Multiple-Models Approach to Violence Risk Assessment Among People With Mental Disorder

• NCJ Number: 205951
• Journal: Criminal Justice and Behavior Volume: 31 Issue: 3 Dated: June 2004 Pages: 324-340
• Author(s): Steven Banks; Pamela Clark Robbins; Eric Silver; Roumen Vesselinov; Henry J. Steadman; John Monahan; Edward P. Mulvey; Paul S. Appelbaum; Thomas Grisso; Loren H. Roth
• Date Published: June 2004
• Length: 17 pages
• Publication Format: Article
• Publication Type: Test/Measurement
• Language: English
• Country: United States of America

Annotation

This article reports on a demonstration of how multiple actuarial models can be combined to produce risk assessments for violence that are significantly more accurate than a single actuarial model alone; by using many different actuarial models, more participants can be categorized into groups with exceedingly high and low rates of violence.

Abstract

The approach described combined the results of five prediction models generated by the iterative classification tree (ICT) methodology developed in the MacArthur Violence Risk Assessment Study. From the data collected in the MacArthur Violence Risk Assessment Study, the ICT method generated two different risk-assessment models. One model was unrestricted in terms of eligible risk factors and achieved an area under the ROC (receiver operating characteristic) curve of .82 (Steadman et al., 2000). Another model restricted eligible risk factors to those commonly available in hospital records or capable of being routinely assessed in clinical practice; it yielded an area under the ROC curve of .80 (Monahan et al., 2000). In expanding a two-models approach to a multiple-models approach, the principal methodological challenge was the combining of the results of the various models. This challenge was addressed by constructing 10 ICT models, each of which featured a different risk factor as a starting point in building the tree. Each of the 10 models was developed by using the 106 clinically feasible risk factors described in Monahan et al. (2000). The risk groups produced by each of the 10 models was divided into 3 categories: low-violence risk, average-violence risk, and high-violence risk. A composite risk score was then computed for each participant by summing across the 10 models. Ultimately, it was found through stepwise logistic regression analysis that only 5 of the models were required to achieve the same results as with the 10 models. Although the multiple-models approach is more complex in its methodology than using only one model, it minimizes the problem of data overfitting that can result when a single best-prediction model is used. 7 tables, 1 figure, 2 notes, and 17 references