Schizophrenia and ViolenceSystematic Review and Meta-Analysis

Seena Fazel, G. Mitu Gulati, Louise Linsell, John Geddes, Martin Grann•View original

OverviewBalancedalloy voice

Let's start with the tension. For a long time, the expert line was simple: people with schizophrenia weren't more violent than anyone else. That view felt humane, it pushed back against stigma, and it seemed to protect access to care. Then, over the last two decades, large registry studies—millions of people followed across years—began telling a different story. Not a tabloid story. A modest association. People with schizophrenia and related psychoses did seem to show more violence than the general population. Cue the controversy. Advocacy groups in the UK called the belief a myth. In the US, the National Alliance on Mental Illness emphasized that violent acts were "exceptional." Clinicians worried that highlighting any link would feed discrimination. Policymakers, who have to write rules about involuntary care and public safety, were left staring at a moving target. Into that mess stepped a careful synthesis by Seena Fazel and colleagues in 2009. Their goal was straightforward and ambitious: pool the best available evidence on violence in schizophrenia and other psychoses, figure out why past studies disagree, and, crucially, ask what actually drives the risk. If we're going to have a public conversation about violence and mental illness, it needs to be anchored in what the data really say—not in wishful thinking or fear. They built the evidence base the unglamorous way, which is the right way. They searched Medline, Embase, and PsycInfo all the way back to 1970, cast a wide net for anything touching on psychosis and violence, and chased down grey literature and non-English reports. When details were missing, they wrote to original authors. They followed established guidelines for observational meta-analysis and kept the inclusion criteria tight: studies had to compare people with schizophrenia or related psychoses to a general population group, and they had to report violence specifically—not just "any conviction." Both case-control snapshots and cohort designs were in, but homicide got its own dedicated review because it's relatively rare and easy to distort. On the analysis side, they did what you hope an honest team will do. They used random-effects models because they expected studies to differ. They measured heterogeneity with the usual suspects: Cochran's Q and the I-squared statistic, and they didn't stop at pooling a single number. They ran metaregression—think of it as asking, "Do the answers change depending on who was studied, where, and how outcomes were measured?"—testing gender, diagnosis subtype, data source, region, study period, study design, and, most importantly, the presence of substance abuse. Then they did something that sounds small but matters: they acknowledged up front that many studies counted violent acts that happened before a formal diagnosis, which could inflate any apparent association. File that under "limitations you can trust." Now, what did they find when they put it all together? Across twenty studies, there were 18,423 people with schizophrenia or other psychoses. About one in ten had been violent—1,832 people, to be exact. In the general population comparison group—1,714,904 individuals—about two in a hundred had been violent, 27,185 in total. That gap translated to a pooled odds ratio of roughly four under a random-effects model. In plain language: on average, people with schizophrenia and related psychoses were about four times as likely to be violent as their neighbors. If you forced the data into a fixed-effects model, which downplays differences between studies, the estimate came down to about 2.9. Big number, then smaller number. The truth likely sits in a range, not a single point. And there was a lot of variation between studies—an I-squared around eighty-eight percent—so we shouldn't pretend otherwise. Breathe for a second here. A fourfold increase sounds dramatic. But remember two things. First, the baseline risk of violence in the general population is low, so even multiples can leave most people in either group never violent. Second, averages hide important patterns, and this literature is all patterns. Take gender. In men, the pooled odds ratio looked a lot like the overall one: around four, with similarly high variability. In women, the relative risk was higher—around seven to eight in random-effects models and about six to seven in fixed-effects models. That doesn't mean women with schizophrenia are more violent than men in absolute terms; it means that when violence does occur, the increase over women in the general population is steeper. It's a relative measure. Useful, but easy to misread. What about how violence was measured—official records versus self-report or informants? You might expect court or police records to show bigger differences, maybe because they capture more serious events. The pooled picture, though, didn't show a robust split. In some single-sex comparisons, register-based outcomes were a bit higher—for men, around 4.1 versus 3.0 for self-report—but both estimates came wrapped in wide variability. The takeaway there is less about which source "wins" and more about the association being visible across very different ways of counting. So where does the heterogeneity—the wild swing from "no effect" to "huge effect" across studies—actually come from? This is the pivot of the whole review. When Fazel and colleagues lined up all the study features in a metaregression, only one factor consistently explained the differences: comorbid substance abuse. Not the country. Not whether the study was case-control or longitudinal. Not even diagnosis subtype when schizophrenia was separated from other psychoses. Substance misuse changed the picture. Quantitatively, the contrast is stark. In eleven studies that reported violence with and without substance abuse, the odds ratio with comorbidity was about 8.9. Without comorbidity, it dropped to about 2.1. That's the same diagnosis, the same era, often the same registry—two very different risk profiles depending on whether alcohol or drugs were in the mix. Restrict the analysis to men and you see the same shape: roughly 12.2 with comorbidity versus 2.8 without. In women, data were thinner and one study reported very large numbers, which should be read cautiously, but the direction matched: higher risk when substance misuse was present. Here's a useful way to center that finding. In studies that reported both diagnoses in the same population, people with substance abuse disorders—without psychosis—had an odds ratio for violence of about 5.5. People with psychosis—without substance abuse—had an odds ratio around 3.3. In other words, substance misuse on its own carried a larger relative risk than psychosis on its own. Put them together and the risk climbed well beyond psychosis alone. That doesn't absolve psychosis; it reframes the mechanism. The excess risk appears to be mediated, to a large degree, by co-occurring substance use. Let's talk about the most extreme outcome—homicide. It's rare, which means the numbers are small and uncertainty tends to loom larger. But five studies had enough data to analyze. Among people with psychosis, about 0.3 percent were convicted of homicide; in the general population, it was about 0.02 percent. That translated into an odds ratio near 19.5 under random-effects. It's a big multiplier on a tiny base rate. When the comparison was substance abuse alone, the absolute risk was also about 0.3 percent, with an odds ratio around 10.9. That pairing is easy to over-interpret. The headline is not "psychosis doubles homicide risk over substance abuse" or vice versa. It's that both conditions, in different ways, are associated with steep relative increases in a very rare event, and that the absolute probability remains low. If your brain is already leaping to public policy, hold on. Before we talk implications, we need to square with the caveats. Heterogeneity was high—often north of eighty percent—which means the pooled estimates are averaging across genuinely different contexts and methods. Some studies counted violence that occurred before a formal diagnosis, which will tend to inflate the association if early symptoms and substance use were already present. In the subset of longitudinal designs—where violence was assessed after diagnosis—the risk estimates tended to be lower. Geography matters too: most data came from high-income countries like the US and the Nordics, places with strong registries and particular criminal justice practices, so it's risky to generalize to low-income settings. And publication bias? When they ran Egger's test on the overall dataset, it hinted at small-study effects, with a p-value around 0.04, and the signal was stronger when they analyzed gender-specific results. Funnel plots were less conclusive. None of this erases the association; it just reminds us to read it with calibration. There's one more scale that matters here: population impact. Even with elevated individual risk, how much violence in society is actually attributable to psychosis? In the handful of studies that reported enough data to calculate it, population-attributable fractions were modest—under ten percent, with several estimates between about three and eight percent. That's not nothing. It also means the vast majority of violence is not linked to psychosis, which is critical when we think about stigma and about where prevention dollars go. So, what do we do with a conclusion that is both clear and nuanced? The clear part is this: across countries, designs, and outcome definitions, people with schizophrenia and other psychoses have a higher risk of violence than their peers, and that elevation is largely driven by comorbid substance misuse. The nuanced part is the "largely." Psychosis without substance abuse still showed a small-to-moderate increase in risk—those odds ratios around 2 are not zero. But the steep climbs—the numbers that make some studies look scary—are overwhelmingly tied to alcohol and drugs in the mix. That points to a very practical implication. If you want to reduce violence associated with severe mental illness, you don't start with blanket assumptions about psychosis. You start with substance use. Integrate substance-use treatment into psychosis care as a core component, not an optional referral. Build services that actually retain people with dual diagnoses—because if you've ever worked in a clinic, you know that handoffs between mental health and addiction services are where good plans go to die. At a population level, the same levers that reduce alcohol misuse and problem drug use—pricing, availability, evidence-based treatment access—are violence prevention tools here too. There's also a communication lesson. Saying "there's a link" without context risks feeding the very stigma that keeps people from seeking help. But pretending there's no link at all misleads the public and ignores an actionable risk factor. A better message is honest and specific: most people with schizophrenia are not violent; when violence risk rises, it's usually because substance misuse is present; and effective treatment can bring that risk down. If you're wondering what to watch for next, keep an eye on three fronts. First, longitudinal cohorts that pin down timing—onset of symptoms, initiation of substance use, first violent episode—because sequencing matters for prevention. Second, studies from low- and middle-income countries, where patterns of substance availability and social support differ. Third, work that separates types of substance use; alcohol is common and strongly associated with violence, but stimulants, opioids, and polysubstance profiles may carry different risks. But that's the brief coda. The story we can tell today, thanks to Fazel and colleagues, is grounded and actionable. There is an association between psychosis and violence. It varies by sex and outcome source, but is robust across methods. The most powerful dial on that association is comorbid substance misuse, which turns a twofold bump into something much steeper. Homicide sits at the extreme—high relative risks riding on tiny absolute probabilities—and substance use looms large there too. For public health, the path forward is not a mystery. It's integration. Treat the psychosis. Treat the substance use. Do both well, together. And talk about the data in a way that reduces harm rather than amplifying fear.

Then, over the last two decades, large registry studies—millions of people followed across years—began telling a different story. Not a tabloid story. A modest association.

People with schizophrenia and related psychoses did seem to show more violence than the general population. Cue the controversy. Advocacy groups in the UK called the belief a myth.

In the US, the National Alliance on Mental Illness emphasized that violent acts were "exceptional." Clinicians worried that highlighting any link would feed discrimination. Policymakers, who have to write rules about involuntary care and public safety, were left staring at a moving target.

Into that mess stepped a careful synthesis by Seena Fazel and colleagues in 2009. Their goal was straightforward and ambitious: pool the best available evidence on violence in schizophrenia and other psychoses, figure out why past studies disagree, and, crucially, ask what actually drives the risk. If we're going to have a public conversation about violence and mental illness, it needs to be anchored in what the data really say—not in wishful thinking or fear.

They built the evidence base the unglamorous way, which is the right way. They searched Medline, Embase, and PsycInfo all the way back to 1970, cast a wide net for anything touching on psychosis and violence, and chased down grey literature and non-English reports. When details were missing, they wrote to original authors.

They followed established guidelines for observational meta-analysis and kept the inclusion criteria tight: studies had to compare people with schizophrenia or related psychoses to a general population group, and they had to report violence specifically—not just "any conviction." Both case-control snapshots and cohort designs were in, but homicide got its own dedicated review because it's relatively rare and easy to distort.

On the analysis side, they did what you hope an honest team will do. They used random-effects models because they expected studies to differ. They measured heterogeneity with the usual suspects: Cochran's Q and the I-squared statistic, and they didn't stop at pooling a single number.

They ran metaregression—think of it as asking, "Do the answers change depending on who was studied, where, and how outcomes were measured?"—testing gender, diagnosis subtype, data source, region, study period, study design, and, most importantly, the presence of substance abuse. Then they did something that sounds small but matters: they acknowledged up front that many studies counted violent acts that happened before a formal diagnosis, which could inflate any apparent association. File that under "limitations you can trust."

Now, what did they find when they put it all together? Across twenty studies, there were 18,423 people with schizophrenia or other psychoses. About one in ten had been violent—1,832 people, to be exact.

In the general population comparison group—1,714,904 individuals—about two in a hundred had been violent, 27,185 in total. That gap translated to a pooled odds ratio of roughly four under a random-effects model. In plain language: on average, people with schizophrenia and related psychoses were about four times as likely to be violent as their neighbors.

If you forced the data into a fixed-effects model, which downplays differences between studies, the estimate came down to about 2.9. Big number, then smaller number. The truth likely sits in a range, not a single point.

And there was a lot of variation between studies—an I-squared around eighty-eight percent—so we shouldn't pretend otherwise.

Breathe for a second here. A fourfold increase sounds dramatic. But remember two things.

First, the baseline risk of violence in the general population is low, so even multiples can leave most people in either group never violent. Second, averages hide important patterns, and this literature is all patterns.

Take gender. In men, the pooled odds ratio looked a lot like the overall one: around four, with similarly high variability. In women, the relative risk was higher—around seven to eight in random-effects models and about six to seven in fixed-effects models.

That doesn't mean women with schizophrenia are more violent than men in absolute terms; it means that when violence does occur, the increase over women in the general population is steeper. It's a relative measure. Useful, but easy to misread.

What about how violence was measured—official records versus self-report or informants? You might expect court or police records to show bigger differences, maybe because they capture more serious events. The pooled picture, though, didn't show a robust split.

In some single-sex comparisons, register-based outcomes were a bit higher—for men, around 4.1 versus 3.0 for self-report—but both estimates came wrapped in wide variability. The takeaway there is less about which source "wins" and more about the association being visible across very different ways of counting.

So where does the heterogeneity—the wild swing from "no effect" to "huge effect" across studies—actually come from? This is the pivot of the whole review. When Fazel and colleagues lined up all the study features in a metaregression, only one factor consistently explained the differences: comorbid substance abuse.

Not the country. Not whether the study was case-control or longitudinal. Not even diagnosis subtype when schizophrenia was separated from other psychoses. Substance misuse changed the picture.

Quantitatively, the contrast is stark. In eleven studies that reported violence with and without substance abuse, the odds ratio with comorbidity was about 8.9. Without comorbidity, it dropped to about 2.1.

That's the same diagnosis, the same era, often the same registry—two very different risk profiles depending on whether alcohol or drugs were in the mix. Restrict the analysis to men and you see the same shape: roughly 12.2 with comorbidity versus 2.8 without. In women, data were thinner and one study reported very large numbers, which should be read cautiously, but the direction matched: higher risk when substance misuse was present.

Here's a useful way to center that finding. In studies that reported both diagnoses in the same population, people with substance abuse disorders—without psychosis—had an odds ratio for violence of about 5.5. People with psychosis—without substance abuse—had an odds ratio around 3.3.

In other words, substance misuse on its own carried a larger relative risk than psychosis on its own. Put them together and the risk climbed well beyond psychosis alone. That doesn't absolve psychosis; it reframes the mechanism.

The excess risk appears to be mediated, to a large degree, by co-occurring substance use.

Let's talk about the most extreme outcome—homicide. It's rare, which means the numbers are small and uncertainty tends to loom larger. But five studies had enough data to analyze.

Among people with psychosis, about 0.3 percent were convicted of homicide; in the general population, it was about 0.02 percent. That translated into an odds ratio near 19.5 under random-effects. It's a big multiplier on a tiny base rate.

When the comparison was substance abuse alone, the absolute risk was also about 0.3 percent, with an odds ratio around 10.9. That pairing is easy to over-interpret. The headline is not "psychosis doubles homicide risk over substance abuse" or vice versa.

It's that both conditions, in different ways, are associated with steep relative increases in a very rare event, and that the absolute probability remains low.

If your brain is already leaping to public policy, hold on. Before we talk implications, we need to square with the caveats. Heterogeneity was high—often north of eighty percent—which means the pooled estimates are averaging across genuinely different contexts and methods.

Some studies counted violence that occurred before a formal diagnosis, which will tend to inflate the association if early symptoms and substance use were already present. In the subset of longitudinal designs—where violence was assessed after diagnosis—the risk estimates tended to be lower. Geography matters too: most data came from high-income countries like the US and the Nordics, places with strong registries and particular criminal justice practices, so it's risky to generalize to low-income settings.

And publication bias? When they ran Egger's test on the overall dataset, it hinted at small-study effects, with a p-value around 0.04, and the signal was stronger when they analyzed gender-specific results. Funnel plots were less conclusive.

None of this erases the association; it just reminds us to read it with calibration.

There's one more scale that matters here: population impact. Even with elevated individual risk, how much violence in society is actually attributable to psychosis? In the handful of studies that reported enough data to calculate it, population-attributable fractions were modest—under ten percent, with several estimates between about three and eight percent.

That's not nothing. It also means the vast majority of violence is not linked to psychosis, which is critical when we think about stigma and about where prevention dollars go.

So, what do we do with a conclusion that is both clear and nuanced? The clear part is this: across countries, designs, and outcome definitions, people with schizophrenia and other psychoses have a higher risk of violence than their peers, and that elevation is largely driven by comorbid substance misuse. The nuanced part is the "largely." Psychosis without substance abuse still showed a small-to-moderate increase in risk—those odds ratios around 2 are not zero.

But the steep climbs—the numbers that make some studies look scary—are overwhelmingly tied to alcohol and drugs in the mix.

That points to a very practical implication. If you want to reduce violence associated with severe mental illness, you don't start with blanket assumptions about psychosis. You start with substance use.

Integrate substance-use treatment into psychosis care as a core component, not an optional referral. Build services that actually retain people with dual diagnoses—because if you've ever worked in a clinic, you know that handoffs between mental health and addiction services are where good plans go to die. At a population level, the same levers that reduce alcohol misuse and problem drug use—pricing, availability, evidence-based treatment access—are violence prevention tools here too.

There's also a communication lesson. Saying "there's a link" without context risks feeding the very stigma that keeps people from seeking help. But pretending there's no link at all misleads the public and ignores an actionable risk factor.

A better message is honest and specific: most people with schizophrenia are not violent; when violence risk rises, it's usually because substance misuse is present; and effective treatment can bring that risk down.

If you're wondering what to watch for next, keep an eye on three fronts. First, longitudinal cohorts that pin down timing—onset of symptoms, initiation of substance use, first violent episode—because sequencing matters for prevention. Second, studies from low- and middle-income countries, where patterns of substance availability and social support differ.

Third, work that separates types of substance use; alcohol is common and strongly associated with violence, but stimulants, opioids, and polysubstance profiles may carry different risks.

But that's the brief coda. The story we can tell today, thanks to Fazel and colleagues, is grounded and actionable. There is an association between psychosis and violence.

It varies by sex and outcome source, but is robust across methods. The most powerful dial on that association is comorbid substance misuse, which turns a twofold bump into something much steeper. Homicide sits at the extreme—high relative risks riding on tiny absolute probabilities—and substance use looms large there too.

For public health, the path forward is not a mystery. It's integration. Treat the psychosis.

Treat the substance use. Do both well, together. And talk about the data in a way that reduces harm rather than amplifying fear.

Related lectures