Measurement of predictive validity in violence risk assessment studies: a second-order systematic review

The objective of the present review was to examine how predictive validity is analyzed and reported in studies of instruments used to assess violence risk. We reviewed 47 predictive validity studies published between 1990 and 2011 of 25 instruments that were included in two recent systematic reviews. Although all studies reported receiver operating characteristic curve analyses and the area under the curve (AUC) performance indicator, this methodology was defined inconsistently and findings often were misinterpreted. In addition, there was between-study variation in benchmarks used to determine whether AUCs were small, moderate, or large in magnitude. Though virtually all of the included instruments were designed to produce categorical estimates of risk - through the use of either actuarial risk bins or structured professional judgments - only a minority of studies calculated performance indicators for these categorical estimates. In addition to AUCs, other performance indicators, such as correlation coefficients, were reported in 60% of studies, but were infrequently defined or interpreted. An investigation of sources of heterogeneity did not reveal significant variation in reporting practices as a function of risk assessment approach (actuarial vs. structured professional judgment), study authorship, geographic location, type of journal (general vs. specialized audience), sample size, or year of publication. Findings suggest a need for standardization of predictive validity reporting to improve comparison across studies and instruments.