Details of the Structural Equation Analyses
The reported analyses were performed using variance-covariance matrices with pairwise deletion of missing values. Two alternative methods of dealing with missing values were used: listwise deletion and estimation using expectation-maximization method. All three methods yielded very similar results.
In the analyses, three indicators per latent construct were used, unless a construct was indicated by a single item. In the latter cases, the measurement error of a single indicator was provided to the program, assuming reliability of 0.80. In cases of multiple indicators, using the accepted approach of parceling (Bandalos, 2002; Stacy, Bentler, & Flay, 1994), each latent variable was indicated by one third of the items that make up the scale. Item parcels were constructed so that their kurtosis statistic would be minimized. To achieve this, items with largest kurtosis were combined with items with smallest kurtosis, and so on. In spite of these, many of the observed variables were nonnormally distributed due to the nature of the phenomena under investigation (victimization). To overcome this violation of SEM assumptions, we employed a maximum-likelihood estimation method with robust standard errors (Hu & Bentler, 1999) together with the Satorra-Bentler rescaled chi-square statistic (Satorra & Bentler, 1994, 1999). This statistic compensates for multivariate nonnormality of variables. Difference between two scaled chi-squares doesn’t distribute as chi-square. Therefore, in computing significance of differences between models, we used the Satorra-Bentler scaled difference test (Satorra & Bentler, 1999).
Following recommendations of Hu and Bentler (1999), we report fit indexes of two types: Non-Normed Fit Index (NNFI, also known as TLI) and Comparative Fit Index (CFI), and two indexes of misfit: Root Mean-Square Error of Approximation (RMSEA) and Standardized Root Mean-Square Residual (SRMR). NNFI