We analyzed both the number of HED and AP over time simultaneously using mixture analysis.8 We added social desirability as a covariate to ensure that derived groups were equivalent in social desirability. Results indicated that, based on the Bayesian Information Criterion (BIC), the model with four latent classes was the best fitting, most parsimonious model (BIC2 = 5654.72, BIC3 = 5604.54, BIC4 = 5432.33 and BIC5 = 5465.26 for two-, three-, four-, and five-latent class models, respectively). The four latent classes were very well separated (Entropy = .99) and the average posterior probability for the most likely class exceeded .98. Entropy values approaching one are considered to indicate well-separated classes (Celeux & Soromenho, 1996). Figure 3 clearly illustrates that a considerable number of students (classes 1 through 3) continued to engage in HED (Figure 3a) and to report AP (Figure 3b) throughout the observed period.9 In the present study, we decided to focus on those who improved versus those who did not, primarily to increase power and improve accuracy of parameter estimates in detecting predictors and moderators in subsequent logistic
