Maartje Boer

CHAPTER 7 204 The problematic SMU sum-score follows a Poisson distribution (Figure 7.1), which does not allow for ordinary LCGA (Reinecke, 2006a). Therefore, we compared the model fit of Poisson and zero-inflated Poisson growth models. We present our findings using the more parsimonious Poisson models, because zero-inflation parameters were not significant and from three classes onwards model fits were comparable (Appendix, Figure A7.1). The trajectories of problematic SMU were estimated in parallel with trajectories of SMU frequency. These co-trajectories were estimated without any covariates, which facilitates interpretation (Van de Schoot et al., 2017). The model specifications are available in the Appendix (Figure A7.2). The number of classes was established based on the model fit and classification accuracy (Van de Schoot et al., 2017). Model fit was evaluated based on the Bayesian Information Criterion (BIC). We used the Lo-Mendell- Rubin adjusted Likelihood Ratio Test (LMR-LRT) and Bootstrap Likelihood Ratio Test (BLRT) to indicate whether a class solution improved model fit compared to a class solution with one class less ( p < 0.050). Classification accuracy was evaluated based on the average class membership probability of each class, with values close to 1 indicating good classification. Also, Entropy with values of 0.700 or higher were considered as adequate (Reinecke, 2006a). As typical for latent class analysis, the model selection was based on a trade-off between all of the above-mentioned criteria (Jung & Wickrama, 2008). The percentage of missing data on the study variables for this part of the analysis ranged from 6.6% (problematic SMU T2) to 65.8% (SMU frequency T4), which was mostly related to dropout. Little’s Chi-square test for missing data was significant ( ꭓ 2 (118) = 262.144, p < 0.001), which means that we cannot assume that data were completely missing at random. Consequently, listwise deletion of cases with one or multiple missing values may bias results (Enders & Bandalos, 2001). However, in our analysis, we aimed to limit the bias that is associated with missing data by conducting the LCGA using full information maximum likelihood with robust standard errors (MLR), which retains all 1,419 respondents. Predictors of Trajectories Based on the latent class solution from the LCGA, we created a nominal class variable that denotes the most likely class membership for each respondent.