Given the high costs of violence perpetration to patients, victims and society,Reference Dolan, Loomes, Peasgood and Tsuchiya1,Reference McCollister, French and Fang2 preventing violent outcomes and improving risk management is a priority for clinical services.Reference Barbui and Saraceno3 Triangulated evidence shows a higher risk of violence among individuals with psychotic disorders than among those without.Reference Whiting, Gulati, Geddes and Fazel4–Reference Whiting, Gulati, Geddes, Dean and Fazel6 Absolute risks of violence can be as high as 26% over 12 months in first-episode psychosis,Reference Whiting, Lennox and Fazel7 but are more typically less than 10% within 5 years of diagnosis by clinical services.Reference Fazel, Wolf, Palm and Lichtenstein5 Identification of risk factors – particularly modifiable ones – in people with psychotic disorders is a next step in developing targeted interventions,Reference Troquete, Van Den Brink, Beintema, Mulder, Van Os and Schoevers8 and could help develop more precise risk assessment tools that allow for risk stratification. Such tools are common in forensic mental health and criminal justice to aid clinical decision-making, but are of varying accuracy.Reference Ramesh, Igoumenou, Vazquez Montes and Fazel9 Updated evidence on risk factors is particularly relevant for treatment allocation in the context of limited resources, for example with the reduction of available psychiatric hospital beds in the UK and USA that has continued in recent years.Reference Tyrer, Sharfstein, O'Reilly, Allison and Bastiampillai10
A 2013 systematic reviewReference Witt, Van Dorn and Fazel11 of risk factors for violence in psychosis outlined a range of replicated risk markers, including criminal history, psychopathological symptoms and treatment-related factors. However, the previous review is now more than a decade old with its search ending in 2011, and many new investigations have since been published. The previous review also included a majority of cross-sectional studies where the temporal relationship between the studied factor and violent outcome is not clear. In this update, we have focused on longitudinal studies to improve the quality of the evidence. In addition, we have conducted separate analyses restricting studies to those using more severe violent outcomes, and those where a majority of participants were recruited from inpatient settings. The latter can inform how to prevent and manage inpatient violence, which has been reported to occur in 21–32% of admitted individuals with psychotic disorders.Reference Broderick, Azizian, Kornbluh and Warburton12 We consider risk factors in a broad way to include descriptive, causal and predictive associations.Reference Shmueli13 We expected to replicate the strongest risk factors in the previous review – including criminal history and comorbid substance use disorders – while drawing on a decade of new evidence on emerging risk factors for violence in psychosis, including the misuse of individual substances.
Method
The study was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.
Protocol
The review methods are based on a previous review from our group,Reference Witt, Van Dorn and Fazel11 published via open access in 2013. One deviation from this protocol was to limit the inclusion to longitudinal designs, which were included in the original protocol alongside cross-sectional designs. We further excluded studies in selected (e.g. solely offender) populations. There were no other material deviations from the original protocol.
Search strategy
We implemented the same search strategy as the 2013 review,Reference Witt, Van Dorn and Fazel11 using the following search term to identify studies examining psychiatric disorders and various violence measures: (schiz* AND (viol* OR aggress* OR crim* OR offend* OR danger* OR hosti*)) OR (psych* AND (viol* OR aggress* OR crim* OR offend* OR danger* OR hosti*)) OR (mental* AND (viol* OR aggress* OR crim* OR offend* OR danger* OR hosti*)) (Supplementary Table 1 available at https://doi.org/10.1192/bjp.2024.120). We conducted the search in five databases – CINAHL, Embase, Global Health, PsycINFO and PubMed – to identify papers published from 1 January 2012 until 30 June 2022. We thus complemented the search from the original review, which searched evidence up to 31 December 2011. The extraction of references from the databases was carried out on 30 June 2022. We also conducted a manual search of reference lists in included or related papers. We translated non-English language publications using Google Translate and asked a native speaker for clarification where necessary. Search and eligibility assessment was carried out by the first author (T.L.). A second reviewer was not considered for systematic eligibility assessment, in line with the previous review.
Study eligibility
We considered studies where the following apply: (a) at least 95% of the participants were diagnosed with schizophrenia, schizoaffective disorder, delusional disorder, other schizophrenia-spectrum disorder, schizotypal disorder, bipolar disorder or other affective psychosis, excluding drug-induced psychosis (e.g. if 94% of patients in a study had schizophrenia spectrum disorder and 6% had drug-induced psychosis or a non-psychotic disorder, then that study would be excluded); (b) diagnosis was made using DSM14 or ICD15 criteria; (c) individuals were aged 15 years or above; and (d) the study employed a design where risk factors preceded the outcome (e.g. cohort studies, nested case–control, randomised controlled trials (RCTs) and prediction studies); (e) further, we excluded studies that only considered repeat violence as an outcome – that is, any study examining violent outcomes in a selected population with violence histories. As in the previous review, bipolar disorder was included as psychosis is a feature in most individuals with bipolar disorder,Reference Chakrabarti and Singh16 and to incorporate risk factors for violence in affective psychosis. We extracted information on each criterion (a)–(e) – the first criterion we noticed that made the study ineligible for inclusion was reported as the reason for that paper's exclusion.
Data extraction
Data from included papers was extracted by the first author (T.L.). A second author (S.L.) independently extracted a randomly selected 20% subset of included studies to assess extraction accuracy. For both risk factors and violent outcomes, we extracted information on whether the variable was categorical or continuous. Risk factor definitions across studies were standardised where possible to ensure they could be pooled. In keeping with previous work,Reference Witt, Van Dorn and Fazel11 we classified risk factors into the following broad and distinctive domains: criminal history, negative symptoms, neuropsychological, positive symptoms, premorbid, psychopathological, sociodemographic, substance misuse, suicidality and treatment-related.
To ensure comparability with previous work and allow for data pooling, we converted all effect estimates to odds ratios, using methods described in previous publications.Reference Witt, Van Dorn and Fazel11,Reference Favril, Yu, Hawton and Fazel17 Hazard ratios and probit regression coefficients cannot strictly be converted to odds ratios. However, probit and logistic regression often give similar results, and hazard ratios can be relatively similar to odds ratios if the outcome event is rare over follow-up.Reference VanderWeele18 We therefore included these effect estimates in the main analysis. As a sensitivity analysis, we excluded all effect estimates measured by hazard ratios or probit regression coefficients.
Study quality was assessed by co-author A.P. using an adaptation of the Newcastle–Ottawa ScaleReference Luchini, Stubbs, Solmi and Veronese19 where we scored quality of exposure ascertainment separately for each risk factor domain included in the study. The overall quality score was then produced by summing the quality score for each risk factor domain included with the other subscores (selection, comparability and outcome). The score was expressed as a percentage of the maximum quality score available given the risk factor domains included in the study.
Statistical analyses
We assessed extraction accuracy by calculating interrater reliability for the extraction of the proportion of violent patients using Krippendorff's alpha.Reference Zapf, Castell, Morawietz and Karch20
We pooled odds ratios using random-effects models owing to the heterogeneity in the design and predictors/outcomes of the studies. We only considered risk factors that occurred in at least three independent samples in the main analysis, but reported results for risk factors occurring in two samples in the supplement. We chose to pool the most minimally adjusted effect estimate available, as this was the most consistently reported and comparable.Reference Witt, Van Dorn and Fazel11,Reference Favril, Yu, Hawton and Fazel17 Some included papers were based on the same original study. If a risk factor occurred in more than one paper that used the same study population, we included the risk factor that derived from the paper with the largest sample size to avoid double counting.Reference Witt, Van Dorn and Fazel11 All analyses in the current paper were carried out at the level of independent samples rather than publications.
We additionally pooled odds ratios within each risk factor domain. If one paper contributed more than one risk factor per domain, we included the risk factor with the highest absolute z-score. The z-score takes into account the size of the effect estimate (strength of association) and its s.d. (precision).Reference Witt, Van Dorn and Fazel11
We assessed heterogeneity using the I 2 statistic to quantify the proportion of the variance in the risk factor effect estimates due to between-study differences rather than random sampling error. To investigate the sources of between-study variability, we ran meta-regression models for risk factors that occurred in at least seven samples and had an I 2 of ≥75%. We considered the following between-study factors in the meta-regression models: whether the study setting was in Europe or not (binary), whether a majority of study participants were recruited from an inpatient setting or not (binary), percentage of men (continuous) and whether the violent outcome was based on forensic care status/criminal record or not (binary).
Sensitivity analyses
To account for different settings and violent outcome types across samples, we conducted analyses including risk factors from only the following: (a) samples where violence was defined by conviction/arrest (violence and/or homicide) or forensic psychiatric care (which typically requires a criminal offence); or (b) samples where 95% or more of the population was recruited from inpatient settings. We also conducted analyses where we excluded risk factors with effect estimates reported as HRs and probit regression coefficients. We further restricted analyses to those papers that were deemed to have a quality score of ≥75%. Finally, we conducted publication bias analyses for risk factors that occurred in at least seven independent samples using Peters’ regression test.Reference Peters, Sutton, Jones, Abrams and Rushton21
Data management and analyses, including effect estimate conversions, were carried out in R version 4.3.0 (R Core Team, Vienna, Austria; see https://www.R-project.org/).22
Results
Study characteristics
We identified 79 988 publications from the five listed databases after removal of duplicates; 288 remained after a title and abstract screening. After full-text screening, 47 studies were included in this update, representing 41 independent samples (Supplementary Fig. 1, Supplementary Tables 2 and 3).Reference Fazel, Wolf, Palm and Lichtenstein5,Reference Arango, Barba, González-Salvador and Ordonez23–Reference Volavka, Van Dorn, Citrome, Kahn, Fleischhacker and Czobor68 Four studies stratified their findings by gender and one further study was stratified by diagnostic category (schizophrenia v. bipolar disorder), and thus each contributed two independent sets of risk factors to the analyses.Reference Fazel, Wolf, Palm and Lichtenstein5,Reference Fazel, Gratin, Carlstrom, Lichtenstein and Langstrom36,Reference Sariaslan, Lichtenstein, Larsson and Fazel54,Reference Witt, Hawton and Fazel62,Reference Witt, Lichtenstein and Fazel63 Fourteen studies overall were overlapping, some of which incorporated aforementioned gender- or diagnosis-stratified samples.Reference Baird, Shaw, Hunt, Kapur, Appleby and Webb24–Reference Beaudoin, Potvin, Giguère, Discepola and Dumais26,Reference Buchanan, Sint, Swanson and Rosenheck28,Reference Fazel, Gratin, Carlstrom, Lichtenstein and Langstrom36–Reference Fazel, Wolf, Larsson, Lichtenstein, Mallett and Fanshawe38,Reference Singh, Grann, Lichtenstein, Långström and Fazel55,Reference Thomas, Leese, Walsh, McCrone, Moran and Burns60,Reference Witt, Hawton and Fazel62–Reference Wootton, Buchanan, Leese, Tyrer, Burns and Creed64,Reference Yesavage66,Reference Yesavage67 Overall, 564 separate effect sizes were extracted, representing 387 unique risk factors after standardisation. Of these, 30 were examined in at least three independent samples, and an additional 29 factors occurred in two. The publication years ranged from 1983 to 2022, with the majority of samples originating from Europe (k = 20, 49%), followed by the USA (k = 10) and the UK (k = 5). Median sample size was 404 (Table 1). Ten samples (24%) recruited more than 95% of their participants from inpatient settings. The majority measured their violent outcome as a physical assault on another person (k = 12, 29%) or as a conviction for a violent crime (k = 12, 29%; Table 1).
Table 1 Characteristics of independent samplesa
PANSS, Positive and Negative Syndrome Scale.
a. For overlapping samples, information from the sample with the greatest sample size is presented.
We found good interrater reliability of the extraction of the proportion of violent individuals (Krippendorff's alpha 0.77).
Table 2 shows the effect estimates of risk factors that were reported in ≥3 independent samples, pooled over the individual risk factor. Fig. 1 illustrates the effect estimates when these risk factors were pooled over the risk factor domain. No risk factors in the neuropsychological domain were present in two or more samples, and hence were not included in the analyses. We allowed individual Positive and Negative Syndrome Scale (PANSS) items to be included as separate risk factors.
Table 2 Pooled odds ratios of risk factors for violence in psychosis
PANSS, Positive and Negative Syndrome Scale; SES, socioeconomic status.
a. Recent: within past year.
Fig. 1 Pooled odds ratios for violence in psychosis by overall risk factor domain.
Criminal history domain
Odds ratio estimates were similar across criminal history factors (Table 2), and were all associated with an increased risk of violence in people with psychosis. The most commonly reported criminal history risk factor was ‘violence history’ (k = 15), with an odds ratio of 2.91 (95% CI = 2.06, 4.10). After pooling across all criminal history factors, the overall domain was associated with more than a three-fold increased risk (odds ratio 3.50, 95% CI = 2.37, 5.16; Fig. 1). A history of non-violent crime and a family history of offending behaviour were also risk factors and were reported in two independent samples (Supplementary Table 4).
Negative and positive symptom domains
Risk factors present in three or more independent samples were the negative symptom score on the PANSS (odds ratio 1.10, 95% CI = 0.95, 1.27) and positive symptom score on the PANSS (odds ratio 1.12, 95% CI = 0.45, 2.82) (Table 2), although neither showed a clear association with violence. In the positive symptom domain, paranoia and hostility were reported in two independent samples (Supplementary Table 4), with paranoia showing the greatest point estimate (odds ratio 3.07, 95% CI = 0.88, 10.69).
Premorbid domain
Two premorbid factors were associated with an increased risk of violence: recent violent victimisation (odds ratio 5.81, 95% CI = 3.45, 9.78) and a history of parental violent crime (odds ratio 1.37, 95% CI = 1.15, 1.63; Table 2). Overall, the premorbid domain was associated with an increased risk of violence (odds ratio 1.79, 95% CI = 1.16, 2.77; Fig. 1). In addition, parental bereavement, ever being non-violently victimised, childhood abuse and a parental history of substance misuse were associated with violence risk, but were reported in only two independent samples (Supplementary Table 4).
Psychopathological domain
A diagnosis of schizophrenia was associated with an increased risk of violence compared with other psychotic disorders in a given study (odds ratio 1.63, 95% CI = 1.17, 2.28; Table 2), as was comorbid diagnosis of personality disorder (odds ratio 2.30, 95% CI = 1.71, 3.09) and number of past admissions to hospital (odds ratio 2.65, 95% CI = 1.45, 4.84). There was also an association for the overall psychopathological domain (odds ratio 1.66, 95% CI = 1.05, 2.62; Fig. 1). A lack of insight, comorbid diagnosis of antisocial personality disorder, a diagnosis of bipolar disorder, younger age at psychosis onset and traumatic brain injury were associated with violence in two independent samples (Supplementary Table 4).
Sociodemographic domain
No educational qualifications versus any (odds ratio 1.46, 95% CI = 1.26, 1.69), Black and minority ethnicity (odds ratio 1.72, 95% CI = 1.08, 2.74) and male gender (odds ratio 1.67, 95% CI = 1.06, 2.64) were associated with higher violence risk (Table 2), as was the overall sociodemographic domain (odds ratio 1.61, 95% CI = 1.07, 2.44).
Substance misuse domain
All factors in the substance misuse domain were linked with an increased risk of violence. Substance misuse, drug misuse (current or recent), alcohol misuse and a history of alcohol misuse had similar effect sizes (odds ratios ranging from 1.61 to 2.41). The overall substance misuse domain was associated with an increased risk of violence (odds ratio 2.36, 95% CI = 1.99, 2.80; Fig. 1). Cannabis use history and recent alcohol misuse were reported in two samples, with the highest odds ratio estimated for a history of cannabis use (odds ratio 3.34, 95% CI = 2.32, 4.82; Supplementary Table 4).
Suicidality domain
Self-harm history was associated with an elevated violence risk (odds ratio 1.74, 95% CI = 1.01, 2.98; Table 2; Fig. 1). Unintentional self-harm, reported in two samples, was also associated with violence (odds ratio 5.50, 95% CI = 4.26, 7.08; Supplementary Table 4).
Treatment-related domain
Antipsychotic medication was associated with a reduced risk of violence (odds ratio 0.51, 95% CI = 0.27, 0.96; Table 2). Treatment adherence had a protective association, though not statistically significant. Antidepressant treatment was also associated with a protective effect (odds ratio 0.80, 95% CI = 0.66, 0.97; Supplementary Table 4) although it was only reported in two samples (Supplementary Table 4).
Sensitivity analyses
When restricting analyses to those samples where a majority of patients were from inpatient settings (k = 10 or 24% of the overall sample), risk factors occurring in two or more of these samples all had non-significant associations apart from the number of previous admissions, reported in two independent samples (odds ratio 2.07, 95% CI = 1.22, 3.53; Table 3). Overall, point estimates were also reduced as compared to the main analysis. Meanwhile, only including samples from community settings did not materially change results (Supplementary Table 5). When restricting analyses to samples where violence was defined by conviction or arrest for violence/homicide (k = 14), there were 11 factors reported in k ≥ 3 independent samples, and findings were similar to the main analyses (Supplementary Table 6). There were no material changes when excluding risk factors where associations were reported as hazard ratios or probit regression coefficients (Supplementary Table 7). Results were also similar when restricting analyses to studies with independent samples that had quality scores ≥75% (Supplementary Table 8). We found no evidence of publication bias.
Table 3 Pooled odds ratios of risk factors for violence in psychosis in inpatient settings
PANSS, Positive and Negative Syndrome Scale; BPRS, Brief Psychiatric Rating Scale.
Meta-regression
For the majority of included risk factors, heterogeneity as measured by the I 2 statistic was high. Meta-regression was conducted for risk factors that occurred in at least seven independent samples and that had an I 2 ≥ 75%. These risk factors include unemployment, self-harm history, alcohol misuse, male gender, substance misuse, drug misuse and violence history. We only found statistically significant results for male gender. In univariate meta-regression, we found that if the study reported the outcome as violent arrest or violent conviction, then being male was statistically significantly associated with a greater risk of violence compared to when the study used a less severe definition of violence. This was also the case for samples from Europe compared with other regions. In samples where there was a higher percentage of men, being male was associated with a lower violence risk as compared to samples where there was a lower percentage of men. None of these sample characteristics retained statistical significance when simultaneously entered into a multivariable meta-regression model.
Discussion
In this updated systematic review and meta-analysis of risk factors for violence in psychosis based on longitudinal studies, we identified 41 independent samples comprising 203 297 individuals. We examined 30 individual risk factors that were reported in at least three independent samples. This synthesis identified novel risk factors and validates associations identified in previous work, providing information on several potentially modifiable risk factors. Importantly, the focus on longitudinal studies reduces the likelihood of reverse causation, meaning that findings can inform risk stratification and help identify potential treatment targets.
We found that criminal history was the risk factor domain with the strongest association with violence, followed by the substance misuse domain. In relation to criminal history, both previous violent and non-violent crime were risk factors – possibly because these measure a general propensity for criminality, or because engaging in criminality introduces individuals to social contexts and networks that increase the risk of subsequent violence. There are many explanations for the importance of comorbid substance misuse as a risk factor domain. Intoxication leads to poorer impulse control, which has been found to be a strong risk factor for violence in cross-sectional studies.Reference Witt, Van Dorn and Fazel11 Individuals with schizophrenia (and other severe mental illnesses, including bipolar disorder) may also self-medicate with substances to manage symptoms,Reference Gregg, Barrowclough and Haddock69 possibly affecting treatment adherence and effectiveness,Reference Lindsey, Stewart and Childress70 while being an indicator of more severe symptomatology. Drug misuse may additionally be an entry route into criminality and expose individuals to violent environments.Reference Lambe, Cooper, Fazel and Freeman71,Reference Zhong, Yu and Fazel72 Another important risk factor domain is suicidality. It is likely that common processes underpin violence and suicidality – for example, they may both be outward expressions to regulate intense internal states.Reference Lambe, Cooper, Fazel and Freeman71 The strength of the association for these and other studied risk factor domains were broadly similar to the previous review,Reference Witt, Van Dorn and Fazel11 validating the evidence using longitudinal designs.
An unexpected finding in this review was the lack of a clear association between overall positive symptom scores and violence, despite previous research finding that positive symptoms are an important risk factor for violence.Reference Coid, Kallis, Doyle, Shaw and Ullrich73 However, the CI for our result was wide (odds ratio 1.12, 95% CI = 0.45, 2.82), and with a high I 2 statistic. We also found a modest, although statistically non-significant, association for negative symptom score (odds ratio 1.10, 95% CI = 0.95, 1.27). While this may be consistent with the null association found in the previous review (odds ratio 1.00, 95% CI = 0.90, 1.20),Reference Witt, Van Dorn and Fazel11 another possibility is that negative symptoms are a marker of disease severity and partial response to treatment. It is also possible that other symptoms associated with violence, such as hostility, are misclassified as negative symptoms owing to their overlap with blunted affect and asociality.Reference Correll and Schooler74
Antipsychotic treatment was statistically significantly associated with reduced violence (odds ratio 0.51, 95% CI = 0.27, 0.96), in line with the protective effect expected from interventions that reduce psychotic symptoms. We also found that treatment with an antidepressant (reported in two independent samples) was a protective factor (odds ratio 0.80, 95% CI = 0.66, 0.97). It is possible that the affective component of a psychosis is being treated with the antidepressant, including anger, which has been shown to be one mechanism for violence in schizophrenia.Reference Coid, Kallis, Doyle, Shaw and Ullrich73 Alternatively, being under medical treatment may be a marker of closer contact with clinical services. Another novel finding is that cannabis use (reported in two independent samples) was identified as a risk factor for violence in the current review with an odds ratio of 3.34 (95% CI = 2.32, 4.82), whereas it had a null association in the previous review. This suggests that cannabis use disorder could be considered as part of an individualised violence risk assessment.Reference Dellazizzo, Potvin, Athanassiou and Dumais75,Reference Dellazizzo, Potvin, Beaudoin, Luigi, Dou and Giguère76 It is also possible that cannabis drives at least part of the overall association of substance misuse with violence, given that many studies did not separate out the effect of cannabis misuse from that of other substances. Previous literature has found that cannabis use is associated with earlier onset of schizophrenia and other psychotic disorders,Reference Hjorthøj, Compton, Starzer, Nordholm, Einstein and Erlangsen77–Reference Di Forti, Sallis, Allegri, Trotta, Ferraro and Stilo79 and that earlier onset of psychosis is associated with greater risk of violent crime.Reference Moulin, Alameda, Framorando, Baumann, Gholam and Gasser80 We further identified traumatic brain injury as a potential risk factor, which was not reported in the previous review. While organic brain disorders have been implicated in violence,Reference Fazel, Philipson, Gardiner, Merritt and Grann81 further research is necessary – especially as we only identified this risk factor in two independent samples.
Clinical implications
Our results confirm the importance of previous crime and previous or current substance misuse, and also identify potential novel risk factors including cannabis use. These findings can inform more precise stratification of violence risk for patients with psychotic disorders. However, the clinical impact of accurate risk prediction models depends on whether effective interventions are available. Many of the identified risk factors, such as comorbid substance misuse, are potential targets for clinical intervention if found to be causal. Others may be markers of modifiable risk factors. For example, criminal history could be a marker of pro-criminal beliefs that may be reduced using cognitive or behavioural interventions.Reference Lambe, Cooper, Fazel and Freeman71 Further research is recommended to test the impact of different types of interventions in individuals with psychosis.
We have included multiple study settings, and it is possible that certain risk factors vary in their association depending on the context in which the assessment is made. When we restricted our analysis to inpatient settings, risk factors were attenuated and became statistically non-significant. Patients from inpatient settings are likely to be a selected sample owing to greater severity of symptoms and may have more similar risk factor profiles, leading to lower variance in violence risk. Consequently, more in-depth risk assessment may be necessary.
Strengths and limitations
We have focused on longitudinal studies, which reduces the likelihood that reverse causation explains the results. However, several limitations should be noted. First, we relied on minimally adjusted effect estimates, given that covariate adjustment varied across studies. This may help explain why the effect estimates are quite similar within certain domains (notably the criminal history domain). Different types of criminality often co-occur within individuals – the presence of a given type may therefore be a marker of another, obscuring the relationship of specific risk factors with violence. We did not have individual participant data, which may have allowed us to provide consistently adjusted effect estimates.Reference Riley, Lambert and Abo-Zaid82 Second, causal inferences are not possible, given that we extract minimally adjusted effect estimates and include a mix of study designs with different aims (descriptive, causal and predictive). Nonetheless, our results may offer insight into risk factors that are important to consider in predictive and causal models, and – if found to be causal – that are modifiable in the sense that they can be improved through clinical intervention. Both types of risk factors are important – causal factors would allow for interventions to mitigate risk, while predictive ones would provide better risk stratification. Third, our decision to pool individual risk factors and risk factor domains needs to be considered in the context of substantial heterogeneity, for example, in terms of study designs and measures. However, pooling can provide a broad overview of a large body of evidence.Reference Witt, Van Dorn and Fazel11 Fourth, some risk factors were only measured in a cross-sectional setting, notably those based on detailed neuropsychological or cognitive tests or biomarkers, which meant that they were not included in our review. Fifth, we did not use a second reviewer for screening, given the large number of references at the screening phase and because we followed a previous protocol. However, the original search and the updated one were carried out by independent reviewers. Sixth, included studies predominantly come from high-income Western countries and further research is needed to better understand risk factors in other regions.
In this updated systematic review and meta-analysis using longitudinal designs, we have validated the importance of comorbid substance misuse and criminal history as risk factors for violence in psychosis, while identifying potential novel risk factors such as cannabis misuse that require replication. If found to be causal, several of these factors represent modifiable intervention targets.
Supplementary material
Supplementary material is available online at https://doi.org/10.1192/bjp.2024.120.
Data availability
Data availability is not applicable to this article as no new data were created or analysed in this study.
Author contribution
T.L., R.Y. and S.F. designed the study. T.L. carried out the search and screening with support from R.Y. and S.F. T.L. carried out data extraction and S.L conducted quality control. A.P. assessed the quality of included studies with support from T.L., R.Y. and S.F. T.L. carried out data management and analysis. T.L. drafted the manuscript. All authors provided detailed comments on the manuscript and contributed to the interpretation of results.
Funding
This work was supported by the National Institute for Health and Care Research (NIHR) Applied Research Collaboration Oxford and Thames Valley at Oxford Health NHS Foundation Trust, Oxford, UK, and by the Wellcome Trust (#202836/Z/16/Z). S.F. is also funded by the National Institute for Health and Care Research (NIHR) Oxford Health Biomedical Research Centre, Oxford, UK.
Declaration of interest
None.
Transparency declaration
T.L. acts as guarantor and affirms that the manuscript is an honest, accurate and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.
Analytic code availability
The analytic code for running the meta-analysis is available upon request.
Research material availability
Tabular data is available upon reasonable request.
Open access
eLetters
No eLetters have been published for this article.