Negative binomial models (Cameron & Trivedi, 1998) were used to examine the effects of DRD4 status, intervention status, and the DRD4 status × intervention status interaction on participants’ substance use. Negative binomial models were used because the substance use measure consists of count data, which does not approximate a normal distribution. Although Poisson models can address this issue, negative binomial models were preferred because of the observed inequality between the mean (M = 1.31) and variance (Var = 5.72) of the substance use data at follow-up. This inequality is termed overdispersion and violates an assumption that must be met for Poisson models to be used (Cameron & Trivedi, 1998). Negative binomial models allow for overdispersion in the dependent measures, rendering them appropriate for the analyses. All the models were analyzed using STATA 12.0 with robust standard errors for the estimates (StataCorp, 2011). The negative binomial analyses comprised two models: (a) a main-effects model involving DRD4 status, intervention status (SAAF–T or control), and gender, and (b) tests of two- and three-way interactions involving the predictors in the main effects model. In
