Survey Description and Sample Selection

The realized divorce decision may deviate from the expected decision due to the onset of a work-limiting disability. Without additional constraints, the effect of dis- ability on divorce is ambiguous because a disability decreases both the value of the current marriage and the value of outside alternatives. However, the framework of Becker, Landes, and Michael 1977 suggests that the effect of disability on divorce is nonnegative, as the decline in the value of marriage would exceed the decline in the value of outside alternatives. 2 The magnitude of the effect depends in part on the deviation between the expected and realized values of marriage. To characterize this deviation due to a disability, three assumptions are made: No one is initially disabled, the likelihood that person becomes disabled is zero, and the incidence of disability for person is . In this j i d i case, expected marital value is ∗ ∗ d ϕ 1,0 + 1−d ϕ 0,0; i i and if person becomes disabled, realized marital value is . Thus, the dif- ∗ i ϕ 1,0 ference between the expected and realized value is given by, ∗ ∗ 1−d [ ϕ 0,0−ϕ 1,0 ] . i This term is positive because and . According to Becker, ∗ ∗ 0 ≤ d ≤ 1 ϕ 0,0 ϕ 1,0 i Landes, and Michael 1977, as this term increases, so too does the effect of dis- ability on divorce. This term has two implications for the effect’s magnitude. First, the expected value of marriage reflects the likelihood of disability onset, so unanticipated disabilities, measured by the incidence of disability , have a greater effect on marital value than d i anticipated disabilities. Second, the effect of disability on divorce increases with dis- ability severity, measured by the change in marital value . Ac- ∗ ∗ [ ϕ 0,0−ϕ 1,0 ] cording to the equation above, disability incidence and disability severity interact, so disability’s greatest effect on divorce occurs when disability has a large effect on marital value and is unanticipated. While the effect of disability on divorce depends on the decline in marital value, the divorce decision ultimately depends on whether the decline leaves the value of marriage lower than the value of outside alternatives. Thus, given the same incidence and severity of disability, lower-quality marriages are more likely to dissolve than higher-quality marriages.

III. Data

A. Survey Description and Sample Selection

The empirical objective is to estimate the dynamic associations between disability, earnings and divorce. The data come from the Survey of Income and Program Par- 2. “A larger deviation between actual and expected values, such as actual and expected earnings or fe- cundity, raises the probability of dissolution. The reason is that the gain from becoming divorced from marrying someone else increases by more than the gain from remaining married to the same spouse Becker, Landes, and Michael 1977.” ticipation SIPP, which is representative of the noninstitutionalized U.S. population. To increase the sample size, SIPP data are pooled from panel years 1990, 1991, 1992, 1993, and 1996. Subsequent panels in 2001 and 2004 are not used because they do not differentiate work limitations by severity. Demographic information, marriage histories, and divorce histories come from topical module two of the panel, which corresponds with the panel’s eighth month. In regards to disability, respondents first report whether they have a health con- dition that limits the kind or amount of work they can do and, if so, the month and year in which it began. Work-limited respondents then report whether the condition prevents them from working at a job or business and, if so, the month and year in which they became unable to work. The SIPP measure of disability, defined by the ability to work, is similar to the measure used by Charles and Stephens 2004 using the Panel Survey of Income Dynamics. Because the precise timing of the work limitation or prevention is essential for the analysis, observations are dropped if the dates of disability onset are missing. Of the 5,020 disabled male respondents between the ages of 30 and 54, 662 respondents 13.2 percent have missing disability onset dates. In regards to marriage, respondents report the number of previous marriages and the dates of marriage, separation, and divorce. The module collects information for up to three marriages, so respondents who report four or more marriages are dropped. This restriction eliminates less than 1 percent of the sample. To examine the effect of disability on earnings, the data are matched to the De- tailed Earnings Record DER. The data, which come from the Social Security Ad- ministration, report earnings on an annual basis from 1978 to 2004. Earnings are measured as the sum of taxable wage earnings, taxable self-employment earnings, and income deferred to 401k or similar type accounts. 3 When these data are re- quired, the sample is necessarily restricted to SIPP respondents matched to admin- istrative records, which depends on whether the respondent consented to the match. 4 The dynamic association between disability, earnings and divorce is examined using event-study methods. Under this method, the association is measured longi- tudinally by the change in earnings and divorce around the time of disability onset. Because of the methodology, and because the disability and marriage data are ret- rospective, two restrictions to the sample are imposed. First, the disabled sample is restricted to males whose first disability occurs within seven years of the survey year. This ensures at least five years of earnings before disability onset—the earliest SIPP year is 1990, and the earliest year of earnings is 1978, leaving five years of predisability earnings for those disabled in 1983. This restriction also reduces mea- surement error in the dates of marriage, divorce, and disability onset, because the period of recall is shorter. Second, disabled respondents are dropped if the first onset of disability occurs during the same year in which the survey was conducted. This is because, for these respondents, the divorce outcome is not observed during the full calendar year of disability onset, as the data used for the analysis are entirely 3. Self-employment income comes from tax return data, and wage income and deferred earnings come from W2 form data. Wage and deferred earnings are not capped at the Social Security taxable maximum. 4. Among the final sample of males, 88.5 percent are matched to administrative records. retrospective. 5 These two restrictions reduce the disabled sample from 4,358 to 2,196. Additionally, the sample is restricted to males whose age is 30 to 54 during the reference year. For the disabled, the reference year is the year of disability onset; for the nondisabled, it is the sixth year before the survey. The final sample consists of 30,809 nondisabled respondents and 1,951 disabled respondents. As mentioned, the analysis relies solely on retrospective data on disability and marriage and does not use data recorded during the SIPP panel. The first reason is that few transitions from nondisabled to disabled occur during the panel, as some panels are less than three years in length. Additionally, if disability has a lagged effect on divorce, then some divorces may not be observed. The second reason is sample attrition, where a large number of respondents exit the survey nonrandomly with each successive wave. 6 Sample attrition poses serious econometric problems if attrition is related to life altering events such as disability onset or divorce.

B. Sample Summary