Maartje Boer

THE COURSE OF PROBLEMATIC SMU 205 7 In addition, we computed respondents’ average level of subjective wellbeing, self-control, and social competencies using their responses from T1 until T4. These person-specific means denote the time-invariant (i.e., trait-like, stable) part of adolescents’ level of subjective wellbeing, self- control, and social competencies. Subsequently, we conducted multivariate multinomial regression analysis to predict class membership with the subjective wellbeing, self-control, and social competencies person-specific means, while controlling for demographic characteristics. In doing so, we specified the measurement error of the class variable using the logits for the classification probabilities for the most likely latent class membership as obtained from the LCGA. This model specification takes into account the uncertainty that is associated with the classification, which improves the accuracy of the multinomial regression estimates (Asparouhov & Muthén, 2014). There was one missing observation for immigrant background. For the predictors, the percentage of missing data ranged between 6.6% (attention deficit T2) and 65.9% (life satisfaction T4), which was mostly related to dropout. Gender, educational level, and the class variable were complete. The missing data were not found to be completely missing at random (ꭓ 2 (604) = 802.317, p < 0.001), which means that retaining all respondents is required to limit possible bias (Enders & Bandalos, 2001). However, with the present multinomial model, MLR-estimation did not retain all respondents. Therefore, for the multinomial analysis, we imputed missing values usingmultiple imputation with chained equations (Royston & White, 2011). In this procedure, missing values were estimated based on predictive mean matching with ‘five nearest neighbors’, whereby missing values were imputed based on the observed data on the study variables. Imputations were computed in Stata 13.0 (StataCorp, 2013) and exported to Mplus 8.6 (L. K. Muthén & Muthén, 2017b) to conduct the multinomial analysis. Preregistration The subsample selection and analytical approaches were preregistered. In order to improve the analytical approach, wedeviated fromthepreregistration by defining the predictors of the trajectories using the person-specific