Survival analyses indicated that parents of probands have a shorter latency to divorce than parents of controls (Breslow = 10.28, p <.01; see ). Follow-up chi-square analyses revealed that the proportion of proband parents experiencing divorce (22.7%) was significantly greater than control parents (12.6%) by the time their children were 8-years-old, χ2 (488) = 8.03, p <.01. The proportion of families experiencing divorce after youth were 8-years-old did not differ between probands (15.3%) and controls (10.7%), χ2 (488) = 2.15, p =.14.
Survival curves displaying latency to divorce between parents of adolescents and young adults with and without ADHD.
Next, using data from the subset of proband families eligible for Cox regression analyses (n=147), relative risk ratios (likelihood of divorce given presence of specific risk factor independent of other risk factors) were computed for the predictors of divorce. Paternal antisocial behavior evinced the largest relative risk ratio (5.53). Variables with moderate risk included maternal (2.70)/paternal (2.24) divorce history, maternal (2.37)/paternal (2.53) substance use, maternal depression (2.17), and child race/ethnicity (2.64). Risk of divorce appeared minimal for years married (1.38), maternal (.77)/paternal (.75) education, child age (.56), ADHD severity (1.32), and ODD/CD severity (1.43).
Prospective Cox regression analyses using the subset of proband families with complete childhood data (n=147) found that maternal and paternal level of education, paternal lifetime antisocial disorder, child age at STP, child race/ethnicity, and baseline child ODD/CD behavior each uniquely predicted the latency to divorce in families of probands (see ). The rate of divorce increased in families with mothers who had less education, fathers with more education and antisocial behavior, and younger, racial/ethnic minority children attending the STP with elevated ODD/CD behavior problems. Years of marriage before birth of proband, maternal and paternal marital history, maternal and paternal lifetime history of substance abuse disorder, maternal lifetime history of depression, and proband ADHD behavior at baseline failed to uniquely predict rate of divorce in the regression.
Statistics for Cox regression model assessing prospective prediction of latency to divorce among parents of youth with ADHD.
Hazard ratios (likelihood of divorce given presence of specific risk factor in context of other risk factors) highlighted the clear risk of divorce between proband parents if fathers had a lifetime history of antisocial behavior (see ). Hazard was also elevated for proband parents of racial/ethnic minority children and children with higher ODD/CD ratings. Of note, confidence intervals were quite large for the paternal antisocial behavior, child race/ethnicity, and child ODD/CD hazard ratios. Therefore, the reliability of the mean hazard ratios is questionable for these variables. Hazard was relatively small for the other significant predictors of latency to divorce (i.e. maternal/paternal education, child age at STP).
Post hoc analyses were conducted to examine an alternative explanation for, and otherwise explicate, the Cox regression results discussed above. First, we explored whether discrepancy in parent education levels predicted latency to divorce in lieu of the unique contribution of individual parent education levels. Parent education discrepancy was computed by subtracting the highest level of education attained by mothers from the highest level of education attained by fathers. When substituting parent education discrepancy for maternal/paternal education variables, results indicated that differences in parent education level uniquely predicted rate of divorce (B=.48, SE=.15, Wald=10.02, p <.01). Unique and nonsignificant predictors included in the remainder of the Cox regression model were no different from those reported in when parent education was entered separately. Thus, parent education discrepancy seems to be a more parsimonious explanation for how education influences the occurrence/rate of divorce between parents of children with ADHD.
Because maternal and paternal education, paternal antisocial behavior, and child ODD/CD variables were significant predictors of divorce latency and strongly correlated with other predictor variables included in the model (r >.30; see above), another set of secondary analyses sought to rule out the effects of multicollinearity in our Cox regression analyses. Multicollinearity was tested by running a series of models with highly correlated predictors excluded one at a time. Multicollinearity was assumed to be present if the standard deviations of the point estimates for the predictors remaining in the model changed substantially in the absence of the related variable withheld from the analyses. Tests revealed that standard deviations of maternal and paternal education, paternal antisocial behavior, and child ODD/CD were undisturbed by removing correlated variables from the model, thereby suggesting that our results were not an artifact of, or influenced by, multicollinearity among the predictor variables.
Next, we explored the influence of missing data on the results of the Cox regression analyses. Predictor variables with the highest rates of missing data were maternal and paternal education (17.0% missing), maternal depression (21.8% missing), and paternal antisocial behavior (24.8% missing). While withholding maternal and paternal education from the full model, the remaining variables from the available cases (n=171) continued to significantly predict rate of divorce (e.g., paternal antisocial, child age, child race/ethnicity, and child ODD/CD were unique, statistically significant predictors at p <.05). While withholding maternal depression, the remaining variables from the available cases (n=165) continued to significantly predict latency to divorce (e.g., maternal/paternal education, paternal antisocial behavior, child age, race/ethnicity, and ODD/CD were unique, statistically significant predictors at p <.05). However, when paternal antisocial behavior was withheld from the model, results of the Cox regression analyses conducted with variables from the available cases (n=155) yielded slightly different results. Paternal education, child age and race/ethnicity still significantly predicted rate of divorce, but maternal education and child ODD/CD were no longer statistically significant predictors. Overall, it does not appear that missing data has limited conclusions drawn from the Cox regression analyses conducted with a sample including only cases with complete data.
Lastly, we submitted the original Cox regression model to a logistic regression analysis in order to investigate whether significant predictors of divorce rate would also predict the occurrence of divorce. Results of the logistic regressions indicated that paternal antisocial behavior (B = 2.43, SE =.70, Wald = 12.15, p <.01, Hazard = 11.36), non-Caucasian descent of the child (B = 1.83, SE =.84, Wald = 4.78, p <.05, Hazard = 6.21), and elevated child ODD/CD behavior ratings (B = 1.72, SE =.88, Wald = 3.86, p =.05, Hazard = 5.61) each uniquely increased risk of experiencing divorce (Overall Model χ2 (14) = 40.28, p <.01). Unlike the results of the Cox regression, only trends for statistical significance emerged for maternal/paternal education level and baseline child age. No predictors failing to uniquely predict rate of divorce in the Cox regression model significantly predicted occurrence of divorce in the logistic regression analyses. Results of the logistic regression analyses, thus, generally corroborated the findings of the Cox regression analyses.