Elien Neimeijer

50 room, seclusion in a locked room (designed for this purpose), and involuntary medi- cation (the administration of rapid tranquillisation via intramuscular injection against a client’s will). A total of 425 coercive measures were used during the study period, involving 92 participants. The number of coercive measures per participant varied from 0 to 116 ( M = 1.70, SD = 8.30). Statistical analyses First, assumptions were checked (missing data, multivariate outliers). We addressed missingness of the data using Little’s MCAR test. We also examined multivariate and influential outliers using visual inspection of the data as well as examining values for Cook’s distance and Mahalanobis distance. Further, we examined associations between group climate, coercive measures, and aggressive incidents using bivariate correlation analyses (Pearson’s r ). Pearson’s correlations of r = .10 - .30 are seen as small, r = .30 - .50 are seen as moderate, and r > .50 are seen as large (Cohen, 1992). Subsequently, we tested the hypotheses through multilevel structural equation modelling (MSEM), due to the nested data structure (clients were nested within groups), using Mplus software version 6.11 (Muthén & Muthén, 2017). We followed the procedures outlined by Hox (2010). First, intraclass correlation coefficients (ICCs) were calculated to examine between-group variability (i.e., the degree of non-independence in the data) (Raudenbush & Bryk, 2002). ICCs greater than zero are indicative of nested data structures, in which case multilevel analysis is warranted (Byrne, 2012). Then, the covariance matrix was decomposed into a pooled within- and between-level covariance matrix. The pooled within-level covariance matrix was used to examine the within-level part of the model, and the pooled between-level covariance matrix was used to examine the between-level part of the model. Next, a multilevel structural equation model was fitted in which the within- and between-level models were estimated simultaneously using the ‘type = two-level’ option in Mplus. Maximum likelihood (ML) was used to estimate all models. We followed the guidelines on using the MSEM framework to test multilevel mediation as outlined by Preacher, Zyphur, and Zhang (2010), as well as the provided Mplus syntax. We hypothesised a direct effect of group climate on aggressive incidents. More spe- cifically: a negative association between aggressive incidents and support, atmosphere and growth. Also, we expected a positive association between aggressive incidents