Violence Risk Assessment: What It Can Predict and Where It Fails
Author: Ukrainian Psychological Hub · Published: October 11, 2026 · Editorial Policy
Violence risk assessment is a structured evaluation of the likelihood, nature, circumstances, and possible consequences of future interpersonal violence over a specified period. Clinicians, forensic psychologists, correctional professionals, and multidisciplinary teams use it to organize evidence and inform prevention, supervision, treatment, discharge planning, and sometimes legal decisions. The central scientific question is whether an assessment improves decisions for a defined person, population, setting, outcome, and time horizon. A result that ranks people by relative risk does not establish whether any particular individual will commit violence.
Research supports useful but limited prediction. A systematic review of 50 studies involving 10,460 people discharged from forensic psychiatric hospitals found that pooled area-under-the-curve estimates for violent recidivism were approximately 0.64–0.72 across instruments with sufficient independent validation; 49 of the 50 studies were judged at high risk of bias. Calibration and other practically important measures were seldom reported. The evidence therefore supports structured assessment as an aid to formulation and decision-making, with explicit attention to uncertainty and potential harm from errors. Forensic psychiatric systematic review, The Lancet Psychiatry.
What Is Violence Risk Assessment?
A violence risk assessment estimates the risk of a specified form of physical aggression or violent offending under identifiable circumstances. A competent assessment names the behavior being considered, the prediction period, the population to which the evidence applies, the information available, and the purpose of the opinion. “Risk of an aggressive incident on an inpatient ward during the next 24 hours” is a different research and practice question from “risk of a new violent conviction during the next three years after release.”
Risk assessment is an evidence-integration process. It commonly brings together documentary history, an interview, observed behavior, corroborated collateral information, relevant clinical and social circumstances, empirically supported risk and protective factors, and—when appropriate—a structured assessment instrument. The output may be a structured professional formulation, a relative-risk category, a validated numerical probability estimate, or a combination of these. These outputs must be labeled according to what they actually represent.
The American Psychological Association’s Specialty Guidelines for Forensic Psychology recommend appropriate methods, adequate scientific foundations, corroboration where feasible, attention to individual differences, and clear separation of observation, inference, and conclusion. The broader psycholegal process is explained in Forensic Psychological Assessment: Purpose, Methods, and Limits.
The question that must be answered first
An assessment cannot be interpreted until the decision problem is specified. Who is being assessed and by whom? What counts as violence? Against whom, in which environment, and during what follow-up period? Does the question concern an imminent inpatient event, community violence, a repeat conviction, intimate partner violence, or a targeted threat? What prevention or supervision options are available? These details define the assessment's relevance and the costs of false alarms and missed events.
A violent charge, a police record, a witnessed aggressive act, an allegation, a clinical symptom, and a proven legal finding represent different classes of information. Arrest or conviction records may be affected by detection and justice-system practices, whereas behavioral reports can vary in completeness. Reliable assessment identifies what each item shows, whether it is corroborated, and how measurement limitations affect the conclusion.
What Can Violence Risk Assessment Predict?
Differences in risk between groups
The best-supported function of many established instruments is discrimination: people who experience the specified outcome during follow-up tend, on average, to receive higher baseline scores or ratings than people who do not. Discrimination can be above chance without being strong enough to justify an individual restriction of liberty. It also cannot show why violence occurred.
A major BMJ meta-analysis of 73 samples found substantial variation in predictive accuracy. Among its violence outcomes, the median positive predictive value was 41%, whereas the median negative predictive value was 91%. These are historical pooled descriptive results across different instruments, settings, and follow-up periods, not universal accuracy rates for today's assessments.
Imminent aggression in a particular institution
Short-term tools addressing observable warning signs over hours or days can be more discriminating in their specific clinical environment than instruments addressing years of future community behavior. A systematic review of forensic psychiatric hospital studies reported median AUC values of 0.83 for imminent instruments and 0.68 for longer-term instruments in its studied samples. This comparison does not establish that either tool can be transported to a different population or used for a different endpoint.
A separate systematic review and meta-analysis of imminent aggression assessment identified research on instruments such as the Brøset Violence Checklist and Dynamic Appraisal of Situational Aggression. Their intended task is repeated monitoring of near-term behavior in relevant settings. Predicting a ward incident is not equivalent to predicting an individual's long-term capacity for violence.
Identification of people at relatively low observed risk
In some validated populations, assessment can help identify a large group with a low rate of recorded violent outcomes during the defined interval. That may inform decisions about proportionate services, subject to the consequences of missing an event. A validation study of the Oxford Mental Illness and Violence tool, OxMIV reported a negative predictive value greater than 99% at a specified threshold in an external validation sample. Its very low event rate was essential to interpreting that result; the estimate does not transfer to other settings automatically.
The central benefit is often a clearer basis for support and reassessment. Low assessed risk means a lower estimated probability within defined assumptions. It never guarantees safety; a high score, similarly, is never a declaration of inevitable violence.
What the Statistics Actually Mean
Discrimination and the AUC
The area under the receiver operating characteristic curve, or AUC, summarizes ranking performance across thresholds. An AUC of 0.70 means that, in the sample and outcome studied, an individual who later had the event would receive a higher predicted score than an individual who did not in about 70% of randomly formed outcome-positive/outcome-negative pairs, with ties handled by the calculation. It does not mean the model is correct about 70% of all people, and it is not a probability that a named person will offend.
The 2023 forensic psychiatric review found that discrimination was widely reported while calibration, sensitivity, and specificity were frequently absent. Practitioners therefore need the actual outcome rates and operating characteristics for their setting, rather than an isolated AUC or a publisher's generic label of predictive validity.
Calibration: does a stated probability match observed events?
Calibration asks whether predicted probabilities agree with observed outcome frequencies in an appropriate population over the stated period. When a model estimates a 10% one-year risk for many comparable people, good calibration requires the observed frequency in that probability range to be around 10% over that year. A model can rank people reasonably well and still overestimate or underestimate absolute risk across the board or within specific subgroups.
Calibration must be checked after meaningful changes in location, time, selection rules, service provision, offense recording, and base rates. A risk category such as “high” does not necessarily contain an empirically validated numeric probability. Never convert an ordinal judgment into a percentage merely because a reader expects a number.
Sensitivity, specificity, and the error trade-off
Sensitivity measures the proportion of people with the later observed outcome classified as positive at a selected threshold. Specificity measures the proportion without the outcome classified as negative. Raising the threshold often reduces false positives at the cost of missing more true outcomes; lowering it typically catches more events while incorrectly flagging more non-events. The suitable balance depends on what action follows a result, the burden that action imposes, and whether its benefit has been demonstrated.
Positive predictive value, or PPV, is the fraction of those flagged who later have the outcome. Negative predictive value, or NPV, is the fraction of those not flagged who remain outcome-free. Unlike sensitivity and specificity, PPV and NPV change markedly when the outcome becomes more or less common in the assessed population.
Why the base rate changes everything: a worked example
Consider an explicitly hypothetical sample of 1,000 people, of whom 100 would have the specified outcome within a year. Suppose a method has 70% sensitivity and 80% specificity in that same population. It would flag 70 of the 100 people with the later outcome and miss 30. Among 900 who do not have the outcome, it would incorrectly flag 180 and correctly classify 720. Thus 70 of 250 positive classifications would be true positives, giving a PPV of 28%; 720 of 750 negatives would be true negatives, giving an NPV of 96%.
These figures are an arithmetic teaching example, not the performance of any named instrument. They show why “high risk” can generate many false alarms even when a method has useful discrimination. If the outcome rate falls while sensitivity and specificity remain unchanged, the proportion of flagged people who actually experience the event also falls. Legal and clinical readers should request absolute counts and denominators, not just percentages.
Confidence intervals and uncertainty
Predictive performance estimates are subject to sampling error, study design, missing information, outcome misclassification, changes in treatment, and changes in population. Reports should provide uncertainty intervals, clear definitions of the endpoint, external validation, the chosen threshold, and the prevalence of the event. A narrow confidence interval around a biased estimate does not make the underlying model appropriate for an individual case.
How Is a Violence Risk Assessment Conducted?
1. Define purpose, mandate, setting, and time horizon
The assessor identifies the referral question and distinguishes the clinical task from any legal question. Examples include near-term aggression management, planning supervised community discharge, deciding what services may reduce identified needs, or answering a court's authorized forensic question. An assessor's clinical opinion cannot independently determine legal responsibility, dangerousness under a statute, commitment eligibility, or a release order.
2. Gather and test information from multiple sources
Sources may include structured interview, health records, verified offense history, contemporaneous observations, collateral reports, relevant substance-use history, treatment response, and environmental information. Contradictory evidence should be evaluated rather than silently harmonized. Important allegations should retain their status as allegations; missing records and refusal to participate must be reported accurately without treating either as proof of dangerousness. The APA forensic guidelines specifically emphasize multiple sources and disclosure of significant information limits.
3. Identify risk factors, protective factors, and their relevance
A risk factor is a characteristic statistically associated with an outcome in relevant research. It may be historical, current, or situational. Protective factors include credible resources or circumstances that may reduce opportunities for violence or help interrupt a plausible pathway, such as reliable support, safe housing, engagement with services, and effective de-escalation arrangements when demonstrably relevant. Their presence is not an arithmetic guarantee of safety.
Historical indicators, such as a documented pattern of prior violence, may improve discrimination but cannot be changed. Dynamic circumstances, such as current intoxication, escalating interpersonal conflict, or loss of supervision, may change and become management targets. Evidence for the predictive meaning of a change score is a separate question from evidence that a baseline score predicts an outcome.
4. Choose a method validated for this task
An actuarial instrument combines predefined information using a published scoring or weighting rule. Structured professional judgment, or SPJ, provides a systematic process for considering relevant factors, their relevance to the person, plausible violence scenarios, and responses. Unstructured judgment relies more heavily on unaided clinical interpretation. Method selection depends on validity for the population, target behavior, horizon, available records, evaluator training, and purpose.
These methods have their own dedicated comparison in Actuarial vs. Structured Professional Judgment in Violence Risk Assessment. The present article addresses the overall validity and limits of risk assessment rather than duplicating the full instrument-by-instrument comparison.
5. Formulate plausible scenarios and practical interventions
A useful formulation specifies potential patterns of future behavior, the situations that could elevate concern, warning signs and uncertainties, and possible protective steps. An example might describe escalating threats in the context of a specific conflict together with access to a potential victim and gaps in support. Such a scenario is a conditional professional judgment supported by identified evidence, not a finding that an attack will occur.
Interventions may include treatment for an identified condition, better continuity of care, conflict de-escalation, targeted social support, safety planning, monitoring or environmental modifications. Their proportionality should be assessed independently of the score. Where liberty, housing, employment, custody, or community access may be affected, the justification for restrictions requires special scrutiny.
6. Communicate uncertainty and plan review
The resulting report should state who was assessed, what evidence was available, why the selected tool was suitable, which factors were observed versus inferred, the exact outcome and period under consideration, significant limitations, and the meaning of any category or probability. Updates should follow material changes in circumstances. An actionable plan identifies who is responsible for review and what event triggers it. See Forensic Psychological Reports: How Opinions, Data, and Limitations Are Communicated.
Which Instruments Are Used?
HCR-20 and structured professional judgment
The Historical, Clinical, Risk Management-20 family is widely associated with structured professional judgment in violence risk assessment. It organizes relevant historical, current clinical, and future management information and can inform formulation and scenario planning. Its versions and their intended uses matter; research on HCR-20 version 2, numerical totals, or a subscale cannot automatically establish the validity of all judgments and management stages in version 3.
A critical review of 206 HCR-20 papers noted that research frequently evaluates summed scores or predictive discrimination while providing comparatively little direct evidence that the full structured professional judgment process reduces violence. A 2026 systematic review of longitudinal HCR-20 properties reported that dynamic scores can change with treatment, often only modestly, and found little evidence that score changes reliably predict subsequent violent recidivism. Score change, treatment progress, and risk reduction therefore demand distinct evidentiary claims.
VRAG and other actuarial tools
The Violence Risk Appraisal Guide, or VRAG, is associated with actuarial prediction research in specified populations. The 2023 review found a pooled AUC of about 0.69 for VRAG in qualifying external validations of violent recidivism among forensic psychiatric discharge samples. That result tells the reader about discrimination across the contributing samples; it does not establish calibration for every institution or make the instrument suitable for every referral question. Systematic review and meta-analysis.
The Psychopathy Checklist–Revised and screening variants sometimes appear in violence prediction research, but psychopathy is a distinct assessed construct rather than an interchangeable measure of violence, legal guilt, or imminent threat. Measures designed primarily for sexual recidivism, general recidivism, juvenile contexts, or psychopathy should not be casually repurposed as general violence-risk tests.
Short-term instruments
The Brøset Violence Checklist and the Dynamic Appraisal of Situational Aggression are examples relevant to monitoring imminent inpatient aggression when used within their validated contexts. Such tools depend on repeated observation and the local prevention process. A short-term warning indicator should prompt proportionate care and reassessment, not an automatic inference about years of future behavior. NICE guidance NG10.
Prediction models such as OxMIV
OxMIV illustrates a probability-oriented approach developed in defined severe mental illness populations. An external validation in first-episode psychosis services reported an AUC of 0.75, and at a 10% predicted-risk cutoff found sensitivity of 71%, specificity of 66%, PPV of 22%, and NPV of 95% for recorded violence within 12 months. These values make the false-positive trade-off tangible: most people above that threshold did not have the recorded outcome within the study period.
This result applies to the population, methods, and endpoint studied. It neither labels people with psychosis as dangerous nor provides a universal clinical recommendation to apply OxMIV to everyone. In particular, clinical and forensic services should assess whether a model has been calibrated and evaluated for their current population.
Where Violence Risk Assessment Fails
Individual events are uncertain
A statistical relationship cannot reveal the exact circumstances of every future decision, exposure, interaction, or intervention. Human behavior depends on changing situations and opportunities. A person can receive a higher risk estimate and never commit violence; another may be assigned lower risk and later commit an act. Predictive validity is a population-level performance property, not foreknowledge about an individual.
The predicted outcome may be only a partial measure of violence
Many long-term studies use arrests or reconvictions as the outcome because these records are available. Yet recorded offenses reflect detection, reporting, legal definitions, charging, and conviction. Some harmful incidents are never officially recorded, while allegations and convictions have different evidentiary meanings. Using administrative records is sometimes necessary but changes exactly what the model is validated to predict.
Populations and institutions differ
A tool developed in a secure forensic hospital may perform differently in general outpatient psychiatry, a juvenile service, a probation population, a different country, or a community with different rates of recorded violence. Selection processes can also make apparent performance look better than it is in broader practice. The Lancet Psychiatry systematic review found substantial concentration of evidence in high-income countries and limited female-only validation data.
Risk labels conceal uncertainty
Two people classified as “high risk” can differ greatly in background risk, available supports, plausible scenarios, and potential response to intervention. Categories may have locally defined thresholds, may vary between instruments, or may represent professional judgment rather than a numerical prediction. Words such as “high,” “moderate,” and “low” require an explicit operational definition.
Treatment changes what is being predicted
Risk assessment is often used to guide intervention. If a plan prevents the anticipated event, an assessment can appear to have been a false alarm even when a real vulnerability was successfully addressed. Conversely, lack of services can change the outcome. This feedback complicates validation: observational prediction accuracy cannot by itself establish whether a risk-management program causes a reduction in violence.
A systematic review of assessment tools and subsequent management found inconsistent implementation and insufficient evidence to conclude that tools themselves directly reduce violence or reoffending. Sound practice therefore evaluates both the accuracy of predictions and the effects, burdens, and fidelity of interventions.
Dynamic scores do not always track actual risk reductions
A patient may report fewer symptoms, a rating may decrease, or an assessor may judge management conditions improved. Those changes can be meaningful in care, but they do not automatically imply an empirically quantified drop in recidivism probability. The 2026 HCR-20 longitudinal review is important precisely because it investigated the validity of change over time as its own question.
False Positives, False Negatives, and Consequences
A false positive is a classification that flags a later outcome that does not occur within the defined time frame. It may lead to unnecessary restrictions, stigma, intensive surveillance, or diversion of scarce services. A false negative is a classification that misses an outcome later observed; it may result in inadequate preventive attention or victim protection. Both errors matter, and their social consequences are unequal across interventions and settings.
Decisions must therefore consider the expected benefits and burdens of acting on a score. An optional, low-burden service offered to many people may justify a different threshold from involuntary confinement. The scientific evidence on a tool's accuracy never replaces the legal and ethical requirements for the actual intervention.
Risk Factors, Mental Illness, and Stigma
A psychiatric diagnosis identifies a clinical syndrome under established diagnostic criteria. It is neither a crime nor a general verdict of dangerousness. A diagnosis, symptom, past event, trait, substance-related factor, psychometric score, and observed behavior should be documented separately. The presence of psychosis or a personality disorder alone cannot support an individualized finding of future violence.
Historical violence and particular contextual conditions can be more directly relevant to some risk models than broad diagnostic labels. Even then, associations are probabilistic and their explanatory meaning depends on other factors. Services should recognize trauma exposure, disability, communication needs, and the possibility that prior system contact changes recorded outcomes. Readers seeking the separate clinical evidence can consult Schizophrenia and Violence: What Research Really Shows.
Treatment planning should address the person's actual clinical and social needs while preserving dignity. When discussing substance use, prior victimization, housing insecurity, or treatment interruption, practitioners should avoid presenting the factor as a singular cause of offending or as proof of blame.
Cultural Fairness, Race, Gender, and Disability
Fairness involves more than a single comparison of AUC scores. Evaluators should examine the accuracy and calibration of a model across relevant groups, differences in false-positive and false-negative rates, data completeness, whether recorded outcomes are influenced by policing or access to care, and the practical distribution of restrictions and benefits. Satisfactory performance in one subgroup cannot be assumed in another.
A 2024 systematic review of disparate impacts found mixed and methodologically limited evidence about how criminal-justice risk instruments affect racial and ethnic disparities in legal decisions. It reported studies showing no worsening or reductions in disparities as well as a smaller number finding adverse impacts. Its findings mainly concerned pretrial instruments and cannot be generalized wholesale to clinical violence assessment. Both unsupported claims of universal bias and unsupported claims of fairness exceed the evidence.
Culture, language, neurodevelopmental differences, disability, and communication style may affect how an interview or behavioral observation is understood. The APA guidelines require attention to population-specific validity and individual differences; NICE NG10 warns against negative assumptions based on ethnicity, religion, or unfamiliar cultural practices. Appropriate accommodations and transparent reasoning are parts of validity, not secondary courtesy.
Violence Risk Assessment vs. Threat Assessment
General violence risk assessment usually concerns a defined class of aggressive or violent outcomes over a prediction horizon. Threat assessment of targeted violence may focus on evolving behavior toward a particular target, such as escalating fixation, communications, preparation, access, and opportunities for intervention. The two practices can share evidence but need different decision questions, sources, and management arrangements.
An alarming statement alone is not proof that someone will carry out an attack. Absence of an explicit threat also cannot establish safety. Individual conduct, context, corroboration, and alternative explanations must be weighed. The separate topic of targeted violence and threat assessment belongs to its own dedicated assessment framework, rather than being inferred from a general recidivism score.
Violence Risk Assessment in Courts and Correctional Decisions
Courts and administrative authorities may consider risk evidence in sentencing, civil commitment, bail, parole, or release decisions, depending on the jurisdiction and governing law. These processes involve legal thresholds, procedural rights, admissibility standards, statutory questions, and value judgments about acceptable restrictions. A psychologist contributes a scientific opinion within their competence; a court or other authorized decision-maker determines the legal outcome.
Any forensic report should specify the exact legal referral question and avoid substituting a clinical risk category for a statutory standard. In the United States, requirements differ among federal, state, and local settings, and they differ across countries. The APA Specialty Guidelines for Forensic Psychology address appropriate methods, legal relevance, limits of inference, and truthful communication; those professional standards do not replace jurisdiction-specific law.
Where a risk tool is used in a liberty-affecting process, evaluators should state how the result was obtained and whether the outcome definition, follow-up window, threshold, population, and relevant error rates fit the actual legal question. A high score cannot prove guilt, criminal responsibility, an intention to harm, or a legally sufficient basis for detention. Risk scores are not substitutes for evidence about a charged act, proof beyond a reasonable doubt, or judicial findings.
Does AI Make Violence Prediction More Accurate?
Automated prediction systems can process large datasets, but greater computational complexity is not itself evidence of better prediction. Accuracy depends on representative outcomes, valid labels, transportability, independent validation, calibration, reasonable thresholds, and real-world clinical or legal utility. The same base-rate problem and potential for biased administrative labels apply to conventional scores, machine-learning models, and AI-supported systems.
A 2026 systematic review and critical appraisal examined 38 studies of machine learning for violence prediction. Only eight studies reported calibration, three investigated external validation, and 31 studies (82%) were judged at high risk of bias. The authors concluded that the models in the reviewed literature had poor clinical utility and could not be recommended for implementation without stronger validation. These findings support requiring independent evaluation, well-calibrated probabilities, and demonstrated benefits before adopting an AI-supported violence risk score. [2026 systematic review, Journal of Criminal Justice].
Automated systems also require transparency about training data, implementation, human oversight, changing populations, and error reporting. A prediction that improves benchmark AUC but worsens absolute probability estimates or operational decision quality may be a poor choice. Using generative text to describe someone as violent adds another layer of unvalidated inference; a narrative that sounds convincing is not a calibrated risk estimate.
What Good Practice Looks Like
The NICE guideline on managing violence and aggression recommends a multidisciplinary approach, attention to previous episodes and verifiable context, consideration of certain structured instruments for inpatient settings, individualized prevention strategies, and regular review. These recommendations reflect health-service practice in their applicable setting, rather than a global legal rule.
Good practice links the prediction question to a proportionate preventive response, documents competing explanations, corroborates consequential claims, and allows the formulation to change when evidence changes. Its goal is to improve safety and care while reducing unjustified restrictions. Responsible programs measure not only how well a score ranked past outcomes, but whether the resulting decisions produced defensible benefits.
Common Myths About Violence Risk Assessment
“A high-risk score means the person will become violent.”
A score or category expresses estimated risk relative to a defined benchmark. Many individuals classified above a threshold do not experience the outcome during follow-up. Positive predictive value and the event base rate are essential to interpretation.
“A person with a mental illness is automatically dangerous.”
Clinical diagnoses are not interchangeable with criminal behavior, verified threats, or case-specific risk. Appropriate assessment evaluates relevant behavior and context and uses the diagnosis only when its specific clinical features are relevant and supported by evidence.
“If an assessment has a good AUC, it provides reliable individual probabilities.”
AUC assesses ranking, while calibration assesses agreement between estimated and observed probabilities. An instrument can show moderate discrimination and poor calibration; some instruments provide ordinal judgments without validated individual percentage estimates.
“If a treatment improves a dynamic score, future violence has been prevented.”
Change in an assessment score and change in actual outcome probability require separate tests. Prevention also requires effective implementation and outcome evidence, not simply completed forms.
“Risk assessment can decide whether someone is guilty.”
Criminal guilt is a legal determination based on evidence about a charged offense under the governing jurisdiction's rules. Future-risk assessment concerns a different time frame and evidentiary problem.
Frequently Asked Questions
How accurate is violence risk assessment?
Accuracy depends on the population, tool, outcome, threshold, and prediction period. In a large 2023 systematic review of forensic psychiatric discharge populations, pooled AUCs for selected tools ranged approximately from 0.64 to 0.72, while many reports lacked calibration statistics and had important bias concerns. That result offers a research benchmark, not an individualized accuracy percentage. Systematic review.
Can psychologists predict exactly who will commit violence?
They can describe evidence-based group-level associations and develop conditional individual formulations, but cannot identify future perpetrators with certainty. A professional opinion must be confined to its factual and methodological basis.
What is the difference between a risk factor and a cause?
A risk factor is associated with later outcomes in a studied population. Causation requires stronger evidence that changing the factor itself changes the outcome. Some associated variables are proxies for unmeasured contexts or selection processes. Individual guilt or intent cannot be inferred from a group association.
Does a low-risk rating guarantee that someone is safe?
No. It means an outcome was judged less likely under a defined method and set of circumstances. Assessment should specify the horizon and events that would trigger reassessment.
What is the difference between risk assessment and a clinical diagnosis?
Diagnosis concerns whether criteria for a clinical condition are met. Risk assessment concerns the likelihood and circumstances of a specified future behavior. One does not automatically establish the other, and neither supplies a legal verdict.
How often should risk be reassessed?
The schedule depends on the decision context and how quickly relevant circumstances can change. Immediate ward assessments may be repeated frequently; longer-term plans require review after material changes in evidence, environment, treatment, supports, or the legal mandate. NICE NG10.
Can violence risk tools be used for children or teenagers?
Only tools and procedures with evidence relevant to the developmental group and the intended outcome should be used. Adult norms and adult risk categories should not be automatically transferred to young people. Youth assessment must consider development, context, accommodations, and potential consequences.
Can a violence risk score justify involuntary detention?
A score does not itself establish the legal criteria for involuntary detention. Jurisdiction-specific law, procedural safeguards, admissible evidence, and the authorized decision-maker govern such decisions. The assessment may inform a lawful process within its validated limits.
Related Articles
References
American Psychological Association. (2013). Specialty guidelines for forensic psychology. American Psychologist, 68, 7–19. https://doi.org/10.1037/a0029889
Campbell, D., Cohen, M., Macpherson, G., & O’Rourke, S. (2026). The longitudinal properties of the HCR-20: A systematic review. Aggression and Violent Behavior, 86, 102111. https://doi.org/10.1016/j.avb.2025.102111
Challinor, A., Ogundalu, A., McIntyre, J. C., Bramwell, V., & Nathan, R. (2021). The empirical evidence base for the use of the HCR-20: A narrative review of study designs and transferability of results to clinical practice. International Journal of Law and Psychiatry, 78, 101729. https://doi.org/10.1016/j.ijlp.2021.101729
Dickens, G. L., O’Shea, L. E., & Christensen, M. (2020). Structured assessments for imminent aggression in mental health and correctional settings: Systematic review and meta-analysis. International Journal of Nursing Studies, 104, 103526. https://doi.org/10.1016/j.ijnurstu.2020.103526
Douglas, K. S. (2014). Version 3 of the Historical-Clinical-Risk Management-20 (HCR-20V3): Relevance to violence risk assessment and management in forensic conditional release contexts. Behavioral Sciences & the Law, 32, 557–576. https://doi.org/10.1002/bsl.2134
Fazel, S., Singh, J. P., Doll, H., & Grann, M. (2012). Use of risk assessment instruments to predict violence and antisocial behaviour in 73 samples involving 24,827 people: Systematic review and meta-analysis. BMJ, 345, e4692. https://doi.org/10.1136/bmj.e4692
Kozhevnikova, S., Yukhnenko, D., Scola, G., & Fazel, S. (2026). Machine learning for violence prediction: A systematic review and critical appraisal. Journal of Criminal Justice, 105, 102697. https://doi.org/10.1016/j.jcrimjus.2026.102697
Lawson, S. G., Narkewicz, E. L., & Vincent, G. M. (2024). Disparate impact of risk assessment instruments: A systematic review. Law and Human Behavior, 48, 427–440. https://doi.org/10.1037/lhb0000582
National Institute for Health and Care Excellence. (2015). Violence and aggression: Short-term management in mental health, health and community settings (NG10). https://www.nice.org.uk/guidance/NG10/chapter/recommendations
Neil, C., O’Rourke, S., Ferreira, N., & Flynn, L. (2020). Protective factors in violence risk assessment: Predictive validity of the SAPROF and HCR-20V3. International Journal of Forensic Mental Health. https://doi.org/10.1080/14999013.2019.1643811
O’Shea, L. E., & Dickens, G. L. (2018). Use of risk assessment instruments to predict violence in forensic psychiatric hospitals: A systematic review and meta-analysis. European Psychiatry, 52, 47–53. https://doi.org/10.1016/j.eurpsy.2018.02.007
Ogonah, M. G. T., Seyedsalehi, A., Whiting, D., & Fazel, S. (2023). Violence risk assessment instruments in forensic psychiatric populations: A systematic review and meta-analysis. The Lancet Psychiatry, 10, 780–789. https://doi.org/10.1016/S2215-0366(23)00256-0
Singh, J. P., Desmarais, S. L., & Van Dorn, R. A. (2013). Measurement of predictive validity in violence risk assessment studies: A second-order systematic review. Behavioral Sciences & the Law, 31, 55–73. https://doi.org/10.1002/bsl.2053
Viljoen, J. L., Cochrane, D. M., & Jonnson, M. R. (2018). Do risk assessment tools help manage and reduce risk of violence and reoffending? A systematic review. Law and Human Behavior, 42, 181–214. https://doi.org/10.1037/lhb0000280
Whiting, D., Mallett, S., Lennox, B., & Fazel, S. (2023). Assessing violence risk in first-episode psychosis: External validation, updating and net benefit of a prediction tool (OxMIV). BMJ Mental Health, 26, e300634. https://doi.org/10.1136/bmjment-2022-300634