62 Chapter 3 2.4 Data analysis All analyses were performed R version 3.6.1 (R Core Team, 2018). Descriptive statistics were obtained for the total sample and for the two conditions (Take it Personal! and control group) separately in terms of demographics (age, sex, total IQ) and all outcome measures. Any demographics that differ between conditions at baseline were added to the statistical models as covariates (see paragraph 3.1). Little’s MCAR test indicated that missing values at follow-up occurred at random, warranting the use of a multiple imputation strategy for analyses according to intention-to-treat principles. To assess the effect of Take it Personal! on emotional and behavioral problems, four multilevel analyses were conducted—one per problem domain. Anxious, withdrawn, rule-breaking, or aggressive behavior at follow-up were entered as dependent variable, each with three predictors: dummy coded covariate sex (0 = male, 1 = female), dummy coded between subjects factor condition (control = 0, Take it Personal! = 1) and the problem domain at baseline (anxious, withdrawn, rule-breaking, aggression). Each multilevel model included a random intercept for treatment centers, and thus controls for data clustering within treatment centers. In addition, to examine if emotional and behavioral problems moderated the program’s effect on substance use frequency, four separate multilevel analyses were conducted, each with substance use frequency at follow-up as the dependent variable. The variable substance use frequency was constructed as each youngster’s most frequently used substance(s) at baseline compared to (the average of) that/those substance(s) at follow-up (cf. Schijven et al., 2020a). This was done because Take it Personal! addressed the use of those substance(s) that was/were most problematic for the individual. Substance use frequency at baseline, sex, condition and one of the four behavioral problems at baseline (anxious, withdrawn, rule-breaking, aggression) were added as predictors. To evaluate moderation effects, each model also included a two-way interaction term for condition with behavioral problem score at baseline. Similar to models evaluating effects on behavioral problems, the models that evaluate moderation by behavioral problems included random intercepts that account for clustering within treatment centers. All continuous predictor variables were mean-centered. To obtain model parameters from the multilevel models, we used functionality from R package lme4 (Bates et al., 2015). Satterthwaite’s method was used to evaluate p-values, for which significance level was set at p < 0.05.
RkJQdWJsaXNoZXIy MTk4NDMw