Martijn van Teffelen

Hierarchical structure of hostility 31 2 Table 2 Spearman’s rho correlations between uni- and multidimensional hostility constructs STAXI-2T AQH FOA α STAXI-2T .88 AQH .54* .86 FOA .60* .43* .93 PID-5H .79* .51* .65* .88 Note . * p < .001. Cronbach’s α is reported in the diagonal. Then, to examine the hierarchical structure of hostility a PCA was run using the ‘Bass- Ackwards’ method. We evaluated multivariate normality and linearity by inspecting Mahalanobis distance. We observed two multivariate outliers who were removed from the analysis. We observed non-normality on eighteen FOA items (i.e., skewness values smaller or larger than three standard errors) (Curran et al., 1996). Of these eighteen items fifteen extremely violated the normality assumption even after inverse-transformation (i.e., 1/ x ) and were removed from further analyses to maintain model robustness. The removed items are shown in Appendix A. Factor loadings are presented in Appendix B. Appendix C shows the decision process for principal component extraction. First, one unrotated principal component was extracted, followed by the extraction of successively (i.e., two, three, etc.) more Varimax rotated principal components. This was then repeated until one of the factors was either too specific to be interpreted (e.g., containing one item) or was no longer interpretable (e.g., by containing items that show hardly any content similarity). The first unrotated principal component accounted for 30% of the total variance. The first ten eigenvalues were: 16.16, 3.90, 2.68, 2.12, 1.61, 1.35, 1.24, 1.20, 1.13, and 1.06. Then, successively larger solutions (i.e., two, three, etc.) were examined. Inspection of the 6-principal component solution showed that the last factor consisted of the two items: “I resent being told what to do, even by people in charge” and “I feel annoyed when not given recognition for doing good work”. Thus, the 6-principal component solution was interpreted as not meaningful, resulting in a 5-principal component solution as base of the hierarchical model, accounting for 49% of the variance. The hierarchical 5-principal component model is shown in Figure 1. Correlations between the component loadings and the original scales are shown in Table 3. Rotated component loadings and item content are shown in Appendix B. All principal components were labeled according to what was most common to all these items. The first component (P1.1) was labeled Hostility and demonstrated significant positive associations to the original hostility scales ranging from r = .66 (AQ-H) to r = .88 (PID-5H). The principal components in the two-factor solution were labeled Hostile Cognition (P2.1) and Aggressive Behavior (P2.2). Hostile Cognition related most strongly to the total scores of the STAXI-2T, AQH and PID-