Using a reassessment framework to determine critical case management needs: DRAOR improves on LS/RNR’s predictive discrimination of short-term recidivism

ABSTRACT

Two core tasks within community corrections are estimating recidivism risk and providing supportive case management to supervised clients. Recent studies demonstrated that regularly reassessing dynamic risk factors enhances the predictive accuracy of case management tools that identify immediate client needs. Using a large sample (N = 2076) of adults on parole in an Australian jurisdiction, we evaluated whether reassessed versus initial scores from Level of Service/Risk, Need, Responsivity (LS/RNR) and Dynamic Risk Assessment for Offender Re-entry (DRAOR) better discriminated recidivists from non-recidivists. Using Cox regression survival analyses, we modelled risk for general and violent recidivism. Updated DRAOR scores incrementally predicted recidivism beyond initial scores, particularly showing that DRAOR Acute scores can identify needs related to immediate recidivism risk. Models combining LS/RNR criminal history scores with updated DRAOR Acute scores demonstrated the greatest predictive validity, suggesting case managers must consider acute dynamic risk factors within the context of long-term risk. We suggest that attending to criminal history, criminogenic needs and acute dynamic risk factors are each necessary for effective case management, with specific attention toward regularly reassessed acute dynamic risk factors among people identified with higher long-term risk and needs.

KEYWORDS:

• DRAOR
• reassessment
• LS/RNR
• dynamic risk
• recidivism

People re-entering the community following incarceration often encounter challenges and stressors (Zamble & Quinsey, 1997). Internationally, many jurisdictions support the re-entry process through supervised parole or other community-based orders (e.g. Corrections Act, 1986 [Victoria, Australia]; Corrections and Conditional Release Act, 1992 [Canada]; Parole Act, 2002 [New Zealand]). The role of community corrections involves two related goals, specifically (1) community safety through preventing or reducing recidivism and (2) assisting re-entry needs. In other words, establishing recidivism risk and engaging in (timely) supportive management of people under correctional supervision are essential components of community corrections supervision. These are also foundational goals of risk assessment (Andrews & Bonta, 2010; Heilbrun, 1997).

Although predicting criminal behaviour and identifying needs of clients under correctional supervision are inter-related aims, using a single method for both tasks is not often practically achievable. Long-term static risk factors are easily scored and accurate predictors of recidivism (see Andrews et al., 2006; Bakker et al., 1999), but these factors are not amenable to change through intervention or case management. Although risk prediction without identifying and addressing needs may be informative for system-wide decision-making such as setting levels of supervision contact, it offers no useful solutions for guiding intervention when working directly with individuals. The goal for corrections agencies should be to facilitate desistance from crime through accurate identification of individual needs, building understanding of clients’ life circumstances, and providing supportive and targeted rehabilitation opportunities.

Risk assessment or case management tools that merge goals by assessing needs to set a risk level are commonly used in community corrections (for examples, see Desmarais & Singh, 2013; Olver et al., 2014b), but need-based measures may not always match or improve upon predictive validity demonstrated by static risk measures (see Caudy et al., 2013; Olver et al., 2014b). Further, with the exception of a few tools (see STABLE-2007 and ACUTE-2007; Hanson et al., 2007; and Dynamic Risk Assessment for Offender Re-entry (DRAOR), Serin, 2007), the timeframe between assessments may be meaningfully longer than the change in needs (i.e. a client’s circumstances and risk factors may meaningfully fluctuate multiple times between formal assessments). As such, some need-based dynamic risk tools may not efficiently predict recidivism nor best indicate imminent needs. Researchers must compare tools and tool components to observe if there are potential inefficiencies for recidivism prediction or meaningful differences in change, by examining whether there are potential increases in prediction accuracy associated with reassessment.

In the present study, we compared models that variously used assessment information identifying static risk based on criminal history, dynamic risk factors reassessed every 6 or 12 months, dynamic risk and strength factors reassessed more frequently (every 3 months), and acute risk and destabilising factors reassessed approximately weekly. The aim in model building was to consider the specific purpose of each component and balance the simultaneous goals of strong predictive discrimination, comprehensive assessment and up-to-date information for case management. Using a large parole dataset from Australia, we analysed risk scores from the Level of Service/Risk, Need, Responsivity (LS/RNR; Andrews et al., 2008) and DRAOR (Serin, 2007). All scores were rated by supervision officers within routine practice, and we drew recidivism events from a police database of criminal charges.

Static and dynamic risk factors

Conceptually, risk factors for recidivism are often separated into static and dynamic risk factors (Andrews & Bonta, 2010). Static factors are unchangeable demographic (e.g. age) and criminal history factors. These factors predict general recidivism well and, as they are often readily available to corrections agencies, have historically been used in correctional risk assessment tools to maximise prediction of future criminal activity (Andrews et al., 2006). Although the strong predictive validity of these factors is important, static factors do not represent theoretical drivers of behaviour and cannot be targeted in intervention, providing little utility for management.

Dynamic risk factors change, and a defining feature is that change is related to a change in recidivism risk. Dynamic factors that can be ethically and effectively targeted through intervention are named criminogenic needs (Andrews & Bonta, 2010). Commonly, risk assessments are designed to measure the Central Eight static risk (criminal history) and criminogenic need areas (antisocial personality pattern, antisocial cognition, antisocial associates, family, school/work, recreation, substance use; Andrews & Bonta, 2010).

Reassessment

When increases in the presence or observed influence or intensity of dynamic factors relate to a higher likelihood of outcome, these changes should guide case managers’ intervention. Because dynamic factors change, evaluating dynamic risk factors requires a research design that incorporates repeated assessment. Recent studies have demonstrated the importance of reassessment by revealing that the most recent assessment of dynamic risk factors is a more accurate predictor of recidivism than the initial assessment (Babchishin & Hanson, 2020; Coupland & Olver, 2020; Hanson et al., 2021; Howard & Dixon, 2013; Lloyd et al., 2020a; Olver et al., 2014a).

Among different analytical techniques used to evaluate prediction using dynamic risk tools, Cox regression survival analyses (Cox, 1972; see Singer & Willett, 2003) offers several advantages. Survival analysis is appropriate for corrections research because it allows participants to record varying follow-up times and varying times to recidivism while accounting for participants that do not experience recidivism. The model also offers flexibility. Some researchers enter two scores (e.g. pre- and post-programme scores, or equivalent pre-programme and change scores) as predictors recorded prior to the start of the follow-up timeframe (e.g. Coupland & Olver, 2020; Olver et al., 2014a), whereas others include initial and updated assessments throughout the follow-up timeframe during which community-based risk changes alongside assessed scores (see Babchishin & Hanson, 2020; Davies et al., 2021; Hanson et al., 2021; Howard & Dixon, 2013; Lloyd et al., 2020a). Models with time-varying predictors also have the advantage of accounting for irregular reassessments. Regardless, studies with a sufficient sample size have consistently shown greater discrimination for updated scores after accounting for prior scores. For example, Hanson et al. (2021) demonstrated incremental predictive validity of updated scores using the Sexual Offender Treatment Intervention and Progress Scale (McGrath et al., 2012) for any recidivism (n_recidivists= 159), but the number of recidivists was too small to analyse violent (n_recidivists= 30) or sexual recidivism (n_recidivists= 13). Several studies also showed the predictive validity of reassessed dynamic factors incremental to static factors (see Coupland & Olver, 2020; Davies et al., 2021; Lloyd et al., 2020a).

Assessment of risk and needs in community corrections

Correctional agencies must choose from a vast number of potential risk tools available for implementation (see Bourgon et al., 2018; Desmarais & Singh, 2013; Risk Management Authority, 2021). Further, when an agency uses a suite of tools each with different purposes, combining information across tools presents an additional challenge (Coulter et al., 2022). Most tools rely on the same finite pool of risk factors (Kroner et al., 2005), predict recidivism to a similar degree (Yang et al., 2010), and may often contain overlapping rather than additive information, even when the item definitions, assessment timeframes and risk estimates have different intended purposes. Using a combination of tools to best maximise both prediction of recidivism and timely management of criminogenic needs while reducing redundancy is important to ensure that practitioners can confidently identify (1) who is at risk, (2) what core personal, interpersonal and community features are driving identified risk and (3) which client needs require immediate attention. Although a single risk tool may address each of these concerns, it is unlikely it would do so through equivalent and equally weighted content and assessment schedules.

In community corrections in Victoria, Australia, supervision officers use both the LS/RNR and DRAOR for different purposes. Both assess generally similar risk domains but using different definitions, timeframes and context (see Andrews et al., 2008; Hanby, 2013). Specifically, the LS/RNR aims to provide a comprehensive overview of rehabilitation needs, estimate long-term risk of recidivism, and set levels of service. By contrast, DRAOR aims to estimate short-term, imminent likelihood of recidivism and guide case managers to focus on the most currently critical areas of an individual’s re-entry process.

LS/RNR

The LS/RNR is an actuarial risk assessment tool that comprehensively assesses the Central Eight risk and needs factors (Andrews & Bonta, 2010). It is designed to assist professionals (e.g. case managers/clinicians) to engage in management and rehabilitation planning for their adult and older adolescent clients within the criminal justice system. The LS/RNR is a recent addition to the Level of Service suite of tools (e.g. Level of Service Inventory-Revised, Andrews & Bonta, 1995; Level of Service/Case Management Inventory [LS/CMI], Andrews et al., 2004; Youth Level of Service/Case Management Inventory [YLS/CMI], Hoge & Andrews, 2002). LS/RNR’s general risk/needs items are identical to LS/CMI items, except LS/RNR excludes the case management sections. This allows LS/RNR users to assess an individual’s risk and needs but apply existing case management policies and protocols in their organisation. For evaluations focused on predictive accuracy, the LS/RNR and LS/CMI are interchangeable due to fully shared risk/need item content.

Internationally, LS/RNR and LS/CMI scores demonstrated moderate to high predictive accuracy in general populations of individuals under correctional supervision (Andrews et al., 2012; Guay & Parent, 2018; Olver et al., 2014b) and a small correlation between risk score and recidivism in a sample of people on probation (Jimenez et al., 2018). Andrews et al. (2012) aggregated results across five datasets to demonstrate that the LS/CMI and YLS/CMI scores showed high predictive accuracy for both females and males in predominantly Canadian jurisdictions. In Australia, Gordon et al. (2015) demonstrated that LS/CMI scores showed moderate predictive discrimination (Area Under the Curve [AUC] = .66) for recidivism within 12 months. We are unaware of published research investigating the incremental prediction of other general recidivism risk assessments beyond LS/RNR scores.

DRAOR

DRAOR is a structured case management tool designed for assessing adults on community supervision. The core purpose of DRAOR is to structure practitioners’ individualised assessment and intervention response to changing risk, strength and destabilising life factors among clients already identified (e.g. from a more comprehensive tool) as at risk and requiring intervention such as community supervision services. Practitioners use discretion when determining intervention strategies following a DRAOR assessment because intervention decisions should be a person-centred choice from multiple possible options. On their own, DRAOR scores do not indicate overall recidivism risk and, thus, neither provide actuarial estimates nor guide summary judgements of overall recidivism risk. However, Coulter et al. (2022) demonstrated one methodology for combining DRAOR and static risk scores to derive actuarial estimates using information from both scores.

Whereas LS/RNR is designed as a comprehensive measure of criminogenic needs, DRAOR is inexhaustive and relatively quick to score so case managers will have timely alerts about changes in factors relevant to recidivism or re-entry success. DRAOR contains three subscales of stable dynamic risk factors (Stable), acute dynamic risk factors (Acute) and strength factors (Protect). Most published research has focussed on DRAOR’s implementation in New Zealand (Davies et al., 2021; Lloyd et al., 2020a; Polaschek & Yesberg, 2017; Scanlan et al., 2020; Yesberg et al., 2015; Yesberg & Polaschek, 2015). In New Zealand, DRAOR scores discriminated general recidivism to a small to moderate degree using Rice and Harris (2005) guidelines (AUCs = .57–.70) within a routine parole sample (Lloyd et al., 2020a), high-risk male samples (Davies et al., 2021; Yesberg & Polaschek, 2015) and a female sample (Scanlan et al., 2020). The present study is the first to examine DRAOR in Australia. Further research is required to test the claims of risk protection using DRAOR’s Protect subscale, but one study suggested that these items may operate similarly to reverse-worded risk items (Lloyd et al., 2020b).

Both Lloyd et al. (2020a) and Davies et al. (2021) demonstrated that the most updated DRAOR score was incrementally related to recidivism after accounting for the baseline (i.e. first) or prior averages of DRAOR scores. In these datasets, the factor structure was invariant across time, suggesting changes in scores represent changes in constructs not scoring drift. Further, models with the greatest predictive discrimination entered scores from both a jurisdiction-specific (internally developed) static risk tool and updated DRAOR scores. In the present study, there is no equivalent jurisdiction-specific static risk measure, so we considered the LS/RNR Criminal History score separately from the other LS/RNR subscales.

The present study

The present study’s purpose was threefold. First, we examined whether DRAOR scores were related to future recidivism in a jurisdiction where DRAOR was recently implemented (in 2018) and had no previous evaluation. Within this aim, we tested whether most recent DRAOR and LS/RNR scores were more strongly related to recidivism than initial scores. Second, considering the goal of efficient prediction, we tested whether DRAOR scores demonstrated incremental prediction by discriminating recidivism beyond LS/RNR scores. Third, because reassessment can be a resource-intensive task, we aimed to determine an optimal model of baseline or reassessed LS/RNR and DRAOR subscales that provided sufficient predictive accuracy, would not impose unnecessary demands on the agency, and retained the conceptual and practical differences in reassessment schedules.

We used a dataset that contained all adults who had an actively managed parole order in Victoria, Australia over 2 years. Corrections officers rated all assessments in routine practice. As a field-based study, our analyses measured LS/RNR’s and DRAOR’s actual predictive performance, as opposed to their potential performance (see Edens & Boccaccini, 2017). Field studies are typically limited in testing interrater reliability and consistency, whereas ecological validity and generalisability are notable advantages. Our outcome was criminal charges recorded in the same state, examined by general or violent outcomes.

Method

Participants

The dataset contained assessment and recidivism data for all people released from prison into the community on a parole order in the state of Victoria if the parole order was actively managed between 1 July 2018 and 30 June 2020 (N = 2088). We removed two people who were released from custody after 30 June 2020. There were 28 additional people who recorded a recidivism event prior to a record of a DRAOR score. Of these 28 people, we removed ten from analysis but retained 18 who had a record of a DRAOR score prior to at least one type of recidivism, allowing sample size to differ across outcome: n = 2076 for violent recidivism, 2063 for general recidivism (criminal charges excluding parole breaches) and 2058 for any recidivism (charges including breaches).

Individuals retained for analysis (N = 2076; 99.5% of the targeted population) were predominately male (88.9%; 11.1% female). Individuals’ ethnicity was recorded as non-Indigenous (93.1%), Aboriginal and/or Torres Strait Islander (5.9%), or unknown (1.1%). Age at the start of study (1 July 2018) ranged from 18 to 89 (M = 39.8, SD = 12.6). With overlapping categories when there were multiple convictions, we categorised current convictions as sexual (12.7%), violent (62.9%), general (93.1%) and breaches (23.6%). Follow-up observation in the community ranged from 1 to 730 days (M = 240.9, SD = 200.3, median = 175.1).

Measures

Recidivism

Using official police records, we defined recidivism as the first criminal charge of each type (i.e. violent, general, or any charge including breaches) during community supervision, retaining the most serious when there were multiple charges related to the same criminal incident occurring the same day. Charges may not have led to a criminal conviction. Using categories from the Crime Statistics Agency (2021), we defined violent recidivism as any crime against the person (including sexual recidivism), general recidivism as any crime except breaches of parole orders and any recidivism as all charges (including breaches). We censored follow-up observation differently depending on type of recidivism (e.g. censoring observation at the time of a general charge when predicting general recidivism, but retaining further observation until a subsequent violent charge when predicting violent recidivism). The date of recidivism was the date the charged crime(s) took place rather than the date police laid the charge(s).

LS/RNR

The LS/RNR is a case-management and treatment planning tool to assess people supervised by correctional services aged 16 and upwards (Andrews et al., 2008). The LS/RNR contains five sections. Section 1 is a general risk and needs section whereas the remaining four sections focus on risk, needs and responsivity domains specific to the client. Section 1 contains 43 items in eight subcomponents. Each subcomponent is based on the Central Eight risk factors (Criminal History, Education/Employment, Family/Marital, Leisure/Recreation, Companions, Alcohol/Drug Problem, Procriminal Attitude/Orientation, Antisocial Pattern; Andrews & Bonta, 2010). Differing across items, corrections staff record scores either on a 4-point scale (where 0 denotes ‘A very unsatisfactory situation with a very clear and strong need for improvement’ and 3 denotes ‘A satisfactory situation with no need for improvement’) or as a dichotomous yes (1)/no (0). Up to four items may be omitted if there is insufficient information to score the items. Subcomponent total scores are calculated by dichotomising the 4-point scale items, summing item scores, then summing the overall risk score. Except for Criminal History, we analysed the overall risk score. No LS/RNR records in the present study were missing overall risk scores. As implemented in Victoria, Australia, staff score LS/RNR during incarceration prior to release, supervision officers again score LS/RNR within 28 days of release, and continue to re-score LS/RNR every 12 months. Corrections staff in Victoria receive MultiHealth System’s approved LS/RNR training.

The Criminal History subcomponent contains eight static items that assess the client’s previous (and current) youth and adult arrests, charges, convictions, incarcerations and institutional misconducts. All items are dichotomous, and scores range from 0 to 8.

DRAOR

DRAOR (Serin, 2007) contains 19 items on three subscales. Each subscale theoretically represents a different temporal or directional relationship to recidivism. The Stable subscale contains six dynamic risk items (peer associations, attitudes towards authority, impulse control, problem-solving, sense of entitlement, attachment with others) that are relatively enduring but are theoretically changeable through intervention. The Acute subscale contains seven dynamic risk items (substance abuse, anger/hostility, opportunity/access to victims, negative mood, employment, interpersonal relationships, living situation) that may change much more rapidly (i.e. day-to-day or week-to-week). Although most Acute items measure the current status of criminogenic need domains also assessed on LS/RNR, other items (opportunity, negative mood and living situation) are life destabilisers or case management needs, but not explicitly criminogenic needs. The Protect subscale contains six items (responsiveness to advice, prosocial identity, high expectations, costs/benefits of remaining crime-free, social support, social control) designed to assess the individual’s internal and external strengths. Case managers rate items on the Stable and Acute subscales as not a problem (0), slight/possible problem (1), or definite problem (2) and rate Protect items as not an asset (0), slight/possible asset (1), or definite asset (2). The presence of Stable and Acute items should be related to a higher likelihood of recidivism, whereas the presence of Protect items should be related to a lower likelihood of recidivism.

As implemented in Victoria, Australia, supervision officers consider DRAOR items at each contact with a client on a parole order. Policies require DRAOR Acute items to be rescored after every client contact (approximately weekly to fortnightly), whereas DRAOR Stable and Protect items must be rescored between 21 and 28 days following release and then every 3 months. Supervision officers may update Stable and Protect items more frequently whenever appropriate. No DRAOR item scores were missing in the present study.

Approximately 90% of corrections staff in the present sample received training and certification directly from a co-developer of the DRAOR scoring manual and training programme. The remaining staff received training from senior corrections staff who completed a ‘train-the-trainer’ programme delivered by the same co-developer. All staff completed certification requirements following training and prior to on-the-job DRAOR scoring.

Demographic information

The databases contained official record information about participants’ age, gender, ethnicity, date of release/start of parole and previous custodies.

Procedure

We obtained ethics approval from institutional review boards at the following organisations: Victorian Department of Justice Human Research Ethics Committee and Swinburne University Human Research Ethics Committee. The context of this study was the case management of supervised individuals. Supervision officers ceased management and recording LS/RNR and DRAOR assessments at the expiry of participants’ parole order, so we censored participants and removed follow-up time and criminal charges after parole expiry. During the study period, 11 individuals were incarcerated and subsequently re-entered the community on a new order (i.e. a new re-entry sequence). Following others (Davies et al., 2021; Howard & Dixon, 2013; Lloyd et al., 2020a), we retained all sequences (see Table 1). See Tables Tables S1–S6 in the online supplemental materials for analyses where we retained only one sequence per person.

Table 1. Sample sizes by outcomes and predictors.

Outcome	Frequencies
	Recidivism events (Base rate)	LS/RNRs	DRAORs	Participants	Sequences
Any recidivism	169 (8.2%)	4511	32,979	2058	2067
General recidivism	134 (6.5%)	4548	33,528	2063	2073
Violent recidivism	54 (2.6%)	4603	33,933	2076	2087

Note: Number of recidivism events, LS/RNRs, and DRAORs are based on sequences. LS/RNR = Level of Service/Risk, Need, Responsivity (Andrews et al., 2008); DRAOR = Dynamic Risk Assessment for Offender Re-entry (Serin, 2007); Sequence = A period of re-entry into the community following prison.

We retained a continuous time data structure, entering and updating all assessment scores (LS/RNR and DRAOR scores) on the days they occurred while omitting any scores on or after the date of charges to ensure prospective prediction. We defined initial, baseline assessments as scores recorded on the day of release from prison. If there was no assessment on the day of release, we used the most recent pre-release assessment when available, or, when not available, we used the first assessment recorded in the community. As shown in Table 2, across the full sample the initial LS/RNR was typically recorded 46 days prior release, and the initial DRAOR was typically recorded approximately on the day of release. Structuring the analysis around continuous time since release allowed analyses to account for some participants being at risk in the community prior to their initial DRAOR assessment.

Table 2. Descriptive statistics and Kendall’s Tau correlations for baseline LS/RNR and DRAOR assessments.

Assessments	Descriptive statistics					Correlations with DRAOR subscales
	N	M	SD	Median	Range	Stable	Acute	Protect
LS/RNR domains
Criminal history	2087	4.97	2.51	6.00	0–8	.307***	.220***	−.215***
Education employment	2087	3.94	2.61	4.00	0–9	.220***	.239***	−.183***
Family marital	2087	2.03	1.23	2.00	0–4	.247***	.242***	−.245***
Leisure recreation	2087	1.45	0.76	2.00	0–2	.259***	.253***	−.244***
Companions	2087	2.90	1.30	3.00	0–4	.320***	.241***	−.250***
Alcohol drug problem	2087	2.85	2.43	2.00	0–8	.292***	.236***	−.190***
Procriminal attitude	2087	2.06	1.25	2.00	0–4	.379***	.257***	−.309***
Antisocial pattern	2087	1.63	1.07	2.00	0–4	.362***	.285***	−.289***
Total	2087	21.84	9.37	22.00	1–42	.385***	.316***	−.297***
DRAOR subscales
Stable	2087	4.95	2.78	5.00	0–12	–	.445***	–.436***
Acute	2087	3.50	2.36	3.00	0–13	–	–	–.337***
Protect	2087	5.26	3.27	6.00	0–12	–	–	–
Days between
LS/RNR and prison release	2087	24.76	257.39	−46.00	−1134–2138	–	–	–
DRAOR and prison release	2087	119.67	268.64	0.00	−48–2255	–	–	–

Note: Days were rounded to the nearest whole number; N = person sequences. LS/RNR = Level of Service/Risk, Need, Responsivity (Andrews et al., 2008); DRAOR = Dynamic Risk Assessment for Offender Re-entry (Serin, 2007). Variance inflation factor (VIF) tests revealed no concerns about multicollinearity, with DRAOR Stable VIF = 2.03, DRAOR Acute VIF = 1.56, DRAOR Protect VIF = 1.58, and LS/RNR VIF = 1.45.

*** p < .001.

Plan of analysis

Following others (see Davies et al., 2021; Lloyd et al., 2020a; Stone et al., 2022), we retained the temporal distinctions intended for each DRAOR subscale and analysed each separately. We also separately analysed the LS/RNR Section 1 overall risk score and the Criminal History subcomponent risk score. We calculated descriptive information (mean, median, standard deviation, range) for the baseline DRAOR subscales, overall LS/RNR risk score and Criminal History score, as well as the number of days between prison release and the baseline assessment. We also calculated correlations between DRAOR subscales and LS/RNR scores using Kendall’s tau.

We used Cox regression survival analyses with continuous time intervals to determine if higher (compared to lower) scores on each predictor discriminated likelihood of outcome. First, using each definition of recidivism as a separate outcome, we entered each baseline predictor separately in univariate models. Second, we added reassessment scores to the models with statistically significant results in the first step. Comparison of reassessment models indicated which tool components best identified more critically immediate case management needs. Third, using baseline scores when reassessment scores did not add incremental prediction in the second step (but otherwise using reassessment scores), we analysed multivariate models that first used either the LS/RNR overall risk score or LS/RNR Criminal History subcomponent score, then incrementally added the DRAOR subscales. Comparison of multivariate models indicated the strongest and most efficient prediction model. We did not analyse baseline DRAOR subscale scores with reassessed LS/RNR scores because this model does not reflect the reassessment frequency of these tools in practice.

Model fit and effect size statistics

To determine the strongest prediction models, we calculated Akaike’s information criterion (AIC), Bayesian information criterion (BIC) and two effect size statistics. AIC and BIC each measure the difference between predicted and observed values while incorporating penalties based on the number of predictors in the model. AIC and BIC values are uninterpretable, but models with lower AIC or BIC values fit the data better than models with higher values. Burnham and Anderson (2004) suggest that differences of at least 10 AIC or BIC values across models are meaningful, but differences of 4–7 may indicate modest support.

We report Heagerty and Zheng’s (2005) weighted average c index (concordance). A summary of the model’s discrimination, c is the probability that a randomly selected recidivist will have a higher risk score (or a lower protective score) on the measure than a randomly selected non-recidivist, while taking follow-up time into account by comparing only recidivists and non-recidivists observed in the community for the same length of time. The c index can be interpreted similarly to AUCs, with values of .56, .64 and .71 associated with small, moderate and large effects respectively (Helmus & Babchishin, 2017).

We also report Xu and O’Quigley’s (1999) R², an analogue of the proportion of variance explained. When R² is 0.0, there is no relationship between the observed order of recidivism events and variance in predictor scores, whereas when R² is 1.0, the first recidivism event is associated with the highest observed predictor score, the second recidivism event is associated with the second highest observed predictor score, and so on, indicating a perfect ordering of the timing of recidivism events and variance in predictor scores. Values are proportions, such that R² = .60 indicates 60% of the variation in the order of recidivism events is explained by the predictor. Because individuals who did not experience the outcome (i.e. recidivism) are excluded when examining the order of outcome events, this R² measure can only be interpreted similarly to a traditional linear R² value (i.e. Cohen, 1988) when all participants have experienced the outcome event. This is not the case in the present study. We conducted analyses using IBM SPSS (Version 26.0) and R (Version 4.0).

Results

Descriptives

Table 2 presents descriptive statistics and correlations for baseline assessment scores. Examining the mean DRAOR subscale scores, supervision officers identified the typical participant as having 3–5 criminogenic needs, 3–5 strength factors and 2–3 acute risk destabilisers. Mean DRAOR subscale scores were lower compared to people on parole in New Zealand (see Davies et al., 2021; Lloyd et al., 2020a), but this sample’s recidivism base rates during parole were also low (see Table 1). There were statistically significant moderate correlations between all LS/RNR and DRAOR subscale scores.

Univariate models

Table 3 displays results from univariate Cox regression models using each case management tool component assessed at baseline to predict each recidivism outcome. All but two models showed statistically significant discrimination of recidivism. DRAOR Protect discriminated any recidivism outcomes (including breaches) but did not discriminate violent or general recidivism (excluding breaches). Models using LS/RNR Criminal History scores generated the strongest effects across all three recidivism outcomes.

Table 3. Univariate Cox regression models with baseline LS/RNR, criminal history, or DRAOR subscales as predictors.

Models	B (SE)	Exp(B) [95% CI]	c index	95% CI	R^2	AIC/BIC
Any recidivism^a
1. LS/RNR	.11*** (0.01)	1.11 [1.09, 1.14]	.73	[0.69, 0.77]	.68	2116.691/2119.821
2. LS/RNR Criminal history	.48*** (0.05)	1.62 [1.46, 1.79]	.74	[0.70, 0.77]	.69	2109.343/2112.473
3. DRAOR Stable	.13*** (0.03)	1.14 [1.08, 1.20]	.61	[0.57, 0.66]	.09	2220.751/2223.881
4. DRAOR Acute	.14*** (0.03)	1.15 [1.08, 1.22]	.60	[0.56, 0.65]	.06	2222.056/2225.186
5. DRAOR Protect	−.06* (0.03)	0.95 [0.90, 0.99]	.55	[0.50, 0.60]	.02	2235.668/2238.798
General recidivism^b
6. LS/RNR	.11*** (0.01)	1.11 [1.09, 1.14]	.74	[0.70, 0.78]	.60	1666.292/1669.190
7. LS/RNR Criminal history	.47*** (0.06)	1.60 [1.43, 1.79]	.74	[0.71, 0.78]	.62	1664.177/1667.074
8. DRAOR Stable	.13*** (0.03)	1.14 [1.07, 1.21]	.62	[0.57, 0.66]	.06	1750.333/1753.231
9. DRAOR Acute	.12*** (0.04)	1.12 [1.05, 1.20]	.59	[0.54, 0.64]	.06	1755.963/1758.861
10. DRAOR Protect	−.04 (0.03)	0.96 [0.90, 1.01]	.54	[0.49, 0.59]	.00	1764.033/1766.931
Violent recidivism^c
11. LS/RNR	.11*** (0.02)	1.12 [1.08, 1.16]	.75	[0.68, 0.82]	.55	653.716/655.705
12. LS/RNR Criminal history	.49*** (0.09)	1.63 [1.37, 1.93]	.76	[0.70, 0.81]	.62	651.736/653.725
13. DRAOR Stable	.12* (0.05)	1.13 [1.02, 1.24]	.59	[0.50, 0.68]	.04	692.689/694.678
14. DRAOR Acute	.13* (0.06)	1.14 [1.02, 1.27]	.60	[0.52, 0.68]	.07	693.025/695.013
15. DRAOR Protect	−.05 (0.04)	0.95 [0.87, 1.04]	.53	[0.44, 0.61]	.02	696.987/698.976

Note: LS/RNR = Level of Service/Risk, Need, Responsivity (Andrews et al., 2008); Criminal History = Criminal History sub-component score from LS/RNR; DRAOR = Dynamic Risk Assessment for Offender Re-entry (Serin, 2007); CI = Confidence interval; c index = concordance (Heagerty & Zheng, 2005); R²= Xu and O’Quigley’s (1999) R², values should not be interpreted in line with Cohen (1988); AIC = Akaike information criterion; BIC = Bayesian information criterion (calculated using the number of events as the sample size; see Volinsky & Raftery, 2000).

^a First criminal charge (including breaches) following prison release, 2067 person sequences (n = 34,790 person periods, 169 recidivism events).

^b First criminal charge (excluding breaches) following prison release, 2073 person sequences (n = 35,426 person periods, 134 recidivism events).

^c First violent criminal charge (including sexual violence) following prison release, 2087 person sequences (n = 35,973 person periods, 54 recidivism events).

* p < .05. ** p < .01. *** p < .001.

Multivariate models

Table 4 displays results from models examining whether updating dynamic predictor scores through the most recent reassessments showed incremental predictive discrimination over the baseline assessment on the same measure. For all three recidivism outcomes, change in DRAOR Stable and Acute scores demonstrated incremental prediction of recidivism over the first recorded DRAOR score. The reassessment effect was also statistically significant for LS/RNR and DRAOR Protect when examining any recidivism (including breaches of parole). Although updates to LS/RNR overall risk scores slightly improved discrimination of any recidivism (but not general or violent recidivism), differences in AIC values indicated that baseline LS/RNR scores (AIC = 2116.691) still generated a better fitting model than updated LS/RNR scores (AIC = 2127.851; see Burnham & Anderson, 2004). Thus, timelier information through LS/RNR reassessment did not generate information that more accurately discriminated individuals who did and did not recidivate during parole.

Table 4. Multivariate Cox regression models with baseline and most recent LS/RNR or DRAOR subscales as predictors.

Models	B (SE)	Exp(B) [95% CI]	c index	95% CI	R^2	AIC/BIC
Any recidivism^a
1. LS/RNR Baseline	.07*** (0.02)	1.08 [1.04, 1.12]	.73	[0.70, 0.77]	.66	2114.336/2120.596
LS/RNR Updated	.04* (0.02)	1.04 [1.00, 1.08]
2. DRAOR Stable Baseline	.01 (0.04)	1.01 [0.93, 1.10]	.63	[0.58, 0.67]	.19	2205.600/2211.860
DRAOR Stable Updated	.17*** (0.04)	1.19 [1.09, 1.29]
3. DRAOR Acute Baseline	.02 (0.04)	1.02 [0.95, 1.10]	.66	[0.62, 0.71]	.25	2183.546/2189.806
DRAOR Acute Updated	.23*** (0.03)	1.25 [1.17, 1.34]
4. DRAOR Protect Baseline	.02 (0.04)	1.02 [0.94, 1.10]	.57	[0.52, 0.62]	.04	2231.504/2237.764
DRAOR Protect Updated	−.10* (0.04)	0.91 [0.84, 0.98]
General recidivism^b
5. LS/RNR Baseline	.08*** (0.02)	1.08 [1.04, 1.13]	.75	[0.71, 0.78]	.58	1665.137/1670.933
LS/RNR Updated	.04 (0.02)	1.04 [1.00, 1.08]
6. DRAOR Stable Baseline	.03 (0.05)	1.03 [0.94, 1.13]	.63	[0.58, 0.68]	.23	1742.000/1747.795
DRAOR Stable Updated	.14** (0.05)	1.16 [1.06, 1.26]
7. DRAOR Acute Baseline	.01 (0.04)	1.01 [0.93, 1.10]	.66	[0.61, 0.70]	.29	1729.751/1735.547
DRAOR Acute Updated	.21*** (0.04)	1.24 [1.15, 1.33]
8. DRAOR Protect Baseline	.01 (0.04)	1.01 [0.93, 1.09]	.57	[0.51, 0.62]	.03	1763.488/1769.284
DRAOR Protect Updated	−.07(0.04)	0.93 [0.86, 1.02]
Violent recidivism^c
9. LS/RNR Baseline	.08** (0.03)	1.08 [1.02, 1.15]	.75	[0.69, 0.82]	.60	653.898/657.875
LS/RNR Updated	.04 (0.03)	1.04 [0.98, 1.11]
10. DRAOR Stable Baseline	.01 (0.07)	1.01 [0.88, 1.16]	.61	[0.52, 0.69]	.27	689.181/693.159
DRAOR Stable Updated	.16* (0.07)	1.17 [1.03, 1.34]
11. DRAOR Acute Baseline	.03 (0.06)	1.03 [0.91, 1.17]	.66	[0.57, 0.74]	.31	683.086/687.064
DRAOR Acute Updated	.21*** (0.06)	1.24 [1.11, 1.38]
12. DRAOR Protect Baseline	−.002 (0.06)	1.00 [0.88, 1.13]	.55	[0.46, 0.63]	.04	697.796/701.774
DRAOR Protect Updated	−.07 (0.07)	0.93 [0.82, 1.06]

Note: LS/RNR = Level of Service/Risk, Need, Responsivity (Andrews et al., 2008); DRAOR = Dynamic Risk Assessment for Offender Re-entry (Serin, 2007); CI = Confidence interval; c index = concordance (Heagerty & Zheng, 2005); R²= Xu and O’Quigley’s (1999) R², values should not be interpreted in line with Cohen (1988); AIC = Akaike information criterion; BIC = Bayesian information criterion (calculated using the number of events as the sample size; see Volinsky & Raftery, 2000).

^a First criminal charge (including breaches) following prison release, 2067 person sequences (n = 34,790 person periods, 169 recidivism events).

^b First criminal charge (excluding breaches) following prison release, 2073 person sequences (n = 35,426 person periods, 134 recidivism events).

^c First violent criminal charge (including sexual violence) following prison release, 2087 person sequences (n = 35,973 person periods, 54 recidivism events).

* p < .05. ** p < .01. *** p < .001.

Having established that changes in DRAOR Stable and Acute scores were related to the timing of recidivism, we present models examining if updated DRAOR scores incrementally discriminated recidivism beyond initial LS/RNR scores in Table 5 and LS/RNR Criminal History subcomponent scores (the strongest and most efficient predictor component) in Table 6. DRAOR Stable added incremental predictive discrimination for any (including breaches) and general (excluding breaches) recidivism outcomes when first accounting for LS/RNR Criminal History, but DRAOR Stable did not add to the prediction of general recidivism when first accounting for LS/RNR total scores. The dynamic risk domains of both LS/RNR and DRAOR Stable are intended to measure criminogenic needs, and these results suggest that their content domains likely overlap. DRAOR Protect did not add incrementally beyond LS/RNR Criminal History or total scores when predicting all recidivism outcomes.

Table 5. Multivariate Cox regression models with baseline LS/RNR and most recent DRAOR subscales as predictors.

Models	B (SE)	Exp(B) [95% CI]	c index	95% CI	R^2	AIC/BIC
Any recidivism^a
1. LS/RNR Baseline	.10*** (0.01)	1.12 [1.08, 1.13]	.73	[0.70, 0.77]	.69	2113.242/2119.501
DRAOR Stable Updated	.07* (0.03)	1.08 [1.01, 1.14]
2. LS/RNR Baseline	.10*** (0.01)	1.10 [1.08, 1.12]	.75	[0.72, 0.79]	.69	2095.224/2101.483
DRAOR Acute Updated	.15*** (0.03)	1.17 [1.10, 1.24]
3. LS/RNR Baseline	.11*** (0.01)	1.11 [1.09, 1.14]	.73	[0.69, 0.77]	.68	2118.681/2124.941
DRAOR Protect Updated	−0.001 (0.03)	1.00 [0.95, 1.06]
General recidivism^b
4. LS/RNR Baseline	.10*** (0.01)	1.12 [1.08, 1.14]	.74	[0.70, 0.78]	.61	1665.070/1671.303
DRAOR Stable Updated	.06 (0.03)	1.06 [0.99, 1.14]
5. LS/RNR Baseline	.10*** (0.01)	1.10 [1.08, 1.13]	.76	[0.72, 0.80]	.62	1655.342/1661.138
DRAOR Acute Updated	.13*** (0.03)	1.14 [1.07, 1.22]
6. LS/RNR Baseline	.11*** (0.01)	1.12 [1.09, 1.14]	.74	[0.70, 0.78]	.61	1667.615/1673.411
DRAOR Protect Updated	.02 (0.03)	1.03 [0.96, 1.09]
Violent recidivism^c
7. LS/RNR Baseline	.11*** (0.02)	1.11 [1.07, 1.16]	.75	[0.68, 0.82]	.57	654.687/658.665
DRAOR Stable Updated	.06 (0.05)	1.06 [0.95, 1.18]
8. LS/RNR Baseline	.10*** (0.02)	1.11 [1.07, 1.14]	.77	[0.70, 0.83]	.61	649.873/653.851
DRAOR Acute Updated	.14* (0.05)	1.15 [1.03, 1.28]
9. LS/RNR Baseline	.11*** (0.02)	1.12 [1.08, 1.16]	.75	[0.68, 0.82]	.55	655.525/659.503
DRAOR Protect Updated	.02 (0.5)	1.02 [0.93, 1.13]

Note: LS/RNR = Level of Service/Risk, Need, Responsivity (Andrews et al., 2008); DRAOR = Dynamic Risk Assessment for Offender Re-entry (Serin, 2007); CI = Confidence interval; c index = concordance (Heagerty & Zheng, 2005); R²= Xu and O’Quigley’s (1999) R², values should not be interpreted in line with Cohen (1988); AIC = Akaike information criterion; BIC = Bayesian information criterion (calculated using the number of events as the sample size; see Volinsky & Raftery, 2000).

^a First criminal charge (including breaches) following prison release, 2067 person sequences (n = 34,790 person periods, 169 recidivism events).

^b First criminal charge (excluding breaches) following prison release, 2073 person sequences (n = 35,426 person periods, 134 recidivism events).

^c First violent criminal charge (including sexual violence) following prison release, 2087 person sequences (n = 35,973 person periods, 54 recidivism events).

* p < .05. ** p < .01. *** p < .001.

Table 6. Multivariate Cox regression models with baseline criminal history and most recent DRAOR subscales as predictors.

Models	B (SE)	Exp(B) [95% CI]	c index	95% CI	R^2	AIC/BIC
Any recidivism^a
1. LS/RNR Criminal History	.45*** (0.05)	1.56 [1.41, 1.73]	.75	[0.71, 0.78]	.69	2102.938/2109.186
DRAOR Stable Updated	.09** (0.03)	1.09 [1.03, 1.16]
2. LS/RNR Criminal History	.43*** (0.05)	1.54 [1.40, 1.71]	.77	[0.73, 0.80]	.71	2079.978/2086.232
DRAOR Acute Updated	.18*** (0.03)	1.19 [1.13, 1.27]
3. LS/RNR Criminal History	.47*** (0.05)	1.60 [1.45, 1.77]	.74	[0.70, 0.77]	.69	2110.404/2116.652
DRAOR Protect Updated	−.03 (0.03)	0.97 [0.92, 1.03]
General recidivism^b
4. LS/RNR Criminal History	.44*** (0.06)	1.55 [1.39, 1.74]	.75	[0.72, 0.79]	.62	1661.254/1667.040
DRAOR Stable Updated	.08* (0.03)	1.08 [1.01, 1.16]
5. LS/RNR Criminal History	.43*** (0.06)	1.53 [1.37, 1.71]	.77	[0.73, 0.80]	.65	1647.785/1653.575
DRAOR Acute Updated	.16*** (0.03)	1.17 [1.09, 1.25]
6. LS/RNR Criminal History	.47*** (0.06)	1.60 [1.43, 1.79]	.74	[0.71, 0.78]	.64	1666.173/1671.968
DRAOR Protect Updated	−.002 (0.03)	1.00 [0.94, 1.06]
Violent recidivism^c
7. LS/RNR Criminal History	.46*** (0.09)	1.59 [1.33, 1.89]	.76	[0.70, 0.81]	.62	651.916/655.832
DRAOR Stable Updated	.07 (0.05)	1.08 [0.97, 1.20]
8. LS/RNR Criminal History	.45*** (0.09)	1.56 [1.32, 1.85]	.79	[0.73, 0.84]	.67	645.761/649.737
DRAOR Acute Updated	.16** (0.05)	1.17 [1.06, 1.30]
9. LS/RNR Criminal History	.49*** (0.09)	1.62 [1.37, 1.93]	.75	[0.69, 0.81]	.62	653.723/657.701
DRAOR Protect Updated	−.006 (0.05)	0.99 [0.90, 1.10]

Note: Criminal History = Criminal History sub-component score from Level of Service/Risk, Need, Responsivity (Andrews et al., 2008); DRAOR = Dynamic Risk Assessment for Offender Re-entry (Serin, 2007); CI = Confidence interval; c index = concordance (Heagerty & Zheng, 2005); R²= Xu and O’Quigley’s (1999) R², values should not be interpreted in line with Cohen (1988); AIC = Akaike information criterion; BIC = Bayesian information criterion (calculated using the number of events as the sample size; see Volinsky & Raftery, 2000).

^a First criminal charge (including breaches) following prison release, 2067 person sequences (n = 34,790 person periods, 169 recidivism events).

^b First criminal charge (excluding breaches) following prison release, 2073 person sequences (n = 35,426 person periods, 134 recidivism events).

^c First violent criminal charge (including sexual violence) following prison release, 2087 person sequences (n = 35,973 person periods, 54 recidivism events).

* p < .05. ** p < .01. *** p < .001.

However, across six models predicting all three recidivism outcomes, DRAOR Acute scores incrementally added to LS/RNR Criminal History and LS/RNR total scores. The multivariate models including both LS/RNR and DRAOR Acute achieved the strongest predictive discrimination, especially when restricting LS/RNR to its Criminal History subcomponent, as seen by comparing AIC, BIC, c index and R² values across Tables 5 and 6. For all models reported in Tables 5 and 6, LS/RNR remained a statistically significant predictor when DRAOR subscales were included as predictors.

Discussion

Corrections agencies are tasked with efficiently establishing the risk of recidivism, generating clinical strategies to facilitate crime desistance, and identifying current re-entry problems that need timely support, yet it is unlikely that a single tool can accomplish these goals using the same item definitions, assessment context and reassessment schedules. In this study, we investigated single and combined components from two case management tools to better establish a practical approach. This involved considering the potential dynamic properties of these tools because reassessment during re-entry is the methodology that links the emerging presence of client needs to the imminent likelihood of recidivism.

Our results support several conclusions. First, when efficient prediction is the sole goal, criminal history scores offer a simple and robust approach. Among the univariate models, the LS/RNR Criminal History score represented the least intensive assessment yet the most discriminating component. Still, attention toward dynamic risk factors is required to understand client needs. Fortunately, in comparison to models using LS/RNR Criminal History only (see Table 6), the combination of LS/RNR Criminal History scores with the other LS/RNR subcomponents resulted in model effect sizes that remained almost as equally high, even though the model fit for recidivism prediction was slightly reduced (see Table 5). Given this largely overlapping variance, it may be best to consider measures of criminogenic needs as providing practitioners with a conceptual understanding of static risk rather than a component explicitly necessary for prediction. Further, including reassessed DRAOR Stable scores alongside LS/RNR Criminal History improved prediction for any and general but not violent recidivism and, when included alongside overall LS/RNR scores, DRAOR Stable scores only improved prediction of any recidivism. DRAOR Stable scores were also more strongly correlated with LS/RNR scores relative to DRAOR Acute and Protect scores, suggesting the eight domains of LS/RNR and DRAOR Stable operate similarly.

By contrast, including acute dynamic risk factors alongside LS/RNR scores in the form of DRAOR Acute subscale scores improved prediction accuracy (see Tables 5 and 6). Notably, combining DRAOR Acute scores with either LS/RNR Criminal History or LS/RNR total scores improved model fit, suggesting community-rated acute dynamic factors represent an efficient and meaningful addition to recidivism prediction models. This finding was consistent across all definitions of recidivism, including violent recidivism. Because the model with the strongest predictive discrimination included LS/RNR Criminal History and DRAOR Acute, this suggests case managers should always consider the nature of current acute risks in the context of clients’ criminal history and supplement decision-making based on static risk level with understanding of current community-based functioning.

Of further note, DRAOR Acute’s dynamic risk and life-destabilising factors provided timely clinical information useful to case managers concerned about recidivism imminence (see Table 6). Across all definitions of recidivism, updated DRAOR Acute scores added incremental prediction beyond the first recorded DRAOR Acute score. In other words, more recent changes in acute dynamic risk factors better indicated current recidivism likelihood. Whereas LS/RNR provided the best overall prediction models, reassessment of LS/RNR was either not statistically significant or did not substantially improve model fit, depending on the definition of recidivism. Because the LS/RNR is designed to assess relatively stable, long-term intervention needs, this suggests the LS/RNR works as intended. This also means LS/RNR scores are a less promising candidate for indicating imminently current case management concerns during parole.

The reassessment effect of DRAOR Acute in this study is consistent with prior studies of DRAOR in New Zealand (see Davies et al., 2021; Lloyd et al., 2020a) and evaluations of other dynamic risk assessment tools in other jurisdictions (see Babchishin & Hanson, 2020; Hanson et al., 2021) using similar analytic models. Lloyd, Hanson and colleagues (2020a) discussed several explanations for this incremental predictive validity. Briefly, acute dynamic risk factors may serve as real-time flags, or, because people who stay crime-free during parole generally show greater decreases in dynamic risk, reassessment models may more appropriately downgrade risk among lower risk people. Further, case managers may score DRAOR more accurately as they gather more client information and gain a better understanding of their clients’ functioning in the community.

Limitations and future directions

The present study focussed on one of two broad types of predictive validity. We examined whether LS/RNR and DRAOR scores discriminated among people with different likelihood of recidivism, but we did not address or establish score calibration, or the absolute recidivism rates associated with each score or combination of scores from the risk assessment tools. Further, although studies in New Zealand (Davies et al., 2021; Lloyd et al., 2020a) have demonstrated that the constructs measured by DRAOR do not vary across time, examining measurement invariance across time was beyond the scope of this present study.

The present study demonstrated that DRAOR subscales (particularly DRAOR Acute) showed incremental prediction to LS/RNR scores. However, practitioners also scored each tool on dissimilar timeframes for dissimilar purposes. Thus, our analyses could not fully determine if incremental prediction of DRAOR beyond LS/RNR may be due to either differences in item content or assessment recency. On the one hand, it may be that information assessed using DRAOR Acute is redundant with LS/RNR when assessed on the same schedule. Supporting this, baseline DRAOR subscales showed no incremental prediction over baseline LS/RNR (see Tables S7 and S8 in the online supplement materials). On the other hand, updated LS/RNR scores did not improve prediction beyond baseline LS/RNR scores, whereas updated DRAOR Acute scores improved prediction beyond LS/RNR. This may suggest recency effects underlie the prediction effect associated with DRAOR Acute or may suggest DRAOR Acute item content becomes most relevant once re-entry is underway.

The present study is limited in that it was beyond the scope to conduct subgroup analyses based on characteristics such as sex/gender and race/ethnicity. Given that differences in discrimination may occur across groups (for examples, see Babchishin et al., 2012; Campbell et al., 2019; Olver et al., 2018), future research should attend to establishing measurement invariance and predictive validity across groups. Further, for many reasons, case managers differ in how they view, use and value case management tools (see Miller & Maloney, 2020; Schaefer & Williamson, 2018) and how they intervene and manage their client’s needs (see Viljoen et al., 2019; Vincent et al., 2021). The present study also did not investigate prediction or intervention effects by case manager.

Finally, the present study’s recidivism base rate was low for a parole population (e.g. 8.2% for any recidivism versus 36.8% reported by Lloyd et al., 2020a). There are several possible explanations. First, this parole sample specifically excluded people who had been sentenced to extended supervision orders due to a court process that identified unacceptable levels of risk for committing serious violent (including sexual) crimes (Serious Offenders Act, 2018 [Victoria, Australia]). Second, although we only analysed data when people were on parole because predictor assessments ceased at parole expiry, the recidivism database notably included a meaningful number of post-parole recidivism events (i.e. 6.5% general recidivism rate observed during parole versus 29.9% observed across parole and the first 12 months after parole expiry). This rate discrepancy suggests client management strategies during parole supervision contributed to low recidivism rates among people genuinely at risk for recidivism. Third, delays between recidivism events and when police laid charges may have caused the omission of some recidivism events because our total follow-up period included only 2 years. Further, comparable studies in New Zealand (see Davies et al., 2021; Lloyd et al., 2020a) included people who reached their statutory release date and were released on conditions (see Parole Act, 2002 [New Zealand]). In Victoria, people who reach their statutory release date are not released on parole (Corrections Act, 1986 [Victoria, Australia]), ensuring a comparatively lower risk parole population in our study. Finally, differences in parole board decision thresholds may affect a parole sample’s level of risk which undermines cross-jurisdiction comparisons. Still, despite a low base rate, this study had enough statistical power to detect effects that reached statistical significance with effect sizes similar to previous studies (Davies et al., 2021; Lloyd et al., 2020a).

Practical implications

The results from the present study suggest that a correctional agency can achieve three important practical goals by applying three specific components within risk measures. Specifically, for solely prediction goals (e.g. triaging clients into service pathways to distribute rehabilitation and management resources), a simple and efficient measure of static risk based on criminal history is likely sufficient. Yet, a suite of policies that solely addresses prediction while omitting effective management would be a disservice to people supervised by corrections agencies and may be unethical (Roychowdhury & Adshead, 2014), when resources are available to provide management and intervention services.

Therefore, an agency requires a more comprehensive measure of criminogenic needs (i.e. the seven dynamic factors within the Central Eight) to generate a clinical understanding of the drivers of client risk in the context of intervention. As such, in the present study context, results suggest it was largely reasonable to trade away some prediction accuracy by using the comprehensive LS/RNR need assessment to both set the overall risk level and identify criminogenic needs. However, the decision-making context may determine whether simple efficiency should be prioritised over comprehensiveness.

On the other hand, a reassessment of all Central Eight need factors would be an inefficient and ineffective way to identify current client triggers that need immediate attention and action. Instead, DRAOR Acute items offered a more efficient framework for ongoing reassessment of critical re-entry challenges for case manager intervention.

Disclosure statement

The views expressed are those of the authors and not necessarily those of Corrections Victoria, Victoria Police, or the State of Victoria. Caleb D. Lloyd is a co-author of the 2017 version of the DRAOR scoring manual and co-developer of the DRAOR training programme and training certifications.

Data availability statement

Corrections Victoria and Victoria Police own the data described in this article; we used these data with their permission and can share the data only with written permission from Corrections Victoria and Victoria Police. Analysis code for this study is available by emailing the corresponding author.

Additional information

Funding

This research was supported by an Australian Government Research Training Program Scholarship and by the Department of Justice and Community Safety Victoria.