Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Identifying Causal Risk Factors for Violence among Discharged Patients

  • Jeremy W. Coid,

    Affiliation: Violence Prevention Research Unit, Wolfson Institute of Preventive Medicine, Queen Mary University of London. Garrod Building, Turner Street, London E1 2AD, United Kingdom

  • Constantinos Kallis,

    Affiliation: Violence Prevention Research Unit, Wolfson Institute of Preventive Medicine, Queen Mary University of London. Garrod Building, Turner Street, London E1 2AD, United Kingdom

  • Mike Doyle,

    Affiliation: Institute of Brain, Behaviour and Mental Health, University of Manchester. Oxford Road, Manchester M13 9PL, United Kingdom

  • Jenny Shaw,

    Affiliation: Institute of Brain, Behaviour and Mental Health, University of Manchester. Oxford Road, Manchester M13 9PL, United Kingdom

  • Simone Ullrich

    s.ullrich@qmul.ac.uk

    Affiliation: Violence Prevention Research Unit, Wolfson Institute of Preventive Medicine, Queen Mary University of London. Garrod Building, Turner Street, London E1 2AD, United Kingdom

Identifying Causal Risk Factors for Violence among Discharged Patients

  • Jeremy W. Coid, 
  • Constantinos Kallis, 
  • Mike Doyle, 
  • Jenny Shaw, 
  • Simone Ullrich
PLOS
x

Abstract

Background

Structured Professional Judgement (SPJ) is routinely administered in mental health and criminal justice settings but cannot identify violence risk above moderate accuracy. There is no current evidence that violence can be prevented using SPJ. This may be explained by routine application of predictive instead of causal statistical models when standardising SPJ instruments.

Methods

We carried out a prospective cohort study of 409 male and female patients discharged from medium secure services in England and Wales to the community. Measures were taken at baseline (pre-discharge), 6 and 12 months post-discharge using the Historical, Clinical and Risk-20 items version 3 (HCR-20v3) and Structural Assessment of Protective Factors (SAPROF). Information on violence was obtained via the McArthur community violence instrument and the Police National Computer.

Results

In a lagged model, HCR-20v3 and SAPROF items were poor predictors of violence. Eight items of the HCR-20v3 and 4 SAPROF items did not predict violent behaviour better than chance. In re-analyses considering temporal proximity of risk/ protective factors (exposure) on violence (outcome), risk was elevated due to violent ideation (OR 6.98, 95% CI 13.85–12.65, P<0.001), instability (OR 5.41, 95% CI 3.44–8.50, P<0.001), and poor coping/ stress (OR 8.35, 95% CI 4.21–16.57, P<0.001). All 3 risk factors were explanatory variables which drove the association with violent outcome. Self-control (OR 0.13, 95% CI 0.08–0.24, P<0.001) conveyed protective effects and explained the association of other protective factors with violence.

Conclusions

Using two standardised SPJ instruments, predictive (lagged) methods could not identify risk and protective factors which must be targeted in interventions for discharged patients with severe mental illness. Predictive methods should be abandoned if the aim is to progress from risk assessment to effective risk management and replaced by methods which identify factors causally associated with violence.

Introduction

The development of structured risk assessment for violent and antisocial behaviour has been a major advance in clinical practice, both within the mental health and the criminal justice system. More recently, however, there have been fewer developments in the field, with division between adherents to actuarial methods of assessment and Structured Professional Judgement (SPJ) [14]. A substantial number of risk assessment instruments have been developed, aimed to help clinicians when making decisions, but with little convincing evidence of superiority of any new instruments over their predecessors [5, 6]. Assessment of predictive accuracy using AUC values has been the standard measure of utility and is almost exclusively used in standardisation and validation of both actuarial and SPJ instruments. The risk factors are typically measured once and then used to statistically predict the occurrence of violence in a subsequent time window with its length substantially varying from study to study. This method assures temporal ordering of risk factor (exposure) and violent behaviour (outcome) but does not take into consideration that some risk factors change and may no longer be in operation when the outcome is measured. Furthermore, predictive accuracy rarely exceeds a moderate level of precision [5] and clinicians must expect three out of ten predictions to be incorrect even when using the best performing instruments. Moreover, a reduction in accuracy (shrinkage) typically occurs when an instrument is subsequently administered in a population on which it was not standardised [7]. Most instruments are checklists where a minority of items have predictive power on which their predictive efficacy ultimately depends [8, 9]. Despite their widespread use, only one randomised controlled trial has been carried out so far which has failed to demonstrate a reduction of violence following administration of an SPJ instrument [10]. There is currently no evidence that violence can be prevented using a standardised risk assessment [10, 11].

Taking into consideration these limitations, it is questionable whether statistical prediction is the most appropriate clinical goal when using a risk assessment instrument if the aim is to intervene to prevent violence. Successful management of risk can only be achieved by targeting risk factors causally related to violence. Surprisingly, few studies have attempted to investigate causal associations between risk factors and violence. Risk factors can be classified as static or dynamic. Whilst no change will occur in the static risk (e.g. gender, ethnicity), it is the nature of a dynamic risk factor to change and fluctuate over time, with varying speed. The risk factor may therefore no longer be present in a subsequent time window when violence is measured. In a predictive (lagged) model, where the risk factor occurs some time before the outcome, significant associations may consequentially be missed. To identify associations which can be considered causal, temporal proximity (co-occurrence) between the risk factor and outcome is of utmost importance [1214]. This can be difficult to establish and requires specification of a definite time frame when exposure and outcome occur. However, it may not be possible to establish temporal ordering within this time frame and, therefore, associations detected in this specified time window have to be interpreted based on plausibility and available knowledge of cause and effect.

Furthermore, causal risk factors can be poor predictors. For example, persecutory delusions did not demonstrate an association with violence among discharged patients with psychotic illness when the exposure measured in the past ten weeks was used to predict the outcome in the subsequent ten weeks. When considering their co-occurrence in the same ten-week window, there was a substantial and robust association although the effect was mediated by anger due to the delusional beliefs [1315].

A treatment or risk management intervention which targets a predictor variable, even if it has a high level of accuracy, cannot succeed in preventing violence unless it is also causally related. If the intention is to reduce risk by improving risk management, then causal factors must be accurately identified and subsequently targeted for intervention. To demonstrate the advantages of causal compared to predictive models for the implementation of clinical risk management we carried out a prospective longitudinal study of patients discharged from medium secure services in England and Wales over a twelve-month period. The overall aim of the study was to compare the utility of predictive and causal models when identifying dynamic risk factors that should be targeted in future interventions. To this purpose we (i) examined predictive efficacy of total scores and individual items using two SPJ instruments, one measuring risk, the other protective factors, on violent outcome by using the risk/ protective factors measured as present in the past six months to predict violence in the subsequent six months (lagged model); (ii) we then tested causal models, including the same dynamic variables, in which we considered their co-occurrence within the same six months’ time window (temporal proximity). Since causal pathways to violent behaviour are not always direct, i.e. some associations are explained by a third variable which is associated with both exposure and outcome, we (iii) performed statistical analyses to identify explanatory variables which are the main drivers in the association with violence.

Methods

The method of data collection has previously been described [16]. In brief, a prospective cohort follow-up study was carried out on all forensic patients discharged from 32 National Health Service (NHS) medium secure units across England and Wales (in 26 NHS secondary care Trusts) between 1st September 2010 and the 31st August 2011. These patients had been detained under compulsory orders following violent and criminal behaviour. Patients who were discharged to the community rather than prison were eligible for the follow-up study. ‘Community’ placements for the purpose of the study included accommodation such as independent tenancies, supported accommodation, hostels, open rehabilitation wards, and open psychiatric units.

A link person was identified at each hospital site and a notification system set up so that researchers would be automatically informed when a patient was discharged. Baseline assessments were then conducted for those discharged by interviewing a member of staff who knew the patient well, and a review of clinical and criminal records. The patients were twice followed-up after release at six and twelve months.

The North West England multi-site research ethics committee approved the study. To ensure a total sample of discharges, permission was sought and granted by the UK National Information Governance Board (NIGB) to conduct the study without patient consent under Section 251 of the NHS Act, 2006.

Measures

Demographic and diagnostic information (primary diagnosis) was recorded at baseline (pre-release) for each patient. Measures of risk and protective factors were completed based on information from a collateral interview and file review. The following assessments were carried out:

The Historical Clinical Risk Management– 20 items version 3 (HCR-20v3) is a violence risk assessment covering a broad range of items. It comprises 10 historical (static) factors, 5 clinical items meant to reflect current, dynamic (changeable) correlates of violence, and 5 risk management items which focus on future factors. The items were scored 0 if not present, 1 if partially present, and 2 if definitely present. Total scores range from 0 to 40.

The Structured Assessment of Protective Factors (SAPROF) is a guideline developed to measure protective factors that mitigate future risk of violence. It includes 17 items covering internal, motivational and external factors. Each item is rated on a 3-point scale (0, 1 or 2) as the HCR-20v3. Total scores range between 0 and 34.

Violence was measured using the McArthur community violence instrument (MCVI) which comprises 18 questions focusing on violent incidents not necessarily leading to a caution or conviction. THE MCVI differentiates experiences of violent victimisation and violence perpetration. Violence perpetration was defined by combining ‘violence’ and ‘other aggressive acts’, which could include sexual acts, assaultive acts involving use of a weapon, or threats made with a weapon in the hand, as well as acts of battery, regardless of whether they resulted in injury. Verbal threats alone were not included. As with the HCR-20v3 and SAPROF, details of any violent incidents mentioned in the clinical case files were noted and cross-matched with the information gathered at interview with the link person. In addition, details of criminal convictions in the twelve months post-discharge were obtained from the Police National Computer (PNC). An individual was deemed to have offended if they had committed a recordable offence within the follow-up period and had the offence proven later, either by accepting a caution, warning or reprimand, or by being found guilty in a court of law. MCVI ratings of violent perpetration as described above and PNC violent cautions/ convictions were then combined to a measure of “any” violence.

Ratings of the HCR-20v3 and SAPROF were carried out at baseline, prior to or shortly after discharge, and at 6 and 12-months post discharge (apart from the historical risk factors of the HCR-20v3 which are considered stable and were only assessed at baseline). Measures of violence were taken at 6 and 12-months post-discharge.

Inter-rater reliability was calculated between four raters on 20 cases who had been trained and supervised in the use of all instruments. Raters obtained an IRR of 0.92 on the HCR-20v3 and 0.98 on SAPROF total scores.

Statistical analysis

For the predictive approach (lagged model), risk/ protective factors which occurred in the past 6 months were modelled as statistical predictors for violent outcome occurring in the subsequent 6 months. In the temporal proximity or causal models, both predictors and outcome occurred within the same 6 month time window.

In order to take advantage of the longitudinal design, multilevel modelling was performed. These models account for the dependence of data collected longitudinally by modelling the correlation of repeated measures within the same individual as random effects. Unlike other approaches, such as ANOVA, mixture models do not require that data are complete for individuals at each time point or imputation of data which may result in bias [17].

We firstly assessed associations of individual HCR-20v3 and SAPROF items using the Area under the ROC Curve (AUC), a statistic that provides a measure of discrimination for each item between violent and non-violent cases. As stated above, analyses were performed using a predictive/ lagged and a temporal proximity/ causal modelling approach. We used the Stata command somersd to estimate AUC statistics that allows for clustering to be taken into consideration. However, somersd allows only for one level of clustering and we therefore estimated AUC statistics allowing for repeated measures to be nested within subjects (2-level structure).

We then carried out predictive/ lagged and temporal proximity/ causal logistic regression analyses to obtain estimates of the Odds Ratio (OR) for the effects of each risk/ protective factor on violence. These were 3-level mixed effect models with repeated measures (level 1) nested within subjects (level 2) nested within the NHS Trusts responsible for the secure services from where the patients had been discharged (level 3).

Since information on temporal ordering was not available for the risk/ protective factors and it was not possible to plausibly infer the order, we could not carry out mediation analyses. The final set of analyses therefore aimed to identify variables which explained the association of certain risk/ protective factors with violent outcome. In order to qualify as explanatory variable a risk/ protective factor was required to be associated with both exposure (other risk/ protective factors) and violent outcome. Only if both associations were significant at an alpha level of <0.05 were variables selected and then entered in an adjusted model. We examined the percentage reduction in the baseline odds of each risk/ protective factor after adding each of the potentially explanatory variables into the following equation:

In a final model, all explanatory variables were entered simultaneously. Comparisons between baseline-adjusted and fully adjusted coefficients were used to estimate the extent to which the association between a risk/ protective factor and violent outcome was accounted for by the explanatory variable. Percentages above 100 were possible if the OR in the fully adjusted model changed to a value below or above 1. As above, we again performed 3-level mixed effect models with repeated measures (level 1) nested within subjects (level 2) nested within discharging NHS Trusts (level3).

All statistical analyses were carried out in Stata 14. The statistical programming code can be found in S1 Appendix. A significance level of 0.05 was adopted throughout.

Results

There were 788 patients discharged during the study period, of whom 409 (52%) were discharged to the community. At 6-months post discharge, collateral interview and case note reviews were completed for 387 (95%), and at 12-months for 344 (89%) patients. At 6-months post discharge, 54 (14.0%) patients had committed a violent act, and at 12-months 43 (12.5%) had been violent. It was possible that a patient acted violently in both 6 month time windows.

The mean age of the study participants was 37.8 years (SD = 9.7); 344 (88.9%) men and 43 (11.1%) women; 232 (60.1%) were white, 98 (25.4%) black, 24 (6.2%) South Asian, 24 (6.2%) mixed heritage, and 8 (2.1%) Chinese or of other ethnic origin. Primary diagnoses included 313 (80.9%) schizophrenia/ schizoaffective disorder, 28 (7.2%) bipolar disorder, 21 (5.4%) personality disorder, 3 (3.8%) anxiety disorder, 5 (1.3%) depression, 1 (0.3%) substance use, and 16 (4.1%) other diagnoses.

Predictive accuracy of static risk factors

As shown in Table 1, only 4 of the 10 individual static risk factors of the historic domain of the HCR-20v3 achieved a statistically significant but poor level of accuracy including: other antisocial behaviour, traumatic experiences, violent attitudes and treatment or supervision response. The AUC value for the total score was also significant but predicted violence just above chance. This was reflected by the results of the logistic regression model which identified the same variables as significant predictors of violent behaviour. After adjustment of age, gender, ethnicity and primary diagnosis at baseline, only traumatic experiences and the historical total score remained significant.

thumbnail
Table 1. Predictive accuracy of HCR-20v3 historical items (measured at baseline).

http://dx.doi.org/10.1371/journal.pone.0142493.t001

Dynamic risk and protective factors for violence: predictive vs. causal models

The AUC values of the lagged/ predictive and temporal proximity/ causal analyses for the dynamic items of the HCR-20v3 and SAPROF are reported in Table 2 and Fig 1. In the predictive model, all dynamic items of the HCR-20v3 except symptoms of mental disorder and professional services and plans achieved a significant AUC value although of poor magnitude. In the causal model, only professional services and plans did not reach the threshold of statistical significance. When comparing the 95% confidence intervals of the two models, the AUCs of most of the clinical risk factors were significantly higher (the 95% CIs did not overlap) and of moderate accuracy in the causal model. These included: violent ideation or intent, symptoms of major mental disorder, instability, and the total clinical risk score.

thumbnail
Fig 1. Predictive vs. causal models–accuracy.

White bars reflect the predictive / lagged model, grey bars reflect the temporal proximity / causal model. The error bars represent 95 per cent confidence intervals.

http://dx.doi.org/10.1371/journal.pone.0142493.g001

thumbnail
Table 2. Predictive accuracy of lagged vs. temporal proximity models—AUC values.

http://dx.doi.org/10.1371/journal.pone.0142493.t002

Apart from professional services and plans, all items of the HCR-20v3 risk management scale achieved a significant AUC value in the predictive model (Table 2 and Fig 1). The causal model corresponded to these findings. The AUC value of the total score, though, was significantly higher in the causal compared to the predictive model.

With regard to the protective factors of the SAPROF, predictive and causal analyses yielded similar results regarding the magnitude of the AUC values (Table 2 and Fig 1). Intelligence, secure attachment in childhood, intimate relationship and professional care failed to achieve a significant AUC value in the predictive model compared to intelligence, secure attachment in childhood, intimate relationship, professional care, living circumstances and external control where the AUCs did not reach statistical significance in the causal model. The AUC value of self-control was significantly higher in the causal model.

In the logistic regression model, all HCR-20v3 clinical items were significantly associated with violence in the predictive model with the exception of symptoms of major mental disorder. This risk factor, however, was significant using the causal modelling approach. Apart from treatment or supervision response, all ORs were significantly greater in magnitude in the causal model (Table 3 and Fig 2).

thumbnail
Fig 2. Predictive vs. causal models–strength of association.

White bars reflect the predictive / lagged model, grey bars reflect the temporal proximity / causal model. The error bars represent 95 per cent confidence intervals.

http://dx.doi.org/10.1371/journal.pone.0142493.g002

thumbnail
Table 3. Strength of association in lagged vs. temporal proximity models–odds ratios.

http://dx.doi.org/10.1371/journal.pone.0142493.t003

Personal support, treatment or supervision response, stress or coping, and the total score in the risk management domain were significantly associated with violent outcome in the predictive model (Table 3 and Fig 2). These risk factors were also significant in the causal model in addition with living situation. Stress or coping demonstrated a significantly stronger association with violence in the causal model.

As shown in Table 3 and Fig 2, the majority of protective factors were inversely associated with violence in the predictive model apart from medication, intimate relationships, professional care, and living circumstances. Variables that were not significantly associated with violence in the causal model included intimate relationships, professional care, living circumstances, and external control. The inverse association with self-control was significantly stronger in the causal compared to the predictive model.

Identification of explanatory variables

As shown in Table 4, the items of the HCR-20v3 (historical and dynamic) and of the SAPROF were only moderately correlated. In order to identify risk and protective dynamic factors independently associated with violent outcome, they were entered simultaneously in the logistic regression model. Table 5 shows that in the baseline model, among the 10 dynamic risk factors of the HCR-20v3, only 3 (violent ideation or intent, instability, and stress or coping) demonstrated independent effects on violent behaviour. In model 1, Table 5, the baseline models which resulted in a significant association with violence (Table 3) were next adjusted for the first explanatory variable, violent ideation or intent. Although there was attenuation in the OR of symptoms of major mental disorder, the association was still significant. Inclusion of instability led to substantial attenuation of all ORs in model 2, whereas inclusion of stress or coping in model 3 did not demonstrate explanatory effects in the association between lack of insight, symptoms of major mental disorder, and treatment or supervision response with violent behaviour. Inclusion of all 3 explanatory variables in model 4 resulted in statistically non-significant effects of all variables with a change in ORs of approximately 100%. This indicates that violent/ ideation or intent, instability, and stress or coping fully accounted for the association between lack of insight, symptoms of major mental disorder, treatment or supervision response (clinical), living situation, personal support, and treatment or supervision response (risk management) with violent behaviour.

thumbnail
Table 5. Explanatory variables—HCR-20v3 clinical and risk management variables.

http://dx.doi.org/10.1371/journal.pone.0142493.t005

The same procedure was finally carried out with the items of the SAPROF. As can be seen in Table 6, only self-control demonstrated an independent protective effect on violent outcome (model 0). After adjusting the significant protective factors derived from Table 3, most of these were no longer significant, indicating that self-control was a highly important explanatory factor. Only the effects of empathy, leisure activities, and attitudes towards authority were not explained by self-control. These factors conveyed direct protective effects on violent behaviour.

Discussion

The standard approach to validation of SPJ instruments employs a predictive method, with corresponding statistical tests to demonstrate predictive accuracy. Accuracy is typically measured using the AUC statistic and, less commonly, methods used to evaluate medical screens for diseases [1821]. Using this approach, we were unable to validate either the HCR-20v3 or SAPROF (using their individual items and total scores) in a representative sample of UK patients discharged from medium security who had committed serious offences or had been unmanageable in other hospital services due to their violence. Items in these instruments were poor at discriminating which patients would become violent over the next 12 months and showed low levels of statistical association with outcome. The predictive method did not therefore indicate which risk factors should be targeted for future risk management or treatment interventions if the aim is to prevent violence. We also observed that the HCR-20v3 performed more poorly than in previous studies using the HCR-20v2 [5] despite its recent revision aimed to improve its clinical utility. Similarly, few items in the SAPROF showed any discriminative value in identifying patients who would not become violent, with low levels of association between each item and no reported violence over a 12 month period. Nevertheless, consideration of temporal proximity of risk/ protective factors and violence indicative of a causal relationship demonstrated better clinical utility in terms of accuracy and strength of association with violence. We found that the predictive approach substantially underestimated the effects of violent ideation or intent, symptoms of major mental disorder and instability as risk, and self-control as protective, factors for violent behaviour. Their AUC values in the causal model were significantly higher than in the predictive model and achieved satisfactory accuracy. This was confirmed in logistic regression models testing the strength of association where the same variables conveyed significantly higher risk/ protective effects on violent outcome and should be therefore considered as important factors for intervention.

It is of interest that the majority of the variables that were strongly discriminative and associated with violence stemmed from the clinical domain of the HCR-20v3. Recent research has demonstrated that most changes in HCR-20v2 scores are clinically irrelevant and that only the score of the clinical domain reduced over repeated risk assessments over time [22]. Our findings suggest that this effect continues with the 3rd version of the HCR-20 since the majority of risk management items showed similar discriminative effects and strength of association in the predictive and the causal models. It appears that there are only small changes in most of these items over a follow-up period of 12 months. We also observed the same pattern with the items of the SAPROF and where only self-control demonstrated significantly better discriminative effects and was a stronger protective factor in the causal model. This suggests that the majority of items in the SAPROF are relatively stable over time.

Explanatory factors for violence

We confirmed that the majority of risk and protective factors were in operation during follow-up, significantly associated with violence, and could therefore have exerted causal effects on violence. However, because only violent ideation or intent, instability, and stress/poor coping were independent risk factors, and only self-control was independently protective, it was unlikely that the other risk and protective factors we identified had exerted their effects directly on violence.

Analysis confirmed that poor insight, symptoms of major mental disorder, poor treatment response, low level of personal support, and unsatisfactory living situation all increased risk of violence, but only when accompanied by violent ideation, behavioural instability, and stress. These latter factors appeared to be explanatory factors in our statistical analysis. They should therefore be key targets for intervention in risk management. Nevertheless, all factors were closely linked and violent ideation, behavioural instability, and stress may have been the outcome of combinations of other factors involving deteriorating mental state, lack of insight, poor response to treatment, and lack of support whilst in an unsatisfactory living situation, all of which require preventive intervention.

Similarly, empathy, coping, work, leisure activities, good financial management, motivation for treatment, positive attitudes towards authority, life goals, taking medication, and positive social networks all conveyed protective effects and reduced the risk of violence, but only when accompanied by good self-control. The explanatory effect of just one variable on these other factors was unexpected and requires further consideration. Good self-control would appear a broad and, arguably, an over-inclusive factor to include in a checklist of protective items, potentially overlapping with certain factors we found were no longer associated with violence after adjusting for each other. Most importantly, the converse of good self-control is poor self-control, as indicated by a history of previous violence in this study.

Limitations

Our study relied on data collected from extensive case files. Baseline and follow-up interviews were conducted with collaterals who were professional staff currently working with the patients with good knowledge of their past history, recent mental state and functioning. However, absence of a direct patient interview is a limitation of this approach. Nevertheless, this method was deliberately chosen to prevent attrition and where a substantial subgroup of patients reported they would not have co-operated in face-to-face interviews or follow-up, a proportion of whom were subsequently violent. This method, which did not require patient consent under UK legislation, minimised loss to follow-up and resulting biases.

We identified explanatory variables by entering them simultaneously in the statistical model and, therefore, exceeded the recommended ratio of 1 covariate per 10 events. As a consequence, we cannot rule out the possibility of over-fitting in these models. Larger samples with multiple measures of violent outcome are necessary to confirm our findings.

When investigating relationships between dynamic risk/ protective factors and violence, it is essential to account for the fact that both exposure and outcome are dynamic in nature and fluctuate over time. It has been emphasized that temporal proximity is of utmost importance to uncover such associations [1214]. However, it can be argued that temporal proximity does not necessarily imply temporal ordering and, therefore, the possibility of reversed causality has to be taken into consideration. Criteria of causality were published some decades ago [23] and have been the guidelines in epidemiology. These include temporal ordering of exposure and outcome. However, another criterion is plausibility and (from a clinical perspective) it is more plausible that the risk and protective factors identified in our study lead to an increase/ reduction in violence rather than the reverse pathway.

Implications

To our knowledge, this is the first study to compare predictive and causal models of association between risk and protective factors, measured using SPJ instruments, in a prospective follow-up study. We found that prediction of future violence has very limited usefulness in a clinical risk assessment. Assessment of predictive accuracy cannot facilitate the identification of the most important risk factors that must be targeted to prevent violence. It also questions whether developing new instruments which are validated using predictive methods [8, 9] should continue in future.

Moving to causal models, which in turn lead to improved clinical interventions and ultimately violence prevention, is necessary for improved service provision and clinical performance. Patients in the offender population we studied have shown increasing length of stay. The throughput of patients admitted to secure forensic services in the UK and other European countries has progressively declined over the past two decades. This follows increasing political and other external pressures against discharging offender patients within increasingly risk-adverse European societies. Emphasis on risk without corresponding links to treatment and violence prevention has further contributed to this decline in throughput and increased length of stay, with no available evidence of improved public protection.

Previous studies which have used a predictive approach and relied entirely on retrospective, case register, and cross-sectional studies emphasise the importance of demographic, social, and previous criminal history factors, and report no association [24] or even negative associations between psychotic symptoms [2527] and violent criminal behaviour. Similarly, that the association is largely or entirely explained by comorbid substance misuse or psychosocial adversity and environmental stressors more common in the lives of people with severe mental illness [2830]. These factors undoubtedly increase risk of violence among persons with psychotic illness, together with lack of supervision after discharge, poor treatment compliance and treatment response. However, certain methodological shortcomings in earlier studies, particularly failure to investigate causal associations with dynamic factors, have led to over-emphasis of the role of pre-existing static risk factors. A key finding in this study was that, using the predictive approach, the HCR-20v3 failed to identify the strong, causal links we observed between appearance of symptoms of mental illness following discharge and subsequent violent behaviour. Reliance on findings from these previous studies could result in clinicians missing clear warning signs of future risk and failing to intervene, with serious consequences.

These findings have additional implications for future research. It has previously been observed that predictive accuracy is not influenced by large sample size, whether studies are retrospective, based on review of case files, or prospective, involving interviews by trained researchers [5, 6]. However, future research into causal associations will require adequate statistical power, using more complex methods within longitudinal studies, with multiple repeated measures, and using more appropriate statistical methods. A small number of previous studies have attempted to measure change over time using repeated ratings with actuarial instruments [31, 32] and SPJ [3336]. However, predictive accuracy formed the basis of statistical analysis in these studies.

To manage risk and prevent violence, clinicians need to know which factors are causally related to that risk and ultimately which interventions are most effective. Alternatively, which factors are truly protective and mitigate that risk, and which might constitute the basis of a future intervention. No SPJ instrument currently specifies what interventions should be used on the basis of their ratings and training manuals leave such decisions to clinicians following a formulation [37]. The present study shows that in similar clinical samples, which largely include patients with psychotic illness, treatment interventions aimed to prevent deterioration in mental state, targeting symptoms of psychosis, associated affective symptoms and affect due to psychotic symptoms, together with associated behavioural disturbance, should be the main priority. Signs of increasing violent intent, behavioural instability, and stress/poor coping should result in urgent intervention.

Finally, our findings indicate that when training others in the use of such instruments, trainers will require extensive personal experience of the complex interplay between risk and protective factors, rather than just theoretical knowledge of managing risk, together with clinical skills in identifying risk factors, such as symptoms, which have causative associations with violence. Most importantly, they require experience of how to act effectively to prevent violence before it occurs.

Supporting Information

S1 Appendix. Statistical programming code.

doi:10.1371/journal.pone.0142493.s001

(DOCX)

Author Contributions

Conceived and designed the experiments: JWC CK MD JS SU. Performed the experiments: JWC CK MD JS SU. Analyzed the data: CK SU. Contributed reagents/materials/analysis tools: JWC CK MD JS SU. Wrote the paper: JWC CK MD JS SU.

References

  1. 1. Large MM, Neilssen OB. Probability and loss: two sides of the risk assessment coin. Psychiatr Bull. 2011;35:413–8. doi: 10.1192/pb.bp.110.033910
  2. 2. Cooke D, Michie C. The generalizability of the Risk Matrix 2000: On model shrinkage and the misinterpretation of the area under the curve. Journal of Threat Assessment and Management. 2014;1(1):42–55. doi: 10.1037/tam0000004
  3. 3. Falzer P. Valuing structured professional judgment: predictive validity, decision-making, and the clinical-actuarial conflict. Behav Sci Law. 2013;31(1):40–54. doi: 10.1002/bsl.2043. pmid:23339121
  4. 4. Hart SD, Cooke DJ. Another look at the (im-)precision of individual risk estimates made during actuarial risk assessment instruments. Behav Sci Law. 2013;31(1):81–102. doi: 10.1002/bsl.2049. pmid:23322617
  5. 5. Yang M, Wong SC, Coid J. The efficacy of violence prediction: a meta-analytic comparison of nine risk assessment tools. Psychol Bull. 2010;136(5):740–67. doi: 10.1037/a0020473. pmid:20804235
  6. 6. Fazel S, Singh J, Doll H, Grann M. Use of risk assessment instruments to predict violence and antisocial behaviour in 73 samples involving 24 827 people: systematic review and meta-analysis. BMJ. 2012;345:e4692. doi: 10.1136/bmj.e4692. pmid:22833604
  7. 7. Monahan J, Steadman HJ, Robbins PC, Appelbaum P, Banks S, Grisso T, et al. An actuarial model of violence risk assessment for persons with mental disorders. Psychiatr Serv. 2005;56(7):810–5. pmid:16020812 doi: 10.1176/appi.ps.56.7.810
  8. 8. Kroner DG, Mills JF, Reddon JR. A coffee can, factor analysis, and prediction of antisocial behavior: the structure of criminal risk. Int J Law Psychiatry. 2005;28(4):360–74. pmid:15936077 doi: 10.1016/j.ijlp.2004.01.011
  9. 9. Coid J, Yang M, Ullrich S, Zhang T, Sizmur S, Farrington D, et al. Most items in structured risk assessment do not predict violence. J Forens Psychaitry Psychol. 2011;22(1):3–21. doi: 10.1080/14789949.2010.495990
  10. 10. Troquete NAC, van den Brink RHS, Beintema H, Mulder T, van Os WDP, Schoevers RA, et al. Risk assessment and shared care planning in out-patient forensic psychiatry: cluster randomised controlled trial. Br J Psychiatry. 2013;202(5):365–71. doi: 10.1192/bjp.bp.112.113043. pmid:23520222
  11. 11. Large MM, Ryan CJ. Violence risk assessment has not been shown to reduce violence. Aust N Z J Psychiatry. 2015;49(1):91. doi: 10.1177/0004867414539570
  12. 12. Van Dorn R, Volavka J, Johnson N. Mental disorder and violence: is there a relationship beyond substance use? Soc Psychiatry Psychiatr Epidemiol. 2012;47(3):487–503. doi: 10.1007/s00127-011-0356-x. pmid:21359532
  13. 13. Coid J, Ullrich S, Keers R, Bebbington P, Destavola B, Kallis C, et al. Gang membership, violence, and psychiatric morbidity. Am J Psychiatry. 2013;170(9):985–93. doi: 10.1176/appi.ajp.2013.12091188. pmid:23846827
  14. 14. Ullrich S, Keers R, Coid J. Delusions, anger, and serious violence: new findings from the MacArthur Violence Risk Assessment Study. Schizophr Bull. 2014;40(5):1174–81. doi: 10.1093/schbul/sbt126. pmid:24048345
  15. 15. Appelbaum PS, Robbins PC, Monahahn J. Violence and delusions: data from the MacArthur Violence Risk Assessment Study. Am J Psychiatry. 2000;157(4):566–72. pmid:10739415 doi: 10.1176/appi.ajp.157.4.566
  16. 16. Doyle M, Coid J, Archer-Power L, Hunder-Didrichsen A, Stevenson R, Wainwright V, et al. Discharges to prison from medium secure psychiatric units in England and Wales. Br J Psychiatry. 2014;205(3):177–82. doi: 10.1192/bjp.bp.113.136622. pmid:25012684
  17. 17. Gueorguieva R, Krystal J. More over ANOVA: progress in analysing repeated-measures data and its reflection in papers published in the Archives of General Psychiatry. Arch Gen Psychiatry. 2004;61(3):310–7. pmid:14993119 doi: 10.1001/archpsyc.61.3.310
  18. 18. Borum R, Lodewijks H, Bartel PA, Forther AE. Structured Assessment of Violence Risk in Youth (SAVRY). In: Otto RK, Douglas KS, editors. Handbook of violence risk assessment. New York and Abingdon: Routledge; 2010. p. 63–80.
  19. 19. Douglas K, Reeves K. Historical-Clinical-Risk Management-20 (HCR-20) Violence risk assessment scheme: rationale, application, and empirical overview. In: Otto RK, Douglas KS, editors. Handbook of violence risk assessment. New York and Abingdon: Routledge; 2010. p. 147–86.
  20. 20. Hart SD, Boer DP. Structured professional judgement guidelines for sexual violence risk assessment: The Sexual Violence Risk-20 (SVR-20) and Risk for Sexual Violence Protocol (RSVP). In: Otto RK, Douglas KS, editors. Handbook of violence risk assessment. New York and Abingdon: Routledge; 2010. p. 269–94.
  21. 21. Kropp PR, Gibas A. The Spousal Assault Risk Assessment guide (SARA). In: Otto RK, Douglas KS, editors. Handbook of violence risk assessment. New York and Abingdon: Routledge; 2010. p. 227–50.
  22. 22. O'Shea L, Dickens G. The HCR-20 as a measure of reliable and clinically significant change in violence risk among secure psychiatric inpatients. Compr Psychiatry. 2015;62:132–40. doi: 10.1016/j.comppsych.2015.07.009. pmid:26343477
  23. 23. Hill AB. The environment and disease: association or causation? Proc R Soc Med. 1965;58(5):295–300.
  24. 24. Bonta J, Law M, Hanson K. The prediction of criminal and violent recidivism among mentally disordered offenders: a meta-analysis. Psychol Bull. 1998;123(2):132–42. doi: 10.1037//0033-2909.123.2.123
  25. 25. Peterson JK, Skeem J, Kennealy P, Bray B, Zvonkovic A. How often and how consistently do symptoms directly precede criminal behavior among offenders with mental illness? Law Hum Behav. 2014;38(5):439–49. doi: 10.1037/lhb0000075. pmid:24730388
  26. 26. Prins SJ, Skeem JL, Mauro C, Link BG. Criminogenic factors, psychotic symptoms, and incident arrests among people with serious mental illnesses under intensive outpatient treatment. Law Hum Behav. 2014. doi: 10.1037/lhb0000104
  27. 27. Skeem JL, Winter E, Kennealy PJ, Louden JE, Tatar JR. Offenders with mental illness have criminogenic needs, too: toward recidivism reduction. Law Hum Behav. 2014;38(3):212–24. doi: 10.1037/lhb0000054. pmid:24377913
  28. 28. Coid J, Yang M, Roberts A, Ullrich S, Moran P, Bebbington P, et al. Violence and psychiatric morbidity in a national household population—a report from the British Household Survey. Am J Epidemiol. 2006;164(12):1199–208. pmid:17032695 doi: 10.1093/aje/kwj339
  29. 29. Elbogen E, Johnson S. The intricate link between violence and mental disorder: results from the National Epidemiologic Survey on Alcohol and Related Conditions. Arch Gen Psychiatry. 2009;66(2):152–61. doi: 10.1001/archgenpsychiatry.2008.537. pmid:19188537
  30. 30. Fazel S, Grann M, Carlstrom E, Lichtenstein P, Langstrom N. Risk factors for violent crime in schizophrenia: a national cohort study of 13,806 patients. J Clin Psychiatry. 2009;70(3):362–9. pmid:19284931 doi: 10.4088/jcp.08m04274
  31. 31. Whitwell JL. Longitudinal imaging: change and causality. Curr Opin Neurol. 2008;21(4):410–6. doi: 10.1097/WCO.0b013e32830719d4. pmid:18607200
  32. 32. Howard PD, Dixon L. Identifying Change in the Likelihood of Violent Recidivism: Causal Dynamic Risk Factors in the OASys Violence Predictor. Law Hum Behav. 2013;37(3):163–74. doi: 10.1037/lhb0000012. pmid:23148466
  33. 33. Quinsey VL, Jones GB, Book AS, Barr KN. The dynamic prediction of antisocial behavior among forensic psychiatric patients: A prospective field study. J Interpers Violence. 2006;21(12):1539–65. pmid:17065654 doi: 10.1177/0886260506294238
  34. 34. Brown S, St Amand M, Zamble E. The dynamic prediction of criminal recidivism: a three-wave prospective study. Law Hum Behav. 2009;33(1):25–45. doi: 10.1007/s10979-008-9139-7. pmid:18629620
  35. 35. Jones NJ, Brown SL, Zamble E. Predicting criminal recidivism in adult male offenders. Research versus parole officer assessment of dynamic risk. Crim Justice Behav. 2010;37(8):860–82. doi: 10.1177/0093854810368924
  36. 36. Schlager MD, Pacheco D. An examination of changes in LSI-R scores over time: making the case for needs-based case management. Crim Justice Behav. 2011;38(6):541–53. doi: 10.1177/0093854811402300
  37. 37. Webster CD, Haque Q, Hucker SJ. Violence risk-assessment and management. New Jersey: Wiley-Blackwell; 2013.