The results will be reported as follows. First, we report the summary of logistic regression for predicting the verdict (first Subsection). Second, we report the summary of multiple regressions for predicting the sentence in years (second Subsection).
Logistic regression analysis for centrality measures predicting verdict
A first logistic regression analysis was conducted to discriminate the verdict, using the Wald statistic to select the significant variables in terms of their prediction ability. The independent variables for this step are betweenness, closeness and degree centrality. The final model used closeness and degree centrality variables to predict the verdict variable (see Table 2).
A test of the full model against a constant only model was statistically significant, indicating that the predictors as a set reliably distinguished between guilty and innocent (X2 23.121, p < .003 with df = 8). Nagelkerke’s R2 of .585 indicated a moderately relationship between prediction and grouping. Prediction success overall was 90.9% (35.7% for guilty and 99% for innocent).
As can be seen in Table 2, the Wald criterion demonstrated that closeness centrality, p < .001, and degree centrality, p < .01, made a significant contribution to prediction. In the final model the classification was improved by eliminating the regression constant, and the betweenness centrality variable, p = .655, was also dropped due to its high residual probability. The odds ratio for closeness centrality is almost 0 (OR < .0001), indicating that the higher is this measure, the less likely is a guilty verdict. For degree centrality the odds ratio is 1.14, meaning that the higher is this indicator, the more likely is a guilty verdict.
The independent variables in the second regression logistic analysis were the betweenness, closeness and degree centrality indicators, as was the case in the first regression, with the addition of the authority centrality, eigenvector centrality, hub centrality, and out-degree centrality measures. For this logistic regression the Wald statistic was applied in iterative steps to select the significant variables that discriminate the verdict, eliminating the non-significant variables (p ≥ .05). The final model used the out-degree centrality and a constant to predict the verdict variable (see Table 4).
A test of the full model against a constant only model was statistically significant, indicating that the predictors as a set reliably distinguished between guilty and innocent (X2= 7.228, p < .300 with df = 6). Nagelkerke’s R2 of .375 implies that the model explained 37.5% of the relationship between prediction and grouping. Prediction success overall was 91.8% (42.9% for guiltye and 99% for innocent).
As can be seen in Table 4, the Wald criterion demonstrated that only out-degree centrality made a significant contribution to prediction, p < .001. In the final model the classification was improved by eliminating the authority centrality variable, p = .451, betweenness centrality, p = .329, degree centrality, p = .454, eigenvector centrality, p = .854, hub centrality, p = .329, were also dropped due to their high residual probability. Thus, for out-degree centrality the odds ratio is 1.34, meaning that the higher is this indicator, the more likely is a guilty verdict.
Then, by comparing the percentage correct classification in the first and second logistic regression model, we can report that the second one provides a better prediction of the verdict.
Multiple regression analysis for centrality measures predicting the imposed sentence
An alpha level of .05 was used. The means, standard deviations and intercorrelations between the variables was presented in Table 1f.
A first linear regression was then run to predict the imposed sentence based on the three independent variables: betweenness centrality, closeness centrality and degree centrality. As can be seen in Table 3, the analysis of the data using a linear regression technique revealed that the combined predictors explained 51.5% of the variance in sentence in years, R2 = .515, adjusted R2 = .521, F (3.107) = 37.95, p < .0001. Betweenness centrality, β = .003, p < .001, and closeness centrality, β = 1, 118.800, p < .01, were significant predictors of the length of the sentence; degree centrality, β = −.003, p = .905, is not a statistically significant predictor of sentence length. Thus, the regression coefficients indicate that betweenness and closeness centrality increase the length of the sentence.
A second linear regression was run to predict the imposed sentence length. This included the same two sets of centrality measures as independent variables: betweenness, closeness and to degree centrality plus authority centrality, eigenvector centrality, hub centrality, and out-degree centrality. As can be seen in Table 5, the analysis of the data using a linear regression equation revealed that the combined predictors explained 64.4% of the variance in the length of the sentence, R2 = .644, adjusted R2 = .631, F (4, 106) = 49.93, p < .0001. Authority centrality, β = −10, 640, p < .05, betweenness centrality, β = .006, p < .001, eigenvector centrality, β = 21.749, p < .01, hub centrality, β = −35.822, p < .001, were significant predictors of the length of the sentence; closeness centrality, β = 294.998, p = .431, degree centrality, β = −.059, p = .596; out-degree centrality, β = .201, p = .268, were not statistically significant predictors of the imposed sentence length. Thus, the regression coefficients indicate that betweenness and eigenvector centrality increase the length of the sentence while authority centrality and hub centrality reduce it.
Upon comparing the R2 of the first linear regression model (Table 3) with that of the second linear regression model (Table 5), we can then report that the second model provides a better prediction of the sentence in years. In addition, because the error variance in the first regression model, S2 = 1.90, is larger than that of the second regression model, S2 = 1.62, the second model generates errors that are less dispersed. This implies that using the additional set of proposed indicators generates a better prediction.