For a pejorative term without proven clinical utility, psychopathy has generated some very catchy sayings. Some bear little relationship to the research that generated them. ‘Treatment makes psychopaths worse’ is one (see Rice et al 1 ). I fear that without urgent corrective action, ‘Risk assessment doesn’t work for psychopaths’ (see Coid et al 2 ) will be another.
Coid et al compared the ability of three structured risk assessment instruments - the Violence Risk Assessment Guide (VRAG), the Historical, Clinical, Risk Management-20 (HCR-20) and the Offender Group Reconviction Scale-II (OGRS-II) - to predict violent offences by released prisoners in different diagnostic groups. They defined one such group, ‘psychopathic personality’, using a score of over 30 on the Psychopathy Checklist-Revised (PCL-R). For most instruments and groups, Coid et al found moderate levels of predictive accuracy. For the 5.7% of the sample scoring over 30 on the PCL-R, however, no risk assessment instruments performed better than flipping a coin. The authors see major implications for risk assessment. They state that new actuarial tools may be required.
A better conclusion would be that if you define a group using a high score on one instrument that predicts violence, other such instruments will struggle to predict violence in that group. Originally designed to measure a psychological construct, psychopathy, the PCL-R has proved to be one of several instruments that consistently predict violence better than chance (area under the curve (AUC) 0.65-0.75; see Singh et al 3 ). The VRAG and the HCR-20 are others. The other instruments could only have successfully predicted violence among Coid et al’s ‘psychopathic personalities’ if structured risk assessment instruments could be applied serially with increasing success.
We know that they cannot. When Seto 4 combined the results of using instruments sequentially to predict serious offending, also in ex-prisoners, he did no better than he had using one instrument alone. These data, and others suggesting the particular items on a scale are less important than the constructs, such as past behaviour and substance use, that the items represent, 5 have led some to suspect that a ceiling effect may apply to the prediction of violence in psychiatric and other populations. 6 Efforts to improve the accuracy of structured risk assessment instruments are probably better directed at reducing the quantity of missing data than at adding new instruments. 7
I have a wager for Coid et al: try the process in reverse. Select the 5.7% of the sample with the highest HCR or VRAG scores and test whether the PCL-R is predictive in these groups. My five pounds says it will not be, and for the same reason. More is not necessarily better. Or, once you have squeezed the fruit, there usually isn’t much point squeezing it again.