The APA published a Task Force report, “Clinical Aspects of the Violent Individual,” in 1974 (1). Since then, the assessment of violence risk by psychiatrists has assumed increased prominence (2, 3). At the same time, significant changes have taken place both in the contexts in which psychiatrists assess risk and in the techniques that help them do so.

Although violence risk assessment has become more prominent in the last 40 years, assessing the various forms of clinical risk has always been integral to psychiatry. It is a necessary part of providing safe and effective outpatient psychiatric care. It is essential element also of proper decision making around hospital admission and discharge and of providing a safe environment to patients and those who care for them. Risk assessment and risk management are practiced in a diverse array of settings, including emergency rooms, hospital consultation services and outpatient clinics. Like other aspects of psychiatric practice, risk assessment is a potential source of legal ziability.

Psychiatric risk assessments are also used by courts to help them make a range of important decisions, including the involuntary commitment of patients to hospitals and those concerning child custody. In criminal cases, psychiatric judgments inform decisions concerning the placement of mentally disordered defendants prior to trial and the sentences the same defendants receive if convicted (4). Often, as with civil commitment proceedings and the appointment of a conservator, psychiatrists have to provide evidence on risk while continuing to provide care to the patient. As with other aspects of risk assessment in legal settings, the needs of courts can present ethical challenges to clinicians which are different from those encountered in other areas of psychiatric practice.

This document reviews the changes that have occurred since 1974 in the context in which psychiatric risk assessment is conducted, the processes by which risk is assessed, the accuracy that can be expected of psychiatrists seeking to assess risk and the ethics of risk assessment. The general principles outlined here are intended to provide a background to the detailed descriptions of risk assessment and management covering particular patient groups, such as children and adolescents, and particular areas of psychiatric practice, such as psychiatric report writing (5), that have been published since 1974.

Temporal trends in the locus of psychiatric care in the United States have had a profound effect on the context of and demands on violence risk assessment (6). When most psychiatric care was provided in closed psychiatric institutions, much of the focus of violence risk assessment was on the risk of release of patients, risk of increasing freedom to leave restricted settings or the risk of violence perpetration within institutions. Risk assessment was grounded in the knowledge of the patient populations served, the staff capabilities and he hospital medical record—all imperfect, but, arguably, predictably so.

As deinstitutionalization advanced with the depopulation of state hospitals from their peak census of 550,000 in 1955, patients moved from hospitals to the community or, as some have argued, were “trans-institutionalized” to other institutions such as adult care facilities and criminal justice settings (7, 8). Community mental health centers were often unprepared to accept the responsibility for the most impaired patients. Many followed a demedicalized “social service model” with little capacity to provide comprehensive psychiatric evaluations or effective care. The staffing patterns of mental health centers reflected this medical de-emphasis, with marked reductions in psychiatrist and nursing manpower. This new context of violence risk assessment relied on information from a markedly deprofessionalized workforce, found patients in largely poor, often criminogenic and predatory neighborhoods and relied on diffuse decentralized record keeping.

At the same time private and public hospital bed capacity continued to shrink—with state and public hospital capacity for a growing population declining to well under 100,000 beds by the turn of the 21st century. Indeed private psychiatric bed capacity also shrank as mental health managed care approaches dramatically reduced private insurance reimbursement and the demand for hospital admission (9). Hospital beds and reimbursement for all privately insured psychiatric care contracted. For example from 1988 to 1998 the portion of the insurance dollar spent on mental health care was dramatically reduced from roughly 6%–3%. Limited access to hospitalization, fierce pressure to discharge and limit lengths of hospital stays and discharges to limited privately insured treatment all became critical features of this new risk assessment landscape.

The locus of psychiatric care has shifted also. The growing number of people with mental illness in the criminal justice system has expanded the role of the traditional “correctional psychiatrist.” Once focused on evaluation and treatment, psychiatrists in a range of criminal justice settings now increasingly contribute to diversion programs, including mental health courts, and to the mental health treatment of clients on parole and probation.

Starting in the 1960s, while many of the larger psychiatric hospitals were closing, involuntary commitment laws swung away from a “need for treatment” standard to a narrower “dangerousness” standard. In many states, “dangerous to self or others” became the sole or predominant justification for involuntary hospital commitment and, for a variety of reasons, involuntary commitment became a criterion for admission to both public and private psychiatric hospitals (10). Violence risk thus became a means of accessing services for patients, giving clinicians a reason to lower the threshold at which they detected it. At the same time, countervailing pressures toward shorter hospital stays nudged those managing discharges to accept higher levels of risk for community placement. These countervailing pressures undoubtedly exerted a corrosive effect on the maintenance of consistent standards in violence risk assessment.

The California Supreme Court’s decision in Tarasoff (11) confirmed that harm caused by patients to third parties could result in judgments against psychiatrists, creating further pressure for risk assessments to prevent such harm. In the backdrop to these specific psychiatric legal concerns, the growing trend toward medical malpractice and other litigation, regulatory pressures related to disclosure of protected health information and other confidentiality constraints propelled medical practice toward growing risk aversiveness. Violence risk assessments by psychiatrists take place in a social climate where heightened scrutiny over all decision making is now the norm, a situation which is not set to change soon. One challenge for U.S. psychiatry in the first half of the 21st century is to respond to this climate in a way that continues to ensure the wellbeing of its patients.

General

Much of violence risk assessment in psychiatry is invisible, carried out routinely by clinicians in the course of their work. Violence risk is one of many considerations that inform a range of decisions from admission to hospital to the most appropriate form of outpatient care. Even when risk of harm to others becomes a focus of the doctor’s interaction with his or her patient, the principles underlying its assessment are the same as those underlying psychiatric practice more generally. An evaluation will be based on the result of taking a history and examining a patient’s mental state.

Accurate assessment depends on the availability of accurate information. This will usually include information obtained from collateral sources, such as medical records, informants and, where the police have been involved, police reports. Assessments carried out at the point of admission to hospital are of necessity often limited in these respects, and unresolved issues of risk, like other clinical issues unresolved at the time the patient enters the hospital, require continued attention in the course of an admission. Additional investigation, including psychological testing, may be required. Particularly with regard to specialist areas of practice, such as assessing the risk of sexual offending, it may be appropriate to ask colleagues and specialist services to consult on the case.

Sometimes, psychiatric assessments of violence risk are conducted to address an explicit question, such as whether the risk to others can be managed in a community setting. In such instances the most useful assessments are usually conducted in response to questions that are clear, specific and clinically focused. People’s mental states change, as do the circumstances in which they find themselves, and assessments that focus on the short term are less likely to be rendered irrelevant by subsequent changes in either of these. The most helpful risk assessments not only describe the situation at the time of the evaluation but what can be done to mitigate risk in the future.

Correlation and Cause in Assessing Risk

Psychiatrists assessing violence risk evaluate cases in several different ways. First, they look for the presence of factors associated with violence. Some of what clinicians know about the correlates of violence derives from empirical research. While empirical research can increase confidence that a risk factor is associated with violence, it cannot be relied upon to identify all such risk factors. To be confirmed empirically, risk factors have to occur frequently enough to be studied and be capable of being measured. Some reported risk factors, such as Capgras phenomena (12–14), are uncommon while others concern interpersonal relationships whose complexity renders them difficult to define for research purposes (15).

The correlation-based data available in the criminological and psychiatric literature suggest that risk factors for violence act differently in some respects among people with mental disorders compared with the general population. The tendency for violent acts to be conducted by men is still present but less strong, first offenses occur later and the likelihood of acting violently does not fall off so rapidly with advancing age (16). The protective effect of stable relationships may also be less (17, 18), particularly where someone’s social and occupational functioning is poor (19). In other respects, however, the correlates of violent offending in the general population apply also to people who suffer from mental disorders (20).

Thus crimes of violence are more often committed by younger males and recidivism for violent crimes is less than for property crime. Substance abuse is associated with both violent and nonviolent crime. The more serious the crime, generally speaking, the lower the risk of repetition, although the incapacitating effects of long sentences make direct comparisons of reoffending rates problematic. First offenders, on average, fare better, in terms of reconviction, following conviction than do people with extensive criminal records. Unemployment, living in a high-crime neighborhood and having antisocial peers all add substantially to risk (20).

The last 40 years has seen the publication of empirical studies describing in greater detail than was available previously the risk factors that apply in particular settings and patient groups. Those that have been identified in general psychiatric settings (17, 21, 22) are listed in Table 1.

In other settings additional factors may be important. In emergency rooms and inpatient units an aggressive attributional style (hostile, suspicious, or believing others intend harm), command hallucinations to harm others and a poor therapeutic alliance have been implicated (23–25). Some paraphilias are risk factors for sexual offending (26).

A second way in which psychiatrists assess risk, in addition to looking for risk factors, is by combining their understanding of the patient’s personality, symptoms and environment with their understanding of the likely causes of violence. Where someone suffers from persecutory delusions that concern their spouse, for instance, there will usually be available no empirical data from research conducted on samples of similar patients demonstrating a correlation between continued cohabitation and violence. Yet the clinician’s understanding of the likely causes of violence may still allow him or her to conclude that continued cohabitation presents a risk (27).

Pollock offers one description of the processes involved:

The skillful clinician assessing dangerous behavior formulates and tests a series of clinical hypotheses to define patterns of violence in the individual’s history. Once defined, these patterns can be applied to the explanation and prediction of violence in that individual (28 at 105).

Approaches to risk assessment based on explanations of this type seem to rely heavily on induction, because they require the clinician to draw conclusions about the future from past observations. Future conditions will never exactly mimic the conditions in which behavior has occurred in the past, yet the circumstances of other episodes of violence, whether in the patient’s case or more generally, will usually be relevant and sometimes be critical. Notwithstanding the uncertainty inherent in this process, one task of risk assessment is to determine the relevance of past patterns.

Clinicians trying to work out what might cause future violence are guided also by the understanding of patterns of behavior that they develop in their training and through clinical practice (29, 30). Claims that “causal” ways of thinking are better than correlation-based ones at predicting rare events (31) have not been confirmed by empirical research. Instead, the persistence of causal approaches when clinicians think about risk may relate to the fact that many of the other judgments required in medicine are causation-based also: establishing why someone has symptoms, for instance, or deciding which further investigations are needed to complete an evaluation. Because clinical practice requires each of these judgments to be integrated into a single plan, it may be that clinicians find it helpful to use the same causal heuristic in assessing risk that they use in other aspects of their work.

Structure in Risk Assessment

Both correlation- and cause-based approaches to risk assessment can be structured. Structure can be provided in more than one way. Actuarial instruments such as the VRAG (32) formalize the process by which the simultaneous presence of more than one correlate of violence increases the perception of risk. They do this by rating variables such as poor school adjustment and alcohol problems and combining these mathematically to generate an overall score or category. A different type of instrument relies on “structured professional judgment.” The HCR-20 (33), for instance, encourages the clinician to assess the relevance of a list of preidentified variables but also to take into account other information, including factors he or she considers unique to the case, before allocating a case to a risk category.

Structured approaches are sometimes treated as a distinct category within risk assessment but are better seen as differing from unstructured methods in degree rather than kind (34). Even in the absence of a structured instrument, clinicians use structures derived from their professional training to take a history and to examine the patient. Some of these structures appear in practice guidelines. Structure offers the same advantages to risk assessment that it offers to other areas of clinical practice. It is a means of integrating and communicating information (35). It can be a useful aide memoir, particularly when the clinical question being addressed is unusual (some risk factors apply particularly to sex offenses, for instance: see page 3, column 2). Learning how to structure the clinical approach is a key aspect of clinical training (36).

Examples of structured instruments for the assessment of violence risk are listed in Table 2.

The time taken to complete these instruments varies with the amount of material reviewed. Further information on their use is provided in the references cited in the Table.

Although the majority of research on the Static 99 has been in relation to sexual offending, all of the instruments in Table 2 have been tested in a range of settings and found to predict violence in a range of patient groups. Different instruments have been developed to help assess risk in particular treatment settings, such as inpatient units (50), and specific patient groups (51). The number and scope of these structured approaches to risk assessment has increased substantially since the APA Task Force Report, “Clinical Aspects of the Violent Individual,” was published in 1974 (52, 53). One recent review counted 126 (54).

When structured approaches combine those variables into risk assessment instruments they have the benefit of allowing the evaluator to make the results of his or her evaluation transparent: with the necessary information on how items were completed and combined the reader can see how the conclusions were arrived at. Early suggestions that structured approaches are also more accurate (55, 56) were initially criticized on the grounds that many of the predictions included in early comparisons did not concern violence (57). Subsequent research seemed to confirm, however, that structured approaches perform better, on average, than unstructured ones when violence risk is looked at separately (58), at least when the follow-up period covers months or years. Recent reviews have reached the same conclusion (53).

This combination of transparency and empirically demonstrated accuracy has contributed to an increased use of structured instruments since 1974, particularly by specialist services. The degree to which structured approaches will be used in the future will ultimately depend on whether they benefit care. Several issues warrant further investigation. First, the evaluations of risk that are required in hospital and outpatient practice frequently concern hours, days or weeks. There is a shortage of empirical data to indicate whether the long term accuracy of structured methods is matched by their accuracy over these shorter periods (59). Some of the reasons for the relative lack of research in this area are discussed in Section 4.

Second, researchers seek to limit the amount of missing data in follow-up studies. Clinicians have little option but to work with a lot of it. Their response to not having a collateral account of someone’s criminal record, for instance, is presumably unstructured, yet the performance of clinicians in the emergency room seems not to be very different from that of structured instruments (60, 61). We do not yet know to what extent the performance of structured instruments is adversely affected by the absence of some of the information necessary to complete the items. Third, although the number of validations has grown rapidly in recent years, a clinician will not always have available an instrument that has been shown to be effective in measuring risk in the relevant patient group.

Fourth, events can make the results of using a rating scale no-longer applicable. Physical incapacity (the socalled “broken leg exception” (62) to using a score as an indicator of risk) is an unusual, though often quoted, example. More common events, such as placement in a supervised setting, can be equally important. Finally, structured instruments generate a risk category (typically, low, medium and high) or a score. Clinical risk assessments address aspects of clinical management, for instance whether someone’s violence risk is sufficiently well managed for them to move to a supported apartment or whether it requires that they be admitted to the hospital. There is not yet available a tested and reliable means by which a score can be applied to decisions such as these. The process is further complicated by the fact that placement decisions are usually influenced by many factors, not just risk. Not all services and not all neighborhoods have the resources to support the same type of psychiatric provision, making the treatment implications of a particular score or category still more difficult to describe consistently.

Prediction and management

This section reviews the accuracy of violence risk prediction in mental health settings. Making predictions is not the same as managing violence risk. Managing violence risk involves a range of activities, from prescribing appropriate treatment to ensuring a safe environment, that have little to do with prediction. Studies of predictive accuracy are nevertheless important. They allow new approaches to violent risk assessment to be tested and offer some insight into the likely future role of risk assessment in clinical practice.

Studies of the predictive accuracy of all types of violence risk assessment face obvious methodological difficulties when they are conducted in treatment settings. The accuracy of a prediction is measured in terms of whether the subject of that prediction engages in violence or not. But when a clinician determines that there is a high risk of violence, there is an ethical and clinical obligation to intervene. To the extent to which the intervention is successful, the predictive accuracy will be diminished and the original assessment of high risk will appear to be a “false positive.” (63) Most examinations of predictive validity, therefore, are follow-ups of groups of patients who have already been either admitted to or discharged from the hospital. This section reviews: i) what has been learned about the predictive accuracy of risk assessment since the American Psychiatric Association’s Task Force report of 1974, ii) the practical implications of current levels of accuracy, and iii) the degree of deviation from this level of accuracy in particular situations and with particular patient groups.

Predictive Accuracy of Psychiatric Violence Risk Assessment

The accuracy of validated structured and clinical approaches to the prediction of violence in people suffering from mental disorders, as measured by the “Index of Effectiveness” (which derives from sensitivity and specificity), improved between 1970 and 2000. The improvement is shown in Figure 1 (following page).

The improvement was limited and characterized by a wide range in the published results.

Research since 1974 has reduced the number of patient groups for whom no data are available regarding the accuracy of prediction and it is now clear that comparable levels of predictive accuracy occur in a wide range of circumstances. Thus the HCR-20 has an AUC of 0.72 for men and 0.77 for women in nonforensic populations (65). The AUC, or area under the Receiver Operating Characteristic curve, is a measure of the predictive accuracy of a screening test, in this case a screening test designed to detect violent offending. The AUC varies between 0.5 and 1 and is the probability that a randomly selected offender will have a higher score on the screening test than a randomly selected nonoffender.

The accuracy of structured methods also generalizes outside of North America: a Danish validation of the VRAG reported an AUC of 0.73 (66). Newer instruments perform similarly to older ones (67). A replication of the iterative classification tree, the algorithm on which Classification of Violence Risk (COVR) software (48) operates, was published in 2005 and demonstrated an AUC of 0.63 and 0.70, depending on the outcome measure (68).

The Practical Implications of Current Levels of Accuracy

Translating these figures into numbers that are clinically meaningful is difficult, not least because the clinical judgments that they might inform, such as whether or not to admit a patient to the hospital, derive from many considerations, not just one. One approach is to ask, if a particular instrument was used as a screening test and those identified as likely to be violent were not discharged, over any given period how many patients would need to be detained to prevent one unwanted act?

This statistic, the Number Needed to Detain (NND), is the inverse of positive predictive value (PPV) and analogous to “number needed to treat”. It derives from sensitivity, specificity and base rate (64). The VRAG has a sensitivity of 0.73 and a specificity of 0.63 (69). Used as a screening test where the base rate of violence is 10% and where, as a result, an unselective approach would lead to the detention of 10 people in order to prevent one from acting violently, the VRAG would require the detention of 5 people to achieve the same end.

Assuming this level of predictive accuracy (in terms of sensitivity and specificity) from an instrument or a clinician making a prediction, the NND rises as the base rate of violence in the population falls (see Figure 2). This means that the number of mistakes is higher when the acts that are sought to be prevented are unusual, as is the case with serious violence in most patient populations. At the base rate reported in the Epidemiologic Catchment Area (ECA) study, where 17% of the sample self-reported acting violently in the previous 12 months, the NND to prevent an act of violence is 3.5. Violence as defined by the ECA study did not require injury. When injury is required by a definetion, the base rate of violence falls. In the CATIE study, the 6-month prevalence of assault with a weapon or causing serious injury was 3.6% (71). Here the NND at a sensitivity of 0.73 and a specificity of 0.63 is 15.

This is a different phenomenon from another noted for many years, that whereby clinicians tend to “overpredict” violence. Overprediction probably stems, not from knowledge of base rates, but from clinicians viewing different types of mistakes differently: they would rather keep someone in the hospital who would not act violently than release someone who would. Apart from having less serious consequences, admitting someone who would not act violently is amenable to review and correction. Because NND is usually calculated by treating both types of error equally and is lowest when this is done, any tendency to overpredict in clinical settings will raise the NND. In everyday practice, the NND is probably higher than Figure 2 indicates also because admission to hospital, whether voluntary or compulsory, prevents only some violent incidents, not all of them. Inpatient violence remains a significant clinical problem.

An additional consequence of the relationship between NND and prevalence is that at low base rates prevalence becomes more important, and the psychometric qualities of a test less important, in determining the number of correct predictions. At the 3.6% base rate in the CATIE study a 20% increase in sensitivity, all other things being equal, reduces the NND from 15 only to 13.

Particular Situations and Particular Patient Groups

A single approach to risk assessment can have different predictive accuracy in different patient populations (72, 73), raising the possibility that higher levels of accuracy might be achieved if risk assessments were restricted to particular situations. A number of studies have investigated whether accuracy varies with the time over which violence is sought to be predicted and, in particular, whether short term predictions are more accurate than long term ones (74–76). The question is complicated by methodological factors: long term follow-up studies detect more violence and some measures of predictive accuracy, such as positive predictive value, rise with the base rate of violence in the population. Studies that measure predictive accuracy using techniques that are unaffected by base rates, such as the AUC, suggest that, at least for periods longer than three months, shorter term predictions are not more accurate than longer term ones (59).

From a clinical perspective, more important than predictive accuracy over months or years is the accuracy of risk assessments that cover the coming days or weeks (77, 78). There are no data from community settings to describe the accuracy of such assessments. This is partly a consequence of the relatively low incidence of serious violence over short time periods and the consequent need for very large samples in order to demonstrate statistically significant correlations. Hospital studies suggest that for inpatient settings the predictive accuracy of structured approaches is maintained when the risk period is reduced to weeks or months. They suggest also that symptoms of mental illness, including delusions and hallucinations, are then of particular importance (79).

It is possible also that particular types of offending might “breed true”, and that an individual’s behavior on one occasion can be used to predict the nature and severity of future violence. In principle, this could provide an opportunity to manage risk by identifying particular situations and potential victims. Criminological research demonstrates little of the “specialization” in offending that would allow this, however (80–82), and patients who offend after discharge from maximum secure hospitals seem to display the same differences in conduct over successive offenses as the general population (83). Finally, it is not clear that the limited degree of specialization that does exist helps. While sexual offenders specialize more than most (84), the accuracy of risk assessments in this population is similar to that for other people who commit crimes (81).

Assessing the risk of violence is a common clinical task, particularly in the context of hospital admission and discharge. These areas of psychiatric practice can present distinct ethical challenges because they require clinicians simultaneously to take into account interests that sometimes conflict, such as encouraging patient autonomy and ensuring the safety of others. In many settings these challenges coincide with others that are also integral to the provision of quality care. In applying civil commitment criteria, for instance, clinicians are expected to take into account the wellbeing of both the patient and other people and to do so without unnecessarily restricting the patient’s freedom.

In some instances, however, violence risk assessments take place in settings different from those in which psychiatry is usually practiced. When risk assessments form part of court ordered evaluations they are often conducted without the consent of the person being evaluated and outside a therapeutic relationship. Even here, a risk assessment will sometimes be the first step toward someone receiving necessary treatment (85). At other times, as is the case in evaluations for sexually violent predator (SVP) commitment, the potential benefit to the individual will be less clear. Violence risk assessments are also requested as part of attempts to improve the health of high risk groups. These attempts include the making of residential services or financial support dependent on medication compliance. Here the ethical situation is complicated by the controversial nature of the balance that is being sought between clinical benefit and coercion.

The ethical dilemmas facing psychiatrists practicing in legal settings remain the subject of extensive debate (86–92). Important factors for the clinician to consider when asked to assess risk of harm to others in a legal environment include whether there is evidence that the person suffers from a mental disorder, how likely it is that the person will be able to receive treatment following the evaluation and whether this is a case where mental health variables, such as delusions, seem to contribute significantly to the possibility of violence. It will also usually be relevant whether anything can be done medically to ameliorate the risk (91, 93). Ethical guidelines do not preclude evaluations that may contribute to an outcome, such as a longer sentence, that the person being evaluated regards as unfavorable provided that the purpose of the evaluation has been explained to him or her in advance (94, 95). The task of showing proper respect for the person being evaluated demands careful reflection, however, and cannot be achieved solely by adherence to rules such as this (96, 97). The American Academy of Psychiatry and the Law publishes guidelines for forensic psychiatric practice that apply to risk assessment in legal settings (98).

In more traditional clinical settings a number of steps will usually assist the clinician seeking to meet his or her ethical obligations in the practice of violence risk assessment. First, he or she should be aware of the relevant literature, including that relating to the utility and limitations of structured instruments designed to help assess violence risk. Second, the clinician should conduct an appropriately thorough assessment, gathering all available relevant information. Third, he or she should approach the risk assessment with objectivity and honesty, showing proper respect for all of the parties involved. Finally, an ethical approach to risk assessment requires the clinician to demonstrate an appropriate degree of modesty regarding the level of accuracy that can reasonably be expected of psychiatric assessments of violence risk.

The increased prominence of violence risk assessment in psychiatry since the APA published its Task Force report in 1974 has not altered the need for risk assessment to be practiced in a way that respects other aspects of clinical care. Nor has the development of new methods of assessing risk removed some longstanding challenges faced by those seeking to manage risk. Predicting human behavior is in many respects difficult, and predicting violent behavior is not an exception to this rule.

This does not mean that no improvements to current practice are possible. Research identifying dynamic risk variables that are responsive to intervention has the potential to benefit patients (99, 100). It may also be that, in future, structured approaches will allow clinicians to develop treatment approaches that more effectively reduce risk and that changes measured using such instruments will be used to guide treatment and monitor progress (101, 102). At present, the principle challenges to the more widespread adoption of structured methods relate to clinicians’ need to respond to changing circumstances and a changing clinical picture, as well as to the ability of instruments to provide information in a form that can be integrated into the complex judgments that inform clinical management.

While research will continue to advance, the pervasive influence of base rates means that, at the base rates of violence seen in most clinical settings and for the foreseeable future, no technique will be available to identify those who will act violently that will not simultaneously identify a large number of people who would not. Ultimately, the criterion governing the adoption of new approaches to violence risk assessment in psychiatry should be the same as that governing the adoption of new approaches elsewhere in medicine: whether they benefit clinical care.