Analyses used MPlus v.7.2 [61]. Because adolescents’ T2 alcohol problems was a count outcome, we compared preliminary models with different distributional assumptions including zero-inflated poisson and zero-inflated negative binomial. Comparison of relative fit indices (AIC, BIC, and −2 log likelihood) supported modeling T2 alcohol problems using zero-inflated poisson (monte-carlo integration) and, therefore, global fit indices were unavailable. Structural equation modeling tested hypothesized relations using MLR and FIML to estimate missing data on endogenous variables. We estimated paths from GABRA2 and covariates (ancestry, gender, age, and FHD) to mediators (impulsivity and sensation seeking factors) and T2 outcomes (hyperactive–inattentive symptoms, conduct problems, and alcohol problems). Paths were estimated from mediators to outcomes. Models were estimated with and without controlling for T1 symptoms. 95 % asymmetric confidence limits tested mediation [62]. The type = complex function was used to adjust SEs because participants were nested in families. To correct for multiple testing, we calculated false discovery rate-corrected p values (q values) using the QVALUE software [71]. Analyses suggested that our sample was adequately powered (0.80) to detect small effect sizes (f2 0.027), which are typical in candidate gene research.
