Maartje Boer

SMU AND ADHD-SYMPTOMS 135 5 measured across all three waves. Therefore, measurement invariance analyses were conducted prior to the analyses, using Mplus 8.1 (L. K. Muthén & Muthén, 2017b). For each measure, this was done by means of multigroup confirmatory factor analysis (CFA) on the data structured in long format ( n = 1,629), where groups were indicated by waves. Measurement invariance was imposed by constraining the loadings and intercepts of the items to be equal across all waves, after which model fit was evaluated. For SMU problems, thresholds instead of intercepts were constrained to be equal, because this scale consists of binary items. Measurement invariance analyses for SMU intensity, attention deficits, impulsivity, and hyperactivity were carried out using Maximum Likelihood estimation with robust standard errors (MLR), which corrects for the somewhat skew distributions of these measures. For SMU problems, Weighted Least Square Means and Variance Adjusted (WLSMV)-estimation was used, which is recommended for categorical items (L. K. Muthén & Muthén, 2017b). For each multigroup CFA, overall model fit was evaluated using the Comparative Fit Index (CFI; > 0.9 = acceptable; > 0.95 = excellent), Tucker Lewis Index (TLI; > 0.9 = acceptable; > 0.95 = excellent), and Root Mean Square Error of Approximation (RMSEA; < 0.08 = acceptable; < 0.05 = excellent) (Van de Schoot et al., 2012). We subsequently evaluated whether removing the equality constraints on the loadings and intercepts/ thresholds would significantly improve model fit based on change in CFI (increase of ≥ 0.010) and RMSEA (decrease of ≥ 0.015) (F. F. Chen, 2007). In measurement invariance analyses, evaluation of model fit using Δ CFI and Δ RMSEA are preferred over χ ²-difference tests, because the latter is sensitive to large sample sizes (F. F. Chen, 2007; Cheung & Rensvold, 2002). Table 5.1 shows that when measurement invariance over time was imposed, the overall model fits of the multigroup CFA models for SMU intensity, SMUproblems, attentiondeficits, and impulsivitywere all acceptable to excellent. Model fits did not significantly improve when equality constraints on the item loadings and intercepts or thresholds were released. This means that measurement invariance was established for these four measures, and that we can make meaningful conclusions about their longitudinal relations (Van de Schoot et al., 2012). The overall model fit for hyperactivity was relatively low (CFI = 0.874, TLI = 0.879, and RMSEA = 0.122), and measurement invariance was not established ( Δ CFI = 0.019). However, additional analyses (results not shown) showed that measurement invariance was only related