Violence Prediction and Risk Analysis - Pacific Institute for the Study of Conflict and Aggression

Evolution, Current Status, and Guidelines

Subsumed under rubrics such as acting out, assaultive behavior, violence, aggression, and dangerousness, the history of attempts within the mental health-law arena to determine the likelihood of future harm to victims has been fraught with difficulty. A key problem had to do with the definition of violence, ranging in the early literature from axe murders to bad check writing. Another basic issue was the typical failure to specify a time period during which the violence could occur. Still another problem was the use of weak or invalid predictor variables (e.g., low self-esteem, hostility towards the examiner, obtaining a high color to movement ratio on the Rorschach Test, enuresis as a child). Choosing limited criterion variables such as psychiatric hospitalization and/or arrests to assess whether violence took place, often without checking with persons who knew the predictee, insured a lower rate of accuracy.

Clinicians often ignored the base rate of violence for the group of which the predictee is a member. Yet base rates are important, if not critical, in rendering violence predictions. Until approximately 1980, a variety of investigative reports showed that practicing professionals, especially those using clinical judgment and clinical opinion as the primary tools to determine the violence potential of the predictee, were more often wrong than correct. The few empirically based measures of the time yielded poor accuracy, so much so that replicated studies showed that one could predict violence more accurately by simply flipping a coin.

Thus the first generation of violence prediction methods, from an applied perspective, consisted of unstructured clinical opinion, a method which did not come close to meeting the rigorous standards demanded in forensic settings and situations. A second generation consisting of structured clinical opinion did not fare much better because of low reliability and validity of the methods employed. A proliferation of research in the 1980s provided the basis for the third generation of violence prediction methods--empirically guided evaluation. A variety of findings were established relating to history (e.g., multiple, recent violence; a past history of different kinds of violence, reinforcing results from violence, child abuse, violent parent or sibling models), opportunity factors associated violence (e.g., recent purchase of a lethal weapon, cessation of psychotropic medication, release into the community) and triggering stimuli (e.g., substance intoxication, breakup of the central love relationship). Many if not most forensic professionals currently utilize empirically guided evaluation methods for a variety of reasons. The methods are based on investigation, are indeed more accurate than clinical opinion, and are usually well-received by referral sources and the criminal court. The webpage on empirically based methods of prediction lists some of these third generation measures.

A fourth generation consisting of pure actuarial measures began to appear in the mid-1990s. These measures are also listed on the webpage on empirically based methods. These scales yielded quantitative degrees of certainty for violent recidivism of various sorts, ranging from three months to 10 years. As such, the actuarial measures represented an advancement in violence prediction. Researchers were no longer vexed by violence representing a low base rate phenomenon. Rather, the problem has been resolved through the use of statistical methods such as Receiver Operating Characteristics (ROC), first developed in communications technology and in signal detection theory in psychophysics, that looks at the tradeoff between hit rate and false alarm rate in predicting violent events. ROCs permit an estimate of the true accuracy of a test, yielding an effect size that is unaffected by different selection ratios and base rates, and has been utilized extensively in the creation of actuarial devices to predict violence. Some test developers allow the forensic professional to clinically adjust the actuarial measure (e.g., due to the predictee developing a sudden debilitating illness, incorporating verified violence which did not result in arrest or conviction), so long as the modifications are slight and do not violate underlying statistical assumptions.

A fifth and most recent generation of violence prediction methods consists of combinations of the above, and is commonly advocated by leading forensic practitioners who routinely predict violence in the course of their forensic work. A variety of actuarial and other empirically based methods are employed along with reporting the findings for each measure in the risk analysis report. Unless the report stops here, the evaluator may opine on the overall risk of violent recidivism within a given time period, citing the basis for the proffered conclusion.

A projected sixth generation may consist of violence prediction measures which attempt to reflect the real-life clinical thinking and the overall, sometimes mind-boggling complexity in individual cases. The classification tree approach of decision analysis rather than a main effects regression approach may provide one such opportunity. As recently proposed, congruent with the findings from the MacArthur Violence Risk Assessment Study, the classification tree approach can employ two decision thresholds for identifying high- and low-risk cases. In this way, conclusions can be directly tied into proposed treatment and risk management strategies.

Practice guidelines for violence prediction in forensic settings and situations which evolved from several decades of investigation include the following:

Be prepared to demonstrate that you are professionally competent to render a violence prediction and that no issues of ethics are involved.
Start with the most valid information first; do not be biased by recent data or by hindsight bias.
Remember that the perpetrator, victim and context must all be analyzed in order to understand a particular violent event.
Retain validated decision rules even when tempted to abandon them for a particular case.
Remember to rule out or account for unintentional distortion as well as deliberate deception on the part of the predictee for both the time of the evaluation and previous violence. You must also take into account possible deception and distortion by witnesses and other third parties if they are utilized in your database.
Never predict violence in the absence of previous significant threatened, attempted or consummated violence. Most extant actuarial devices assume a previous history of violence and hence are really measures of violence recidivism.
The sequence of violence baseline behavior, acceleration stage, the violent act, recovery, return to baseline should be analyzed for all previous acts of violence in order to determine commonalities and differences, as well as trends over time.
Think in terms of base-rates unless using actuarial devices that do not rely on base rates for prediction.
Do not fall prey to illusory associations between evaluation data and supposed violence potential.
Consider triggers to possible future violence based on previous triggering stimuli for an individual.
Consider opportunities for future violence such as availability of weapons, victims, and other factors.
Consider inhibitions to violence based on past inhibitory stimuli and responses for that individual.
List alternative hypotheses to explain exhibited violence and seek evidence for each, especially those contrary to your own interests or the vested interest of your referral party.
Operationalize your conclusions and make them capable of replication by independent examiners.
Present conditional predictions if possible, avoid dichotomous conclusions unless using a classification tree approach.
Limit the temporality of the predictions to avoid the perception of an interminable predictions, which, among other faults, increases your liability.
Specify the factors that formed the proffered conclusions.
Suggest feedback mechanisms to assist future predictors of violence.
Given your database findings, be prepared to fairly and accurately represent the predictee's possible violence potential as well as practical and effective interventions.
Avoid recommending aversive interventions if at all possible. In particular, punitive interventions to violent behavior often lead to aversive cycles between perpetrator, victim and, unfortunately, innocent parties.