Maartje Boer

CHAPTER 4 110 Modelling Two-level regression analyses were conducted on the imputed datasets using Mplus 8.3, with individual-level measures at the first level and country- level measures at the second level. Although the data consist of a three-level structure, where individuals were nested in schools and countries, applying three-level analyses was not feasible because then the number of parameters would exceed the number of country-clusters, which does not provide model identification. In addition, to retain fewer parameters than country-clusters, associations with all six wellbeing outcomes were examined in separate models. Models were estimated using Maximum Likelihood estimation with Robust standard errors to account for the skewed distribution of the wellbeing outcomes. Figure 4.1 illustrates our analytical model, which was examined using a stepwise procedure. In our first model (denoted as M1 a ), on the individual- level, we examined associations between intense and problematic SMU and life satisfaction (while controlling for gender, age, and family affluence) without any country variation. On the country-level, we tested associations between mobile internet access and the country-level prevalence of intense and problematic SMU (M1 a ). We extended this model with a random slope (S1) for intense SMU, which means that its association with life satisfaction was allowed to vary across countries (M1 b ). Subsequently, we added a random slope (S2) for problematic SMU (M1 c ). Next, we added two cross-level interactions that examined whether the association between intense SMU and life satisfaction varied by the country-level prevalence of intense SMU (M1 d ) and problematic SMU (M1 e ). Finally, we added two additional cross-level interactions that examined whether the association between problematic SMU and life satisfaction varied by the country-level prevalence of intense SMU (M1 f ) and problematic SMU (M1 g ). These steps were repeated for the other five wellbeing outcomes (M2 a-g to M6 a-g ). Interpretations After the stepwise analyses were conducted, for each wellbeing outcome, we selected the model with the best model fit for further interpretation. As a result, random slopes and cross-level interactions were only interpreted when they improved model fit. Model fit was evaluated using the Bayesian