We used a hierarchical generalized linear model (HGLM) framework to address potential dependencies within these data. HGLM permits use of all data and properly models the correlated observations within each person, unlike ordinary least squares regression. Time points (level-1) were nested within individuals (level-2), which accounted for the association across the three time periods. We used HGLM with a binomial link, as opposed to HLM, because the dependent variable consisted of dichotomous data, which violated assumptions of normality (Raudenbush & Bryk, 2002). Specifically, the level-1 model is a Bernoulli function that models the probability of using a specific substance. The full information maximum likelihood estimation procedure (FIML) was used to address missing values at level-1 of the HGLM because FIML produces less biased estimates than does listwise case deletion or mean substitution (Acock, 2005).
