Nodewise Predictability in Cross-Sectional Data Does Not Outperform Mechanical Totals in Predicting Sexual Reoffending.

Abstract

This study compares the predictive accuracy of sexual reoffending using dynamic risk factors' sum score (mechanical totals) and nodewise predictability, a model accounting for their interrelationships. Dynamic risk factors of North American men (N = 5,315) were measured by the STABLE-2007. The area under the curve (AUC) of both methods was determined by splitting the dataset at a [20:80] ratio, repeated over 300 iterations with random training and test samples. Mechanical totals' predictive accuracy outperformed nodewise predictability (AUC_mechanical = 0.67, SD = 0.04; AUC_nodewise = 0.50, SD = 0.03; t[299] = 80.2, Cohen's d = 4.63, p < .001). This suggests that the conventional approach to predicting sexual reoffending is superior to a model considering dynamic risk factors' interrelationships at the group level. Future research should explore whether nodewise predictability's accuracy improves by incorporating temporal effects, subject variances, and centrality indices of individualized networks.