Martijn van Teffelen

Chapter 6 136 Statistical analyses Statistical analyses were similar to Study 1, except for a few notable differences. First, the hypotheses that CBM-I leads to greater increases in benign bias and greater reductions in hostile bias, state hostility, general psychiatric symptoms, and trait hostility were assessed using twelve mixed regression models. Condition, time, and condition by time interaction indicators were entered as predictors. Mixed regression was opted for given its ability to handle missing data and modelling error terms, leading to an increase in statistical power (Baayen et al., 2008). Within each regression model, repeated measures were clustered in participants. AR1 was selected as covariance structure, as -2 log likelihood testing revealed that it was most parsimonious. As VDT scores were highly skewed, we used Poisson regression. 4 This is in line with analytic practices in de VDT literature (DeWall et al., 2013). VDT post-score was entered as dependent variable. VDT pre-score and condition were entered as independent variables. For this specific analysis, VDT pins were imputed following the multiple imputation method. Specifically, pre-test VDT scores were used to predict post-test VDT scores in five pooled imputations. Moreover, we calculated a reliable change index (RCI) for both HIB measures (WSAP-H and SIP-AEQ) (Jacobson & Truax, 1991). Second, to examine effects of intervention on working alliance in subsequent psychotherapy the following analyses were run. One independent samples t -test on working alliance was run to test intervention differences. Then, we examined correlations between change scores of the outcome measures and working alliance. As VDT scores were highly skewed, we used Poisson regression to regress working alliance on VDT post-intervention scores, corrected for baseline instead of examining their correlation. Analyses were conducted following the intent-to-treat principle. Moreover, to test confounding influences from active alcohol or drug use during CBM-I sessions, analyses were performed with and without participants who did not complete at least 75% of sessions while not under the influence of alcohol or drugs. Results Means and standard deviations of study variables are presented in Table 2. Tests of baseline differences revealed that variables did not differ significantly at baseline between conditions, except of atypical antipsychotic use. Overall, this indicates that random allocation was successful. Comparing our baseline values to other studies revealed that hostility levels in this sample are comparable to or larger than other studies using clinical samples (Bach et al., 2016; Coccaro et al., 2017; Dillon et al., 2016; Hornsveld et al., 2009; Lievaart et al., 2016). In total, 14.32% of values were missing. Missed sessions per conditions are shown in Figure 1. Number of missed sessions did not differ per condition 4 Pre-registration file stated that we would analyze VDT scores using mixed regression. However, VDT-scores were extremely skewed.