Elien Neimeijer

53 also used (Cheung & Rensvold, 2002): comparative fit index (CFI), Tucker-Lewis index (TLI), the root mean square error of approximation (RMSEA), and the standardised root mean square residual (SRMR). The SRMR at the within-group level (SRMRW) and be- tween-group-level (SRMRB) were examined. Modification indices, giving the expected drop in chi-square if the parameter in question is freely estimated, were used to improve model fit. A non-significant Chi-square value is considered to indicate an exact fit to the data. The following fit values indicate a good fit to the data: TLI > .95; CFI > .95; RMSEA ≤ .05; SRMR ≤ .08, and values of TLI > .90; CFI > .90; RMSEA ≤ .08 are indicative of acceptable model fit (Cheung & Rensvold, 2002). Results Preliminary analyses There was a very small proportion of missing values on the self-reported group climate data, ranging from 0% to 4.8% per variable. Little’s MCAR test ( χ 2 (63) = 54.70, p = .763) was not significant, indicating that missing values were completely at random. Missing values were imputed through expectation maximization. Two cases were identified as in- fluential outliers regarding the variable coercive measures. These cases displayed a coer- cive measures score of 116 and 49, respectively, while the scores of the sample excluding these outliers had a range of 0-18 (M = 2.66, SD = .30). Also, values for Cook’s distance (43.8 and 17.9, respectively) and Mahalanobis distance (186.7 and 32.4) as indicators for multivariate outliers were very high for these cases compared to the means of these values in the sample (Cook’s distance M = 0.18 and Mahalanobis distance M = 1.99). The analyses reported below were run with and without the outliers and results indicated that they impacted the results significantly, particularly the parameter estimates of the between-level models. For example, the ICC of the variable coercive measures was ex- tremely low (.01) with outliers present compared to the ICC value without outliers (.30). Also, between-level models with outliers indicated standardised beta’s greater than 1, as well as very large standard errors of standardised estimates of associations involving the variable coercive measures. Therefore, these two cases were removed from the dataset. Table 1 shows the means, standard deviations, and ICCs of the study variables as well as the correlations among these variables.