
CHILDREN       = number of children 18 years or               0.957
                 younger.
HUSB VETERAN   = 1 if husband is a military veteran;          0.448
               = 0 otherwise.
Sample Size                                              30,567
                Sample
               Standard
Variable       Deviation
WIFE DISAB1      0.297
WIFE DISAB2      0.244
HUSB DISAB1      0.336
HUSB DISAB2      0.265
HUSB AGE        15.537
WIFE AGE        15.017
HUSB EDUC        2.725
WIFE EDUC        2.307
HOME OWN         0.425
MORTGAGE         0.500
HUSBAND WHITE    0.423
WIFE WHITE       0.312
CHILDREN         1.179
HUSB VETERAN     0.497
Sample Size
                        Simultaneous Probability
                            Model: Wives [a]
Variable            Disability Limits Work
CONSTANT                    -0.308          (1.57)
WIFE AGE                     0.012          (3.63) [*]
WIFE EDUC                   -0.081          (5.27) [*]
WIFE AGE*WIFE EDUC           0.001          (1.33)
CHILDREN                    -0.046          (3.85) [*]
WIFE WHITE                  -0.267          (8.22) [*]
HOME OWN                    -0.198          (6.00) [*]
MORTGAGE                    -0.172          (5.74) [*]
HUSB DISAB1                  0.341          (8.07) [*]
HUSB DISAB2
CHI SQUARE                2319.29           (8 d.f.)
Variable            Disability Prevents Work
CONSTANT                     -0.339           (1.50)
WIFE AGE                      0.013           (3.35) [*]
WIFE EDUC                    -0.118           (6.24) [*]
WIFE AGE*WIFE EDUC            0.001           (2.65) [*]
CHILDREN                     -0.026           (1.68)
WIFE WHITE                   -0.259           (6.65) [*]
HOME OWN                     -0.230           (6.09) [*]
MORTGAGE                     -0.212           (5.80) [*]
HUSB DISAB1
HUSB DISAB2                   0.320           (8.08) [*]
CHI SQUARE                 2211.97            (8 d.f.)

(a.)Figures in parentheses are absolute t statistics. The variables HUSB DISAB1 and HUSB DISAB2 are obtained from the first stage reduced form for husbands, Equation 7. Standard errors are corrected to account for the variance covariance matrix that results from the two-stage procedure.

(*.)Significant at the 0.05 level.

                        Simultaneous Probability
                          Model: Husbands [a]
Variable            Disability Limits Work
CONSTANT                    -0.707          (3.98) [*]
HUSB AGE                     0.015          (5.06) [*]
HUSB EDUC                   -0.105          (7.91) [*]
HUSB AGE*HUSB EDUC           0.001          (3.48) [*]
HUSB VETERAN                 0.041          (2.02) [*]
HUSB WHITE                   0.040          (1.56)
HOME OWN                    -0.153          (4.72) [*]
MORTGAGE                    -0.202          (6.94) [*]
WIFE DISAB1                  0.198          (4.21) [*]
WIFE DISAB2
CHI SQUARE                2816.43           (8 d.f.)
Variable            Disability Prevents Work
CONSTANT                     -0.454            (2.11) [*]
HUSB AGE                      0.010            (2.86) [*]
HUSB EDUC                    -0.173           (10.27) [*]
HUSB AGE*HUSB EDUC            0.002            (6.38) [*]
HUSB VETERAN                  0.007            (0.29)
HUSB WHITE                    0.047            (1.41)
HOME OWN                     -0.126            (3.29) [*]
MORTGAGE                     -0.197            (5.26) [*]
WIFE DISAB1
WIFE DISAB2                   0.256            (5.41) [*]
CHI SQUARE                 2792.09             (8 d.f.)

(a.)Figures in parentheses are absolute t statistics. The variables WIFE DISAB1 and WIFE DISAB2 are obtained from the first stage reduced-form probit for women, Equation 8. Standard errors are corrected to account for the variance covariance matrix that results from the two-stage procedure.

(*.)Significant at the 0.05 level.

                        Alternative Samples and
                           Specifications [a]
                              Low                      Working
Model                     Earners [b]   Young [c]      Age [d]
Disability limits work
[[delta].sub.1]              0.228         0.471           0.209
                            (2.36) [*]    (1.97) [*]      (4.01) [*]
                            [0.079]       [0.025]         [0.031]
[[delta].sub.2]              0.339        -0.040           0.336
                            (3.88) [*]    (0.15)          (7.41) [*]
                            [0.112]      [-0.002]         [0.038]
Husband's wage
Wife's wage
Disability prevents work
[[delta].sub.1]              0.319         0.043           0.313
                            (3.12) [*]    (0.10)          (5.84) [*]
                            [0.093]       [0.001]         [0.019]
[[delta].sub.2]              0.359        -0.334           0.316
                            (3.31) [*]    (1.20)          (7.24) [*]
                            [0.086]       [0.001]         [0.018]
Husband's wage
Wife's wage
Sample size               3711          3077          26,960
                          Retirement        Model with
Model                      Age [e]      Wages: Husbands [f]
Disability limits work
[[delta].sub.1]              0.182             0.266
                            (1.43)          (5.87) [*]
                            [0.066]           [0.038]
[[delta].sub.2]              0.451
                            (3.74) [*]
                            [0.144]
Husband's wage                          -0.562 (14.11) [*]
Wife's wage
Disability prevents work
[[delta].sub.1]              0.156             0.215
                            (1.27)          (5.85) [*]
                            [0.050]           [0.013]
[[delta].sub.2]              0.363
                            (3.63) [*]
                            [0.105]
Husband's wage                             0.065 (1.21)
Wife's wage
Sample size               3607            26,960
                             Model with
Model                     Wages: Wives [f]
Disability limits work
[[delta].sub.1]
[[delta].sub.2]                 0.341
                             (8.77) [*]
                               [0.038]
Husband's wage
Wife's wage               0.390 (8.03) [*]
Disability prevents work
[[delta].sub.1]
[[delta].sub.2]                 0.333
                             (7.56) [*]
                               [0.017]
Husband's wage
Wife's wage                1.009 (14.23) [*]
Sample size               26,960

(a.)Variance covariance matrix corrected to account for two-stage estimation. Figures in parentheses are absolute t statistics. Figures in brackets are estimated partial effects.

(b.)Couples with annual earnings below $15,000.

(c.)Both spouses younger than 30 years.

(d.)At least one spouse younger than 65 years.

(e.)Both spouses 65 years or older.

(f.)Restricted to the working age population.

(*.)Significant at the 0.05 level.