Generally, risk assessments in forensic psychiatry are based on a rating scale or a check-list, in which items reflecting risk (and sometimes also protective) factors are systematically weighted and added. The results of one or several such scales are subsequently evaluted by a clinician, who can override or modify the results from the instruments in a final evaluation documented for the court, hospital record or similar. In Sweden, the global risk is assessed as Low, Medium or High (for unspecified crimes during an unspecified time-frame).
Scientific support for the methods are derived from longitudinal studies (prospective or retrospective), comparing assessments to actual outcomes. Assessments of risk for relapse in violence or criminality typically have Areas Under Receiver Operating Characteristics (ROC) Curves (AUC) between 0.70 and 0.80, considered modest or moderate predictive validity. To AUCs, other statistics, such as Diagnostic Odds Ratios (DOR) and Number Needed to Detain (NND), are added for further detail. The overall consensus seems to be that a. these instruments perform better than chance alone and b. that there is no clear evidence that one method is better than the other.
Some aspects are clearly missing in this picture. First, the strongest predictor of future human behaviour is the previous behaviour of the individual. Who among us will be slim and fit a month from New Year's Eve? Probably those who were last year. This is the backbone of every psychiatric risk assessment instrument. The relevant question for science would be "does this instrument/method add predictive value to that derived from the individual's history of previous behaviours?". Why keep repeating that methods are better than random when this can be achieved by common sense alone?
In the whole Swedish population, the risk of being convicted for a violent crime is <4%. After one conviction, the risk for another conviction is still less than 50%, but after two convictions, a majority will go on to reoffend (59%), climbing to 68% after three convictions, >80% after seven and >90% after 11 (Örjan Falk and coworkers, here). These figures may solve the old ethical dilemma whether preventive efforts can be directed at individuals who have not yet offended. It may suffice to do something about those who have.
Second, the time-frame is neglected. Often, the studied time-frame is determined by convenience: researchers take what they can get from data-sets. But clearly, the risk for reoffending varies according to situational factors, such as relapse in substance abuse or unstable social relations. And preventive efforts are best motivated when relapse comes in a reasonable brief time-frame, in some way corresponding to the “normal” prison sentence the offender would have received. It hardly seems justified to allocate large ressources to prevent a threat five years down in time, while 20 years of efforts are clearly appropriate to prevent a murder.
Third, the type of outcome should be specified. A high risk for threats is not easily compared to a medium or low risk for murder.
Therefore, my advice is:
- for legislators and courts: base risk management decisions on what is known about previous violent behaviours.
- for health care staff: assess risk in a here-and-now perspective based on both actuarial and clinical factors with your patient's best interest as goal (which includes not to commit violent crimes!), do your utmost to help and prevent misfortunes of every kind, but desist from speculating about future behaviour.
- for scientists in the field: develop evidence-based treatments that prevent violence and study risk in fine-grained, clinically meaningful models.