Model of Divorce Results

from six years before onset to two years before, the percent married declines by only 0.007 percentage points. The decline in the percent married at the time of disability onset may reflect a decrease in marriage formation, an increase in divorce, or both. To disentangle these two mechanisms, Panels B and C of Figure 2 plot the rate of marriage formation and divorce, respectively. The marriage formation rate reflects the percent of individuals who enter a marriage during the calendar year, conditional on not being married during the previous calendar year. The divorce rate reflects the percent of individuals who exit a marriage during the calendar year, conditional on being married during the same calendar year. Divorce is defined as either separation or legal divorce, whichever occurs first, and does not reflect separations due to the death of a spouse. 9 Among the work prevented, the rates of marriage formation and divorce exhibit downward, albeit noisy trends. However, only the divorce rate changes discretely at the time of disability onset. In fact, the highest rate of divorce occurs in the year following disability onset, reaching 4.7 percent. Thus, the discrete decline in marriage among the work-prevented, illustrated in panel A, is driven predominately by divorce.

B. Model of Divorce

The graphs show that the dynamic effect of disability on divorce is concentrated among the work-prevented. To quantify the effect, the following event-study model of divorce is estimated: g g g g Y = α + βX + ∑ γ D + ∑ δ D + ε . it it i t it it g t The unit of observation is person and calendar year. The variable is an indicator Y it whether individual in year exits a marriage through separation or divorce, which- i t ever occurs first. The vector of coefficients represents the effects of covariates β . Demographic covariates include age and its square in years, race white versus X it other and educational attainment no high school diploma, high school diploma, and college degree. To control for different age profiles by race and education, age and its square are interacted with the indicators of race and educational attainment. Marriage covariates, intended to control for marriage quality, include marriage tenure and its square in years as well as indicators for the second and third marriage. The term is an error that is correlated within person over time. ε it The index corresponds to the two disability groups: the work-limited and the g work-prevented. The indicators controls for baseline differences between the dis- g D i ability groups and the nondisabled; it equals one for those belonging to disability group and zero otherwise. The variable allows for divorce among the disabled g g D it to change relative to baseline; it equals one for those belonging to disability group in period , defined relative to the reference year, and zero otherwise. The model g t therefore allows for the rate of divorce to vary over time by disability severity. This event-study type model is preferred, as it is similar to the model of Charles and Stephens 2004. 9. As Espinosa and Evans 2008 conclude, the death of a spouse may have a causal effect on the health of males. The divorce model is estimated using both disabled and nondisabled respondents. Initially, the sample of disabled respondents is restricted around the time of disability onset, from three years before to three years after. 10 The index references three t periods relative to the reference year: Years −2 and −1, Years 0 and 1, and Years 2 and 3. The nondisabled sample is then restricted to cover the range of ages and years spanned by the disabled sample—ages 27 to 57 and years 1980 to 1998—up to five calendar years after the reference year. Among the nondisabled, the sixth calendar year after the reference year corresponds with the year of the survey, which is not observed in retrospect. The estimates of the model are presented in Column 1 of Table 3. The baseline difference in divorce rates, in Year −3, is positive for both disability groups, but is statistically significant only for the work limited. As time elapses, the divorce rate among the work limited declines, reaching a level comparable to the nondisabled by Years 2 and 3. In contrast, the divorce rate among the work prevented increases then decreases, peaking to 1.78 percentage points in the year of and immediately after disability onset Year = 0,1. Although the estimate suggests that disability onset precipitates divorce, the coefficient is statistically insignificant at the 5 percent level. Statistical significance, however, appears to be a matter of statistical power. A less-parameterized alternative to the event-study model controls for disability-spe- cific trends in divorce, with deviations from trends in the years before and after disability onset. To properly estimate the trends in divorce, the sample of disabled respondents is expanded to include six years before to three years after disability onset. The estimates of this alternative model, which includes the same set of con- trols as the event-study model, are presented in Column 2 of Table 3. As shown, the onset of a work-preventing disability increases the rate of divorce by an estimated 1.77 percentage points, which is similar to the 1.78 percentage point increase from the event-study model. 11 Moreover, the estimate is statistically significant at the 5 percent level. The trends in divorce, measured by the interaction of disability status and “Year,” are negative, but statistically insignificant at the 5 percent level. The negative trend may reflect that, as marriages dissolve, average, unobserved marital quality increases, decreasing the divorce rate over time.

C. Heterogeneous Effects by Disabling Condition