Violence risk assessment: A brief review, current issues, and future directions.

Abstract

A little over 11 years ago, the Journal of Interpersonal Violence invited its contributors to provide a brief response to questions for a special edition on the 20th anniversary of the journal (Mills, 2005). One of the questions was, What is the most important thing we need to learn in the next 10 years? Given my opportunity to review the literature for this article and the relative time that has passed, I revisited my answers to those questions, and they will serve to frame this brief review. Topics I identified then included (a) the identification that unstructured clinical methods were grossly inaccurate in risk assessment and that there was an emerging debate between structured clinical judgment (now referred to as structured professional judgment [SPJ]) and actuarial approaches to risk assessment, (b) the need to advance our knowledge of risk communication and risk management within the risk assessment process, (c) the importance of measuring dynamic change in risk assessment, and (d) the need to recalibrate frequently used risk instruments to meet specific population needs.The field of criminal and violence risk assessment has grown considerably over the past decade, as evidenced by a surge in risk assessment instruments. In a study involving 2,135 mental health professionals drawn from around the world, respondents reported over 400 instruments being used in violence risk assessment within the past 12 months of the survey (Singh et al., 2014). Among the most commonly used methods, approximately half were actuarial and half were SPJ. Although it is self-described as a measure of a personality construct, also included in the list was the Psychopathy Checklist-Revised (PCL-R; Hare, 2003). During these intervening years, the field has benefited from numerous meta-analytic investigations related to criminal and violence risk assessment that have served as sign posts of the underlying scientific findings over the years. For example, Yang, Wong, and Coid (2010) compared nine commonly used risk instruments and found they predicted recidivism equally well for the most part (with the exception of Factor 1 of the PCL-R); however, there was considerable heterogeneity between studies. For example, only 25% of the heterogeneity in the effect sizes found for the analysis was attributed to differences between the risk instruments whereas the majority of variance was accounted for by the methodological differences between the studies. In another meta-analysis that compared a number of instruments, Campbell, French, and Gendreau (2009) examined 88 studies and found that the effect sizes related to various instruments generally did not significantly differ. The Violence Risk Appraisal Guide (VRAG; Quinsey, Harris, Rice, & Cormier, 2006) was the measure that was most predictive of violent recidivism and those measures containing dynamic risk factors performed better than measures based solely on static risk factors.In their systematic review, Singh, Grann, and Fazel (2011) reported that the median area under the curve values for many of the commonly used risk instruments examined were well within the moderate to high range of predictive accuracy. A more recent meta-analysis by Fazel, Singh, Doll, and Grann (2012) included 73 samples and found that measures that were specifically designed to predict violence were more accurate in their predictions than instruments designed to predict general or sexual reoffending. Taken together, these meta-analyses indicate that the predictive accuracy for most instruments are generally in the moderate to high range and are similar across instruments and samples.Other meta-analyses have examined instruments related to specific types of violence. For example, in a meta-analytic study of risk assessment methods used to predict recidivism among sexual offenders Hanson and Morton-Bourgon (2009) found that actuarial instruments were more accurate than SPJ, and further that clinically adjusting actuarial risk resulted in significantly poorer predictive accuracy. …