Models with 2–6 classes were fit. To facilitate model convergence and identification, the variance of the quadratic term was fixed to 0 within each class for all models. A 5-class solution was selected for additional analyses: Although a 6-class solution provided minor improvements in AIC and BIC, the likelihood ratio test indicated that it did not provide a significantly better fit (Table 3). Furthermore, the sixth class reflected a difference in severity, but not shape, from another class. We did not pursue additional refinement of parameter estimates and variances, as our aim was to explore differences in alcohol outcomes across classes rather than to examine whether within-class variation differentially predicted alcohol outcomes. The variances of the intercept and slope terms were estimated to be different from 0 (p<0.001).
