Maartje Boer

CHAPTER 2 38 regression analyses, we controlled for gender, educational level, age, and ethnic background. To facilitate interpretability, estimates were transformed into odds ratios (ORs) that denote the extent towhich the odds of, for example, mental health problems increase with the number of endorsed problematic SMU criteria. Good criterion validity of the test score interpretations was established when a higher number of endorsed criteria was associated with higher probabilities of mental, school, and sleep problems. Predictors of Problematic SMU Following the validation steps, we examined which demographic characteristics (gender, educational level, age, and ethnic background) predicted a higher number of endorsed problematic SMU criteria. Given that this problematic SMU outcome was considered as a count variable with a high number of zero counts (Figure 2.1), we conducted the analysis using a zero-inflated negative binomial model. We selected this model because it showed better model fit than a zero-inflated Poissonmodel (chi-bar-square(1) = 428.71, p < 0.001). Furthermore, the zero-inflated negative binomial model showed better fit than an ordinary negative binomial model ( z = 3.24, p = < 0.001). The model was interpreted using incidence rate ratios (IRRs), which denote, for example, how much higher the number of endorsed problematic SMU criteria is expected to be for girls relative to boys. IRRs were calculated using boys (gender), highly educated adolescents (educational level), 12-year- olds (age), and native adolescents (ethnic background) as the reference categories. Mplus 8.3 (L. K. Muthén & Muthén, 2017b) was used to conduct the EFA, CFA, and measurement invariance analysis, using Weighted Least Square Means and Variance Adjusted estimation with a probit regression link and theta parameterization. This estimation method was selected because it provided all fit indices for categorical data that were required for model evaluations. The LCA was also conducted using Mplus 8.3, but with Maximum Likelihood estimation with robust standard errors and a logit regression link, as is common for LCA. Stata 14.2 (StataCorp, 2015) was used to conduct Velicer’s MAP analysis using the minap package (Soldz, 2002). Analyses related to IRT, criterion validity, and associations between demographic characteristics and problematic SMU were also performed with Stata with the default Maximum