Person-year data were fitted to discrete time logistic regression models using PROC SURVEYLOGISTIC (SAS v9), which adjusts for complex survey designs, such as data clustered in families, by using a robust variance estimator. After an initial analysis that only included gender, AA race/ethnicity, Hispanic status and age at last interview as covariates, a model that included all covariates was fitted to the data to obtain an adjusted odds-ratio. In addition, we tested for interactions with sex (female as reference group) and race (European-American as reference group) in separate models that only accounted for demographic variables. Significant interaction terms were further adjusted for other covariates. The models fitted to the data are described below. For Model 1, SI and SA were the dependent variables and early substance use was the independent variable; in other words, we examined whether early substance use “predicted” onset of SI and SA, as has been frequently done in prior longitudinal studies. We elected to focus on early substance use for Model 1 because, as shown in Supplemental Figures, the median age of onset of SI and
