Abstract
The present study featured an investigation of the predictive properties of risk and change scores of two violence risk assessment and treatment planning tools—the Violence Risk Scale (VRS) and the Historical, Clinical, Risk–20, Version 2 (HCR-20)—in sample of 178 treated adult male violent offenders who attended a high-intensity violence reduction program. The cases were rated on the VRS and HCR-20 using archival information sources and followed up nearly 10 years postrelease. Associations of HCR-20 and VRS risk and change scores with postprogram institutional and community recidivism were examined. VRS and HCR-20 scores converged in conceptually meaningful ways, supporting the construct validity of the tools for violence risk. Receiver operating characteristic curve analyses demonstrated moderate- to high-predictive accuracy of VRS and HCR-20 scores for violent and general community recidivism, but weaker accuracy for postprogram institutional recidivism. Cox regression survival analyses demonstrated that positive pretreatment and posttreatment changes, as assessed via the HCR-20 and VRS, were each significantly associated with reductions in violent and general community recidivism, as well as serious institutional misconducts, after controlling for baseline pretreatment score. Implications for use of the HCR-20 and VRS for dynamic violence risk assessment and management are discussed.
Appraisals of violence risk provide important information regarding security level, release planning, and applications for preventative detention, but perhaps most important, in the prevention of future harm through informing risk management and intervention strategies (Douglas & Kropp, 2002). With a large volume of validated instruments available, evaluators are faced with the task of selecting which established measures to use. Advances in risk assessment over the past 20 years have involved incorporating dynamic risk variables, which have been defined as “changeable or potentially changeable factors (such as substance abuse, impulsivity, and criminal attitudes) that can be influenced or changed by psychological, social, or physiological means such as treatment interventions” (Wong & Gordon, 2006, p. 283). Some examples of prominent measures commonly used with violent offenders that use dynamic risk factors include, but are not limited to, the Violence Risk Scale (VRS; Wong & Gordon, 1999-2003), the Historical, Clinical, Risk management–20, Version 2 (HCR-20; Webster, Douglas, Eaves, & Hart, 1997) and 3 (Douglas, Hart, Webster, & Belfrage, 2011), the Level of Service Inventory–Revised (Andrews & Bonta, 1995), and the Level of Service/Case Management Inventory (Andrews, Bonta, & Wormith, 2004). Given the focus of the HCR-20 and VRS on dynamic violence risk assessment and management, the present study examined dynamic violence risk featuring these two tools on a sample of adult male-treated violent offenders.
The HCR-20 and VRS: A Brief Overview
Briefly, the VRS is designed to appraise violence risk, identify targets for violence reduction treatment, and to assess changes in risk from treatment or other change agents. As its items can be summed to yield a total score linked to risk categories (low, medium, high) and violent recidivism estimates, the VRS would be considered an empirical actuarial tool per Hanson and Morton-Bourgon (2009). In their meta-analysis of violence risk measures, using multilevel modeling procedures to control for sample, setting, and study characteristics, Yang, Wong, and Coid (2010) reported a mean area under the receiver operating curve (AUC) value of .65 for VRS total scores (k = 4, n = 1,148) in the prediction of violence. The HCR-20, in turn, is a 20-item structured professional judgment (SPJ) measure consisting of 10 historical (i.e., static/stable), 5 clinical (i.e., dynamic items focusing on recent functioning), and 5 risk (i.e., dynamic items focusing on future risk and circumstances) items. As an SPJ measure, in clinical use, one does not assign numeric ratings, sum the items, or link these to recidivism estimates; rather, there is a detailed multistep process of gathering risk information from multiple sources, obtaining item ratings, and evaluating the presence and relevance of risk information to inform an estimate of low, medium, or high as it pertains to case prioritization, imminence, and severity. HCR-20 Version 2 has been translated into more than 20 languages and is one of the most frequently used violence risk assessment tools in the world, while Version 3 (Douglas et al., 2011) has marshalled strong psychometric support and has recently supplanted Version 2. The Yang et al. (2010) meta-analysis also found the HCR-20 demonstrated strong predictive accuracy for violent recidivism (AUC = .71, k = 16, n = 4,161).
Considerations and Recent Findings in Dynamic Violence Risk Assessment With the HCR-20 and VRS
In principle, changes in dynamic risk factors should be linked to changes in recidivism, and while many risk assessment tools include dynamic risk factors, Douglas and Skeem (2005) note that much of the extant research has examined associations between violent recidivism and dynamic risk factors measured at a single time-point. To make claims of dynamism requires a minimum of two ratings at different time-points. Thus, unless associations are examined between changes on the dynamic variables with changes in subsequent recidivism, it remains unclear whether dynamic risk factors are actually dynamic (Douglas & Skeem, 2005).
Preliminary attempts to address this gap in the literature have examined dynamic variables at multiple time-points. For instance, Belfrage and Douglas (2002) found that the HCR-20 clinical and risk management scores rated on a sample of forensic psychiatric inpatients showed movement from pretreatment to posttreatment. The authors did not examine whether the change on the dynamic variables represented a change in recidivism rates as they did not have a follow-up period, the treatment program length was highly variable, and the treatment program did not have violence reduction as its specific focus. As such, it was difficult to ascertain whether the observed movement on the dynamic factors were risk relevant or whether they related to a different aspect of the treatment program. Belfrage, Fransson, and Strand (2004) subsequently examined movement on the HCR-20 domains in a small subsample of 13 Swedish forensic inpatients. Although, the scores on the item and subscale domains did not change significantly between assessments, the authors noted that rates of institutional violence had declined markedly over the study period. As such, they posited that increased knowledge of the patients’ risk factors by the staff coupled with increasing clinical attention toward risk management, could have contributed to the observed reductions in inpatient violence.
Similarly, Wilson, Desmarais, Nicholls, Hart, and Brink (2013) examined changes in dynamic risk for institutional violence using a sample of 30 forensic inpatients. Over the course of 1 year, the HCR-20 was rated four times (i.e., every 3 months) and reliably predicted inpatient violence at each follow-up. Dynamic scores varied slightly over the four assessments (i.e., by less than a point in total or Clinical + Risk [C + R] scores), although the authors did not examine whether changes on dynamic risk factors were associated with subsequent changes in aggression owing to the small sample size. Two subsequent studies have examined changes in the HCR-20 Version 3 and associations with outcome, each of which found the Clinical and Risk (C and R) domains to change significantly from pretreatment to posttreatment. In a Canadian sample of 99 forensic inpatients, Hogan and Olver (2016) found the positive changes measured on the HCR-20 C, R, and combined C+R domain to have significant concurrent associations with decreased inpatient aggression. Moreover, in an outcome study of 108 forensic patients in the Netherlands, De Vries Robbé, de Vogel, Douglas, and Nijman (2015) found changes on the R domain and total score to be significantly associated with reductions in long-term violence in the community.
Further lines of research have examined the association between change scores on the VRS and its sexual violence counterpart, the Violence Risk Scale–Sexual Offense version (VRS-SO; Wong, Olver, Nicholaichuk, & Gordon, 2003) with possible changes in general, violent, and sexual community recidivism after controlling for pretreatment risk (Beggs & Grace, 2011; Lewis, Olver, & Wong, 2013; Olver, Nicholaichuk, Kingston, & Wong, 2014; Olver, Wong, Nicholaichuk, & Gordon, 2007; Sowden & Olver, 2017). Lewis et al. (2013) examined the predictive properties of VRS risk and change scores in a sample of 150 adult male high-risk violent offenders who completed a high-intensity CBT-based violence reduction program. VRS change scores were negatively associated with violent recidivism at both total follow-up and 3-year fixed follow-up. Furthermore, offenders with high change scores (greater than 7 points of change) had lower rates of recidivism (23.1%) than offenders with low change scores (56.7%; less than 3 points of change). Olver, Lewis, and Wong (2013) also found that VRS dynamic change scores added incrementally to the prediction of community violence and any violence (i.e., institutional and community) after controlling for Psychopathy Checklist–Revised total score.
In an Australian sample of 82 violent offenders who had attended a prison-based violence reduction program, O’Brien and Daffern (2017) found VRS scores to change by approximately one standard deviation from pretreatment to posttreatment in a subsample of 61 men who had completed the program, although changes were not associated with reductions in recidivism; of note, treatment noncompleters, who could be well assumed not to have made risk change, were excluded from the outcome analyses. Finally, in a New Zealand sample of 121 treated high-risk prisoners who had attended a high-intensity violence reduction program, and a sample of 154 comparison controls, Polaschek, Yesberg, Bell, Casey, and Dickson (2016) found that prerelease VRS ratings of risk mediated the association between treatment completion and decreased violent reconviction; that is, the treatment group, by virtue of lowered VRS scores had decreased rates of violence postrelease. In a direct examination of VRS change and recidivism in the treated sample, a nonsignificant inverse association was reported between VRS change and future violence (eB = .888), controlling for pretreatment score (Yesberg, 2015).
Context of the Present Study
A growing body of research has featured the examination of dynamic violence risk through measuring risk and change across multiple time-points in response to possible change agents, using putatively dynamic tools such as the HCR-20 and VRS. Since Douglas and Skeem’s (2005) recommendations more than 10 years ago, there has been an increase in applied lines of research, but further cross-validation work on the change properties of the HCR-20 and VRS is needed. To our knowledge, changes in risk measured by the HCR-20 have yet to be examined in a prison-based correctional sample. As such, the present study is an examination of dynamic violence risk, specifically, through an exploration of the convergent and predictive properties of HCR-20 and VRS risk and change scores in a high-risk sample of treated adult male violent offenders.
Method
Participants
Participants were 178 adult federal offenders who attended a high-intensity violence reduction program, the Aggressive Behavioral Control (ABC) Program, at the Regional Psychiatric Centre (RPC), a multilevel security correctional mental health facility operated by the Correctional Service of Canada (CSC) in Saskatoon, Saskatchewan, Canada. Participant files were randomly drawn from all consecutive admissions (approximately 75% of cases) to the ABC Program from 1995 to 2004. The sample did not overlap with previous examinations of the VRS (e.g., Lewis et al., 2013) from the ABC program. On average, the men were 32 years of age (SD = 9.2) on admission to the ABC program and 33 years (SD = 9.0) at time of release to the community, spending an average of 6 months (SD = 1.9) in the ABC program. Most (57%) of the men were of Canadian indigenous ancestry, while the remainder were White (37%) or other (6%) ethnic decent. Most of the men in the sample were single (55%), 25% married/common-law, 20% divorced/separated, and 1% widowed. The mean education level was grade 9.5 (SD = 2.1) with an average reading level of Grade 10 (SD = 3.1). Roughly three quarters of the men were each diagnosed with a personality disorder (76%) in general, antisocial personality disorder (73%) in particular, or a substance use disorder (75%), while less than a third (30.3%) were diagnosed with a major nonsubstance-related mental illness (e.g., psychosis, depression). The men had serious criminal histories with the average age at first violent conviction being 18 years (SD = 4.6) and an average of 4.6 (SD = 4.1) prior violent convictions. The men also had serious institutional histories, with 88% having a history of institutional misconduct and 71%, serious misconduct (e.g., fighting and assaults). Most of the men (95%) had an index conviction for a violent offense (e.g., assault, homicide, robbery) and were serving a mean determinate sentence of 6 years.
Aggressive Behavior Control Program
The ABC Program was established in 1993 at the RPC as a 6- to 8-month, high-intensity, cognitive–behavioral therapy program with the goal of reducing violent reoffending in male offenders with extensive histories of violence and/or histories of serious institutional misconduct. The program was interdisciplinary in nature, and utilized psychoeducation, relapse prevention skills, as well as individual and group therapy. The program subscribed to the principles of risk (high-intensity services for high-risk offenders), need (target dynamic risk factors linked to criminal behavior for intervention), and responsivity (adapt service delivery to the cognitive capabilities, motivation, and personal background and characteristics of clientele; see Andrews, Bonta, & Hoge, 1990). The program was divided into three phases which were integrated with a modified application of the transtheoretical model of change (Prochaska, DiClemente, & Norcross, 1992).
Phase one moved offenders through the Precontem-plation and Contemplation stages where the focus was to increase insight into their violence, identify treatment targets, and increase the men’s motivation and treatment engagement. Precontemplation is characterized by a lack of awareness about the problem area and no intention to change, while Contemplation is characterized by recognition that the behavior is maladaptive; however, there has yet to be an attempt to use skills or strategies to change the behavior. In phase two, the goal of treatment was to teach skills that could be used when the men reached the Preparation and Action stages, for instance, learning to examine and challenge destructive behavior patterns linked to violent offending and developing skills in cognitive restructuring, emotional regulation, and behavior management. Preparation is characterized by awareness of problem areas with concordant use of cognitive and behavioral skills and strategies to attempt to manage the area; however, these changes are very recent and lapses tend to be frequent. Action is characterized by active engagement in alternate behavior over a sustained period of time relative to their lifetime functioning with lapses being very infrequent. The final phase entailed the development of relapse prevention skills that could be used when the men reached the Action and Maintenance Stages. Offenders began relapse prevention and release planning, as well as consolidating, reinforcing, and generalizing the skills and strategies learned in the program. In the Maintenance stage, the individual has sustained positive behavior changes over an extended period of time across a variety of contexts.
Outcome research has shown ABC program completion to be linked to reductions in violence risk (Wong, Gordon, & Gu, 2007), institutional misconduct (Wong et al., 2005), violent recidivism (Lewis et al., 2013), recidivism for gang affiliated offenders (Di Placedo, Simon, Witte, Gu, & Wong, 2006), and recidivism for high-psychopathy offenders (Wong, Gordon, Gu, Lewis & Olver, 2012).
Measures
Violence Risk Scale
The VRS consists of 6 static and 20 dynamic items, each of which is rated on a 4-point (0, 1, 2, 3) ordinal scale; higher item ratings on the static items indicate a more serious history of violence, while higher item ratings on the dynamic items indicate links to increased violence risk. Although the VRS can be rated at a single time-point, it is intended to be a repeated measures tool to be rated across multiple time-points (e.g., pretreatment and posttreatment). A modified application of the stages-of-change (SoC) model, as outlined in the ABC program description, is used to evaluate changes in risk on each of the 20 dynamic items. Each dynamic item with a 2 or 3 rating is considered criminogenic (i.e., linked to violence risk) and is given a baseline SoC rating at pretreatment; items with 0 or 1 ratings are not typically given SoC ratings. The SoC is then rerated at posttreatment for each 2- or 3-rated items. Progression, in the direction of improvement, from one stage to the next on an item is given a 0.5-point deduction, two-stages, a 1.0-point deduction, and so on; the one exception is progression from Precontemplation to Contemplation, which is given no point deductions since there are no risk relevant behavioral changes taking place. The change ratings are summed across all dynamic items to generate a change score which is subtracted from the pretreatment total to generate a posttreatment score. In the present study, very strong interrater reliability was obtained for the VRS on 20 randomly selected double coded cases with the following intraclass correlation coefficient (ICC) values (two way, mixed effects, single rater, consistency agreement): static ICCc,1 = .98, dynamic (pretreatment) ICCc,1 = .98, dynamic (posttreatment) ICCc,1 = .96, dynamic (change) ICCc,1 = .79, total (pretreatment) ICCc,1 = .98, total (posttreatment) ICCc,1 = .98.
Historical, Clinical, Risk, Management–20
The HCR-20 consists of 20 items organized into three domains that correspond to past (i.e., Historical such as previous violence, substance use problems, prior supervision failures), recent functioning (i.e., Clinical such as lack of insight, negative attitudes, impulsivity), and future (i.e., Risk such as plans lack feasibility, lack of personal support, exposure to destabilizers). Each item is rated on a three-level scheme of absent, possibly/partially present, and present, which are sometimes numerically coded on a 0-, 1-, 2-point rating scale. As an SPJ tool, the HCR-20 item ratings are not intended to be summed to generate numeric scores in clinical practice. However, numeric scores (in addition to SPJ summary risk ratings) are typically used to evaluate the tool’s psychometric properties in research (see Douglas et al., 2014). As the third version of the HCR-20 was not released until after the start of the current program of research, the second version of the tool was used. In the present study, acceptable interrater reliability was obtained for the HCR-20 on 20 randomly selected double coded cases with ICC values (two way, mixed effects, single rater, consistency agreement) computed for continuous scores and kappa for SPJ ratings: Historical ICCc,1 = .95, Clinical (pretreatment) ICCc,1 = .67, Clinical (posttreatment) ICCc,1 = .81, Clinical (change) ICCc,1 = .80, Risk Management (pretreatment) ICCc,1 = .71, Risk Management (posttreatment) ICCc,1 = .63, Risk Management (change) ICCc,1 = .61, total (pretreatment) ICCc,1 = .93, total (posttreatment) ICCc,1 = .94, SPJ (pretreatment) Κ = .71, SPJ (posttreatment) Κ = .52.
Outcome Measures
Community Recidivism
Community recidivism was defined as a conviction for a new criminal code violation on release to the community. We chose convictions given that it is a robust operationalization of this criterion commonly used in the risk assessment field (see Bonta, Rugge, & Dauvergne, 2003). There were few instances in this broadly high-risk sample in which men may have been arrested or charged for a violent offense but were not also convicted. Violent community recidivism included any criminal conviction against the person with the potential for physical or psychological harm (e.g., sexual assault, robbery, uttering threats, murder) per the definition in the HCR-20 manual. General community recidivism involved any conviction for a new offense, violent, or nonviolent. Recidivism was analyzed in a binary (yes–no, recidivist–nonrecidivist) manner, including the date of new recidivism events. The conviction date were used as the offense date for survival analysis given that this was the most consistent and reliable source of information available. If there was evidence of a lengthy period of remand prior to the conviction date, the time in remand was subtracted off the survival time.
Institutional Recidivism
Institutional recidivism was defined as a formal charge initiated by the institution for misconduct. Only institutional misconducts that occurred postprogram but prior to release were examined. Serious misconduct charges were explicitly labeled as such on the individual’s file by the charging officer and involved major rule violations and possible threats/harm to staff, inmates, or security (e.g., fighting, assaults). Any misconduct included any formal institutional charges (serious or minor) documented postprogram on file. Institutional recidivism was also analyzed in a binary (yes–no, recidivist–nonrecidivist) manner, including the date of new institutional offense.
Procedure
The University of Saskatchewan Behavioural Research Ethics Board (certificate No. BEH 12-68) provided ethical approval for the present study and agency approval was provided by CSC National Headquarters. This study was archival in nature. All study measures were coded from comprehensive institutional file information; no interviews were conducted. Community recidivism data were obtained from the Canadian Police Information Centre (CPIC), a nationwide criminal record database maintained by the Royal Canadian Mounted Police. Institutional recidivism data were collected from formal institutional misconduct records retrieved from CSC’s electronic filing system, the Offender Management System (OMS). To prevent criterion contamination, electronic copies of all relevant file documents for coding were retrieved from OMS by two research assistants who did not complete ratings of any study measures. Materials for coding the HCR-20 and VRS were saved in separate electronic pretreatment and posttreatment folders for each participant. Raters scored all measures on documents contained in the pretreatment folder prior to accessing those in the posttreatment folder to complete posttreatment ratings. This sequence ensured that raters were not exposed to posttreatment information while completing pretreatment ratings, and all risk and change ratings were completed on the study measures while blind to all postprogram criterion information. Consistent with the sequence observed in clinical practice, pretreatment ratings are made without knowledge of posttreatment information, and posttreatment ratings are completed (typically by the same evaluator) while cognizant of pretreatment risk and program performance, but without knowledge of postprogram institutional or community outcomes. Community recidivism data were extracted from CPIC, and institutional recidivism data from OMS, once all risk ratings had been completed and were coded by the first author.
The raters were trained on the risk measures in group format by the second author (a registered psychologist with clinical and research experience with the measures) using sample cases drawn from the VRS training materials developed by the instrument authors; the same training cases were used to train raters on the VRS and HCR-20. Each rater then coded the same five cases from the current pool and the ratings were reviewed with the raters for coding fidelity. Twenty randomly selected files were later double coded to establish interrater reliability after the raters had commenced autonomous coding, the results of which were reported in the instrument descriptions previously. Each rater was responsible for coding all measures on a given case, with the order of study measures counterbalanced.
Data Analytic Plan
The analyses proceeded in several phases. Given the volume of data and myriad analytic possibilities, we strove to prioritize what in our view were a combination of the most practical, parsimonious, and rigorous tests of the predictive properties of HCR-20 and VRS risk and change scores, balanced with the power constraints of a reasonable but modest sized sample. First, however, we computed a convergent validity correlation matrix to examine linear relations among the HCR-20 and VRS static, dynamic, total score, and change scale components. Although the HCR-20 is an SPJ tool and the VRS, an actuarial tool, both measures were designed to assess risk and to inform treatment planning and risk management efforts. Thus, we anticipated positive moderate to large correlations (i.e., .30-.50, per Cohen, 1992) between scores on structurally and conceptually similar scale components (e.g., VRS pretreatment dynamic and HCR-20 summation of C and R scales). Given that the C and R scale are dynamic, all analyses focused on their summation (i.e., C + R).
Second, we examined the predictive accuracy of VRS and HCR-20 risk scores for the four community (violent and general) and institutional (major and any) recidivism outcomes though receiver operating characteristic (ROC) curve analyses. ROC analyses generate an AUC statistic ranging from 0 to 1.0 representing the probability that a randomly selected recidivist has a higher score on a given measure than a randomly selected nonrecidivist. For instance, an AUC = .75 would be interpreted to mean that there is a 75% probability that a randomly selected recidivist would have a higher risk score on a given measure, than a randomly selected nonrecidivist. Values of .50 represent chance level predictive accuracy, values of .556 to <.639 correspond to small effects, .639 to <.714 moderate, and .714 and higher as large in magnitude predictive effect sizes (Rice & Harris, 2005).
Third, we examined the incremental predictive validity of static and dynamic risk scores on the two instruments for each of the four recidivism outcomes through Cox regression survival analyses. Cox regression is a survival analytic technique that adjusts and controls for individual differences in follow-up time and thus can be used to examine the prediction of recidivism over time by combinations of covariates. For community recidivism analyses, a survival time variable was created using the duration from the point of release to the date of first conviction for a given offense category (violent vs. any). For violent reconviction, time spent in custody for nonviolent offenses was subtracted off the follow-up time to obtain a more accurate projection of time spent in the community prior to recidivism. For nonrecidivists, the total follow-up time from the point of release to the CPIC data collection date was employed. If individuals died prior to the data collection date, the date of death was used as the end date. For institutional recidivism analyses, the time at program discharge up to the point of first institutional offense of a given category (serious vs. any) was employed; for nonrecidivists, the total time duration from the point of program discharge to the release date was employed.
The remaining analyses used survival analytic techniques to examine the association of risk changes, as measured by the VRS and HCR-20, to possible decreases in community and institutional recidivism. The change analyses begin through using Cox regression survival analysis, entering the pretreatment total risk score (i.e., static + dynamic) followed by the change score, representing risk change, and examining their unique associations with community or institutional recidivism over time. For analyses using SPJ summary risk ratings, indicator contrasts were performed comparing movement from one SPJ risk level with the next (with no change as the reference group), controlling for pretreatment SPJ rating. Given that higher risk offenders have more room for movement, and hence potential for change, yet are still more likely to reoffend than lower risk offenders, who remain low-risk irrespective of any changes that they make, it is essential to control for baseline risk at pretreatment. If risk and change are incrementally predictive, risk scores should show unique positive associations with outcome, while change should show unique inverse associations with outcome. Cox regression survival analysis generates a hazard ratio (eB) a measure of effect size that represents the proportionate increase in the hazard of recidivism per 1-unit change in the predictor variable; values above 1.0 indicate positive predictor criterion associations (e.g., risk score and recidivism), while values below 1.0 indicate increases in the predictor (e.g., treatment-related change) to be associated with decreases in the criterion. The final set of change analyses were Kaplan–Meier survival analyses to visually illustrate the risk-change recidivism associations demonstrated in Cox regression to aid interpretation. In short, risk and change scores for the VRS and HCR-20 were dichotomized using mean splits to create four risk-change groups (i.e., high-risk vs. low-risk × high change vs. low change); trajectories of community and institutional recidivism were then examined among the four risk-change groups.
Results
Risk Profiles and Sample Description
The bottom row of Table 1 reports basic descriptive statistics for the VRS and HCR-20 measures. At both pretreatment and posttreatment, the mean total score on the VRS fell in the high-risk category. A pretreatment VRS mean total score of 57.8 corresponds to the 84th percentile of the validation sample (Wong & Gordon, 2006), or approximately 1 standard deviation above the mean. The proportion of offenders in each VRS risk bin was near identical to those reported in for the New Zealand high-risk specialized treatment units program (Polaschek & Kilgour, 2013). The SPJ ratings of the HCR-20 similarly placed most of the offenders in the high-risk category at both pretreatment (n = 137, 77%) and posttreatment (n = 117, 64%), while much smaller proportions were observed at medium (n = 36, 20.2% and n = 37, 32%, respectively) or low (n = 5, 2.8% and n = 7, 3.9%, respectively) risk. The mechanical tally of the HCR-20 total at pretreatment (M = 28.3, SD = 5.5) was more than a full standard deviation above that reported in Douglas, Yeomans, and Boer’s (2005) federal correctional sample of violent offenders (M = 20.1, SD = 7.9), and was at approximately the 90th percentile of that sample. VRS change scores were moderate in magnitude, while HCR-20 clinical, risk management, dynamic (C + R management), and total change scores were moderate to large in magnitude.
Convergent Validity Correlation Matrix of VRS and HCR-20 Risk and Change Scores. a
Note: N = 178. VRS = Violence Risk Scale; HCR-20 = Historical, Clinical, Risk–20 Version 2; pre = pretreatment; post = posttreatment; H = Historical; C + R = summation of Clinical and Risk scales; SPJ = structured professional judgment; NS = not significant. Values in parentheses on principal diagonal are Cronbach’s alpha coefficients except (—) = cannot be computed.
All p < .001 except for *p < .05 and **p < .01.
Convergent Validity of the VRS and HCR-20
Table 1 also reports the convergent validity correlations for scores on the two risk measures. Cronbach’s alphas are presented on the principal diagonal. The VRS and HCR-20 showed strong convergence, with large positive correlations observed, particularly for respective pretreatment and posttreatment score on the respective static, dynamic, and total scores of the instruments (rs = .61-.80, p < .001) indicating the measurement of a common underlying construct of violence risk. The VRS and HCR-20 change ratings were also significantly positively correlated (r = .63, p < .001), indicating, in turn, the measurement of pre–post risk change. HCR-20 SPJ risk categories showed good convergence with continuous VRS scores, as well as the VRS risk bands for pretreatment (Cramer’s V = .54, p < .001) and posttreatment (Cramer’s V = .67, p < .001) ratings.
Prediction of Community and Institutional Recidivism
For community recidivism analyses, the sample size was 155 participants as 23 individuals were either never released (n = 16), died before release (n = 5), or were deported (n = 2). For institutional recidivism analyses, all 178 participants were included; however, many participants had short institutional follow-up periods before their release, and any cases are removed from Cox regression models that were either released or removed for one of the other aforementioned reasons prior to the first institutional recidivism event. The mean community recidivism follow-up length was 9.7 years (SD = 2.6), with a range of 0.1 to 13.8 years. In this sample, 60% (n = 93) had at least one new violent conviction and 78.7% (n = 122) had at least one new conviction (i.e., any reconviction). The mean institutional recidivism follow-up length was 29.7 months (SD = 40.3) with a range of 0 to 163.7 months; the mean institutional follow-up length for the released cases was shorter at 18.0 months (SD = 18.3). The difference between maximum institutional and community follow-up lengths relates to different offenders having different release and discharge dates. In this sample, 30.9% (n = 55) had at least one posttreatment serious misconduct and 55.6% (n = 99) had at least one new misconduct in general.
Table 2 reports findings from ROC analyses for the VRS and HCR-20 in the prediction of community and institutional recidivism. Both tools, static, dynamic (pretreatment and posttreatment) and total scores (pretreatment and posttreatment) predicted both sets of community outcomes, with broadly moderate range accuracy for community violent recidivism and high accuracy for general (i.e., any) new criminal conviction on release. An exception was that the HCR-20 dynamic (C + R) ratings showed considerably lower predictive accuracy for violence at pretreatment but markedly improved accuracy at posttreatment after the items had been rerated to take into account treatment change. In general, posttreatment ratings for dynamic and total scores for both instruments had higher AUC magnitudes at posttreatment, after adjusting the risk ratings to take into account treatment changes. HCR-20 SPJ ratings significantly predicted both community outcomes, although three out of the four effects (pretreatment and posttreatment) were lower for the SPJ ratings than the numeric summation of scores. The instruments fared less well in the prediction of institutional recidivism; AUCs were all small in magnitude and not much better than chance in the prediction of any new institutional misconduct. Only the VRS static items and posttreatment ratings on the VRS and HCR-20 dynamic and total scores attained significance in the prediction of serious institutional misconducts.
ROC Analyses: Predictive Accuracy of VRS and HCR-20 Risk Scores for Community and Institutional Recidivism Outcomes.
Note: VRS = Violence Risk Scale; HCR-20 = Historical, Clinical, Risk-20 Version 2; AUC = area under the receiver operating curve; CI = confidence interval; pre = pretreatment; post = posttreatment. n = 155 for community recidivism analyses, n = 178 for institutional recidivism analyses.
p < .06. *p < .05. **p < .01. ***p < .001.
Table 3 reports incremental validity analyses examining the relative contributions of static and dynamic scores on the VRS and HCR-20 in the prediction of the four recidivism outcomes through Cox regression survival analysis, which controls for individual differences in follow-up time. A given regression model represents the unique contributions of each predictor variable in the prediction of recidivism over time, postrelease within the community (for community recidivism) or posttreatment within the institution (for institutional outcomes). In the prediction of community violent recidivism, VRS pretreatment and posttreatment dynamic scores (Models 1 and 2) were incrementally predictive, but the static scores were not; for the prediction of general recidivism, both sets of scale components were uniquely predictive of outcome. The HCR-20 Historical items uniquely predicted violent recidivism over time at pretreatment, while the dynamic (C and R) did not (Model 3); however, HCR-20 posttreatment dynamic ratings uniquely predicted community violent recidivism as did the Historical items (Model 4). In addition, both Historical and dynamic HCR-20 scores (pretreatment and posttreatment) were incrementally predictive of general recidivism.
Cox Regression Survival Analysis: Incremental Predictive Validity of VRS and HCR-20 Static and Dynamic Items for Community and Institutional Recidivism.
Note. VRS = Violence Risk Scale; HCR-20 = Historical, Clinical, Risk–20 Version 2; SE = standard error; CI = confidence interval; pre = pretreatment; post = posttreatment. Significant p values in bold font. n = 155 for community recidivism analyses, n = 161 for serious institutional recidivism, and n = 170 for any institutional recidivism.
p < .05. **p < .01. ***p < .001.
For the prediction of institutional recidivism, the VRS static items were consistently uniquely predictive of new serious as well as any new institutional misconduct in general across all models. Although VRS pretreatment dynamic scores did not uniquely predict either institutional outcome (Model 5), the more proximal- and treatment-adjusted posttreatment dynamic ratings significantly incrementally predicted both outcomes (Model 6). Finally, the HCR-20 H items uniquely predicted both sets of institutional misconducts (as with the VRS static factors), but only the posttreatment rated C and R items attained significance in the prediction of serious institutional misconduct (Model 8).
Dynamic Violence Risk: Association of Changes in Risk to Changes in Recidivism
Cox Regression Survival Analysis
Significant differences were observed from pretreatment to posttreatment on the dynamic components of both risk measures, approaching three quarters of a standard deviation for the VRS (d = .70, p < .001) and a full standard deviation for the HCR-20 (d = .92, p < .001). The next set of analyses examined the extent to which changes in risk assessed from pretreatment to posttreatment on the HCR-20 and VRS were associated with reductions in recidivism after accounting for baseline risk level. The results of Cox regression survival analyses examining these associations are reported in Table 4.
Cox Regression Survival Analysis of Dynamic Violence Risk: Associations Between VRS and HCR-20 Change Scores With Community and Institutional Recidivism.
Note. VRS = Violence Risk Scale; HCR-20 = Historical, Clinical, Risk–20 Version 2; SE = standard error; CI = confidence interval; pre = pretreatment; post = posttreatment; C + R = summation of HCR-20 Clinical and Risk Scales; SPJ = structured professional judgment. Significant p values in bold font. n = 155 for community recidivism analyses, n = 161 for serious institutional recidivism, n = 170 for any institutional recidivism for continuous change score analyses, and n = 1 for SPJ change analyses. SPJ analyses conducted using indicator contrasts with no change in SPJ summary risk rating as the reference group; VRS and HCR-20 change scores otherwise treated as continuous covariates.
p < .05. **p < .01. ***p < .001.
First, VRS change scores were significantly associated with reductions in community violent (eB = .93) and general recidivism (eB = .89), controlling for pretreatment total score (Model 1). That is, each 1-point in change on the VRS dynamic items, representing risk reduction, is associated with 7% and 11% decreases in the hazard of violent and general recidivism, respectively, controlling for pretreatment risk level. VRS pretreatment total scores remained uniquely predictive of all recidivism outcomes, meaning that higher risk men at pretreatment still posed a greater probability to reoffend in the community than lower risk men, even though their risk had been reduced. Similarly, HCR-20 numeric change scores were significantly associated with reductions in community violent (eB = .89) and general recidivism (eB = .90) after controlling for pretreatment total score (Model 2); that is, each 1-point increase in HCR-20 change score was associated with an 11% and 10% decrease in the hazard of a conviction for a violent or any new offense in general, respectively. Positive changes in HCR-20 SPJ risk category (i.e., changing one risk level vs. remaining within the same risk band) was significantly associated with decreased general recidivism but not violent recidivism, as demonstrated through indicator contrasts using no change as the reference group. 1 Finally, in terms of institutional recidivism analyses, risk changes on the VRS and HCR-20 scores, and HCR-20 summary risk ratings, were uniquely associated with a reduction in new serious institutional misconducts arising posttreatment (Models 4, 5, and 6, respectively). None of the change prediction models for the reduction of any new institutional misconduct attained significance.
Kaplan–Meier Survival Analysis
We extended the Cox regression analyses through conducting a series of Kaplan–Meier survival analyses on 2 (high- vs. low-risk) × 2 (high vs. low change) groups examining group differences in trajectories of the three recidivism outcomes, where incremental change effects were observed. Individuals scoring at or above the mean on a given measure were placed in the “high group,” while those scoring below the mean were placed in the “low group” Four risk-change groups were each created as follows (respective ns for the VRS and HCR-20 in parentheses): high-risk high change (ns = 65 and 49), high-risk low change (ns = 42 and 50), low-risk high change (ns = 26 and 19), and low-risk low change (ns = 22 and 37). The high-risk low change group was the focus of particular scrutiny; specifically, if changes among high-risk offenders are most predictive of decreased recidivism (per the risk and need principles), then there should be substantive differences between high-risk men who evidence meaningful change and high-risk men who do not. By contrast, there would be little reason for there to be much in the way of differences between low-risk groups, irrespective of any changes they may make. The results are presented graphically in Figure 1.
Kaplan–Meier survival analysis: Trajectories of recidivism as a function of risk and change group for the VRS (A, violent; B, general; C, major institutional) and HCR-20 (D, violent; E, general; F, major institutional).
With respect to community violent recidivism, for the VRS (Figure 1A), the high-risk low change group had significantly higher and faster rates of violent reconviction than the low-risk low change and low-risk high change groups (both p ⩽ .001), but not significantly different from the high-risk high change group, χ2(n = 107) = 2.62, p = .106. For the HCR-20 (Figure 1D), the high-risk low change group had significantly higher and faster rates of violent reconviction than each of the three remaining groups, including the high-risk high change group χ2(n = 99) = 7.87, p = .005. Interestingly, for the HCR-20 violent recidivism analyses, there were no significant difference in rates of violent failure between the high-risk high change group and the low-risk high change groups χ2(n = 69) = 0.37, p = .545.
In terms of community general recidivism, for the VRS (Figure 1B), the high-risk low change group had significantly higher and faster rates of any new reconviction than each of the three remaining groups, including the high-risk high change group, χ2(n = 107) = 6.27, p = .012. The high-risk high change group also had significantly steeper recidivism trajectories than both low-risk groups (p < .001). For the HCR-20 (Figure 1E), the high-risk low change group had significantly faster and higher rates of general recidivism than the two low-risk groups (p = .05 and < .001 for high and low change, respectively), but was only different from the high-risk high change group at χ2(n = 99) = 2.99, p = .084.
Finally, in terms of risk-change associations with serious (major) institutional misconduct, for the VRS (Figure 1C), the high-risk low change group had higher and faster rates of serious misconduct posttreatment than all three groups (p < .001 for both low-risk groups), including the high-risk high change group, χ2(n = 107) = 7.21, p = .007. The high-risk high change group had significantly higher rates of serious misconduct than the low-risk high change (p = .026), but not the low-risk low change group (p < .068). For the HCR-20 (Figure 1F), the same pattern of associations was observed, with the high-risk low change group having higher rates of serious institutional misconduct than both low change groups (p ⩽ .001) and the high-risk high change group, χ2(n = 99) = 4.98, p = .026. The high-risk high change group had higher rates of serious institutional misconduct than both low-risk groups (p = .041 and .054, for high and low change, respectively). There were no significant differences between the two low-risk groups, irrespective of change, across either instrument for any of the community or institutional recidivism outcomes.
Discussion
The present study featured an examination of dynamic violence risk, via the HCR-20 and VRS, in a treated sample of high-risk violent federal offenders who had attended a violence reduction program and followed up in the community nearly 10 years postrelease. The cohort had many similarities to high-risk samples used to examine the two measures and their change correlates. The average VRS score at pretreatment was a full standard deviation above the mean compared with the normative sample (Wong & Gordon, 2006), but very much consistent with independent samples of treated high-risk violent offenders in New Zealand (Polaschek & Kilgour, 2013), Canada (Lewis et al., 2013), and the United Kingdom (Sheldon & Krishnan, 2009). Similarly, more than three quarters of the sample was classified as high-risk on the HCR-20 at pretreatment, and approximately two thirds as such at posttreatment, with the sample scoring at approximately the 90th percentile on the HCR-20 mechanically derived total score relative to a Canadian federal sample of violent offenders (Douglas et al., 2005).
Convergent and Predictive Validity: Replication of Previous Findings
As reported elsewhere (Dolan & Fullam, 2007), scale components of the VRS and HCR-20 showed good convergence in the measurement of (dynamic) violence risk, including C + R summation (pretreatment and posttreatment) and VRS dynamic scores, change scores across both tools, and HCR-20 SPJ ratings with dimensional VRS scores. Moreover, consistent with extant meta-analytic findings (Yang et al., 2010), both tools significantly predicted violent and general community recidivism with moderate to high-predictive accuracy in this correctional sample, although predictions of institutional misconduct (particularly general misconduct) were lower. The weaker predictive efficacy for institutional misconducts, especially any new disciplinary infraction, we anticipate is likely attributable to the much lower bar to get in trouble for a variety of indiscretions (e.g., possession of an unauthorized item), as opposed to more serious infractions (e.g., fights/assaults/threatens) versus formal criminal convictions for new offenses occurring in the community. Finally, dynamic scores for the HCR-20 and VRS, particularly those assessed at posttreatment, were incrementally predictive of community violent and general recidivism as well as serious institutional misconducts. This makes sense conceptually, given that: (a) this represents the most recent assessment and that which has closest temporal proximity to the outcome; (b) there may be more comprehensive information available at posttreatment, thereby yielding higher quality assessments; and (c) given that this is a treatment sample that presumably changed their risk from pretreatment to posttreatment after participating in a high-intensity violence reduction program.
Cross-Validating Dynamic Predictive Properties: Emergent New Findings From a Correctional Sample
Perhaps the most novel contribution of the present study was the examination of dynamic violence risk, assessed at pretreatment and posttreatment, and linking these changes to important correctional outcomes, extending previous works from Sweden, Canada, Netherlands, New Zealand, and Australia. The sample demonstrated changes overall from their participation in the ABC Program. On each of the VRS and HCR-20, dynamic scores changed by close to a full standard deviation from pretreatment to posttreatment, representing reductions in violence risk. The results of Cox regression survival analyses, examining community and institutional recidivism over time, demonstrated VRS and HCR-20 risk and change scores to be uniquely significantly associated with most of these outcomes. For the VRS, the observed hazard ratios indicated a 7% decrease in odds of violent reconviction, 11% decrease in the odds of any reconviction, and 13% decrease in odds of a new serious institutional misconduct, for every 1-point increase in dynamic change score, controlling for baseline risk (pretreatment score). Similarly, for the HCR-20 (numeric ratings), the respective hazard ratios represented decreases of 11% (violent recidivism), 10% (general recidivism), and 14% (serious institutional misconduct) per 1-unit improvement in dynamic change score (C and R), controlling for baseline risk. That the pretreatment risk score remained significant across all analyses demonstrates that risk level at the beginning of treatment retains important relevance (i.e., high-scoring men on the VRS and HCR-20 are still higher risk than low scoring men), but that dynamic changes in risk have prognostic relevance, and ultimately strengthen appraisals of future risk.
Implications for the Use of SPJ, Mechanical, and Actuarial Measures in Violence Risk Assessment
Hanson and Morton-Bourgon (2009) note that when ratings are summed to yield numeric scores (e.g., adding the items of SPJ tools), but they are not linked to recidivism estimates per se, this characterizes what they termed mechanical tools. In the present sample, numeric ratings of the HCR-20, that is, when it was used as a mechanical tool, generated larger effect sizes for all prediction analyses than SPJ ratings. Although the differences in prediction effect sizes for SPJ versus mechanical ratings of the HCR-20 were not particularly large, they were consistently higher and much more frequently statistically significant for the mechanical ratings, which cannot be ignored. Part of this is likely attributable to greater variance of risk and change scores when using the longer mechanical scale, as opposed to 3-point SPJ ratings. However, we also cannot rule out the possibility that numerically based ratings were more sensitive to detecting the nuances of treatment-based changes across individual items, and may have performed slightly better in the present correctional sample.
Our suggestion is to use tools clinically as they are intended to be used. Although the HCR-20 performs well as a mechanical tool across repeated validation studies (see Douglas et al., 2014), there are no user guidelines within the manual for clinical applications of numeric scores and the instrument developers are quite explicit not to use the tool in a mechanical (adding the items) or actuarial (linking score summations to recidivism estimates) manner (Douglas et al., 2011). In the present study, SPJ ratings, particularly those adjusted at posttreatment, still fared well in the prediction of community recidivism, and a decrease in one full SPJ risk category on the HCR-20 was associated with more than a 40% decrease in the hazard of violent or general community recidivism after controlling for pretreatment SPJ risk level.
The VRS is an empirical actuarial risk assessment tool, with items summed to yield risk scores, that are arranged into risk bands linked to violent and general recidivism estimates (see Wong & Gordon, 2006). Future work on the VRS will involve integrating risk and change information in a more systematic manner across multiple samples, such as through logistic regression modeling and using this to expand and refine the current risk categories and recidivism estimates to improve applications and clarity of risk communication. For instance, Olver et al. (2018) employed logistic regression modeling on a large multisite sample of treated sexual offenders, using VRS-SO risk and change score information to generate sexual recidivism estimates over fixed 5- and 10-year follow-ups. This permitted the use of odds ratio information from multiple predictors to adjust risk estimates as a function of specific risk scores and treatment change.
Importantly, the same line of research also found that the same posttreatment score could generate different recidivism estimates depending on how much it had changed from baseline, particularly for high-risk cases. As an illustration, a posttreatment score of 50 will be associated with different observed or estimated rates of recidivism, depending on whether the individual changed 0 points from baseline (pretreatment score 50), 5 points (pretreatment score 55), or 10 points (pretreatment score 60), with larger changes from baseline being associated with successively lower rates of recidivism. Given the dynamic nature of risk, we believe it is prudent to reassess risk across multiple time-points and to track change cumulatively, especially if there has been a deliberate change agent such as participation in one’s correctional plan or other risk reduction regime.
Conclusions, Limitations, and Future Directions
Ultimately, Mills (2017) reminds us that a risk instrument is not a risk assessment. While a specific tool may employ the SPJ, mechanical, or actuarial tradition, risk assessments involve gathering information from multiple sources, assessing multiple domains of functioning, using multiple assessment procedures, including the possibility of multiple risk tools (Boer, Hart, Kropp, & Webster, 1997); the accumulation of data are then integrated into a final risk appraisal and case formulation that is used to inform risk management recommendations. Invariably, there is a great deal of judgment involved in the process of integrating the data. Although a tool generates a score or is linked to a group recidivism estimate, this does not mean the evaluator is slavishly or rigidly bound by this estimate; it is ultimately one more piece of data that goes into the final formulation and associated recommendations.
There are some potential limitations of the present investigation which merit discussion. First, this sample included male violent offenders receiving services in a correctional treatment facility; as such, the current findings may not generalize to other offender populations, such as female offenders or individuals within forensic or civil psychiatric inpatient or outpatient settings. A second potential limitation is that for the majority of the analyses, simple binary recidivism variables were used to explore treatment-related changes in recidivism rates; using alternative recidivism outcomes such as crime severity estimates or aggregate sentence length, that is, indexes of harm reduction, could provide invaluable information on treatment-related change. A third, related limitation is the use of official criminal records (CPIC) to provide an index of recidivism; although the existence of a federal reporting system of criminal offenses has merits, inevitably it leaves open the possibility of undetected violent recidivists.
A fourth potential limitation is that although the change score results for the VRS and HCR-20 were encouraging, it should be noted that in the absence of a true control group, statements regarding the causal connection between treatment and reduced recidivism must remain tentative. It is possible that other causal agents could be responsible for the association (e.g., other treatments, aging, etc.). With the comparatively lower interrater reliability observed for HCR-20 ratings, a fifth possible limitation is that this may result in a conservative estimate of its psychometric properties in the current sample, bearing in mind that sufficient true score variance (as opposed to measurement error) must be present to obtain the type of significant predictor–criterion associations observed in the present study. Finally, we conducted a large volume of analyses to maximize transparency, aid interpretability, and minimize the perception of selectively reporting findings. Given the relatively modest sized sample, this does increase the potential for Type I error although this is offset by the a priori focus on a collection of core findings, specifically, the association between risk change and recidivism.
In conclusion, the present study examined the associations of risk and change to a series of recidivism outcomes. The findings provide support for the dynamism of violence risk in a high-risk sample, which would allow for the examination of change in a population where potentially the most treatment-related gains could be made. Given that risk change scores appear to uniquely add to the prediction of recidivism, as noted above, future research will need to examine how best to incorporate change-related information into risk assessments to better inform treatment staff, parole officers, judges, and other decision makers in managing risk and preventing recidivism. Simply relying on posttreatment ratings may not sufficiently capture change information and a more systematic approach to the integration of this information may be needed. The present findings also underscore the importance of assessing risk at multiple time-points, particularly in the context of correctional treatment or other change agents; conclusions from any one risk assessment should be considered to have an expiration date and risk should be reassessed after major life changes or crime-specific treatment programs. In this manner, applied developments in risk assessment science and practice have promise to aid in the rehabilitation and reintegration of violent offenders to reduce violent victimization and promote public safety.