Esther Mertens

80 | Chapter 4 feature of M plus to correct for clustering at school level. This is a conservative clustering procedure to gain unbiased estimates of the standard errors (Muthén &Muthén, 2010). Clustering at class level was not taken into account as class composition was not stable over the years (e.g., Cross et al., 2016). Full Information Maximum Likelihood (FIML) procedures were used to include all participants in the model with Robust Maximum Likelihood estimation (MLR) for parameter estimates as this estimator is robust to non- normality and non-independence of the data (Muthén & Muthén, 2010). For our analyses we used Latent Growth Curve (LGC) models. LGCmodels estimate individual growth curves and use these curves as indicators of latent variables (i.e., intercept and slope) to estimate average group growth trajectories (Muthén &Muthén, 2010). This approach is recommended by Greenberg and Abenavoli (2017) for analyzing universal interventions as these models have the potential to demonstrate (small) changes in the population curve and examine differences in trajectories between potential subgroups. To examine adolescents’ personality traits as moderators of intervention effects we analyzed LGC models with unspecified growth and interaction effects. Not specifying the growth allowed us to examine nonlinear growth. We fixed the factor loading of T1 at 0 and of T4 at 3. The factor loadings of T2 and T3 were determined by the data (Duncan & Duncan, 2004). For the interaction effects, we created three dummy variables, each representing a R&W condition compared to the Control condition, and grand mean centered the scales representing the personality traits. Subsequently, we created interaction effects of the centered trait and each dummy variable (e.g., Extraversion X Light, Extraversion X Standard, Extraversion X Plus). In the LGC model, we regressed the intercept and slope on the three condition dummy variables and the trait. Additionally, we regressed the slope on the interaction effects. A significant interaction effect indicated moderation by that trait. Adolescents’ age and ethnic background were added as covariates. We estimated an LGC model per trait for each outcome separately. In the case of a significant interaction, we conducted multigroup LGC models in which we split the sample in three groups representing adolescents scoring low, average, and high ( M ± 1 SD ) on the concerned personality trait. Growth in these multigroup LGC models was specified based on the growth estimated in unspecified growth LGC models including the three dummy variables and the two covariates. All parameters were constrained to be equal across the groups, except for the slopes regressed on the dummy variables. Using these estimates we calculated effect sizes for the low, average, and high scoring groups by multiplying the rate of change by time span of the intervention as modeled in the LGC models (i.e., factor loading of T4) divided by the standard deviation of the concerned outcome ( d = (slope * duration) / SD; Feingold, 2013).