Risk Factors for Violence in PsychosisSystematic Review and Meta-Regression Analysis of 110 Studies

People with psychosis are statistically more likely to be victims of violence than perpetrators. However, the fear runs the other way, and that gap between perception and reality costs individuals — in stigma, in lost care, and in policies built on the wrong model. So here is the question that actually matters: not whether there is some link between psychosis and violence, but which specific factors drive that link, and crucially, which of those factors can be changed. That is exactly what Katrina Witt, Richard Van Dorn, and Seena Fazel set out to answer. Before their work, the literature was filled with conflicting findings. One narrative review emphasized positive psychotic symptoms and substance misuse. Another added personality disorders and theory of mind. A third, focused on first-episode psychosis, pointed to criminal history and duration of untreated psychosis. Nobody had produced a systematic synthesis that ranked the relative importance of risk factors and sorted them by what clinicians can actually act on. Witt and colleagues fixed that. Their review included one hundred and ten eligible studies representing seventy-three independent samples and forty-five thousand five hundred and thirty-three individuals in total. Of those individuals, eight thousand four hundred and thirty-nine — about eighteen point five percent — were classified as violent. The vast majority, nearly eighty-eight percent, were diagnosed with schizophrenia; the remainder had other psychoses. The average participant age was thirty-five point eight years. This is a clinical population, not a random sample of the public, which is worth keeping in mind whenever you hear that number. The methodological architecture matters here because it is what makes the findings trustworthy. The team searched six electronic databases — CINAHL, EBSCO, EMBASE, Global Health, PsycINFO, and PubMed — from their start dates, some going back to nineteen sixty, plus Google Scholar. They included non-English language studies and dissertations. For pooling, they only meta-analyzed factors reported in three or more primary studies, and they converted diverse statistics into odds ratios with ninety-five percent confidence intervals so everything was comparable. An odds ratio of 2, to put it plainly, means twice the odds of violence compared to those without that factor. The cleverest methodological move was the domain approach. Instead of treating every correlated risk factor as a separate signal, they grouped similar factors into ten domains — criminal history, substance misuse, demographics, premorbid factors, psychopathology, positive symptoms, negative symptoms, neuropsychological factors, treatment-related factors, and suicidality. Where one study reported multiple factors in the same domain, they included only the one with the highest statistical weight. This prevented double-counting. They then used random-effects models to pool estimates and the I-squared statistic to quantify how much disagreement between studies was real rather than due to sampling error. When disagreement was high, meta-regression examined whether study-level characteristics — setting, country, diagnosis, and prior violence — could explain it. Now, to the findings. The clearest signal in the whole analysis came from criminal history. At the domain level, criminal history showed a pooled odds ratio of 3.1, with a ninety-five percent confidence interval running from 2.4 to 4.1. That is the strongest domain-level association in the review. Prior violent victimization, one specific item within that domain, had an odds ratio of 6.1. Demographic factors as a domain came in at a pooled odds ratio of 1.8. Being male was independently associated with violence, with an odds ratio of 1.6. These are static factors. They do not move. They are useful for identifying who sits at an elevated baseline risk — who needs closer monitoring — but you cannot intervene on someone's criminal record. The more consequential findings, for clinical practice, are the dynamic ones. Witt and colleagues identified a set of modifiable risk factors that hold up across the literature. Non-adherence with psychological therapies showed a pooled odds ratio of 6.7 — the largest single treatment-related estimate — though it was based on a smaller number of studies and carries wider confidence intervals. Poor impulse control had a pooled odds ratio of 3.3. Substance misuse as a domain came in at 2.3, and within that domain the picture gets specific: a history of polysubstance misuse had an odds ratio of 10.3, comorbid substance use disorder sat at 3.1, and recent substance misuse — proximal, happening now — had an odds ratio of 2.9. Recent drug misuse and recent alcohol misuse each came in around 2.2. Medication non-adherence had a pooled odds ratio of 2.0, smaller but still significant. The pattern here is coherent. Substance use, poor treatment engagement, and behavioral dyscontrol all amplify violence risk in people with psychosis. Each of these is something a clinician can engage with. They are not destiny. Now here is where the findings push back against the dominant public narrative. Positive psychotic symptoms — hallucinations and delusions — showed surprisingly weak and heterogeneous associations with violence. Higher positive symptom scores had a pooled odds ratio of just 1.2. For auditory hallucinations specifically, the odds ratio was close to 1.1, barely distinguishable from no effect. The image of the dangerous person driven to violence by voices is not what this literature supports. The actual drivers are more mundane and more tractable: drinking, drug use, not taking medication, and not engaging with therapy. That matters enormously for how we talk about this population. Witt and colleagues ran two prespecified sensitivity analyses to test whether these findings were real or artifacts. First, they restricted to the seventy-seven studies — seventy percent of the total — that measured severe violence rather than aggression or hostility more broadly. The pattern of associations did not change materially. Second, they restricted to the thirty-four inpatient studies. Here the substance misuse domain showed a somewhat weaker association, while the psychopathology and positive symptoms domains were somewhat more strongly associated. But the direction of every effect held. Associations differed in strength by setting; they did not flip. That consistency is the review's strongest claim to credibility. The clinical implications are direct. For persistent aggression and violence, the review highlights clozapine as an agent that, in the authors' words, "should be considered carefully." Interventions to improve treatment adherence — including trials of financial incentives — are flagged. Anti-craving agents and antidepressants for substance-use comorbidity are mentioned. Risk instruments that already incorporate suicidality assessment, like the START, get a mention as practically useful given the suicidality domain's stronger association in severe-violence analyses. One finding worth naming carefully: receiving involuntary or leveraged treatment was associated with roughly a fourfold increase in violence risk. The authors are clear that this association is almost certainly explained by confounding — people who receive coerced treatment are already at higher risk — not by the treatment causing violence. This is a good example of why cross-sectional studies, which make up the majority of the literature here, require caution when thinking about causation. And that is the central limitation. Most primary studies were cross-sectional or had short follow-up periods. Unadjusted odds ratios were common. Publication bias was significant for thirty-seven of the one hundred and forty-six factors the team examined. Only two studies were conducted in prisons, limiting what can be said about forensic populations. These are not reasons to distrust the findings — the meta-regression approach handles much of the heterogeneity — but they are reasons to hold the effect sizes with appropriate humility. What the field now needs, and what Witt and colleagues explicitly call for, is large observational studies and clinical trials that use violence as a primary outcome rather than an afterthought. That is a high bar. Trials targeting substance misuse, medication adherence, and impulse control in psychosis exist, but few are powered to detect effects on violence directly. The case this review builds is ultimately about redirection. Criminal history tells you who to watch more closely. Substance misuse, treatment non-adherence, hostile behavior, and poor impulse control tell you what to do about it. Getting that distinction clear — static versus dynamic, background risk versus active target — is what separates useful clinical risk assessment from stigma dressed up as science. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.