Abstract
In this review article, we critically analyze the relationship between change in dynamic risk (i.e., intraindividual change) and recidivism. We start by reviewing the empirical evidence, which indicates intraindividual change is associated with recidivism. However, we highlight how the predictive value of change scores needs careful interpretation. The finding that change scores predict incrementally over baseline scores may simply reflect the improved accuracy of a more recent assessment. Alternatively, the degree of change preceding the reassessment may be relevant in addition to the current level of dynamic risk at reassessment. We propose theoretical reasons why past change may be relevant for prediction beyond current risk scores. Furthermore, the empirical evidence suggests prior change may contain information not solely reflected within the current risk score, but the current evidence has several limitations. Due to the important implications for correctional practice, we encourage further research that more directly examines this question.
Assessing an individual’s risk of recidivism has long been central to correctional psychology, but recently, attention has shifted to assessing change in an individual’s risk of recidivism. This shift has been driven by a more widespread use of risk assessment measures derived from dynamic (i.e., variable) risk factors, rather than measures derived solely from static (i.e., unchangeable) risk factors (Bonta & Andrews, 2016; Serin et al., 2016). Change in risk of recidivism at the individual level, or intraindividual change, can only be assessed using measures derived from dynamic risk factors. Such assessments can have serious consequences in correctional settings. In prison, for example, an assessment of change can affect whether an individual is granted early release (Beggs, 2010); in the community, it can determine the level of supervision or indicate the need for immediate intervention. These decisions rely on the assumption that intraindividual change is associated with recidivism. Until recently, that assumption was largely untested, but support for the association between intraindividual change and recidivism is emerging. As this evidence grows, researchers and practitioners need a clear understanding of how to interpret this research and apply it to practice.
In this article, we analyze the relationship between intraindividual change and recidivism. We start by briefly reviewing the current state of empirical research. Subsequently, we examine the different ways this research can be interpreted and consider the implications those different interpretations have for practice. We focus on the role of intraindividual change in the prediction of recidivism. We question whether it is only the risk score after the period of purported change that is relevant to this prediction or whether the amount of prior score change is additionally relevant. To answer that question, we discuss reasons why intraindividual change might have an incremental predictive value and assess relevant empirical evidence. We finish with a discussion of the implications for practice and research.
The Association Between Intraindividual Change and Recidivism
The association between intraindividual change and recidivism is central to the risk-need-responsivity (RNR; Andrews, Bonta, & Hoge, 1990) model of effective correctional intervention, the dominant model in correctional psychology (Polaschek, 2012). The RNR model was developed in response to claims that rehabilitative efforts did not reduce recidivism (e.g., Martinson, 1974). Initial proponents of the RNR model argued that these claims were based on a misreading of the existing research evidence. Using then new meta-analytic techniques, they highlighted how rehabilitation programs adhering to key principles consistently demonstrated significant reductions in recidivism (Andrews, Zinger, et al., 1990). In the RNR model, the association between change and recidivism is primarily specified in the need principle, which states that interventions should target variables “that, when influenced, are associated with changes in the chance of recidivism” (Andrews, Bonta, & Hoge, 1990, p. 20). These dynamic risk factors, referred to in the context of correctional rehabilitation as criminogenic needs, include both internal features of an individual such as attitudes and external features of the individual’s environment or circumstances such as employment situation.
Later research has further established that rehabilitation programs targeting criminogenic needs are more effective at reducing recidivism than programs targeting noncriminogenic needs (Bonta & Andrews, 2016; Smith et al., 2009). The research usually assumes that intraindividual change is the mechanism driving the reduction in recidivism although, for years, there was little empirical research testing that assumption (Douglas & Skeem, 2005). Two comprehensive reviews (Beggs, 2010; Serin et al., 2013) highlighted the few studies that conducted such tests. For example, Serin et al. (2013) found that only 17 of 378 cognitive skills, violence reduction, and substance abuse studies that measured treatment change directly examined the association between change and recidivism. They suggested there was some evidence in those 17 studies that decreases in well-established dynamic risk factors such as antisocial associates, antisocial personality, and antisocial attitudes were associated with lower recidivism (i.e., 10 of the 17 studies found a significant link). However, their main conclusion was that there was an absence of high-quality studies in the area, and further research was needed to establish the association between change and recidivism.
Since those reviews, the research has grown considerably, in part due to the increasing use of the Violence Risk Scale (VRS; Wong & Gordon, 1999-2003) and Violence Risk Scale: Sexual Offense version (VRS-SO; Wong et al., 2003). These tools were designed for assessing risk—of violence and of sexual violence, respectively—before and after individuals engage in RNR-based programs. Consequently, the tools are well suited for research examining the extent to which pre-post intraindividual change is associated with recidivism. Multiple studies show that change on the VRS (Coupland & Olver, 2018; Hogan & Olver, 2019; Lewis et al., 2013; Olver et al., 2013) and VRS-SO (Beggs & Grace, 2010; Olver et al., 2007, 2014, 2020, 2021; Olver & Wong, 2011; Sowden & Olver, 2017) is associated with recidivism. Limitations of this body of research include that more samples are retrospective, archival file reviews than field-based prospective studies, and that, with the exception of Beggs and Grace (2010), these studies have been conducted by the developers of the tools and with primarily Canadian samples. One prospective field-based Australian study, albeit using a small sample, found change scores on the VRS were not significantly associated with violent recidivism (O’Brien & Daffern, 2017).
Recent research with other dynamic risk measures has also found associations between intraindividual change and recidivism although results have been inconsistent. In juvenile residential placement, Baglivio and colleagues (Baglivio, Wolff, Jackowski, & Greenwald, 2017; Baglivio, Wolff, Piquero, et al., 2017; Baglivio et al., 2018) found some support, but change in most dynamic variables was not associated with recidivism. Several studies of versions of the Historical Clinical Risk Management-20 (Webster et al., 1997) also found conflicting evidence (Coupland & Olver, 2018; de Vries Robbé et al., 2015; Mastromanno et al., 2018). Similar to Baglivio and colleagues’ work, change in dynamic risk during community supervision (Barnes et al., 2016; Cohen et al., 2016; Cohen & VanBenschoten, 2014; Howard & Dixon, 2013; Labrecque et al., 2014; Vose et al., 2013; Wooditch et al., 2014), or following completion of an intervention program (Kroner & Yessine, 2013), has found some support. However, some of those studies (Kroner & Yessine, 2013; Wooditch et al., 2014) found most dynamic risk measures were not associated with recidivism.
The most straightforward method for examining the association between change and recidivism is to calculate and use raw change scores (e.g., posttreatment score minus pretreatment score). However, that approach can produce misleading results. Baglivio, Wolff, Jackowski, and Greenwald (2017) provide the hypothetical scenario of two individuals, one who demonstrates positive change across the treatment and one who makes no change during the treatment (see Panel A of Figure 1, detailed below). Based on change alone, the first individual should have a lower likelihood of recidivism than the second. However, if the first individual was at a much higher risk than the second before the treatment, then despite their changes, they remain at higher risk than the second individual after treatment. Therefore, analyses that do not account for baseline risk may misleadingly suggest that intraindividual change is not associated with recidivism.
Three Hypothetical Patterns of Change in Two Individuals During Psychological Treatment
Consequently, the most common method used in research examining change and recidivism, used in most studies reviewed earlier, is to investigate incremental predictive validity of raw change scores over baseline—usually pretreatment—scores. Baseline risk should include both static and baseline dynamic risk because (a) static and dynamic risk are not multicollinear, so controlling only for static risk is insufficient, and (b) the baseline dynamic score will be used to calculate the raw change score. The value of controlling for baseline risk was illustrated in the meta-analysis by van den Berg et al. (2018) that focused on six samples of individuals convicted for sexual offense. Change scores showed a stronger association with sexual, violent, and general recidivism when the analyses controlled for baseline risk (static and dynamic) than that for the analyses that did not control for baseline.
Another approach for examining the relationship between intraindividual change and recidivism examines the association between reassessment scores and recidivism, again after controlling for baseline scores. Reassessment scores are scores from a second or subsequent assessment with the same dynamic risk measure; a posttreatment score is a common form of reassessment score. Several reassessment studies (Babchishin & Hanson, 2020; Brown et al., 2009; Davies et al., 2022; Hanson et al., 2021; Howard & Dixon, 2013; Jones et al., 2010; Lloyd et al., 2020; Mulvey et al., 2016) have found that reassessment scores demonstrate significant incremental predictive validity over baseline scores although with some inconsistencies. Notably, these results are statistically equivalent to tests of the incremental validity of change scores over baseline scores. A regression model testing the incremental validity of change scores over baseline dynamic scores will produce identical results to a regression model testing the incremental validity of reassessment scores over the baseline scores (Laird & Weems, 2011; Lloyd et al., 2020). This equivalence is because a model that includes a change score and a baseline score is simply a more complex way of organizing the same two pieces of statistical information: the baseline score and the reassessment score (Laird & Weems, 2011). Therefore, these findings can be interpreted in the same way as the studies using change scores and provide further empirical evidence of the significant association between intraindividual change and recidivism.
This brief review suggests an increasing number of studies show that change in dynamic risk factors is associated with recidivism. There are, however, several limitations with this research. The strongest evidence comes from prison treatment studies; Baglivio, Wolff, Piquero, et al. (2017) and Lloyd et al. (2020) have both highlighted the need for further research in community settings. Specific groups, such as juveniles, need further investigation (Barnes et al., 2016; Clarke et al., 2017). In addition, the strength of the relationship between intraindividual change and recidivism remains unclear. The van den Berg et al. (2018) meta-analysis found a small, statistically significant association between intraindividual change and sexual recidivism. Meta-analyses of the link between change and other measures of recidivism (e.g., violent and general recidivism) have not yet been published because of the lack of primary research (see Papalia et al., 2020). Thus, despite progress, further research is needed.
Is Intraindividual Change Itself Relevant to the Prediction of Recidivism?
A key question that arises from the existing research is the extent to which intraindividual change is relevant to the prediction of recidivism. It is tempting to suggest that the answer to this question is simple: Change must be relevant to prediction because research shows change is significantly associated with recidivism. That answer is partially correct; change (and reassessment) scores enhance prediction of recidivism. Predictive accuracy is likely to be poorer when prediction relies solely on a baseline assessment compared with both baseline and either a change or reassessment score. Change is also relevant at the individual level. For example, individuals assessed at the end of treatment on validated measures that show reduced dynamic risk should be less likely to recidivate than they were at the beginning of the treatment; similarly, individuals who are assessed as being at higher risk should be more likely to recidivate.
But at the same time, intraindividual change is not clearly relevant to the prediction of recidivism. We noted earlier the importance of considering the research methods. Most studies have not examined change scores in isolation. Instead, most research has controlled for baseline dynamic risk when predicting recidivism. Therefore, it is not technically accurate, for example, to conclude that individuals who make better progress are less likely to recidivate than individuals who make poorer progress. More accurately, individuals who make better progress are less likely to recidivate than individuals who make poorer progress and started with the same risk score. Because change is only relevant for prediction among individuals who start with the same risk, the latter interpretation states that individuals with higher posttreatment risk scores are more likely to recidivate than individuals with lower posttreatment risk scores. In other words, it is not the amount of change that is relevant to the prediction of recidivism but rather the risk score at reassessment, or the risk score after the period of purported change.
To further illustrate this point, it is useful to return to the scenario outlined by Baglivio, Wolff, Jackowski, and Greenwald (2017) that showed a higher risk, changing individual may remain at higher risk compared with the lower risk, unchanging individual (see also Olver et al., 2021). Previously, we stated that this scenario highlights the need to control for baseline risk when examining the association between intraindividual change and recidivism. However, the scenario also suggests that the risk score after the period of purported change is the relevant factor for the prediction of recidivism, rather than intraindividual change. In other words, the individual who makes positive change during treatment remains more likely to recidivate than the individual who makes no change, not because their pretreatment risk was higher, but because their posttreatment risk is higher.
For correctional practice, the primary implication of this interpretation is that, where optimal prediction is the goal, the risk score after the period of purported change is the most relevant variable. It is important to emphasize that, even if that is the case (i.e., current risk provides optimal prediction), information about change will still be relevant where prediction is not the goal, or not the only goal. For example, program providers will still want information about which risk factors have changed. Similarly, there will still be a need for researchers to include change scores in models they are testing. For example, the only way to establish whether intraindividual change is the mechanism by which rehabilitation programs are effective at reducing recidivism, as is assumed by the RNR model (Andrews, Bonta, & Hoge, 1990), is to test the incremental validity of change (or reassessment) scores over baseline scores. Therefore, although we are suggesting that current research may show the risk score after the period of purported change is the relevant factor for the prediction of recidivism, it is also clear that intraindividual change will continue to be relevant to practice and research in this area. For that reason, researchers should continue to test regression models that examine the incremental validity of change scores (or reassessment scores) over baseline risk.
In summary, we argue that the current prediction of recidivism using intraindividual change scores is best explained by the risk score after the period of purported change. Change is relevant to the prediction of recidivism that individuals who demonstrate risk factor reduction are less likely to recidivate than before. However, individuals who demonstrate risk factor reduction are not automatically less likely to recidivate than those who fail to do so. Rather, because studies control for baseline risk, findings simply show that those with higher dynamic risk after the period of purported change are more likely to recidivate than individuals with lower risk at the same point in time. Thus, as the research currently stands, change is relevant to predicting recidivism, but change alone is not inherently relevant. However, change may predict recidivism in a way that has not yet been explicitly examined in research. Specifically, intraindividual change may be relevant to the prediction of recidivism in addition to the risk score after the period of purported change. This idea is the focus of the rest of this article.
Is Intraindividual Change Relevant for Predicting Recidivism in Addition to the Level of Dynamic Risk After the Period of Purported Change?
To illustrate the possibility that intraindividual change leading up to a reassessment may be relevant to the prediction of recidivism in addition to the risk score at reassessment, consider the three panels of hypothetical pretreatment and posttreatment scores for two individuals presented in Figure 1. As stated previously, Panel A represents a visual illustration of the example provided by Baglivio, Wolff, Jackowski, and Greenwald (2017). Panel B illustrates the approach taken in most research to date, where baseline risk is controlled, and change (or reassessment) scores show a significant association with recidivism. In both Panel A and Panel B, the two individuals have different posttreatment risk scores. Our argument is that, as a result, they present different risks of recidivism; thus, the posttreatment score is the most relevant factor for recidivism prediction. In contrast, in Panel C, the two individuals have the same posttreatment risk score, illustrating a situation where intraindividual change may be relevant to the prediction of recidivism in addition to the risk score after the period of purported change. In other words, do the two individuals present the same risk of recidivism because of their identical posttreatment assessment, or do they present a different risk of recidivism due to differential change in treatment?
In this section, we first consider the theoretical reasons why two individuals may each present a different risk of recidivism despite having the same posttreatment or reassessment risk score and, therefore, why intraindividual change may be relevant to the prediction of recidivism in addition to the risk score at reassessment. We then consider the empirical evidence.
Theoretical Rationale
There are at least two reasons why intraindividual change may be relevant to the prediction of recidivism in addition to the risk score after the period of purported change.
Change Is Continuous
The first, and most likely, reason is that in the gap between reassessment and measurement of the outcome, further change may occur. Douglas and Skeem (2005) noted that dynamic risk assessments do not occur frequently enough to be sufficiently proximal to the outcome for optimal prediction. In other words, levels of dynamic risk factors at reassessment are likely to differ from levels when recidivism occurs. For example, an employed individual completes a rehabilitation program in the community, with employment status included in the posttreatment dynamic risk assessment. Shortly following that assessment, however, the individual loses their job. Before any further reassessment, the individual recidivates. We refer to this as unobserved change because it is not reflected in assessment scores; for the purposes of prediction, the individual would be recorded as employed prior to recidivism when, actually, they were unemployed at the time of recidivism.
Intraindividual change will be relevant to the prediction of recidivism in addition to the risk score after the period of purported change if observed change predicts unobserved change. For example, individuals may demonstrate change during a correctional rehabilitation program in prison. When those individuals are released from prison, those who made more change during treatment might be expected to continue in that direction in the community, and the individuals who made less positive change in treatment may also continue on their trajectory in the community (i.e., continue to maintain their existing dynamic risk factors). Consequently, two individuals assessed to have the same dynamic risk score at the end of treatment would present a different likelihood of recidivism depending on the earlier change they made within treatment.
To illustrate this point, in Figure 2, we present an extension of the hypothetical scenario presented in Panel C of Figure 1. Figure 2 shows two individuals who have the same posttreatment dynamic risk score, but one individual initially assessed as higher risk makes positive change in treatment, whereas the lower risk individual makes no change. Here, however, we have added a projection (the dotted lines) of the possible trajectories following their posttreatment assessment but before the measurement of recidivism. In this example, the individual who made positive progress in treatment continues to make positive progress after treatment (subsequent to the posttreatment assessment), whereas the individual who made no progress in treatment continues to make no progress after treatment.
Hypothetical Patterns of Observed and Unobserved Change of Two Individuals During and After Psychological Treatment and Before the Measurement of Recidivism
The key assumption here is that dynamic risk follows a continuous trajectory; risk scores that have measurably declined will continue to decline after measurement, or vice versa (i.e., a prior increase will precede a further increase). Continuous trajectories of dynamic risk have been briefly considered in previous research. Babchishin (2013; see also Babchishin & Hanson, 2020) and Baglivio, Wolff, Piquero, et al. (2017) used the term trajectories, but both were in the context of observed change rather than unobserved change. For example, Baglivio and colleagues compared the recidivism rates of individuals with different patterns of change across reassessments on a measure of dynamic risk during residential placement. These patterns of observed change were termed risk trajectories. Serin et al. (2013) also alluded to continuous risk trajectories in their review of the rehabilitation literature although without explicitly using the term trajectory. They argued there was a strong need to take a long-term focus when considering intraindividual change, and highlighted the potential benefits of measuring change both during and after the completion of rehabilitation programs, to determine if change was sustained (e.g., maintained or perpetuated, rather than backtracked). However, they did not discuss whether the change measured during rehabilitation programs could provide a reliable estimate of change after the program when later change was not measured.
A possible problem with the assumption of continuous trajectories of dynamic risk, highlighted by Serin et al. (2013), is that prominent theories of desistance and change (e.g., Maruna, 2001; Prochaska et al., 1992) suggest change is often not a linear progression. For example, the transtheoretical model of change by Prochaska et al. (1992), modified for conceptualizing and measuring change in the VRS and VRS-SO, proposed that individuals progress through five stages when changing their behavior: precontemplation, contemplation, preparation, action, and maintenance. Importantly, individuals were frequently expected to regress to earlier stages of change rather than simply progressing through the stages in a continuous, linear fashion. However, although change is not always linear, observed change may still predict unobserved change. A small effect might still be observed if a continuous pattern of change is most commonly found. Relevant empirical evidence is limited, but one small study of 35 prisoners sentenced to life in New Zealand (Yesberg & Polaschek, 2014) observed a continuous trajectory of change during and after intensive psychological treatment for slightly more than half of the sample, with the remainder of the sample having different patterns of change after treatment compared with those during treatment.
At least three additional factors may have an impact on whether observed change can be a marker of later unobserved change: (a) the amount of time between assessments, (b) the follow-up time, and (c) the type of variables being assessed. First, observed change may be less likely to predict unobserved change if the observed change was measured over a short period because change over a short period is arguably more likely to be temporary than change observed over a longer period. Second, observed change may be less likely to predict unobserved change when shorter follow-up periods are used because less unobserved change is likely to occur than over longer follow-up periods. Third, the relationship between observed change and unobserved change may be dependent on the type of dynamic risk factors being measured. Acute dynamic risk factors are hypothesized to change in weeks, days, or hours and signal imminent recidivism; in contrast, stable dynamic variables are slow-changing variables more akin to propensities for criminal behavior (Hanson & Harris, 2000; Hanson et al., 2007; Thornton, 2016). These differences could have several effects on the relationship between prior change and unobserved change. For example, we might expect that unobserved change would be less likely to occur over a short follow-up if the measure of dynamic risk comprises primarily stable variables than if the measure comprised primarily acute variables. Because acute variables are theoretically more volatile, observed change in acute variables might be a less reliable predictor of unobserved change than observed change in stable variables. Note, however, that the distinction between acute and stable dynamic risk factors remains mostly untested (Davies et al., in press).
Possibly the strongest current evidence for continuous risk trajectories was provided by Yesberg and Polaschek (2019), who examined whether change in dynamic risk factors during a psychological treatment program in prison predicted change in dynamic risk during re-entry into the community. Treatment change was significantly and positively associated with change during re-entry on all three measures of dynamic risk, including measures derived from stable dynamic and acute dynamic risk factors. Greater reduction in risk during treatment was significantly associated with greater reduction in risk during re-entry. There was also evidence that intraindividual change during re-entry mediated the relationship between treatment change and recidivism, but the mediation was weak and inconsistent depending on the dynamic risk measure and outcome. An earlier study by Polaschek and Dixon (2001) similarly showed individuals who made positive change in treatment continued on the same trajectory after treatment and were less likely to recidivate than individuals who made less positive change in treatment. However, that study was limited by its sample size of 33 and the use of a self-report measure of dynamic risk.
Change Scores May Reflect Unobserved Variables
A second, more speculative, reason that change might be incrementally relevant for the prediction of recidivism is because change scores may reflect unobserved variables associated with recidivism. We highlighted earlier how research has found that change scores demonstrate incremental predictive validity over baseline dynamic risk scores. Overall, however, predictive accuracy in these studies is low to moderate, consistent with most research examining the prediction of recidivism with dynamic risk factors (Bonta & Andrews, 2016). These results suggest additional variables associated with recidivism are unmeasured. Intraindividual change may be a marker of one or more of these unobserved variables. For example, a single latent variable such as a prosocial attitude toward desistance may generate improvements in several dynamic domains such as employment, interpersonal relationships, and social support. If so, once employment is gained, relationship conflict has stabilized, and new prosocial supports exist, the underlying attitude that led to observed risk reductions may generate additional prosocial behaviors that further protect against stressors and temptations. In this way, intraindividual change could be a proxy for an underlying variable such that underlying attitudes may cause both observed risk reductions and unobserved growth in personal self-efficacy for desistance (i.e., prosocial agency) and commitment to prosocial choices in the face of setbacks.
There is limited literature to support change as a marker of unobserved variables. Ward and colleagues (Ward, 2016; Ward & Beech, 2015; Ward & Fortune, 2016) argued that, although some variables included in risk assessment measures may have a causal role in behavior, most dynamic risk factors are not causal mechanisms, and most causal mechanisms remain unmeasured, hence the suggestion that observing intraindividual change in dynamic factors could be a proxy for unobserved underlying causal processes. If so, the change score needs to reflect an unobserved process not synonymous with the observed dynamic risk score after the period of purported change. Although plausible, the very nature of unobserved variables means they are difficult to test empirically; we are not aware of relevant research of this description.
Summary
Overall, there are plausible reasons why intraindividual change may be relevant to the prediction of recidivism in addition to the dynamic risk score after the period of purported change. In particular, it would seem highly likely that change related to recidivism occurs after dynamic risk assessment ceases, especially when all assessments occurred in prison. But does observed change predict unobserved change? Continuous trajectories of dynamic risk provide one possible explanation but remain largely unexamined. The relationship between observed change and unobserved change may also be affected by additional variables: time between assessments, follow-up length, and type of risk factors being measured. Intraindividual change may also be a proxy measure for other unobserved variables that are associated with recidivism, but this proposition is theoretically underdeveloped and empirically untested.
Empirical Evidence
To examine whether intraindividual change is relevant to the prediction of recidivism beyond the current risk score at reassessment, researchers need to examine intraindividual change controlling for the dynamic risk score after the period of purported change rather than examining intraindividual change and dynamic risk score before the period of purported change (e.g., pretreatment scores). In other words, they need to examine the extent to which intraindividual change demonstrates significant incremental predictive validity over reassessment. Alternatively, because models that test incremental validity of change scores over reassessment scores are statistically equivalent to those testing incremental validity of baseline scores over reassessment scores (Laird & Weems, 2011; Lloyd et al., 2020), researchers could test whether baseline scores demonstrate significant incremental predictive validity over reassessment scores.
To our knowledge, Davies et al. (in press) conducted the only study to date that has explictly explored whether relatively recent changes in risk scores can support prediction in addition to the current risk score at reassessment. They found evidence from a series of Cox regression models that change over time periods of 2 to 6 weeks on a composite measure of theoretically acute dynamic risk factors was significantly associated with recidivism after controlling for the most-recent assessment score. In contrast, on a measure of theoretically stable dynamic risk factors, change scores over the same short time periods were not significantly associated with recidivism after controlling for the most recent assessment score but only after controlling for baseline scores.
Outside of Davies et al. (in press), identifying relevant empirical evidence requires re-examination of studies that aimed to examine different research questions. For example, several studies examined recidivism rates of individuals reassessed into a different dynamic risk category (e.g., high, then medium, showing improvement compared with others at high risk), but the same results can identify recidivism rates of individuals similar at reassessment but differed at baseline (e.g., two individuals at medium risk, one who was previously at high risk). Re-examining recidivism rates in this way, high-risk individuals initially assessed at lower risk had higher recidivism rates than other high-risk individuals, suggesting that change—in this case, increasing risk scores—was relevant in addition to current risk (Cohen & VanBenschoten, 2014). In contrast, results from the study of Cohen et al. (2016) showed no difference in recidivism rates of individuals in the same reassessment category following different initial categories. Results from Vose et al. (2013) add to the mixed picture, showing differing recidivism rates within higher and lower risk categories at reassessment based on prior change, but not within the moderate category. Key limitations are the categorical rather than continuous measurement of risk, substantially varied sample size within risk categories, and our consideration of these data differs from these studies’ original focus.
Stronger evidence can be gleaned from studies that examined whether reassessment enhances recidivism prediction. For example, Lloyd et al. (2020) entered dynamic risk scores from reassessment in regression models predicting recidivism alongside baseline scores, finding only reassessment scores significantly predicted recidivism. These findings indicated reassessment improved prediction but also showed that change was not relevant for prediction after accounting for risk score at reassessment. Davies et al. (2022) replicated these findings using a different sample. However, using a similar approach, Mulvey et al. (2016) found their reassessment scores showed incremental predictive validity over baseline scores for all outcomes and that baseline scores showed incremental validity over the reassessment scores for self-reported crimes (but not re-arrest). Therefore, prior change sometimes incrementally predicted beyond current, reassessed score.
Finally, the publicly available online calculator (Mundt, 2015) for the VRS-SO can be used to obtain recidivism estimates for different combinations of change and posttreatment VRS-SO scores. These estimates use the regression models reported by Olver et al. (2018) and derived from the results of several previous VRS-SO studies (Beggs & Grace, 2010; Olver et al., 2007, 2014; Sowden & Olver, 2017). The calculator does not provide the option to enter posttreatment scores, but a user can implicitly input posttreatment scores by subtracting change scores from pretreatment scores. In Figure 3, we present the relationship of change scores with 10-year sexual recidivism estimates for (a) VRS-SO pretreatment and (b) VRS-SO posttreatment scores. Figure 3 is divided into two panels that show the same data from two different perspectives (e.g., a pretreatment score of 56 and a change score of 6 produce the same recidivism estimate as a posttreatment score of 50 and a change score of 6: 43.7%). Panel A, reproduced from the online calculator when used as intended, shows the estimated recidivism rates for the full range of pretreatment scores for individuals with low (change score of 0), moderate (change score of 3), and high (change score of 6) change scores. Panel A illustrates the finding reported by Olver et al. (2018); controlling for pretreatment scores, VRS-SO change scores are significantly related to likelihood of sexual recidivism, as shown by the clear separation of the three lines. In Panel B, we present the estimated recidivism rates for the full range of posttreatment scores across different change scores. Olver et al. (2018) presented similar analyses but only compared three possible posttreatment scores (see Table 5, p. 952), whereas we present recidivism estimates for all possible posttreatment scores. Similar to Panel A, there is a division between the three lines, indicating that change scores may be relevant for predicting recidivism after accounting for posttreatment scores. Among individuals with the same posttreatment score, those assessed to have made less change during treatment have higher estimated recidivism rates than those assessed to have made more change, consistent with the continuity of change theory outlined earlier.
Estimated 10-Year Sexual Recidivism Rates for VRS-SO Pretreatment and Posttreatment Total Scores With Low, Moderate, and High Change Scores
Overall, there is some empirical evidence that intraindividual change adds to the prediction of recidivism beyond the risk score after the period of purported change. Studies examining the relationship between intraindividual change and recidivism provide clearer evidence that change is associated with recidivism after accounting for baseline risk, whereas it is more equivocal whether change adds beyond current risk. In summary, some studies showed change added incrementally to current scores (Cohen & VanBenschoten, 2014; Mulvey et al., 2016; Olver et al., 2018), some showed no unique association of change with recidivism (Cohen et al., 2016; Davies et al., 2022; Lloyd et al., 2020), and others were mixed depending on risk category (Vose et al., 2013). However, because these researchers conceptualized nonidentical research questions, none used inferential statistics to explicitly test incremental prediction of change. The one study that directly examined this question (Davies et al., in press) found change was relevant to recidivism prediction beyond current score but only for acute yet not stable dynamic risk. Clearly, replication and further research explicitly examining this question is needed before drawing firmer conclusions, including about the strength and direction of any relationship.
Potential Implications for Practice and Research
With such limited empirical research to support the conclusion that intraindividual change improves recidivism prediction beyond risk scores at reassessment, discussion of potential practical implications is somewhat presumptuous. Nonetheless, we briefly highlight some potential impacts for practice and outline how more research can provide further clarification.
If it is established that intraindividual change is relevant to predicting recidivism in addition to current risk, practitioners may need to alter both how they communicate risk and how they respond to changes in risk. For example, risk communication may have in-built assumptions about whether change observed in treatment (or in another context) is expected to continue after treatment. Specifically, communicators may assume that because individuals made positive change in treatment, they are expected to make additional positive change after treatment, engaging in projection (appropriately or inappropriately) to assume that the current risk score is an upper estimate of future risk that is on a downward continuous trajectory. If so, we think this assumption should be more explicitly communicated to decision-makers alongside a discussion of any potential barriers that might prevent the assumed further change. Furthermore, more explicit research that specifies the unique contribution of change beyond current risk would better allow practitioners to communicate more precise estimates about postprogram prognosis because of intraprogram observations. This would avoid the general and imprecise assumption that observed change is expected to progress in the same way into the future without qualifying its degree of likelihood.
In terms of responding to risk, finding that change adds to prediction above current risk would require a practitioner additionally consider the developmental process of risk reduction beyond the focus on reducing risk factors. The specific components of change (e.g., rate of change, linear or nonlinear shape of change, breadth of change across risk domains, etc.) may contain their own unique information about risk and future prognosis. This broadening of consideration would affect many areas of correctional practice, particularly community correctional practice. Community correctional staff must contend with individual and surrounding circumstances that can change quickly and often. In that environment, it would be beneficial to know if a pattern of change observed on a dynamic risk measure should inform how an individual is managed, perhaps signaling the need for an immediate intervention to reduce the likelihood of recidivism. For example, should an individual with consecutive weeks of minor elevations in dynamic risk factors who otherwise remains at low risk continue to be managed in a typical “hands off” way, or do recent changes suggesting steady escalation indicate they should be managed more actively?
From a research perspective, as we have already noted, the priority is further empirical research explicitly examining the different components of the relationship between intraindividual change and recidivism. We recommend a two-step approach: (a) establish that intraindividual change is associated with recidivism after controlling for pretreatment or baseline risk and (b) investigate the extent to which change is associated with recidivism after controlling for posttreatment or current risk. Researchers should pay careful attention to the amount of time between assessments, the follow-up time, the type of risk factors being assessed, and the theory of change embedded within each tool, all of which could affect the nature of the relationship observed. Other methodological factors should be explored, including whether change (or absence of change) in RNR-informed interventions has a stronger unique relationship to recidivism than change in other contexts and whether that relationship is further moderated by the setting in which the intervention takes place (i.e., in an institution or the community). Community-based interventions, including community supervision, are well established as falling within the definition of RNR-informed interventions (Andrews, Zinger, et al., 1990), and when researchers have explicitly compared efficacy across settings, community interventions demonstrate stronger effects on recidivism than institution-based interventions (Andrews, Zinger, et al., 1990; Bonta & Andrews, 2016). Thus, identifying relevant similarities and differences in the relationship between change and recidivism across these settings would be valuable.
Summary and Conclusions
The relationship between intraindividual change and recidivism is central to correctional psychology and many aspects of correctional practice. Consequently, researchers and practitioners need a clear understanding of the nature of this relationship. Our review of the empirical research indicates a growing body of evidence that supports a significant association between intraindividual change and recidivism. This finding has implications for several areas of practice. Most important, this finding supports the assumption that intraindividual change across key risk variables is the mechanism by which rehabilitation programs reduce recidivism. Further research in this area is still needed to enhance the understanding of rehabilitation efficacy toward the development of even more effective programs.
At the same time, in our view, current research has not yet established that change information is the best way to specify multiple-assessment information when predicting recidivism. There is strong evidence that change information alone should not be used for prediction but must account for baseline risk. In this way of thinking, the most relevant factor for recidivism prediction is the resulting risk score after the period of purported change (i.e., initial risk plus change). The relevance of change, then, is primarily an indicator of postchange risk.
But could intraindividual change be relevant to recidivism prediction in addition to dynamic risk after the period of purported change? We have dedicated this article to exploring this question. We highlighted there is a promising theoretical rationale that is underdeveloped and, therefore, underresearched. There is some implicit evidence from previous research (e.g., Cohen & VanBenschoten, 2014; Mulvey et al., 2016; Olver et al., 2018; Vose et al., 2013), but further explicit tests are needed (see Davies et al., in press). There are several potential practice implications if intraindividual change is uniquely relevant to predicting recidivism, but these practices require further evidence before adoption. For now, we believe the current evidence guides practice to focus on the risk score after the period of purported change as the most important information for recidivism prediction, and we caution overinterpreting change scores to indicate a process that is complete, rather than ongoing and evolving in many potential directions. With further research, though, that position may need to be revised.
Footnotes
AUTHORS’ NOTE:
Aspects of the conceptual discussion in this article were previously presented in Simon T. Davies’s doctoral thesis at Victoria University of Wellington (see Davies, 2019).
