Predicting official recidivism among perpetrators of intimate partner violence against women: A random forest model – Psychiatry AI: Real-Time AI Scoping Review

AI Summary

Random forest model including neuropsychological and demographic measures predicted official IPVAW recidivism; feature selection retained age and specific CPT-3 and WCST indicators.
Model performance was modest: AUC 0.65, accuracy 0.70, sensitivity 0.46, specificity 0.76, Cohen's kappa 0.19, better classifying nonrecidivists.
Neuropsychological functioning offers useful predictive information and warrants further investigation; any eventual use should be supportive rather than stand-alone.

TL;DR (AI Generated)

This study aimed to develop and internally validate a random forest model to predict official recidivism among men convicted of IPVAW, and to examine whether performance-based neuropsychological indicators contributed to prediction alongside psychosocial and criminogenic variables.

We also hypothesized that executive functioning and attentional indicators would emerge as relevant predictors of official recidivism, together with substance misuse, childhood maltreatment, prior IPVAW behaviors, and social support.

RESULTS: The final model retained age, Variability and Commissions from the Conners' Continuous Performance Test-Third Edition, and Conceptual Level Responses, Perseverative Errors, and Perseverative Responses from the Wisconsin Card Sorting Test.

Basic summary

Law Hum Behav. 2026 Nov 9. doi: 10.1037/lhb0000689. Online ahead of print.

ABSTRACT

OBJECTIVE: Predicting recidivism in intimate partner violence against women (IPVAW) remains a major challenge in forensic and clinical practice. This study aimed to develop and internally validate a random forest model to predict official recidivism among men convicted of IPVAW, and to examine whether performance-based neuropsychological indicators contributed to prediction alongside psychosocial and criminogenic variables.

HYPOTHESES: We expected the model to show moderate discriminatory capacity. We also hypothesized that executive functioning and attentional indicators would emerge as relevant predictors of official recidivism, together with substance misuse, childhood maltreatment, prior IPVAW behaviors, and social support.

METHOD: The sample comprised 284 adult men court-mandated to attend a community-based intervention program. Official recidivism was defined as any new IPVAW incident and/or breach of judicial conditions during the 12-month follow-up after program completion. Baseline predictors included age, childhood maltreatment, social support, substance misuse, self-reported IPVAW behaviors, and neuropsychological performance. The random forest pipeline was evaluated using repeated nested cross-validation, comprising 10 outer folds repeated 10 times. Feature selection was conducted within each training fold using Boruta, followed by permutation-based importance filtering.

RESULTS: The final model retained age, Variability and Commissions from the Conners’ Continuous Performance Test-Third Edition, and Conceptual Level Responses, Perseverative Errors, and Perseverative Responses from the Wisconsin Card Sorting Test. Performance was modest (AUC = .65; accuracy = .70; balanced accuracy = .61; sensitivity = .46; specificity = .76; Cohen’s κ = .19), with better classification of nonrecidivists; no psychosocial or criminogenic self-report variables were retained.

CONCLUSIONS: Neuropsychological functioning may provide useful predictive information and warrants further investigation as a potential complement to traditional IPVAW risk assessment approaches. However, the model’s modest performance indicates that further validation is required and that any eventual use should be supportive rather than stand-alone. (PsycInfo Database Record (c) 2026 APA, all rights reserved).

PMID:43002507 | DOI:10.1037/lhb0000689

Document this CPD