Predicting Recidivism Among Released Juvenile Offenders in Florida

Abstract

Each year in the United States, as many as 100,000 juvenile offenders are released after completing a residential placement. A significant task for researchers is to identify the factors that explain variations in recidivism. This study considers this by evaluating the predictive validity of the Residential Positive Achievement Change Tool (R-PACT), a fourth-generation risk assessment instrument adopted by Florida for use in all of its juvenile residential facilities. The R-PACT includes a wide variety of static and dynamic risk and needs scales that are used here to predict reoffending among 4,700 released juvenile offenders in Florida. We devote special attention to (1) whether R-PACT scales typically predict reoffending and (2) whether the R-PACT’s predictive validity varies across different subgroups of offenders. In considering these questions, we also consider whether the predictive risk and protective factors in prior research are predictive in the R-PACT as well. The analysis revealed relatively strong support for the R-PACT, but there were nuanced exceptions to that pattern. We discuss the implications these findings have for assessing risk, monitoring progress among residential youth, and predicting reoffending.

Keywords

juvenile reentry risk assessment Residential Positive Achievement Change Tool juvenile justice

The U.S. approach to juvenile crime includes a heavy reliance on incarceration, with roughly 60,000 juvenile offenders confined in a criminal or juvenile justice facility each day (Hockenberry, 2014). A key implication of this pattern involves the daunting challenge of youth reentry (Mears & Travis, 2004). Specifically, because most juvenile residential placements involve short stays of 1 year or less (Hockenberry, 2014), nearly every residential confinement produces a reentry case not long thereafter. Thus, as many as 100,000 juveniles are returned to their communities each year after completing a residential confinement (Snyder, 2004).

Much research has examined how these offenders fare upon release. Although recidivism estimates vary across states and different studies, the general pattern for released youth is that within 2 years, roughly 70% are rearrested for a new offense, 50% receive an adjudication or conviction, and 20% return to a correctional institution (Annie E. Casey Foundation, 2011; Krisberg, 2011; Trulson, Haerle, DeLisi, & Marquart, 2011). Such high recidivism contributes to the perception that there is a revolving door of juvenile justice in which young offenders often reenter the system not long after exiting it (Benda & Tollett, 1999). However, these recidivism rates also reveal genuine variation—although many released juvenile offenders persist with crime and reenter the justice system, many do not. In short, real instances of desistance occur among juvenile offenders and that pattern is confirmed in ethnographic and self-report research that goes beyond official measures of recidivism (Basto-Pereira, Começanha, Ribeiro, & Maia, 2015; Panuccio, Christian, Martinez, & Sullivan, 2012; Steinberg, Cauffman, & Monahan, 2015).

A task for researchers is to explain these variations. Many studies reveal that static background characteristics—especially prior offending history and membership in high-risk demographic subgroups—consistently predict recidivism (Lattimore, Macdonald, Piquero, Linster, & Visher, 2004; Trulson, DeLisi, & Marquart, 2011). Recent research and risk assessment tools also prioritize dynamic attributes—values, social commitments, and interpersonal skills that are shaped by new experiences and that may change during the residential stay (Labrecque, Smith, Lovins, & Latessa, 2014; McGrath & Thompson, 2012). Considering dynamic attributes builds on the idea that juvenile offenders are at a point in life when biological, psychological, and social changes are common, perhaps especially for those with justice system involvement. With that in mind, state-of-the-art risk assessment tools now track static and dynamic risk factors over time to inform service plans and treatment delivery. Prominent such tools include the Youth Level of Service/Case Management Inventory (YLS/CMI), the Structured Assessment of Violence Risk in Youth (SAVRY), and the Psychopathy Checklist (PCL), all of which have received relatively strong support in validation research (Fass, Heilbrun, Dematteo, & Fretz, 2008; Olver, Stockdale, & Wormith, 2009).

It bears emphasizing, however, that other risk assessment tools are prominently used and also require rigorous validation. This includes the Residential Positive Achievement Change Tool (R-PACT), which is the focus of the present study. The R-PACT is a substantively important tool—the state of Florida now uses it in all juvenile residential facilities, which house roughly 3,700 juvenile offenders on a given day (Hockenberry, 2014). The R-PACT is administered multiple times during the residential stay, and data are collected not just on the static risk factors prioritized in prior research but also on dynamic risks involving such things as social relationships, academic and work performance, attitudes, and social skills. Its statewide adoption is a recent development, however, and research is needed to evaluate its predictive validity.

This study provides a needed validation of the R-PACT. We examine its ability to predict reoffending in a sample of 4,700 released juvenile offenders in Florida. The analysis focuses on two questions. First, do R-PACT domain scales significantly predict reoffending? In considering this issue, we especially attend to whether significant effects of dynamic domains are maintained even when accounting for key static background factors. Second, is the R-PACT’s predictive validity similar across different subgroups of offenders, including those that vary in terms of age, sex, race and ethnicity, and prior offending? In considering these questions, this study can reveal whether the R-PACT is accomplishing the risk assessment goals envisioned when it was adopted system-wide by such a large state juvenile justice system. Such research is relevant not just to Florida but also to any jurisdiction that will be adopting similar assessment tools, and it can inform the broader study of juvenile reentry by providing evidence on which risk factors are most consequential. As Taxman and Caudy (2015) have argued, and as we describe below, there still is uncertainty on this issue.

Prior to examining these issues, we first describe prior efforts to predict and understand juvenile recidivism, highlighting the distinguishing factors of recent approaches. We then describe the unique history of Florida, emphasizing its evolving approach to juvenile justice and risk assessment. That evolution has produced in recent years a systematic statewide commitment to assessing the risks and needs of juvenile offenders and better understanding how these things affect recidivism.

Predicting Reoffending

Efforts to predict juvenile recidivism have been undertaken for decades, but the methods used have changed over time. Andrews, Bonta, and Wormith (2006) described this in identifying four generations of approaches, the first of which relied on unstructured professional judgments about a youth’s odds of reoffending. This “gut instinct” approach ultimately gave way to actuarial second-generation approaches that became quite common in the 1980s and 1990s. Second-generation tools rely on empirically based inventories that especially emphasize static risk factors like prior offending history, demographic affiliations, and indicators of a troubled social history. As Andrews and his colleagues (2006) note, this shift was critical—the ability of second-generation tools to predict reoffending is substantially greater than that observed with first-generation tools.

In more recent years, third- and fourth-generation assessment tools have become prominent. Their most distinctive qualities are (1) attention to a wide array of criminogenic risks and needs and (2) a dynamic emphasis that more fully considers offender change. Thus, whereas second-generation tools were mostly concerned with static aspects of an offender’s prior history, third- and fourth-generation tools more fully consider an “offender’s current and ever-changing situation” (Bonta & Andrews, 2007, p. 4), often in reference to the “central eight” areas of risk and need thought to best predict reoffending (Andrews, Bonta, & Wormith, 2006). This list includes risks in the areas of prior offending, antisocial personality, antisocial beliefs, and antisocial peers, along with needs in the areas of family, leisure and recreation, substance use, and employment/education. Fourth-generation tools are noteworthy not just for collecting data on these risks and needs but using that data to inform treatment delivery and assess offender progress (Latessa & Lovins, 2010). In this way, these tools promote adherence to the risk-needs-responsivity (RNR) model in which treatment targets high-risk offenders to address specific needs with evidence-based programming appropriate to offender circumstances.

Risk assessment research now revolves around the leading fourth-generation tools, the most prominent of which is the YLS/CMI. This clinician-rated risk and needs measure is focused especially on the central eight predictors, and it has been adopted for statewide use (at the probation stage) by many states, including Pennsylvania and New York (Wachter, 2015). Validation studies reveal weighted effect sizes for the total YLS/CMI score that often are 0.20–0.30 but sometimes as high as 0.60 (Olver et al., 2009; Schwalbe, 2007). This same research indicates that it has predictive validity for both general and violent reoffending and across samples that differ in sex, race or ethnicity, and national context. Favorable results also have been obtained for other tools, including a youth-adapted version (PCL-YV) of the PCL and the SAVRY (Hilterman, Nieuwenhuizen, & Nicholls, 2014; Olver et al., 2009; Vincent, Chapman, & Cook, 2011).

This research points to the progress in predicting juvenile recidivism, but key issues remain unresolved. Most notably, there are important risk assessment tools that still have received minimal empirical attention in the published literature (see Baird et al., 2013). This includes the Comprehensive Risk and Needs Assessment (a youth-adapted version of the adult COMPAS tool) used by Georgia, the Dynamic Risk Instrument used in Arizona, and the R-PACT used statewide in Florida. This diversity of assessment tools reflects the highly decentralized approach to juvenile justice in the United States, with each state confronting risk assessment somewhat independently (Wachter, 2015). This decentralization complicates the task of assessing risk and predicting recidivism—there are many tools to assess, and although they often originate from a common research literature, differences across them may be relevant to predictive accuracy. Indeed, as a number of comparative risk assessment studies indicate, validation of one tool cannot be seen as validation of others (Drake, 2014; Fass et al., 2008; Olver et al., 2009).

Beyond this need to validate prominently used tools, a second unresolved issue involves the consequentiality of different risk factors. As noted earlier, efforts to predict juvenile recidivism have focused special attention on the central eight risk factors identified by Andrews et al. (2006). However, research that examines these risk factors separately often indicates varied effects among them (Hilterman et al., 2014; Olver, Stockdale, & Wormith, 2014). Prior offending/antisociality, antisocial personality, and needs in the area of education/employment often have the greatest effects; conversely, domains reflecting family needs and problems in the area of leisure/recreation often have smaller effects. Perhaps these patterns will be confirmed in future studies, but it bears emphasizing that meta-analyses on this issue often include many studies that use adult samples. If the focus is specifically on juveniles, different results may emerge regarding the relative potency of the different risk and need areas. Taxman and Caudy (2015) point to just such a pattern in noting the limited support for some central eight criminogenic needs. They highlight the notable “variation in the impact of these known dynamic risk factors across studies, samples, and measurement strategies” (p. 74). They emphasize, however, that the variations are difficult to interpret, given that few studies have assessed the effects of dynamic criminogenic needs while simultaneously controlling for other factors, including indicators of static risk.

Thus, the question of which criminogenic risks and needs are most consequential—and therefore, which are the most fruitful targets for intervention—remains largely unanswered. With that in mind, new evaluations of prominent assessment tools should attend not just to how well these tools perform but also to which risk factors appear to be most important.

Juvenile Risk Assessment in Florida

The research cited above points to the growing priority placed on sophisticated juvenile risk assessment. Such priority can be found in Florida also, with this emerging most notably in the past 15–20 years. The Florida Department of Juvenile Justice (FDJJ) was not created until 1994,¹ and even then, punishment and incapacitation were chief among its priorities (Frazier, Bishop, & Kaduce, 1999). In this way, Florida was consistent with the national context of the 1990s that saw heightened fears over a growing wave of violent “super-predators” (DiIulio, 1995, p. 23). In response, Florida increased juvenile admissions into high security residential facilities and continued to waive a high number of cases to punitive adult courts (Frazier et al., 1999).

Many factors converged, however, to produce a shift in Florida in the early 2000s. Most notably, juvenile crime was down (both nationally and in the state), thus dampening the most pro-punishment tendencies. Also, Florida was marred by high profile abuses of youth in state juvenile justice facilities (DeMuro, 2008). These cases brought dramatic attention to the limitations of the get-tough movement and spurred support for an alternative approach. Indeed, key stakeholders (including elected officials, influential secretaries of the FDJJ, and child advocacy organizations) actively sought to replace the “tough” approach of the 1990s with a “smart” approach that would reduce crime by more fully investing in Florida youth. These goals were communicated directly in two pivotal state reports: the 2008 Report of the Blueprint Commission (a group commissioned by then Republican Governor Charlie Crist) and the 2013 Roadmap to System Excellence (issued by FDJJ Secretary Wansley Walters). Both reports embraced an evidence-based approach emphasizing community prevention, a wider continuum of sanctions and services (to reduce reliance on institutionalization), and adherence to the RNR model for effective intervention.

FDJJ was well positioned to implement these priorities—in Florida’s highly centralized system, it is responsible for everything from probation to residential aftercare—and risk assessment was central to its efforts. Beginning in 2006, a 46-item prescreen version of the PACT was being used at intake for every youth referred to the system. Along with autopopulated data on prior criminal history, the PACT includes static and dynamic risks (covering most of the central eight) assessed by probation officers. PACT scores classify youth into risk groups ranging from low to high risk to reoffend, and these classifications guide decisions about dispositions and treatments. Early validation research confirmed the PACT’s ability to predict reoffending (Baglivio, 2009). However, the PACT’s principal limitation (similar to that for many other risk assessment tools) was that it was not designed to track the risks and needs of residentially placed youth in particular. This is critical, given the unique circumstances of a residential placement, including extensive contact between staff and youth and the wide array of services and training that take place in the residential facility. FDJJ therefore worked with a proprietary vendor to create an adapted PACT tool; it became known as the R-PACT to indicate its use for residential youth, leading the PACT to take on the new label of “C-PACT” to denote its front-end use with “community” youth.

By 2009, all youth residential facilities in the state had received R-PACT training and were fully administering it to residential youth. The first R-PACT assessment occurs within 30 days of admission, with new assessments done every 90 days thereafter; also, an exit R-PACT is conducted prior to release. All administrations involve a semistructured interview led by the case manager, who receives feedback from residential staff, including representatives from the clinical, nursing, direct care, administration, and education departments. As we discuss below, the R-PACT’s over 150 items are organized into domain scales that cover the central eight risk factors, plus additional areas of interest to FDJJ. Youth scores on these scales are used to create treatment plans and assess progress on risk reduction and protective factor enhancement during placement. If progress is not observed, then treatment plans are adjusted by revising the intensity of services or offering different ones.

A central question, however, is whether the R-PACT scales successfully predict future offending. To date, there is limited information on this, although the research is promising. For example, Baglivio, Wolff, Jackowski, and Greenwald (2015) found that changes in R-PACT risk during the residential stay predicted recidivism in the 12 months after release. They also observed partial evidence that increases in risk factors during the residential stay negated the normally beneficial effects on reoffending of being released to an affluent community. Baglivio et al. (2016) also used R-PACT data to reveal that exposure to traumatic experiences in childhood increases reoffending among youth released from a residential placement. As useful as this research is, however, it remains true that a full-scale validation of the R-PACT has yet to appear in the published literature.

The Present Study

This study examines the full scope of the R-PACT to determine its predictive validity. Specifically, 48 of its domain scales are used to predict reoffending in the 12 months after release from a residential program. As noted above, although prior research confirms moderate to strong validity of other fourth-generation risk assessment tools, similar validation is needed for the R-PACT, especially given its statewide adoption in Florida, a state that houses roughly 3,700 juvenile offenders on a given day (Hockenberry, 2014). Such a validation can give insight into how useful the R-PACT is for identifying (1) the offenders most at risk for reoffending and (2) the risks and needs that should be most prioritized in treatment plans.

The analysis focuses on two specific research questions. First, do the R-PACT’s 48 scales—14 static and 34 dynamic—significantly predict reoffending? We first consider this on a bivariate basis but then estimate multivariate models that examine effects of the dynamic scales when controlling for key static predictors, including prior offending. This provides insight on the value-added that comes from the R-PACT’s extensive attention to dynamic qualities that may change during the residential placement. If such attention is warranted, dynamically assessing risk should produce improvements in recidivism prediction that go beyond that available with the established static predictors. Our second question involves the generality of the R-PACT’s scales—does their ability to predict reoffending vary across key subgroups of offenders? We consider this by examining predictive validity across youth who differ in sex, race and ethnicity, age at release, and prior offending.

Data and Sample

Data were drawn from the FDJJ archival data records. The FDJJ maintains a centralized database, the Juvenile Justice Information System, that contains complete information on a youth’s offending, placement, and risk assessment history. This includes results from R-PACT administrations. Data were drawn for all youth who completed a residential placement in the period from July 1, 2010, to June 30, 2011. FDJJ records revealed 5,162 such youths. Of these, 4,736 were administered at least one R-PACT during the residential stay, and 4,700 of these youth were released upon program completion to an address in the state of Florida. These 4,700 youth were the analytic sample for this study (see Table 1 for descriptive statistics). The average youth had spent approximately 250 days in the residential program before being released at a mean age of 17 years and 3 months old. Most subjects (85%) are male, and the sample is marked by racial and ethnic diversity—50% of youth are Black, 36% are non-Hispanic White, and 11% are Hispanic. With respect to prior offending, the modal youth had prior referrals for two felonies and two misdemeanors.

Table 1.

Descriptive Statistics for Reoffending and Demographic Controls.

Variables	Observation	Mean	Min	Max
Reoffending	4,700	0.40	0	1
Demographic controls
Male	4,700	0.85	0	1
White	4,700	0.36	0	1
Black	4,700	0.50	0	1
Hispanic	4,700	0.11	0	1
Other	4,700	0.04	0	1
Age at release	4,700	17.22	10.83	20.91

Measures

Reoffending

Reoffending was assessed for a 12-month period following each youth’s release date. Because some youth turned 18 years of age during this period, both juvenile and adult arrest records were used. Our reoffending variable is a dichotomous measure of convictions in which youth were coded as 1 if they experienced either a juvenile court adjudication (including an adjudication withheld) or an adult conviction for a new law violation that occurred during the year after release; 40% of cases were reoffenders for this measure.

R-PACT domain scales

As noted above, the R-PACT is administered by case managers who lead semistructured interviews with youth and solicit information from staff involved in all aspects of the residential program. This is done for over 150 items relating to 11 domain areas of risk and need: prior offending history, school/vocational training, use of free time, employment, peer relationships, the family, substance use, mental health, antisocial and prosocial attitudes, involvement in aggression, and the use of social skills for such things as controlling impulses and dealing with difficult situations.² Each domain area is represented with multiple scales. This occurs in part because some domains include distinct subareas that are given their own scale but also because most domains include both a risk scale (for factors that should increase offending) and a protective scale (for factors that should decrease offending). Also, many domains include both a static scale to measure youth’s prior history and a dynamic scale to measure youth’s current circumstances. For many domains, the dynamic scales include items in which case managers explicitly assess youth progress on key goals and overall responsiveness to treatment and services. Taken together, there are 48 automatically generated domain scales that we evaluate.³

These 48 scales are shown in Table 2, which distinguishes the 14 static scales from the 34 dynamic ones, specifies the domain areas that have both risk and protective scores, and provides illustrative items for each domain area. (Also, we identify the domain number [e.g., “3A”] to provide further clarity to those who might use R-PACT data in future research). Although the R-PACT’s full list of items is available at the FDJJ’s website (http://www.djj.state.fl.us), we do not provide the exact scoring here because specifications regarding which items are scored and how they are done so is proprietary to Assessments.com, the vendor who worked with FDJJ to create the R-PACT. The only adjustment we made to the original system-generated scales was to standardize them such that all scales have a mean of 0 and a standard deviation (SD) of 1. This promotes a clearer interpretation of the odds ratios we present from logistic regression equations. Those odds ratios reveal changes in the odds of reoffending that come from a one-unit change in the predictor; with standardized scales, this one-unit change reflects a 1 SD unit change in the variable.

Table 2.

Domain Scales in the R-PACT.

Static domain scales (14)		Illustrative items
1A	Prior offending history: R	Age of first offense, felonies, misdemeanors
3A	School history: R and P	Expulsions, dropped out vs. graduated
4A	Historic use of free time: P	Structured and unstructured prosocial activities
5A	Employment history: R and P	Has been employed, positive relationships
6A	Relationship history: R and P	Positive adults in life, pro/antisocial friends,
7A	Family history: R and P	Running away, family criminal history
8A	Alcohol/drug history: R and P	Past use/referrals for alcohol/drugs
9A	Mental health history: R and P	Prior diagnosis, medication, and treatment
Dynamic domain scales (34)
3C	Current school status: R and P	GPA, conduct in school, values education
3D	Current vocational training: R and P	Values training, works with instructors
4B	Current use of free time: R and P	Positive or negative use of unstructured time
5B	Employability: R and P	Employment aspirations, acquired skills
5D	Program supervised tasks: R and P	Performance of tasks, level of cooperation
6B	Current relationships: R and P	Pro/antisocial friends, resists negative peers
7B	Current family relationships: R and P	Visits, conflicts, caretaker accepts antisociality
8B	Current alcohol/drug use: R and P	Current use, progress on substance use goals
9B	Current mental health: R and P	Current complications, progress on goals
10	Current attitudes: R and P	Respect for authority, believes in rules/laws
11	Current aggression: R and P	Temper, verbal/physical aggression
12	Current skills: R and P	Problem solving, consequential thinking, self-monitoring
12A	Dealing with others: R and P	Listens, asks for help, compliments, apologizes
12B	Dealing with situations: R and P	Dealing with embarrassment, accusations
12C	Dealing with emotions: R and P	Knows, expresses, and understands feelings
12D	Controlling impulses: R and P	Uses reframing, replacing, and relaxation
12E	Controlling aggression: R and P	Asks permission, uses self-control, negotiates

Note. R indicates a risk scale and P indicates a protective scale. GPA = grade point average; R-PACT = Residential Positive Achievement Change Tool.

One other point bears emphasizing. Because the R-PACT is administered at multiple points in time (an initial R-PACT at the time of admission, an exit R-PACT near the time of release, and every 90 days in between), youth have multiple R-PACT scores for each scale. All analyses below are for the exit R-PACT administration that occurred near the time of release. These exit data provide the assessment that is most proximate to the recidivism follow-up period. Also, our preliminary analyses with these data revealed, as expected, that when there were differences in prediction between the initial and exit scales, this involved stronger effects for the exit R-PACT.

Demographic controls

All multivariate analyses include controls for the youth’s age, sex, and race–ethnicity. Age was measured in years for the time at release, while sex is coded 1 for males and 0 for females. Race and ethnicity were measured with dummy variables for the following categories: Black, non-Hispanic White, Hispanic, and other (with non-Hispanic Whites serving as the reference category in multivariate models).

Results

Bivariate Results

The analysis began by considering the bivariate relationship between the R-PACT scales and reoffending. This provides a first glimpse into the empirical adequacy of the R-PACT. Logistic regression equations were estimated that took reoffending as the dependent variable and included the domain scale in question as the sole independent variable. Table 3 reports the odds ratios from equations that were estimated for each domain scale. Focusing first on the static domains, the top panel of Table 3 reveals that 11 of the 14 static scales have no significant effect on reoffending. Also, those with significant effects are all risk rather than protective scales, with reoffending increased by risks involving prior offending, school history, and relationships (this latter scale relates to nonfamily relationships with peers and adults). For prior offending, for example, a 1 SD unit increase in that variable is associated with a 33% increase in the odds of reoffending (e.g., [1.33–1] × 100 = 33); with school and relationships, a high-risk history increases the odds of reoffending by 14% and 16%. These patterns notwithstanding, the more general theme is that static history variables are not especially predictive of reoffending.

Table 3.

Bivariate Models Predicting Reoffending From R-PACT Static and Dynamic Scales.

		Risk Scales		Protective Scales
		b	SE	b	SE
Static scales
1A	Record of referrals	1.33***	0.04
3A	School history	1.14***	0.03	0.97	0.03
4A	Historic use of free time			1.00	0.03
5A	Employment history	0.95	0.07	1.04	0.07
6A	History of relationships	1.16***	0.04	1.02	0.03
7A	Family history	1.00	0.03	0.99	0.03
8A	Alcohol/drug history	0.97	0.03	0.97	0.03
9A	Mental health history	0.97	0.03	1.06	0.03
Dynamic scales
3C	Current school status	1.17***	0.04	0.83***	0.03
3D	Current vocational training	1.14*	0.08	0.85*	0.06
4B	Current use of free time	1.04	0.03	0.91**	0.03
5B	Employability	1.14***	0.03	0.85***	0.03
5D	Program supervised tasks	1.02	0.03	0.89***	0.03
6B	Current relationships	1.15***	0.03	0.87***	0.03
7B	Parent/caregiver relationships	1.03	0.03	0.94*	0.03
8B	Current alcohol/drug use	1.10***	0.03	0.92**	0.03
9B	Current mental health	1.05	0.03	0.98	0.03
10	Current attitudes	1.17***	0.03	0.84***	0.02
11	Current aggression	1.17***	0.03	0.87***	0.03
12	Current skills	1.10***	0.03	0.84***	0.03
12A	Dealing with others	1.09**	0.03	0.84***	0.02
12B	Dealing with situations	1.08**	0.03	0.86***	0.03
12C	Dealing with emotions	1.09**	0.03	0.85***	0.03
12D	Controlling impulses	1.06*	0.03	0.85***	0.03
12E	Controlling aggression	1.07*	0.03	0.85***	0.03

Note. R-PACT = Residential Positive Achievement Change Tool.

*p < .05. **p < .01. ***p < .001 (two-tailed).

The bottom panel of Table 3 focuses on the dynamic variables. The odds ratios for these equations point to the predictive value of the dynamic domains—29 of the 34 dynamic scales are significantly related to reoffending. For the risk domains, the highest odds ratios (1.14–1.17) are for school status, employability, relationships, antisocial attitudes, and involvement in aggression. A similar pattern is observed for the protective domains. The five domains just noted each have odds ratios of 0.87 or below (indicating at least a 13% reduction in the odds of reoffending) for their protective scales. Also, protective scales for the six social skills domains have odds ratios in the 0.84–0.87 range. Taken together, these results indicate that for the dynamic domains, being 1 SD above the mean on risk and protective scales often alters the odds of reoffending by roughly 15%.

Multivariate Results for the Dynamic Domains

The next step was to determine whether significant effects of the dynamic scales are maintained even when accounting for key static background variables. This was done for each of the 34 dynamic scales (17 risk and 17 protective). A logistic regression equation was estimated for each scale that took reoffending as the dependent variable and included as independent variables the dynamic scale in question, demographic controls (for sex, race–ethnicity, and age at release), and controls for the three static variables earlier revealed as significant predictors of reoffending (prior offending history, school history, and relationship history).

Below, we provide full equation results (for controls also) for two illustrative scales, but Table 4 first provides the odds ratios that emerged from each scale’s equation. The most notable pattern is that these dynamic scales often predict reoffending even after accounting for the effects of the demographic controls and static predictors. This is true for 7 of the 17 dynamic risk scales, with the highest odds ratios ranging from 1.10 to 1.13, therefore indicating a 10–13% increase in the odds of reoffending for those 1 SD above the mean in risk. Effects of this size were observed for current school status, relationships, current attitudes, and current aggression. For the protective scales, there was a similar but more consistent pattern—14 of the 17 were significantly and negatively related to reoffending. Current attitudes and current social skills were the most consequential variables, with odds ratios of 0.87 indicating 13% reductions in reoffending for those 1 SD above the mean on these measures. A number of other variables produced reductions of 10–12%. This included all of the social skills variables and two measures that capture youth’s commitment to work/education (current school status and performance of program supervised tasks).⁴

Table 4.

Multivariate Models Predicting Reoffending From R-PACT Dynamic Scales.

		Risk Scales		Protective Scales
		b	SE	b	SE
Dynamic scales
3C	Current school status	1.10**	0.04	0.88***	0.03
3D	Current vocational training	1.08	0.08	0.90	0.07
4B	Current use of free time	1.03	0.03	0.94*	0.03
5B	Employability	1.08*	0.03	0.93*	0.03
5D	Program supervised tasks	1.02	0.03	0.90***	0.03
6B	Current relationships	1.12***	0.04	0.91**	0.03
7B	Parent/caregiver relationships	0.99	0.03	0.95	0.03
8B	Current alcohol/drug use	1.07*	0.03	0.93*	0.03
9B	Current mental health	1.05	0.03	0.96	0.03
10	Current attitudes	1.13***	0.03	0.87***	0.03
11	Current aggression	1.11***	0.03	0.91**	0.03
12	Current skills	1.08*	0.03	0.87***	0.03
12A	Dealing with others	1.05	0.03	0.89***	0.03
12B	Dealing with situations	1.05	0.03	0.90***	0.03
12C	Dealing with emotions	1.06	0.03	0.90***	0.03
12D	Controlling impulses	1.02	0.03	0.90***	0.03
12E	Controlling aggression	1.04	0.03	0.89***	0.03

Note. All models include controls for sex, race/ethnicity, age at release, prior offending history, school history, and relationship history. R-PACT = Residential Positive Achievement Change Tool.

*p < .05. **p < .01. ***p < .001 (two-tailed).

To provide a more complete picture of these results, Table 5 shows full regression results for the equations estimated for two illustrative scales. Model 1 is for the dynamic protective scale for current school status, which captures youth’s school performance while in the residential program. High scorers were those who received good grades, avoided removals and suspensions from the academic program, and were perceived as valuing education. Scoring high on this variable produced a 12% reduction in the odds of reoffending, and this occurred despite significant effects of the other variables in the model. For the demographic variables, the odds of reoffending were significantly higher among males and significantly lower among those released at an older age (with each additional year of age reducing the odds of reoffending by roughly 25%). For the static predictors, prior offending history and a history of antisocial relationships continued to significantly affect reoffending.

Table 5.

Full Logistic Regression Results for Two Illustrative Dynamic Scales: Current School Status and Current Social Skills.

		Model 1		Model 2
		b	SE	b	SE
Predictors
3C	Current school status (dynamic protective)	0.88***	0.03	—
12	Current social skills (dynamic protective)	—		0.87***	0.03
	Male	2.54***	0.25	2.56***	0.25
	Black	1.12	0.08	1.13	0.08
	Hispanic	0.87	0.10	0.92	0.10
	Other race/ethnicity	0.96	0.17	0.95	0.17
	Age at release	0.76***	0.02	0.75***	0.02
1A	Prior offending history (static risk)	1.28***	0.04	1.29***	0.04
3A	School history (static risk)	1.05	0.03	1.06	0.03
6A	Relationship history (static risk)	1.12**	0.04	1.12**	0.04
	Intercept	36.06***	16.25	38.45***	17.01

*p < .05. **p < .01. ***p < .001 (two-tailed).

A quite similar pattern is observed in Model 2 for the dynamic protective scale for current social skills. High scorers on this scale consistently used skills in such areas as problem-solving, dealing with difficult situations, and monitoring thoughts that lead to trouble. Additionally, one item in this dynamic scale directly incorporated staff assessments of progress toward social skills goals. Scoring high on this variable produced a 13% reduction in the odds of reoffending, and this once again occurred in the midst of significant effects of the demographic and static predictors. Indeed, the same such variables that were significant in the model for current school status—sex, age at release, prior offending history, and an antisocial relationship history—were significant in this model also with effects of similar magnitude.⁵

And one final issue should be considered. As we earlier noted, in evaluating risk assessment tools, it is important to consider which specific risk factors are most consequential for reoffending. With that in mind, we considered the effects of R-PACT scales in reference to the central eight risk factors prioritized in prior research. Table 6 identifies the R-PACT scales that correspond to central eight risk factors, indicating the effects that were observed. The expected effects emerged in the areas of prior offending history, antisocial personality patterns, antisocial beliefs, antisocial relationships, substance use (dynamic), and education/employment—R-PACT scales corresponding to these central eight risk areas often were related to reoffending. Indeed, the dynamic protective measures of current social skills were among the most consistent predictors of reoffending. The building of skills in this area was explicitly cited by Andrews et al. (2006) as a priority for reversing problems in the area of antisocial personality patterns. It bears emphasizing, however, that in the area of family risks and needs, none of the static and dynamic family scales affected reoffending. Also, for youth’s use of free time, only the dynamic protective scale affected reoffending.

Table 6.

The Risk and Need Areas of Greatest Consequence.

“Central Eight” Risk Factors	Corresponding R-PACT Scales and Their Effects
Prior offending	Offending history: p < .001 (static risk)
Antisocial personality	Current aggression: p < .001 (dynamic risk), p < .01 (dynamic protective)
	Current skills (general): p < .05 (dynamic risk), p < .001 (dynamic protective)
	Current skills (others): ns (dynamic risk), p < .001 (dynamic protective)
	Current skills (situations): ns (dynamic risk), p < .001(dynamic protective)
	Current skills (emotions): ns (dynamic risk), p < .001 (dynamic protective)
	Current skills (impulses): ns (dynamic risk), p < .001 (dynamic protective)
	Current skills (aggression): ns (dynamic risk), p < .001 (dynamic protective)
Antisocial beliefs	Current attitudes: p < .001 (dynamic risk), p < .001 (dynamic protective)
Antisocial relationships	Relationship history: p < .001 (static risk), ns (static protective)
Antisocial relationships	Current relationships: p < .001 (dynamic risk), p < .01 (dynamic protective)
Family	Family history: ns (static risk), ns (static protective)
Family	Current family: ns (dynamic risk), ns (dynamic protective)
Leisure and recreation	Use of free time history: ns (static protective)
Leisure and recreation	Current use of free time: ns. (dynamic risk), p < .05 (dynamic protective)
Substance use	Alcohol/drug history: ns (static risk), ns (static protective)
Substance use	Current alcohol/drug use: p < .05 (dynamic risk), p < .05 (dynamic protective)
Education/employment	School history: p < .001 (static risk), ns (static protective)
	Current school: p < .01 (dynamic risk), p < .001 (dynamic protective)
	Vocational training: ns (dynamic risk), ns (dynamic protective)
	Employability: p < .05 (dynamic risk), p < .05 (dynamic protective)
	Program-supervised tasks: ns (dynamic risk), p < .001 (dynamic protective)

Note. ns = Not significant. For the static scales, the effects shown are from the bivariate analyses reported in Table 3, whereas for dynamic scales, the effects are from the Table 4 multivariate analyses that controlled for static predictors. R-PACT = Residential Positive Achievement Change Tool.

Prediction of Reoffending Across Offender Subgroups

The final step in the analysis was to consider whether the effects of R-PACT scales depended on a youth’s sex, race and ethnicity, age at the time of release, and history of prior offending. To conduct this analysis, we used logistic regression to reestimate bivariate relationships between all scales and reoffending across the following groups (with each group’s percentage representation in the sample in parentheses): males (85%), females (15%), Whites (36%), Blacks (50%), Hispanics (11%), those released at age 16 or below (40%), those released at age 17 or higher (60%), those with two or more prior felonies (53%), and those with 1 or fewer prior felonies (47%). To test for statistically significant differences across the subgroups for a given variable, we computed z-scores that divide the difference in regression coefficients between the groups by the square root of the sum of squared standard errors (see Paternoster, Brame, Mazerolle, & Piquero, 1998). A z-score with an absolute value of 1.96 or higher indicates a difference in coefficients that is significant at a level of p < .05.⁶

Across risk and protective scales for all domains and subgroups, this involved the estimation of more than 400 equations. Full results for these equations are available upon request, and Table 7 summarizes these results, revealing the number of contrasts that yielded a statistically significant difference in coefficients. The most notable pattern involved nonsignificant differences in coefficients for the different groups. For the comparison of males and females, only 2 of the 48 contrasts (4%) yielded a significant difference in coefficients. (Using a p value of .05, this is roughly what would be expected from chance alone). A similar pattern was observed across categories of age at release—there were no instances in which a scale’s effect on reoffending significantly differed between those who were younger and older at the time of release. For prior offending, 4 of the 48 contrasts revealed significantly different effects between youth who differed in the number of prior felonies (0–1 vs. 2 or more), but there was no clear pattern to these contrasts. One of the four involved the static protective measure of alcohol/drug history. It had a more pronounced protective effect for those with 0 to 1 felonies, but no such differences were observed for other scales involving alcohol/drug use. The other three significant contrasts involved somewhat idiosyncratic effects of certain mental health scales.⁷ Thus, the most notable pattern is that R-PACT scales have effects on reoffending that do not systemically depend on youth’s sex, age at the time of release, or prior offending history.

Table 7.

Summary of Subgroup Comparisons for Effects on R-PACT Scales on Reoffending.

Group Comparison		Statistically Significant Contrasts
Sex	Male vs. female	2 of the 48
Race–ethnicity	White vs. black	2 of the 48
	White vs. Hispanic	7 of the 48
	Black vs. Hispanic	4 of the 48
Age at release	≤16 vs. 17 or older	0 of the 48
Prior offending	0–1 felony vs. 2 or more	4 of the 48

Note. Significant contrasts were those in which a test for the equality of regression coefficients had a p value of .05 or less. R-PACT = Residential Positive Achievement Change Tool.

A slightly different pattern was observed for race and ethnicity. As Table 7 shows, the analysis of z-tests revealed mostly similar effects of R-PACT scales across Black, White, and Hispanic youth—only 13 of the 144 contrasts revealed significantly different coefficients. However, the exceptions that emerged often followed a pattern. Specifically, 11 of the 13 significant contrasts involved Hispanics, and each of these involved a lower (often nonsignificant) effect of an R-PACT scale among Hispanics. For example, the protective dynamic measure of current aggression reduced reoffending among both Black and White youth, but there was no such effect for Hispanic youth. A similar pattern was observed for the protective dynamic measure of social skills. This points to the possibility that the R-PACT may have less predictive validity for Hispanics. Two caveats, however, must be emphasized. First, there were fewer Hispanic youth in the sample (522 vs. 1,668 Whites and 2,334 Blacks), and this reduced N seemed to contribute to the nonsignificance of effects (the coefficients for Hispanics often had much higher standard errors). Second, despite these instances in which Hispanics differed, the more typical pattern involved similar effects of R-PACT scales across the different racial and ethnic groups. As noted above, only 13 of the 144 contrasts revealed coefficients that were significantly different from one another.

Last, we should emphasize that for all subgroup analyses, we estimated heterogeneous choice models (Williams, 2009, 2010) that account for possible differences in residual variance across groups (these differences may bias z-tests for binary outcomes). These models revealed no instance in which effects significantly differed across groups. We encountered a minority of instances in which models estimated with the oglm command (Williams, 2010) failed to converge. However, examination of the effects and standard errors in question gave no indication that converged models would deviate from the prevailing pattern in which effects of R-PACT scales are similar across different groups.

Discussion

A key priority in research on youth reentry is to understand the factors that predict recidivism in the period after release. Sophisticated risk assessment tools now abound, and the present study assessed the predictive validity of one such tool: Florida’s R-PACT, which is used in all residential programs in the state. The analysis largely supports the predictive validity of the R-PACT. Below, we describe the specific findings buttressing that conclusion. We also discuss the exceptions to that overall conclusion, describing their implications for risk assessment research and juvenile justice practices.

Three findings emerged in support of the R-PACT’s validity. First, a clear majority of its scales significantly predicted reoffending in the 12 months after release. This largely reflected the significant effects of the dynamic measures of current circumstances rather than effects of static scales capturing a youth’s prior history. For the dynamic scales, significant bivariate effects emerged for 29 of the 34 scales; 21 of these effects maintained their significance when controlling for static risks (including prior offending) that significantly affected reoffending. These significant multivariate effects involved shifts in the odds of reoffending of 6–13% among those who were 1 SD away from the mean on the scale in question.

A second supportive finding involved a pattern in which the risk factors that predict reoffending in prior research were generally predictive in the R-PACT as well. This is consistent with the idea that the R-PACT is accurately measuring these risk factors. The most predictive R-PACT scales were in the areas of prior offending history, antisocial personality patterns, antisocial beliefs, antisocial relationships, substance use (dynamic), and education/employment. All of these risk areas are members of the central eight that have been prioritized in risk assessment research (Andrews et al., 2006).

And third, the vast majority of analyses found that effects of R-PACT scales did not vary across different subgroups of the sample. This pattern of robust predictive validity was especially evident across categories of sex, age at the time of release, and prior offending history. For these offender characteristics, the percentage of effects that significantly varied across groups was around or below what was expected from chance alone when using a p value of .05 for assessing the equality of coefficients.

These findings point to the R-PACT’s usefulness in predicting reoffending, but in each area, there are important nuances that merit further consideration. The first involves the mostly nonsignificant effects of the static scales. Only 3 of the 14 static scales were significantly related to reoffending. The three significant static predictors were prior offending history, troubled school history, and a history of antisocial relationships, but beyond these three, the static history variables in key realms of life—including family history, alcohol and drug history, and mental health history—had no predictive value for this sample. A related finding is that scales for two central eight risk areas had small to nonsignificant effects on reoffending. Specifically, none of the family scales (either static or dynamic) were related to reoffending, while one dynamic measure of youth’s use of free time exerted only a small significant effect (with the other dynamic measure having no such effect). And last, on the question of whether effects of R-PACT scales on reoffending were similar across different subgroups, there were instances in which these effects were significantly lower for Hispanics than for White and Black youth. Importantly, however, this occurred in only around 10% (11 of the 96) of the contrasts for Hispanics.

It is important to consider the implications of these findings. On the one hand, the supportive findings point to the R-PACT’s usefulness in predicting reoffending. This justifies Florida’s decision to adopt the R-PACT for use in all residential programs in the state. But beyond that, these results also offer specific insights on efforts to predict reoffending among institutionalized juvenile offenders. First, these results certainly support the field’s movement toward fourth-generation tools that prioritize dynamically measured risks and needs (Labrecque et al., 2014; McGrath & Thompson, 2012). Most of the dynamic scales had at least some predictive value, and effects were especially pronounced for the R-PACT’s dynamic protective scales, which include items in which staff members explicitly assess youth progress toward treatment and behavior goals. Also, among the dynamic scales, some areas were especially consequential. This included the use of social skills, performance in education/employment, and indications of aggressive behavior or commitment to antisocial attitudes. In these areas, the R-PACT’s scales seemed to capture the aspects of an “offender’s current and ever-changing situation” that were most relevant to reoffending (Bonta & Andrews, 2007, p. 4). The clear implication on these most consequential risk areas is that they should be especially prioritized—both as targets for intervention and as critical sources of information on how much youth are progressing.

Indeed, as an important practical matter, these findings also justify using R-PACT data to inform decisions on when to release residential youth. This, in fact, is the prescribed practice in Florida. By statute, juvenile commitments are indeterminate in length (within certain specified limits⁸), with the time of release based on youth progress on performance goals that are established early in the commitment and modified where necessary (Florida Statutes 985.455(3)). Release decisions ultimately are made by a judge who receives information on youth progress in two ways. First, every 90 days, a youth performance summary is submitted by the residential program. Second, as a youth nears completion of the treatment and performance plan, the program submits a prerelease notification seeking approval for release. Both of these reports explicitly incorporate R-PACT data, including graphical comparisons of a youth’s initial and exit R-PACT scores. Without question, youth who successfully complete performance goals and fare well on R-PACT scales are seen as “more ready” for release, while others are seen as “less ready.” In this way, scores on R-PACT dynamic scales are interpreted by residential programs and judges as a form of “signaling” in which youth are foreshadowing their future odds of reoffending (Bushway & Apel, 2012). The results presented here clearly support this interpretation, and therefore support this continued use of the R-PACT to guide the release process. However, an important caveat should be emphasized: This implication applies most specifically to those dynamic domains that were cited above as being most consequential for reoffending.

This raises an important question: What should be done in reference to the less consequential risk areas that this analysis revealed? To varying degrees, this included substance use history, mental health, the family, and youth use of free time. These risk areas often have had weak effects in evaluations of other tools, including the Level of Service Inventory (LSI; see Olver, Stockdale, and Wormith’s, 2014, recent meta-analysis of 128 LSI evaluations). The question of how to approach them is critical, given that the RNR model explicitly prioritizes the risk factors that are most relevant to reoffending. With that goal in mind, perhaps these risk areas should be deemphasized in the assessment tools themselves, in treatment efforts, and in assessments of youth progress (including those that inform release decisions). Such a de-emphasis would be consistent with the idea that resources must be allocated wisely, and research findings are the ultimate arbiter for determining which risk areas truly are “central.”

Importantly, however, there are two good reasons to proceed cautiously on this issue. First, in reference to the nonsignificant effects of many of the R-PACT’s static scales in particular, there is an important caveat: Because youth scores on static scales guide service plans, nonsignificant effects of these scales on reoffending could reflect the benefits of treatment for these youth. Second, for central eight risk areas marked by weaker effects (even when looked at dynamically), more research is needed on the causal nuances that may be at work. It is possible that the scales for these risk areas exert the expected effects on reoffending under some circumstances but fail to do so under other circumstances. Regarding the family, for example, parental involvement in a youth’s life (including visits during the residential stay) normally should reduce reoffending, but this effect may depend on patterns of offending among the parents themselves—close parental involvement could increase reoffending if the parents have an antisocial history (see Grunwald, Lockwood, Harris, & Mennis, 2010). If so, this points to a process in which a risk area still may be consequential, albeit in a conditional manner. Attention to these types of causal nuances is important. Given the centrality of the family in research both on the original causes of crime (Sampson & Laub, 1993) and on programs for reducing crime among juvenile offenders (Piquero, Farrington, Welsh, Tremblay, & Jennings, 2009), there is a compelling need to better understand how family history and family dynamics affect the odds of juvenile reoffending.

These findings and implications should be seen in the context of this study’s limitations. A first limitation involved the use of an official reoffending measure as the dependent variable. The concern with official measures is that they capture not just whether an offense is committed but also the legal and social processes that affect whether an adjudication occurs. Also, our measure involved a relatively short follow-up period of 12 months, and reoffending was measured dichotomously, therefore obscuring variation among youth in the extent of reoffending. Ideally, the results presented here would be confirmed with nonofficial measures that capture the frequency of reoffending over a longer stretch of time. This likely would make use of self-report methodology in follow-up surveys. However, similar to much prior research, such measures were unavailable in this study.

A second limitation follows from our lack of data on the postrelease circumstances experienced by youth, given that R-PACT data are collected only during the residential stay. Postrelease circumstances of greatest interest likely involve the same domain areas of risk prioritized in the R-PACT; researchers might especially be interested in such things as reintegration into school and the family, peer associations, and the use of social skills. These postrelease data could indicate whether beneficial developments that are captured in dynamic risk and protective scales are maintained or lost after release. Similarly, such data could indicate whether the effects on reoffending of risk and protective factors—as observed during the residential stay—substantially vary across different postrelease circumstances. A key point therefore to emphasize is that a truly dynamic approach to assessing risk among institutionalized offenders should extend data collection beyond the point of release.

A third limitation also relates to this issue of assessing dynamic change and its implications for reoffending. Although the R-PACT contains explicitly dynamic measures, dynamic risks and needs can be assessed in ways that go beyond what was done here. Specifically, all R-PACT data for this study came from the exit R-PACT administered near the time of release. Most residential youth, however, received at least two R-PACT administrations, thus enabling analyses that examine shifts in dynamic risk and protective factors that may occur during the residential stay. Offenders marked by significant shifts in their scores on dynamic measures may be foreshadowing their future odds of reoffending in a way that is not as clearly observed with data from a single point in time. Indeed, change analyses of this kind point to the promise of this approach (Baglivio, Wolff, Jackowski, & Greenwald, 2015; Labrecque et al., 2014; Mulvey et al., 2016). Such an analysis, however, was beyond the scope of the present study because of the need to clearly establish baseline effects for the full array of R-PACT scales. It is hoped that future research can build on the present effort to thoroughly investigate this issue of change.

In concluding, the principal theme to emphasize is that the R-PACT that Florida has so substantially invested in receives favorable support for its ability to predict reoffending among residential youth. It successfully identifies the key risk factors that should be targeted in treatment and investigated over time to assess offender progress. Just as is true for other fourth-generation tools, however, there are key unresolved issues regarding which risk factors matter most and how exactly static and dynamically assessed attributes operate together to affect reoffending. Meeting the challenges of youth reentry will require continued efforts to better understand these nuances for assessing risk and predicting reoffending.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was funded by Contract #X1708 awarded by the Florida Department of Juvenile Justice.

Notes

References

Andrews

D. A.

Bonta

Wormith

J. S.

(2006). The recent past and near future of risk and/or need assessment. Crime and Delinquency, 52, 7–27.

Annie

. Casey Foundation. (2011). No place for kids: The case for reducing juvenile incarceration. Baltimore, MD: Author.

Baglivio

M. T.

(2009). The assessment of risk to recidivate among a juvenile offending population. Journal of Criminal Justice, 37, 596.

Baglivio

M. T.

Wolff

K. T.

Jackowski

Greenwald

M. A.

(2015). A multilevel examination of risk/need change scores, community context, and successful reentry of committed juvenile offenders. Youth Violence and Juvenile Justice, 15, 38–61.

Baglivio

M. T.

Wolff

K. T.

Piquero

A. R.

Bilchik

Jackowski

Greenwald

M. A.

Epps

(2016). Maltreatment, child welfare, and recidivism in a sample of deep-end crossover youth. Journal of Youth and Adolescence, 45, 625–654.

Baird

Healy

Johnson

Bogie

Dankert

E. W.

Scharenbroch

(2013). A comparison of risk assessment instruments in juvenile justice. National Council on Crime & Delinquency. Retrieved from http://www.nccdglobal.org/publications

Basto-Pereira

Começanha

Ribeiro

Maia

Â.

(2015). Long-term predictors of crime desistance in juvenile delinquents: A systematic review of longitudinal studies. Aggression & Violent Behavior, 25, 332–342.

Benda

B. B.

Tollett

C. L.

(1999). A study of recidivism of serious and persistent offenders among adolescents. Journal of Criminal Justice, 27, 111–126.

Bonta

Andrews

D. A.

(2007). Risk-need-responsivity model for offender assessment and rehabilitation (User Report 2007-06). Ottawa: Public Safety Canada.

10.

Bushway

S. D.

Apel

(2012). A signaling perspective on employment-based reentry programming. Criminology & Public Policy, 11, 21–50.

11.

DeMuro

(2008). Boot camps revisited. Journal for Juvenile Justice Services, 22, 49–60.

12.

DiIulio

(1995). The coming of the super-predators. Weekly Standard, 1, 23–29.

13.

Drake

(2014). Predicting criminal recidivism: A systematic review of offender risk assessments in Washington State (Doc. No. 14-02-1901). Olympia: Washington State Institute for Public Policy.

14.

Fass

T. L.

Heilbrun

DeMatteo

Fretz

(2008). The LSI-R and the Compas: Validation data on two risk-needs tools. Criminal Justice and Behavior, 35, 1095–1108.

15.

Florida Statutes Title XLVII. Criminal Procedure and Corrections § 985.455. Other dispositional issues.

16.

Frazier

C. E.

Bishop

D. M.

Lanza-Kaduce

(1999). Get-tough juvenile justice reforms: The Florida experience. Annals of the American Academy of Political and Social Science, 564, 167.

17.

Grunwald

H. E.

Lockwood

Harris

P. W.

Mennis

(2010). Influences of neighborhood context, individual history, and parenting behavior on recidivism among juvenile offenders. Journal of Youth and Adolescence, 39, 1067–1079.

18.

Hilterman

E. L. B.

Nicholls

T. L.

van Nieuwenhuizen

(2014). Predictive validity of risk assessments in juvenile offenders: Comparing the SAVRY, PCL:YV, and YLS/CMI with unstructured clinical assessments. Assessment, 21, 324–339.

19.

Hockenberry

(2014). Juveniles in Residential Placement, 2011. Washington, DC: U.S. Department of Justice.

20.

Krisberg

(2011). Juvenile corrections: An overview. In Feld

B. C.

Bishop

D. M.

(Eds.) The Oxford Handbook of Juvenile Crime and Juvenile Justice (pp. 748–798). Oxford, England: Oxford University Press.

21.

Labrecque

R. M.

Smith

Lovins

B. K.

Latessa

E. J.

(2014). The importance of reassessment: How changes in the LSI-R risk score can improve the prediction of recidivism. Journal of Offender Rehabilitation, 53, 116–128.

22.

Latessa

E. J.

Lovins

B. L.

(2010). The role of offender risk assessment: A policy maker guide. Victim & Offenders, 5, 203–219.

23.

Lattimore

P. K.

Macdonald

J. M.

Piquero

A. R.

Linster

R. L.

Visher

C. A.

(2004). Studying the characteristics of arrest frequency among paroled youthful offenders. Journal of Research in Crime and Delinquency, 41, 37–57.

24.

McGrath

Thompson

A. P.

(2012). The relative predictive validity of the static and dynamic domain scores in risk-need assessment of juvenile offenders. Criminal Justice and Behavior, 39, 250–263.

25.

Mears

D. P.

Travis

(2004). Youth development and reentry. Youth Violence and Juvenile Justice, 2, 3–20.

26.

Mulvey

E. P.

Schubert

C. A.

Pitzer

Hawes

Piquero

Cardwell

(2016). An examination of change in dynamic risk of offending over time among serious juvenile offenders. Journal of Criminal Justice, 45, 48–53.

27.

Olver

M. E.

Stockdale

K. C.

Wormith

J. S.

(2009). Risk assessment with young offenders: A meta-analysis of three assessment measures. Criminal Justice and Behavior, 36, 329–353.

28.

Olver

M. E.

Stockdale

K. C.

Wormith

J. S.

(2014). Thirty years of research on the Level of Service Scales: A meta-analytic examination of predictive accuracy and sources of variability. Psychological Assessment, 26, 156–176.

29.

Panuccio

E. A.

Christian

Martinez

D. J.

Sullivan

M. L.

(2012). Social support, motivation, and the process of juvenile reentry: An exploratory analysis of desistance. Journal of Offender Rehabilitation, 51, 135–160.

30.

Paternoster

Brame

Mazerolle

Piquero

(1998). Using the correct statistical test for the equality of regression coefficients. Criminology, 36, 859–866.

31.

Piquero

A. R.

Farrington

D. P.

Welsh

B. C.

Tremblay

Jennings

W. G.

(2009). Effects of early family/parent training programs on antisocial behavior and delinquency. Journal of Experimental Criminology, 5, 83–120.

32.

Sampson

R. J.

Laub

J. H.

(1993). Crime in the making: Pathways and turning points through life. Cambridge, MA: Harvard University Press.

33.

Schwalbe

C. S.

(2007). Risk assessment for juvenile justice: A meta-analysis. Law and Human Behavior, 31, 449–462.

34.

Snyder

H. N.

(2004). An empirical portrait of the youth reentry population. Youth Violence and Juvenile Justice, 2, 39–55.

35.

StataCorp. (2012). Stata Statistical Software: Release 2012. College Station, TX: StataCorp LP.

36.

Steinberg

Cauffman

Monahan

(2015). Psychosocial maturity and desistance from crime in a sample of serious juvenile offenders (OJJDP Juvenile Justice Bulletin). Washington, DC: U.S. Department of Justice.

37.

Taxman

F. S.

Caudy

M. S.

(2015). Risk tells us who, but not what or how. Criminology & Public Policy, 14, 71–103.

38.

Trulson

C. R.

DeLisi

Marquart

J. W.

(2011). Institutional misconduct, delinquent background, and rearrest frequency among serious and violent delinquent offenders. Crime and Delinquency, 57, 709.

39.

Trulson

C. R.

Haerle

D. R.

DeLisi

Marquart

J. W.

(2011). Blended sentencing, early release, and recidivism of violent institutionalized delinquents. The Prison Journal, 91, 255–278.

40.

Vincent

G. M.

Chapman

Cook

N. E.

(2011). Risk-needs assessment in juvenile justice. Criminal Justice and Behavior, 38, 42–62.

41.

Wachter

(2015). JJGPS StateScan: Statewide risk assessment in juvenile probation. National Center for Juvenile Justice. Retrieved from http://www.jjgps.org/publications

42.

Williams

(2009). Using heterogeneous choice models to compare logit and probit coefficients across groups. Sociological Methods & Research, 37, 531–559.

43.

Williams

(2010). Fitting heterogeneous choice models with oglm. Stata Journal, 10, 540–567.