Maartje Boer

CHAPTER 6 166 Chen, 2007). In case equality constraints deteriorated model fit, modification indices were consulted to find the source of misfit. For all measures, applying equality constraints to the factor loadings did not deteriorate model fit, which suggests that all measures had invariant factor loadings over time. Four out of eight measures had one intercept that was not invariant over time. However, each measure had at least two items where the item intercepts were equal over time, which is sufficient for the purpose of our study, namely comparing effect sizes and latent means over time (Van de Schoot et al., 2012). Plausible Values Third, we calculated plausible values for our measures, which are imputed values that represent the values of latent variables based on a specified factor model using Bayes estimation (Asparouhov & Muthén, 2010c). We followed this method because due to the complexity of our main analyses, it was not feasible to use latent variables in our models. In addition, due to the highly skewed distribution of the sum-score of SMU problems it was not possible to use item sum-scores. Plausible values have been found to accurately resemble covariances between latent variables (Asparouhov & Muthén, 2010c), and as such have been used by researchers to obtain reliable scores for their measures that can be used for subsequent analyses (Ciarrochi et al., 2016; Deutsch et al., 2014; Rhee et al., 2013). For each latent variable, we imputed 20 plausible values based on the factor models as established in our measurement invariance analyses. That is, item factor loadings and intercepts (or thresholds) for which measurement invariance was established were constrained to be equal over time. Our data were not completely missing at random , as the attrition analysis showed that there were small relationships between the observed data and dropout. In that case, retention of dropout cases provides more reliable model estimates than listwise deletion of dropout cases, especially when dropout rates are high (Enders & Bandalos, 2001). Hence, plausible values for complete as well as dropout cases were estimated with a full information approach (Asparouhov & Muthén, 2010a). That is, plausible values of dropout cases could be estimated based on available data from previous and/or subsequent waves. As a result, all respondents ( n = 2,109) were retained in our analyses. All imputations were merged into one dataset for subsequent