As a complementary analysis for GEE to illustrate the AUD variation over follow-up, subgroups were formed on the basis of trajectory analysis of AUD symptom count derived from zero-inflated Poisson mixture model (40), with a cubic trajectory for each group and the number of groups determined by model Bayesian information criterion (BIC). Finally, to discern individual differences in the prospective relationship between acute alcohol responses and future drinking problems, exploratory k-means cluster analyses, which partitioned observations into k clusters with each observation belonging to a cluster with the nearest mean, were conducted (41) to create subgroups that exhibited different relationships between alcohol responses (stimulation, sedation, liking, wanting) and AUD symptoms. The optimal number of clusters was determined by the Variance Ratio Criterion (42). Statistical packages used were Stata 12.0 (Stata-Corp, College Station, Texas) for GEE and trajectory analyses and SPSS 17.0 (SPSS, Chicago, Illinois) for cluster analyses.
