Logistic regression was subsequently employed to establish the statistical significance of the relative contributions of familial liability and maltreatment on antisocial outcome. PROC SURVEY LOGISTIC in SAS was employed to allow control for clustering (by zip code in MOTWIN and by family in COGA). For these analyses, a nested approach to modeling was taken, so that the model fit and accompanying predictive statistics could be compared as variables were added, beginning with demographics only (Model 1) and progressing through models with the addition of the following variables: child welfare contact (Model 2); familial liability (Model 3); both (a full model (Model 4)); and a full model with interaction between familial liability and child welfare contact (Model 5). In addition to calculating odds ratios (OR) and significance levels, the Wald Chi-square for the sandwich estimator was calculated for model fit, along with the max rescaled r-square, and the c statistic corresponding to the receiver operating curve. The c statistics can be thought of similarly to a grading scale, with .70 considered adequate, .8 considered good, etc. [30] Because there were so
