Statistical analyses were performed with STATA 10 GLLAMM [33, 34]. The statistical design was a random coefficient ordinal logistic regression model of alcohol use with the intercept and slope for time allowed to be random for individuals. The dependent variable (alcohol use) was measured at waves 3, 4, and 5 (ages 12, 14, and 15). The predictors were parent disapproval of alcohol use, mother/father closeness, family conflict, and peer alcohol use and these were treated as time-varying. Family variables were lagged by one year (measured at waves 2, 3 and 4; ages 11, 12 and 14). Control variables included fixed effects for sensation seeking, SES, and family structure. Model testing was conducted separately on boys and girls to improve interpretability and because a pooled model (with gender interacted with family variables) assumes that the variance of the random effect for individuals/age is equal between sexes, which was not the case. The unconditional growth models demonstrated higher inter-individual variability in both intercept and slope for girls than boys, and a stronger correlation between the random intercept and slope for girls (compare variance components for Model 1 in Table 2 versus Table 3).
