Police‐initiated diversion for youth to prevent future delinquent behavior: a systematic review

This Campbell systematic review examines the effects police‐initiated diversion programs on delinquent behavior, compared to traditional system processing. The review summarizes evidence from nineteen high‐quality studies, including 13 randomized controlled trials and six quasi‐experimental studies. This review includes studies that evaluated the effects of police‐led diversionary practices compared to traditional processing for youth under 18 years of age. We identified a total of 14 manuscripts representing 19 evaluations. Of these 19 evaluations, 13 used randomized controlled designs (random assignment to conditions) and 6 used quasi‐experimental designs (no random assignment to conditions). Many of these designs included two or more diversionary conditions compared to a common control (traditional processing) producing 31 treatment‐comparison contrasts for analysis. These studies were conducted between 1973 and 2011, inclusively. Most were conducted in the USA (11) with the remaining conducted in Canada (4), Australia (2), and the UK (2). The general pattern of evidence is positive, suggesting that police‐led diversion reduces the future delinquent behavior of low‐risk youth relative to traditional processing. Assuming a 50 percent reoffending rate for the traditional processing condition, the results suggest a reoffending rate of roughly 44 percent for the diverted youth. This overall benefit of diversion holds for the random assignment studies judged to be free from any obvious risks of bias. No meaningful differences were found across types of diversionary programs. Furthermore, we found no evidence to suggest these findings suffer from publication selection bias. Plain language summary Police‐led diversion of low‐risk youth reduces their future contact with the justice system Police‐led diversion of low‐risk youth who come into contact with the justice system is more effective in reducing a youth's future contact with the justice system compared to traditional processing. What is this review about? Youth misconduct and misbehavior is a normal part of adolescence and that misbehavior sometimes crosses the line from disruptive or problematic to delinquent. Nationally representative surveys of youth in the USA have indicated that minor delinquent behavior is normative, particularly for boys. The normative nature of minor delinquent behavior raises the question of how police should respond to minor delinquent behavior in a way that is corrective, but also avoids involving the youth in the criminal justice system beyond what will be effective in reducing future misbehavior. Police diversion schemes are a collection of strategies police can apply as an alternative to court processing of youth. Diversion as an option is popular among law enforcement officers, as it provides an option between ignoring youth engaged in minor wrongdoing and formally charging such youth with a crime. Police‐led diversion has the potential to reduce reoffending by limiting the exposure of low‐risk youth to potentially harmful effects of engagement with the criminal justice system. What is the aim of this review? This Campbell systematic review examines the effects police‐initiated diversion programs on delinquent behavior, compared to traditional system processing. The review summarizes evidence from nineteen high‐quality studies, including 13 randomized controlled trials and six quasi‐experimental studies. This review examined whether police‐led diversion and traditional processing of youth have different effects on rates of official delinquency. What are the main findings of this review? This review includes studies that evaluated the effects of police‐led diversionary practices compared to traditional processing for youth under 18 years of age. We identified a total of 14 manuscripts representing 19 evaluations. Of these 19 evaluations, 13 used randomized controlled designs (random assignment to conditions) and 6 used quasi‐experimental designs (no random assignment to conditions). Many of these designs included two or more diversionary conditions compared to a common control (traditional processing) producing 31 treatment‐comparison contrasts for analysis. These studies were conducted between 1973 and 2011, inclusively. Most were conducted in the USA (11) with the remaining conducted in Canada (4), Australia (2), and the UK (2). The general pattern of evidence is positive, suggesting that police‐led diversion reduces the future delinquent behavior of low‐risk youth relative to traditional processing. Assuming a 50 percent reoffending rate for the traditional processing condition, the results suggest a reoffending rate of roughly 44 percent for the diverted youth. This overall benefit of diversion holds for the random assignment studies judged to be free from any obvious risks of bias. No meaningful differences were found across types of diversionary programs. Furthermore, we found no evidence to suggest these findings suffer from publication selection bias. What do the findings of this review mean? The findings from this systematic review support the use of police‐led diversion for low‐risk youth with limited or no prior involvement with the juvenile justice system. Thus, police departments and policy‐makers should consider diversionary programs as part of the mix of solutions for addressing youth crime. Many of the studies included in the review were conducted in the 1970s and 1980s. Newer high quality studies are needed to ensure that the findings still hold for contemporary juvenile justice contexts. Additional studies are also needed outside of the USA for this same reason. Finally, we recommend that research explore the usefulness of diversion for low‐risk adult offenders. How up‐to‐date is this review? The search for eligible studies was completed in January of 2017, so only studies identifiable through January 2017 were included. This Campbell systematic review was published in May 2018. Executive summary/Abstract Background Overly punitive responses to youth misconduct may have the unintended consequence of increasing the likelihood of future delinquency; yet, overly lenient responses may fail to serve as a corrective for the misbehavior. Police diversion schemes are a collection of strategies police can apply as an alternative to court processing of youth. Police‐initiated diversion schemes aim to reduce reoffending by steering youth away from deeper penetration into the criminal justice system and by providing an alternative intervention that can help youth address psychosocial development or other needs that contribute to their problem behavior. Objectives The objective of this review was to synthesize the evidence on the effectiveness of pre‐court interventions involving police warning or counseling and release, and cautioning schemes in reducing delinquent behavior. Search methods A combination of 26 databases and websites were searched. References of relevant reviews were also scanned to identify studies. We also consulted with experts in the field. Searches were executed by two reviewers and conducted between August 2016 and January 2017. Selection criteria Only experimental and quasi‐experimental designs were eligible for this review. All quasi‐experimental designs must have had a comparison group similar to the police diversion intervention group with respect to demographic characteristics and prior involvement in delinquent behavior (i.e., at similar risk for future delinquent behavior). Additionally, studies must have included youth participants between 12 and 17 years of age who either underwent traditional system processing or were diverted from court processing through a police‐led diversion program. Studies were also eligible if delinquency‐related outcomes, including official and non‐official (self‐report or third‐party reporting) measures of delinquency were reported. Data collection and analysis This study used meta‐analysis to synthesize results across studies. This method involved systematic coding of study features and conversion of study findings into effect sizes reflecting the direction and magnitude of any police‐led diversion effect. There were 19 independent evaluations across the 14 primary documents coded for this review. From this, we coded 67 effect sizes of delinquent behavior post diversion across 31 diversion‐traditional processing comparisons. We analyzed these comparisons using two approaches. The first approach selected a single effect size per comparison based on a decision rule and the second used all 67 effect sizes, nesting these within comparison condition and evaluation design. Results The general pattern of evidence is positive, suggesting that police‐led diversion modestly reduces future delinquent behavior of low‐risk youth relative to traditional processing. Authors' conclusions The findings from this systematic review support the use of police‐led diversion for low‐risk youth with limited or no prior involvement with the juvenile justice system. Thus, police departments and policy‐makers should consider diversionary programs as part of the mix of solutions for addressing youth crime.


Executive summary/Abstract Background
Overly punitive responses to youth misconduct may have the unintended consequence of increasing the likelihood of future delinquency; yet, overly lenient responses may fail to serve as a corrective for the misbehavior. Police diversion schemes are a collection of strategies police can apply as an alternative to court processing of youth. Police-initiated diversion schemes aim to reduce reoffending by steering youth away from deeper penetration into the criminal justice system and by providing an alternative intervention that can help youth address psychosocial development or other needs that contribute to their problem behavior.

Objectives
The objective of this review was to synthesize the evidence on the effectiveness of pre-court interventions involving police warning or counseling and release, and cautioning schemes in reducing delinquent behavior.

Search methods
A combination of 26 databases and websites were searched. References of relevant reviews were also scanned to identify studies. We also consulted with experts in the field. Searches were executed by two reviewers and conducted between August 2016 and January 2017.

Selection criteria
Only experimental and quasi-experimental designs were eligible for this review. All quasiexperimental designs must have had a comparison group similar to the police diversion intervention group with respect to demographic characteristics and prior involvement in delinquent behavior (i.e., at similar risk for future delinquent behavior). Additionally, studies must have included youth participants between 12 and 17 years of age who either underwent traditional system processing or were diverted from court processing through a police-led diversion program. Studies were also eligible if delinquency-related outcomes, including official and non-official (self-report or third-party reporting) measures of delinquency were reported.

Background The problem, condition or issue
Misbehavior is a normal part of adolescence and that misbehavior sometimes crosses the line from disruptive or problematic to delinquent. Nationally representative surveys of youth in the United States have indicated that minor delinquent behavior is normative, particularly for boys (Bongers et al., 2003;Elliott et al., 1983;Moffitt, 1994). The normative nature of minor delinquent behavior raises the question of how police should respond to minor delinquent behavior in a way that is corrective, but also avoids involving the youth in the criminal justice system beyond what will be effective in reducing future misbehavior. Stated differently, what is the right level of response to minor delinquent acts? Overly punitive responses may have the unintended consequence of increasing the likelihood of future delinquency; overly lenient responses may fail to serve as a corrective for the misbehavior. Police diversion schemes are a collection of strategies police can apply as an alternative to court processing of youth. Police-initiated diversion schemes aim to reduce reoffending by steering youth away from deeper penetration into the criminal justice system and by providing an alternative intervention that can help youth address psychosocial development or other needs that contribute to their problem behavior.
Diversion as an option is popular among law enforcement officers, as it provides an option between ignoring youth engaged in minor wrongdoing and formally charging such youth. In 2016/17, approximately 60% of first-arrest juveniles in England and Wales received a caution rather than court-processing (Ministry of Justice, 2017) and an unknown number were diverted prior to any cautioning. Diversion has the potential to reduce reoffending by limiting the exposure of low-risk youth to potentially harmful deviant peers within the criminal justice system. Furthermore, diversion may reduce criminal justice system costs, freeing these resources for higher risk youth. However, some commentators (Ray & Childs, 2015; have noted that diversion may widen the population of youth under the surveillance of the criminal justice system if youth are subsequently punished for failing to meet the terms of their diversion. Consequently, diversion may inadvertently increase youth reoffending. The uncertain potential for diversion to produce both benefits and harms and law enforcement's sustained use of diversion underscores the importance of comprehensively reviewing the effectiveness of these interventions.

The intervention
This review will focus on the pre-charge diversion of youth. Police-led juvenile diversion is a pre-court intervention initiated by police that represents an alternative to court processing or the imposition of formal charges. Examples might involve a caution, a restorative caution, or a final warning or reprimand. Each of these alternatives might be combined with an additional program element such as referral to a treatment service provider. Police-led diversions may be known by many names, such as cautions, final warnings, police-led intervention, police control of juveniles, police-led proactive prevention, police-led diversion, pre-charge diversion or simply as diversion. The essential feature involves police initiating and leading the intervention and the youthful offender receiving a diversionary scheme to avoid a criminal record and any negative consequences that may result from continued formal contact with the criminal justice system (e.g., imposition of formal charges, conviction, etc.).
The participants in a traditional police cautioning scheme include a police officer, the youth in question, and the parents, at a minimum. Victims are not involved nor do police officers routinely receive any training, but solely provide an explanation about the legal and social consequences of continued delinquent behavior. However, variants of this scheme, caution plus and restorative cautioning, involve other interventions and services (Audit Commission, 1996) or involvement of a script with certain questions to structure discussion between an offender and the affected parties and the presence of the victim, in the case of restorative cautioning or conferencing (Wilcox et al., 2004). As for the final warning and reprimand scheme, this involves an assessment-based approach to evaluate the seriousness of the offense and, depending on the gravity of the offense, a reprimand or final warning with referral to a multi-agency team for further assessment and placement in a behavioral treatment program (Holdaway, 2003, p. 352). Wilson and Hoge (2015) articulate two theoretical supports for diversion: labeling theory and differential association theory. Labeling theory posits that the stigmatizing effect of labeling a youth as delinquent may establish expectations for future delinquent acts and alter that youth's social networks toward more deviant peers, thus increasing the likelihood of future deviant behavior (Bernburg et al., 2006). Thus, diverting a low-risk youth not already labeled as "delinquent" may reduce future offending. Sutherland's (1939) differential association theory states a youth learns the values, attitudes, and techniques of criminal behavior through the interaction with delinquent peers. As such, diverting low-risk youth from the juvenile justice system may reduce exposure to deviant peers. In essence, a youth who has engaged in a delinquent act may learn more serious forms of delinquency from others already in the juvenile justice system, including the values and attitudes that support involvement in delinquency.

How the intervention might work
The reintegrative shaming aspect of restorative justice serves as another theoretical mechanism that might explain a positive effect of diversion. Reintegrative shaming emphasizes an intervention strategy where responses to transgressions are de-stigmatizing and inclusive of "a meaningful community-based process that reaffirms the boundaries of acceptable behavior" (Zhang, 2011(Zhang, , p. 2325. These responses should aim to reduce or inhibit new or further stigmatization as a result of contact with the justice system. Forgiveness and non-stigmatization are central principles of reintegrative shaming, as these tenets reinforce another core feature-reintegration. Reintegration concerns efforts to restore offenders (and victims alike) after a transgression and reintegrate them back into the community. This process also implicates communities of care such as significant others, e.g., family members or individuals of import in an offender's life who are central to disavowing unlawful behavior and facilitating forgiveness. Our interpretation of community of care includes that of authority figures such as police officers, whom we understand may not traditionally be viewed as members of an affected person's broader prosocial community. Yet, the interaction between the police officer and the youth provides an opportunity for an authority figure to reinforce appropriate norms by briefly detaining the youth engaged in a problem behavior, thus enacting an element of shame. The shame may be enhanced by the police officer taking the youth to her home and discussing the youth's acts with her parent. Additionally, the element of reintegration emerges with diverting the youth from any further court processing, allowing the youth to return to a state of "good standing" in the community (Hay, 2001;Sherman, 1993).
Referral to needed and effective services remains as a final mechanism that might contribute to the effectiveness of some diversion schemes. As discussed above, some diversion programs involve formal referrals to treatment services or a needs assessment for services. As the first point of contact with the justice system, police officers are poised to intervene early and provide referrals to needed services that may be more beneficial in reducing future delinquent acts than court processing (Butts, 2016;. Diversionary practices, however, may also be harmful in the sense of increasing a youth's propensity to engage in delinquent behaviors. From a deterrence perspective, diversion provides a swift sanction, but the sanction may be too mild to deter a youth from similar (or more severe) behavior in the future. A youth who perceives that he was not held responsible for his actions may think that he "got away with it" and will continue to engage in similar ways.  also suggest a possible "net widening" effect of diversion programs. When a diversion with conditions is used instead of a diversion with 'no further action' and a youth fails to meet the specified conditions, such as attending an appointment with a counselor, the youth may be brought into the criminal justice system as a consequence. Perversely, this may result in a low-risk youth experiencing the negative consequences of juvenile justice system involvement that the diversion was designed to prevent.

Prior reviews
Several meta-analyses of diversion programs exist and these differ from each other and from the proposed review in important ways. In a Campbell Collaboration review, Petrosino et al. (2010) examined the effectiveness of juvenile justice system processing compared to any alternative non-system condition. Their focus was on any form of diversion, whereas our focus is on pre-charge or police-initiated diversion only. Petrosino et al. (2010) found that court processing produced worse outcomes than diversion from the system. Similarly, Wilson and Hoge (2015) examined 45 studies and found that, on average, diversion conditions had lower recidivism rates than formal court processing. Additionally, their analysis showed slightly larger beneficial effects for pre-charge diversion compared to post-charge diversion. However, methodologically stronger research designs failed to find a positive effect for diversion relative to traditional processing. Finally, a meta-analysis completed by Schwalbe et al. (2012) focused on diversion programs with a treatment component such as case management, family treatment, youth court, etc. Although not explicitly stated, these diversion programs likely occurred post-charge. The findings showed a small overall effect favoring these programs, but the effects were not statistically significant except for family treatment. Overall, Schwalbe et al. (2012) did not provide a meaningful examination of police-led diversions, which are the focus of the current review.
Taken as a whole, these meta-analyses provide an equivocal answer regarding the effectiveness of diversion programs. Furthermore, these prior reviews did not specifically focus on pre-charge or police-led diversion; we aim to clearly describe the effect of pre-charge diversion on reoffending outcomes compared to court processing. Based on labeling and differential association theories, we would expect diversion at this stage to be more effective as it avoids any labeling of the youth, even if temporarily via a formal charge and at a minimum, reduces potential exposure to deviant peers in the juvenile justice system. Hence, the purest form of diversion occurs at this stage per the avoidance of any juvenile justice system processing.
Despite the uncertainty regarding the effectiveness of diversionary practices, diversion schemes are used frequently by police officers. Precise estimates of the prevalence of diversion are difficult given that a central feature of much diversion involves no court processing or recording in criminal justice records. Consequently, prevalence estimates of police-led youth diversions are rare or must be extrapolated to national levels from small area studies. Additionally, definitions of diversion vary by jurisdiction, further impeding accurate estimates of prevalence. For example, the option of a 'caution' that exists in England and Wales, South Africa and Australia still forms part of a youth's juvenile record, but is not a formal charge. Therefore, it is debatable whether these constitute a diversion as discussed above. It is also important to note that police-led juvenile diversion is not a disposal option in some jurisdictions. For example, Algeria, Argentina, Italy and Kuwait all forbid police-led diversion, although options for diversion later in the judicial process may exist (Hazel, 2008).
According to a Ministry of Justice study, 18% of youth arrests in England and Wales result in a caution (Ministry of Justice, 2017). In the United States, Puzzanchera and Kang (2008) estimate that a similar number (25%) of youth entering the juvenile justice system are diverted, but much of this is initiated post-charge, rather than by the police officer during the initial interaction with the youth.
Departing from, but building on the work of prior reviews, our systematic review and metaanalysis focuses on pre-charge (police-led) diversion prior to the imposition of formal charges. This narrower focus will help inform police practice related to the use of diversion. Furthermore, we explore the differential effectiveness of the various diversionary schemes, such as a diversion with no further action, restorative caution, or diversion with various therapeutic elements.

Objectives The problem, condition or issue
The objective of this review was to synthesize the evidence on the effectiveness of pre-court interventions involving police warning, reprimand, and cautioning schemes in reducing delinquent behavior. Our specific research questions were: 1. Are police-initiated diversions effective in reducing future delinquent behavior (i.e., additional cautioning, arrest, court appearances, or findings of guilt)?
2. Is effectiveness related to the type of police-initiated diversion used (i.e., traditional cautioning, caution plus, police restorative cautioning, final warning or reprimand)?
3. Is effectiveness related to characteristics of the youth (i.e., age, gender, race/ethnicity, crime committed, and offense history)?

Deviations from protocol
This review has deviated from the published protocol in a few ways. These deviations are listed and explained below and do not impact the results of the review.
• We originally proposed to examine the impact of police-initiated diversion on secondary outcomes such as perceptions of law enforcement, perceived fairness, and satisfaction. We were unable to identify enough of these secondary outcomes in the literature to undertake a meaningful review.
• Our a priori planned moderator analyses for diversion type originally specified interventions such as traditional cautioning, caution plus, police restorative cautioning, and final warning or reprimand. These categories were collapsed to traditional cautioning, caution plus with a service referral, and police restorative cautioning and inform the moderator analysis for diversion type as shown in table 4.

Types of studies
Both experimental and quasi-experimental designs were included. The specific eligibility criteria for each design is detailed below. Qualitative studies were not eligible for inclusion in this review.
• Experimental designs. Eligible experimental designs must have randomly assigned participants to a diversion or a control condition(s). Designs that used a quasirandom assignment procedure, e.g., assignment based on an alternate case basis, were also eligible.
• Quasi-experimental designs. Several types of quasi-experimental designs were eligible; however, all quasi-experimental designs must have had a comparison group that was similar to the police diversion intervention group with respect to demographic characteristics and prior involvement in delinquent behavior (i.e., similar risk for future delinquent behavior). This similarity can be achieved through matching or statistical controls. Matching at the individual level or at the group level was permitted. Statistical control methods could include regression analysis, analysis-ofcovariance, and propensity score modeling, among others. Use of a statistical control method was sufficient for inclusion meaning, we did not exclude studies based on a subjective assessment of the quality of the statistical controls. Rather, any quasiexperimental design that controlled for baseline risk factors, such as age, gender, and prior offense history, was eligible. Quasi-experimental designs were not eligible if the comparison group was comprised of participants who refused participation in a police diversion scheme or who dropped out of a police diversion scheme. Quasiexperimental designs that did not have a comparison group were not eligible.

Types of participants
The population of interest was youth suspected of involvement in a crime or delinquent behavior. Eligible studies must have included participants who were youth between 12 and 17 years of age, inclusively. Participant samples that included a small proportion (i.e., less than 20%) of youth over 17 but less than 22 were also eligible. Participants must also have been apprehended, arrested, or otherwise referred to the juvenile justice system, and either diverted to a police-involved intervention prior to the imposition of formal charges, or in the case of a comparison condition, treated in some other fashion.

Types of interventions
Interventions that were considered eligible must have been initiated and implemented by police officers as identified by the study. This included programs where diversion occurred any time prior to formal charges-whether before or after arrest-but prior to the imposition of formal charges. Interventions that involved court or prosecutorial referrals, even with the inclusion of police officers, were not considered eligible. Findings were considered relevant if measured at the police 'level of referral'. Control conditions were typically 'treatment as usual', which is often a process of laying criminal charges followed by adjudication through the criminal justice system. Also excluded were studies where a disposal of 'no further action' was treated as a control condition.

Primary outcomes
The primary outcome of interest was delinquency. Eligible studies must have reported at least one delinquency-related outcome. This could include official measures of delinquency, such as an arrest, or other measures of delinquent-type behaviors, such as self-report, parent-report, or school records of wrongdoing.

Secondary outcomes
Secondary outcomes of interest could have included self-report measures related to improved relations, such as satisfaction with police or the cautioning process.

Search methods for identification of studies Electronic searches
Four categories of key words were developed for this search. The first category lists key terms and synonyms related to youth and their social status. The second category of key terms are related to pre-court cautioning practices and schemes. The third and fourth categories address the methodology and the measured study outcomes, respectively. Zotero, a reference management software program was used to retrieve, store, and document the search process. Each database had its own file folder within Zotero and was searched individually. Search notes were created for each database and stored in the appropriate file folder. The search notes captured: the date of the search, the database name, the final search string used, the reference yield produced, and a notes field to capture any aberrant issues.
Population youth OR child OR juvenile OR delinquent OR devian* OR student OR adolescent OR "young person" OR "young offender*" OR bully* OR "youthful offender" Treatment diverted OR diversion OR caution* OR "caution plus" OR restorative OR "restorative caution" OR triage OR "final warning" OR reprimand OR "alternative* to custody" OR "precharge" OR "pre-caution" OR "pre-court" OR "pre-custody" OR "alternative program*" OR disposal OR disposition OR liaison OR Police-led OR "police initiated" OR "police control" OR "police diversion" OR police OR "law enforcement" OR "civil citation" Methodology outcome OR evaluat* OR effect OR effectiv* OR experiment* OR quasi OR assessment OR RCT OR "random* control*" Outcome recidivism OR arrest* OR rearrest* OR citation OR offend* OR reoffend* OR conviction OR reconviction OR adjudication OR adjudicated This search strategy was applied to the following databases and websites, which cover both the easily accessible sources as well as the grey literature.

Searching other resources
In addition to searching the electronic resources listed above, we also scanned references of relevant reviews and identified studies. We also consulted with an expert in the field, Peter Neyroud, a lecturer at the Institute of Criminology, University of Cambridge, and received a list of experimental and quasi-experimental studies.

Data collection and analysis Data extraction and management
The second and third author screened titles and abstracts for relevance and then equally divided the smaller pool of resulting studies to review the full-text of the reference for eligibility. At this stage, reviewers also hand searched relevant, but ineligible referencesprior meta-analyses and reviews (see Farrington & Welsh, 2006;Petrosino et al., 2010;Schwalbe et al., 2012;Wilson & Hoge, 2013)-to be exhaustive and cross-reference for additional references not captured in the systematic search. These references were then sorted into two categories-a relevant reference for further review or a reference relevant for background information or context (see Figure 1). Next, each reviewer applied the inclusion criteria noted above to review the full-text of the smaller set of studies. From this, the final eligible sample of studies emerged and the second and third author fully coded each study in this final sample. At this stage reviewers also eliminated studies that were coded as eligible, but upon further review and through consensus meetings were deemed ineligible (see Patrick & Marsh, 2005). All three authors were involved in the coding and double coding of the final set of eligible studies. All three authors also reconciled coding differences through weekly consensus meetings.
We coded eligible studies for study characteristics, intervention/comparison characteristics, outcome characteristics, and effect size data (see coding forms in appendix A). Coding was unique for each eligible study, as the unit of analysis for the meta-analysis is an independent study. In cases where there were multiple publications for the same study, the most complete study was coded as the primary study and all other related publications were coded as crossreferences. Included studies were double-coded with differences resolved through consensus meetings. All coding and data management, including capturing decisions from consensus meetings was managed using FileMaker Pro database software.

Assessment of risk of bias in included studies
Methodological quality and risk of bias was coded as data was extracted for study, intervention/comparison, and outcome characteristics. Specifically, at the study level, we coded for the type of experimental and quasi-experimental design based on assignment (e.g., matching, wait list control, cohort, etc.). Risk of bias was captured by assessing the risk of selective outcome reporting. At the intervention/comparison level, risk of bias was coded based on reported or observed differences between groups at baseline (selection bias) and attrition bias for the primary outcome, in terms of quantity and differential attrition. Finally, at the outcome level risk of bias was based on one item, which captured whether there was potential bias from non-blinding procedures.

Measures of treatment effect
The primary outcome for this review is delinquency and is most often reported on a dichotomous scale, that is, as delinquent or non-delinquent. As such, the effect size of choice for this review was the odds ratio. Odds ratios were computed from any available information such as proportions, percentages, raw frequencies, chi-square and marginal distributions, etc. In the case of quasi-experimental designs with statistical adjustments for baseline differences, the regression coefficient from a logistic regression model was coded as the logged odds ratio along with the reported standard error. Effect sizes based on scaled measures of delinquency were computed as d-type effect sizes and then converted to odds ratios using the logit transformation method (Lipsey and Wilson, 2001). All effect size computations used established equations as implemented in the online effect size calculator available on the Campbell website.

Assessment of heterogeneity
In place of the traditional -statistic, heterogeneity was assessed using τ 2 , which is the random effects variance component used in random effects models. We did not report Q because we used the robust standard errors method of analysis for handling statistically dependent effect sizes (Hedges, Tipton, & Johnson, 2010). This method does not produce a formal test of the statistical significance of τ 2 . When τ 2 is zero, the distribution is homogeneous. As τ 2 increases, it indicates increasing levels of heterogeneity. To get a sense of when τ 2 might be statistically significant, analyses were rerun using David Wilson's meanes macro for Stata, ignoring the statistical dependencies issue. These analyses suggested that with these data, any τ 2 above 0.05 likely reflects statistically significant heterogeneity. This should be treated as a rough guide only and not as a formal indicator of significant heterogeneity.

Assessment of reporting biases
Publication-selection bias was assessed in three ways. First, analyses compared the results from published and unpublished reports. Published documents included peer-reviewed journal articles, books, and book chapters. All other report forms, such as theses, technical reports, government and agency reports, were considered unpublished. Second, we performed a trim-and-fill analysis. Third, we visually inspected a funnel plot.

Unit of analysis issues
Studies based on the same sample were treated as a single study, of which the manuscript with the most complete information was coded as the primary study. The other related documents were consulted for additional information and are part of the reference list of included studies. Publications or documents with multiple independent evaluations within were coded as separate studies, such as studies with RCTs conducted in two different cities.

Data synthesis
Meta-analysis was conducted using random effects models. Primary analyses were performed using the robust standard error method of modeling statistical dependences as implemented in the Stata package robumeta (see http://www.northwestern.edu/ipr/qcenter/RVE-metaanalysis.html for details).

Sensitivity analysis
Our a priori planned moderator analyses included the type of diversion (e.g., diversion only, diversion plus services, and diversion plus police-led restorative cautioning), the type of research design (e.g., experiment versus quasi-experiment), and publication type (i.e., published versus unpublished). Post hoc moderator analyses explored the relationship between other study features such as data collection period and country of intervention, and effect size. Meta-analytic regression was used to complete all moderator analyses using the robust standard errors method.

Results of the search
The search strategy yielded 14,161 references. Removal of duplicates reduced this to 11,979 references. A pre-eligibility screening of titles and abstracts identified 220 references potentially relevant from this reduced yield. During this process an additional 67 references were also tagged as relevant for background information to help contextualize this review. Two coders examined the 220 potentially relevant documents in detail for eligibility. Fourteen documents representing 19 unique studies satisfied our eligibility criteria. As is often the case, several studies that were initially coded as eligible were later recoded as ineligible upon closer inspection. In the case of the current review, one study met the criterion for being a near-miss study. This study is listed in the reference section as marginally eligible, but excluded. Appendix B includes a full list of studies included in and excluded from the meta-analysis.

Included studies
Tables 1, 2, and 3 provide descriptive information on 31 treatment-comparison contrasts detailed across the 19 studies in the 14 publications. As shown in Table 1, evaluation of 10 of the studies was based on technical reports. In several cases, these reports were also published in a peer-reviewed journal, but the associated technical reports were preferred as they provided more detailed information about the study. Three of the studies were based on a Master's (n=1) or a PhD thesis (n=2). The remaining five studies were each published as a peer-reviewed journal article (n=4) or a peer-reviewed agency report (n=1). The publication dates of the studies ranged from 1979 to 2015. Of the 19 studies, 11 were conducted in the United States -of all which were conducted between 1974 and 1997, with the majority taking place in the late 1970s and early 1980s. The four studies that were conducted in Australia took place between 2000 and 2010; one Canadian study took place in the mid-1970s and one in the mid-1990s; and the two UK-based studies were conducted between 2009 and 2011. Most studies were conducted within a single police department area (usually a city; n=15) although four studies covered a wider area such as a county (Koch, 1985) or a region (Cunningham, 2007;Little, 2015).
As shown in Table 2, of the 31 interventions included in the review, 13 examined traditional caution, 14 examined caution with a referral to services, and 4 examined police restorative cautions. Of the 21 studies that included a referral to services, these services were predominantly provided by external agencies (n=20) and, generally were provided by a public or non-profit agency and were not purchased (the exception was , although purchase of services may not always have been reported). The control condition was almost always formal court processing, which equates to a "treatment as usual" condition; one study used probation as the control condition.
The age range of juveniles at the point of entering the study was 10 years to 17 years. Although age range was not reported in six studies, the interventions were limited to juveniles, so all participants would have been between the minimum age of criminal responsibility for the study area and 17 years of age. The gender distribution of samples was typical of the juvenile offending population, with all studies that reported this statistic having a predominantly male sample. The ethnic composition of the study samples was not wellreported with eight studies not including any information about the ethnicity of their samples.
The most common outcome measure was a binary indicator of arrest or police contact, which was used in 23 of the 31 comparisons. Self-reported delinquency was used in ten comparisons -the majority of times this measure was calculated by summing the different offence types they committed. The severity of self-reported delinquency was also measured in one study.
For studies that included police or court records as the outcome measure, follow-up periods were typically in excess of 12 months. All studies that used a self-reported delinquency measure had follow-up periods less than one year. Within studies, follow-up time was consistent across participants in all but three studies: one followed participants to a fixed age (19.5 years) while follow-up periods in two other studies varied across individuals (15-21 months and 15-30 months).
All studies compared treatment effectiveness at the individual level and study designs were consistent within studies when there were multiple comparisons. Studies employed either a quasi-experimental (6) or a randomized controlled design (13). All the quasi-experimental studies matched individuals or statistically controlled for baseline differences between treatment and control groups. All but one of the randomized controlled studies were undertaken without matching.

Risk of bias in included studies
Risk-of-bias assessment is shown in Table 3. For each potential source of bias, we coded each study on the following scale: low-risk, high-risk, and unclear. The dimensions of bias assessed were: (1) selection (were there meaningful differences between groups at baseline?), (2) general attrition (was there a risk of general attrition bias for the primary outcome measure?), (3) differential attrition (was there a risk of meaningful differential attrition bias for the primary outcome measure?), (4) measurement (did the measure used create a risk of bias on the part of the assessor?), and (5) outcome reporting (is there any evidence that the authors have not reported findings for all variables measured as part of this study?).
The internal validity of the studies was generally high. Study designs were used that were capable of ensuring there were few baseline differences between treatment groups. The use of police records as an indicator of reoffending limited the potential for bias in the data collection process and virtually eliminated the potential for differential rates of study attrition across treatment conditions.
However, there were some limitations to the internal validity of the studies. There was no potential for the blinding of treatment condition on the part of the participants. This was a particular concern when the outcome measures were based upon self-reported delinquency.
Theoretically, respondents may have been supportive of a disposal they regarded as lenient and may have biased their self-reported delinquency accordingly. It is noteworthy that participant blinding in this type of intervention is not possible and not desirable, as awareness of being diverted is part of the intervention. Police officers, who were responsible for generating the arrests upon which the outcome measures were based, were also not blind to the treatment group of an individual. Hypothetically, this could have impacted their use of discretion in the decision to arrest a study participant at some future time point after the initial diversion. However, we saw no evidence of participant or administrator bias in any of the study reports.
At the level of treatment assignment, some studies reported failures in the randomisation process when police officers were responsible for randomisation. For example, in the New York City study completed by Dunford et al (1982), the project team found that police officers were funneling offenders whom they regarded as low-risk towards the diversion scheme while funneling higher-risk offenders towards the control condition. Efforts were made to rectify those treatment assignment biases when they were identified and they also undertook baseline comparisons of treatment group characteristics and statistically adjusted for these biases when appropriate.
There was little to no evidence that attrition or loss of study participants over the course of the study was an issue for these studies. We rated only one study as being at risk of general attrition bias and no study as at risk of differential attrition. The latter is relevant for the internal validity of a study, whereas the latter affects the generalizability of any observed findings. What we were not able to assess was the degree to which the selection of youth into each study's sample was representative of all eligible youth.

Synthesis of results
Across the 19 independent evaluations identified across the 14 primary documents coded for this review, multiple comparisons of diversion versus traditional processing were common across these evaluations. This also included comparisons of diversion with no referral to services (simple counsel and release) and diversion with referral to services compared to a traditional processing condition. We coded both of these comparisons when possible, producing 31 comparisons for analysis (see Table 2 for details).
These 31 comparisons came from 19 independent evaluations and in cases with multiple comparisons from a single evaluation, the comparisons had a common control group. That is, many of the evaluations included two or more diversion conditions compared to a single control condition. As discussed in the methods section, we used the robust standard errors method of analysis to accommodate these statistical dependencies.
We coded 67 effect sizes of delinquent behavior post diversion across the 31 diversiontraditional processing comparisons. We analyzed these comparisons using two approaches. The first approach selected a single effect size per comparison based on a decision rule and the second used all 67 effect sizes, nesting these within comparison condition and independent evaluation. The decision rule for selecting effect sizes for the primary analysis was: (1) select the official measures of delinquency (i.e., drop self-report measures), (2) select effect sizes based on dichotomous indicators of delinquency, (3) select effect sizes closest to 12-months post diversion, (4) give preference to measures based on arrest over court appearances or convictions, (5) give preference to regression-adjusted effect sizes over raw effects for quasi-experimental designs. The logic of this decision-rule was to give preference to the most common outcome source (official measures) and most common outcome type (dichotomous) and the most common time post diversion (12-months). These first three are designed to increase the comparability of the effect sizes used in the analysis. The fourth rule gives preference for arrest over court appearances or convictions. The latter two represent deeper penetration into the juvenile justice system and as such are more likely to be affected by other factors, including prior actions by the system. Finally, the decision rule gave preference to regression adjusted effect sizes from quasi-experimental designs as these are adjusted for observed baseline differences between groups and as such are likely be less biased. This produced a single effect size for all but two comparisons (i.e., Little, 2015;Haines et al. 2012, Lewisham evaluation). In these latter two cases, the remaining effect sizes were averaged to produce a single selected effect size per comparison.
The most common method of effect size computation for these selected 31 effect sizes was based on dichotomous outcome data (e.g., 2 by 2 frequency data or proportions/percentages). This was the case for 29 of these 31 effect-sizes. For two of the treatment-comparison contrasts (two independent contrasts from the same study), the effect size was based on a study reported Cohen's d effect size, converted to a logged odds-ratio. For the full collection of 67 effect-sizes, 46 were based on dichotomous outcome data, 16 were based on means and standard deviations, 4 were based on a logistic regression coefficient, 2 were based on hand computations. All of the effect-sizes from the latter two methods were consistent in magnitude with the general distribution of effects (e.g., were not extreme). A meta-regression with robust standard errors compared each method to the null category of dichotomous outcomes. The mean effect-sizes based on means versus dichotomous outcomes were quite similar and not statistically significantly different. The mean for effect-sizes based on logistic regression models was slightly small (less positive result), but not significantly so. However, the mean for the two effect-sizes based on hand computations did differ significantly from the null category of dichotomous outcomes. These two effects were in the direction of diversion being less effective than the alternative so leaving these in the model makes our overall results more conservative.

Overall synthesis results
The results of the effect size analyses are shown in Table 4. The general pattern of evidence is positive, suggesting that police-led diversion modestly reduces the future delinquent behavior of low-risk youth relative to traditional processing. The random effects mean odds ratio across the 31 selected effect sizes and for all 67 effect sizes were roughly the same (mean odds ratio = 0.77 and 0.82, respectively, with 95% confidence intervals of 0.63 -0.95 and 0.66 -1.00, respectively). Both of these mean effect sizes are statistically significant at the .05 level. The size of the effect is modest, however. Assuming a 50 percent reoffending rate for the traditional processing condition, the odds ratio of 0.77 indicates a reoffending rate of 44 percent for the diverted youth.
Figures 2 -4 present the forest plots of the selected effects for random assignment studies and quasi-experimental studies. As can be seen in these figures, the general pattern of effects is to the left of the no-effect value of one, suggesting generally beneficial effects of diversion.

Analysis by risk of bias
The overall analysis ignores research design and risk of bias. To address this, we computed the results separately by design and risk of bias to examine the impact of these factors on the results for the 31 selected effect sizes and all 67 effect sizes (also shown in Table 4). Twentyfour of the 31 contrasts used a random assignment design and the mean odds ratio for this design was similar to the overall result (mean = 0.82, 95% C.I. = 0.66 -1.02). In contrast, the effect was larger (farther from the null value of 1) for the quasi-experimental designs (mean = 0.72, 95% C.I. = 0.40 -1.27). The results are similar for the analysis based on all 67 studies. These two analyses, however, don't fully address risk of bias as several of the random assignment studies had evidence of randomization failures or meaningful differences between groups at baseline. To account for this, we used our ratings for seventeen of the 24 random assignment contrasts coded as having low attrition (general or differential), low risk of selection bias with no indication of randomization failure or unbalanced groups at baseline. The overall mean for these effects is statistically significant and homogeneous (mean odds ratio = 0.81, 95% C.I. = 0.74 -0.88, τ 2 = 0.000). The effects for studies with high risk of bias were roughly the same in most cases, but somewhat larger for studies with an unclear risk of bias. All confidence intervals, however, substantially overlap suggesting that any differences across these categories of risk of bias are not meaningful. Furthermore, we examined 16 comparisons rated as having low risk of bias across all indicators of bias (see Table 3) and observed similar results (mean odds ratio = 0.82, 95% C.I. = 0.68 -0.98, τ 2 = 0.000). The results are similar for the analyses based on all effect sizes, although these analyses have more heterogeneity given the greater diversity in outcomes, increasing the confidence interval such that the mean is no longer statistically significant at a conventional level for the lower risk of bias studies. The effects for studies with high risk of bias are roughly the same in most cases, but somewhat larger for studies with an unclear risk of bias. All confidence intervals, however, substantially overlap suggesting that any differences across these categories of risk of bias are not meaningful.

Analysis by type of diversion program and country
As shown in Table 1, we categorized the diversion programs into one of four types: (1) diversion only (i.e., caution and release), (2) diversion with referral to services, either internal or external to the criminal justice system, (3) diversion with police-led restorative justice, and (4) other form of police-led diversion. Results comparing the first three of these categories are shown in Table 4. We are not showing results for the "other" category, as this would not be meaningful given that only two contrasts from one evaluation were in this category. Furthermore, as the "odd-ball" category it was not of substantive interest.
Although the results appear to vary across diversion type, these differences are not statistically significant. A meta-regression model in robumeta showed that both the difference between diversion only and diversion with referral to services and the difference between diversion only and diversion with restorative justice were not significant (t = 1.41, p = 0.180, t = 0.08, p = 0.939, respectively). This analysis is an indirect assessment of these differential effects given that different studies contributed to each category. However, eight studies compared both diversion only and diversion with referral to services to a common traditional processing condition. A meta-regression model was run testing the effect of referral to services with these studies (not shown in Table 4). The result was not statistically significant, but the effect modestly favored the referral to services condition, contrary to the difference shown in Table 4 (β = -.10, t = -0.71, p = 0.506, τ 2 = 0.0174). Thus, the current evidence does not suggest any meaningful difference across these variations in diversion type and the small-observed differences should not be interpreted as suggestive of differential effectiveness.
We also examined whether country of study affected the findings. Given the small number of non-United States studies, we compared the mean odds ratio for studies conducted in the United States to all studies conducted elsewhere (i.e., Australia, Canada, and the UK). Although the mean odds ratio indicated a modestly large beneficial effect for U.S.-based studies, the difference was small and not statistically significant (a mean odds ratio of 0.76 versus 0.85, respectively). Thus, the current evidence does not provide a basis for concluding that there is a meaningful difference in diversion effects between these different Anglophone countries.

Publication selection bias
We assessed for publication selection bias in three ways, each suggesting that our results are unlikely to be overly influenced by it. First, we compared the results for journal articles relative to other publication or manuscript forms, primarily technical reports and dissertation theses. A full 74 percent of our comparisons came from the latter, in itself a protection against publication selection bias. Table 4 shows the mean odds ratio for journals and for these other manuscript types. Surprisingly, the mean odds ratio for journals (mean = 0.91, 95% C.I. = 0.51 -1.61) was closer to the null value than the other manuscript types (mean = 0.73, 95% C.I. = 0.58 -0.92). With publication selection bias we would expect the opposite.
A second approach to assessing publication bias was a visual examination of a funnel plot ( Figure 5). This plot illustrates a slight asymmetry in the funnel with an absence of effects in the lower left part of the funnel. Given the location of the null value relative to the funnel, however, these absent effects are smaller (larger positive effects) than the mean. This helps explain why our third approach, the trim-and-fill method suggested no missing effects given that it only assessed for missing effects near or greater than one (Note: using Stata's metatrim, this necessitated flipping the sign of the logged odds ratios, changing the direction of the effects).
Although publication selection bias cannot be definitely ruled out, we assess a low risk of our findings resulting from publication selection bias given the large percentage of technical reports in this synthesis.

Summary of main results
The main objective of this review was to synthesize the evidence on the effectiveness of precharge interventions involving police warning or counsel and release, and cautioning schemes in reducing delinquent behavior. Specifically, we aimed to assess the overall effectiveness of police-initiated diversions strategies, compare the relative effectiveness of different types of police-initiated diversion, and finally examine whether effectiveness was related to characteristics of the youth.
The overall results suggest that police-led diversion modestly reduces future delinquent behavior of low-risk youth relative to traditional processing. The subset of studies judged to be free from any obvious risk of bias produced an overall mean effect size that was statistically significant and favoured the diversion condition. Analyses based on less methodologically sound studies produced more varied results that were not always statistically significant, but that had an overall mean effect that was roughly similar to that of the higher quality studies. The size of this effect, however, is modest. The overall mean result across all studies translates into a 6 percentage point reduction in reoffending from a benchmark rate of 50 percent.
We were able to categorize the diversion methods used across the studies into one of three based models: diversion only (often called counsel and release), diversion with referral to services, and diversion with police-led restorative justice. Rather unexpectedly, the effect was modestly larger for the diversion only contrasts, although the differences across the diversion types were not statistically significant. Furthermore, the comparison between diversion only and diversion plus referral-to-services modestly favored the latter in a sensitivity analysis that only included studies that compared both of these diversion types to a common traditional processing condition. As such, the existing evidence provides no basis for concluding that there are meaningful differences across the diversion types.
Finally, there was insufficient information and variability across the studies in terms of youth characteristics to allow for examination of differential effects for different types of youth. Thus we were unable to address this aim of the review.

Overall completeness and applicability of evidence
The positive findings of this review with regard to the benefits of diversion may have limited applicability to current juvenile justice processing. The majority of the evidence is from data collected during the 1970s and 1980s. Analyses showed no clear evidence of differential effectiveness over time, but the number of studies post-2000 (five studies) is too limited to be certain that the positive results found in this review will continue to hold given shifts in traditional-processing of that time period. Similarly, most of the evidence is based on data collected in the United States, with a handful of studies from Australia, the UK, and Canada. Thus, these results are clearly only applicable to these countries and additional evidence from outside of the United States is clearly needed to better inform policy in those contexts.

Quality of the evidence
Overall, the quality of evidence for this review is strong, with a majority of studies using random assignment designs. The availability of official measures of delinquency (e.g., police or court records) reduced problems of attrition and the potential for knowledge of the condition on the part of the youth from affecting outcome measurement. There was no evidence across these studies of selective reporting of outcomes and all studies reported a general measure of delinquency based on official data. Furthermore, the subset of studies judged by us to be at low-risk of bias produced a homogeneous overall effect favoring the diversion condition.

Limitations and potential biases in the review process
In general, the description of the delivery of the interventions was poor. In most cases, the reports did not give a sufficient description of the cautioning process that would facilitate a direct replication of the intervention. Specifically, it was rarely clear what was said to diverted juveniles by way of a caution and there was little information that described how the diversion disposal was perceived by the juvenile. Theoretically, a diversion presented as a 'last chance' could be perceived very differently from a diversion presented as a 'helping hand' or a 'way out'. The inclusion of this information could have benefitted our understanding of study heterogeneity.

Agreements and disagreements with other studies or reviews
The review that most closely matches the aims of our review was Wilson and Hoge (2013) who included both pre-charge and post-charge diversions, but separated these in their analyses. They found a slightly larger beneficial effect for pre-charge compared to postcharge diversions (which we did not test). Wilson and Hoge (2013) did not identify the 11 studies they defined as pre-charge interventions, but our analysis featured nine of the same references, suggesting that both reviews used rigorous search methods. However, it appears that Wilson and Hoge (2013) treated Lincoln, Teilmann, Klein, and Labin (1977) and  as discrete studies although they reported on the same evaluations: this may have led to a small bias in their results. We screened all references in the Wilson and Hoge (2013) paper for eligibility as part of our review. Consequently, as we included fourteen publications, we are confident that our review is an advancement on their work. Our results are also consistent with the broader Campbell systematic review and meta-analysis on any form of diversion by Petrosino et al. (2010). Overall, the findings of these three reviews are consistent in their support for the preventive effects of pre-charge diversion.

Authors' conclusions Implications for practice and policy
The findings from this systematic review support the use of police-led diversion for low-risk youth with limited or no prior involvement with the juvenile justice system. The pattern of evidence clearly favors diversion over traditional processing with no indication that diversion programs are harmful, that is, leads to increased delinquency. Although there are caveats to these conclusions that justify some equivocation in our conclusion (e.g., the age of the evidence, variations in statistical significance across different analyses), the pattern of evidence suggests that police departments and policy-makers should consider diversionary programs as part of the mix of solutions for addressing youth crime.

Implications for research
The main research need that emerges from this review is contemporary random-assignment studies comparing police-led diversion with traditional processing. There are simply too few studies conducted in the last decade to be confident that the positive results will hold for the current context of juvenile justice. Similarly, most of the evidence comes from the United States, indicating a clear need for additional high-quality research from other countries.
Although we did not systematically catalogue non-delinquency outcomes across these studies, they were few. This is unfortunate. The potential benefits of police-led diversion extend beyond reduced-risk of future delinquent behavior and include such possible benefits as improved attitudes toward the police (e.g., perceptions of police legitimacy), and a greater sense of procedural justice and fairness. Additionally, the lack of penetration into the juvenile justice system may help prevent negative school outcomes. We also were unable to capture race and gender effects, as they were largely unreported in the studies. The need for contemporary random-assignment studies also includes greater sensitivity to reporting subgroup data by race and gender and other equity criteria, in order for future meta-analyses to explore these factors as meaningful moderators (see Neyroud & Slothower, 2013).
In the included studies, labelling theory was the dominant theoretical rationale for why diversion might work. This helps explain why much of the research featured in our review is from the 1970s and 1980s; the heyday of labeling theory. Although rarely alluded to in the studies, more contemporary criminological theories also have relevance to diversionary practices: restorative justice and procedural justice in particular. Discretionary police use of diversion could be seen by a youthful offender as an act of 'fairness' that supports the perceived legitimacy of police. Alternatively, the demonstration of a restorative-led diversion may model effective conflict resolution for the youth. Exposure to procedural or restorative justice applied appropriately may form part of the mechanism that explains the reduction in reoffending observed in our review. Future theoretical work should consider applying these theoretical frameworks to diversionary practices to allow for a new wave of empirical research into potential change mechanisms related to diversion.
The role and value of services following diversion remains unclear. Our results found no statistically significant difference in the effectiveness of diversions with and without a referral to services. The diversity of service activities, the limited descriptions of these activities, and insufficient information on the extent of client compliance prevented a detailed analysis of how these factors might have influenced recidivism. Importantly, this prevented distinguishing the effect of diversion from the effect of services. As is common in many evaluation studies that attempt to reflect the realities of juvenile justice, it is the effectiveness of the referral to services that is tested, not the effectiveness of the services themselves. While this is the more valuable approach for decision-making in juvenile justice, it obscures the varied effectiveness of service activities.
The findings of this review are broadly supportive of police-led diversion's role in reducing juvenile reoffending. To date, no review about the effectiveness of police-led diversion in reducing adult reoffending exists. A preliminary scoping search by one of the authors has found far fewer evaluations of police-led diversion schemes for adults than for juvenile offenders. While the factors that influence offending vary between juveniles and adults, they are not entirely independent, suggesting that police-led diversion may have benefits for adults and should be explored further.

Notes
Notes about coding this comparison cnotes

Outcome (Dependent Variable) Coding Form
Code each eligible outcome or dependent variable using the form below. Note that you should code this only once for a variable that is measured at multiple time points. That is, recidivism measured at 3, 6, and 9-months is a single dependent variable. Code the characteristics of the measure using this form and the data for each measurement time point on the effect size forms.

Effect Size Coding Form
Code all effect sizes of interest using the form below, coding each effect size separately (i.e., with a different copy of the form or record in the database). Indicate the study ID, the comparison ID, and the dependent variable ID. Give each effect size within a study a unique idea, numbering sequentially (1, 2, 3 ...). There are several ways to compute effect sizes using the different tabs. ONLY USE ONE METHOD per effect size. If you have the raw means and also a regression coefficient for the same outcome from a model that adjusts for baseline differences, these are two different effect sizes. The different effect size computation methods are: 1. Means and standard deviations 2. Means and standard errors 3. Frequency of failures in each condition 4. Proportion of failures in each condition 5. Logistic regression coefficient for treatment effect dummy code 6. OLS unstandardized regression coefficient 7. OLS standardized regression coefficient 8. Independent samples t-test 9. Chi-square test (2 by 2, df = 1) 10. Point-biserial correlation coefficient 11. Phi correlation coefficient 12. Hand computation (e.g., using the online effect size calculator) (delinquent) AND ("law enforcement" OR diversion) AND (evaluat*) AND (recidivism)

Date:
October 20, 2016 Search terms: (youth OR child OR juvenile OR delinquent OR devian* OR student OR adolescent OR "young person" OR "young offender*" OR bully* OR "youthful offender") AND (diverted OR diversion OR caution* OR "caution plus" OR restorative OR "restorative caution" OR triage OR "final warning" OR reprimand OR "alternative* to custody" OR "pre-charge" OR "pre-caution" OR "pre-court" OR "pre-custody" OR "alternative program*" OR disposal OR disposition OR liaison OR Police-led OR police OR "law enforcement" OR "civil citation") AND (outcome OR evaluat* OR effect OR effectiv* OR experiment* OR quasi OR assessment OR RCT OR "random* control*") AND (recidivism OR arrest* OR rearrest* OR citation OR offend* OR reoffend* OR conviction OR reconviction OR adjudication OR adjudicate) Search terms: (youth OR child OR juvenile OR delinquent OR devian* OR student OR adolescent OR "young person" OR "young offender*" OR bully* OR "youthful offender") AND (diverted OR diversion OR caution* OR "caution plus" OR restorative OR "restorative caution" OR triage OR "final warning" OR reprimand OR "alternative* to custody" OR "pre-charge" OR "pre-caution" OR "pre-court" OR "pre-custody" OR "alternative program*" OR disposal OR disposition) AND (outcome OR evaluat* OR effect OR effectiveness OR experiment* OR quasi OR assessment OR RCT OR "random* control*") AND (recidivism OR arrest* OR rearrest* OR offend* OR reoffend* OR conviction OR reconviction OR adjudication) AND (police OR "law enforcement") Couldn't extract results directly into Zotero, so saved as text files to import into EndNote and transfer to Zotero. (youth OR child OR juvenile OR delinquent OR devian* OR student OR adolescent OR "young person" OR "young offender*" OR bully* OR "youthful offender") AND (diverted OR diversion OR caution* OR "caution plus" OR restorative OR "restorative caution" OR triage OR "final warning" OR reprimand OR "alternative* to custody" OR "pre-charge" OR "pre-caution" OR "pre-court" OR "pre-custody" OR "alternative program*" OR disposal OR disposition OR liaison OR Police-led OR police OR "law enforcement" OR "civil citation") AND (outcome OR evaluat* OR effect OR effectiv* OR experiment* OR quasi OR assessment OR RCT OR "random* control*") AND (recidivism OR arrest* OR rearrest* OR citation OR offend* OR reoffend* OR conviction OR reconviction OR adjudication OR adjudicate)

Date:
October 20, 2016 Search terms: (youth OR child OR juvenile OR delinquent OR devian* OR student OR adolescent OR "young person" OR "young offender*" OR bully* OR "youthful offender") AND (diverted OR diversion OR caution* OR "caution plus" OR restorative OR "restorative caution" OR triage OR "final warning" OR reprimand OR "alternative* to custody" OR "pre-charge" OR "pre-caution" OR "pre-court" OR "pre-custody" OR "alternative program*" OR disposal OR disposition OR liaison OR Police-led OR police OR "law enforcement" OR "civil citation") AND (outcome OR evaluat* OR effect OR effectiv* OR experiment* OR quasi OR assessment OR RCT OR "random* control*") AND (recidivism OR arrest* OR rearrest* (youth OR child OR juvenile OR delinquent OR devian* OR student OR adolescent OR "young person" OR "young offender*" OR bully* OR "youthful offender") AND (diverted OR diversion OR caution* OR "caution plus" OR restorative OR "restorative caution" OR triage OR "final warning" OR reprimand OR "alternative* to custody" OR "pre-charge" OR "pre-caution" OR "pre-court" OR "pre-custody" OR "alternative program*" OR disposal OR disposition) AND (outcome OR evaluat* OR effect OR effectiveness OR experiment* OR quasi OR assessment OR rct OR "random* control*") AND (recidivism OR arrest* OR rearrest* OR offend* OR reoffend* OR conviction OR recondition OR adjudication) AND (police OR "law enforcement")

Database:
ProQuest Dissertations and Theses UK and Ireland

Date:
August 11, 2016 Search terms: (youth OR child OR juvenile OR delinquent OR devian* OR student OR adolescent OR "young person" OR "young offender*" OR bully* OR "youthful offender") AND (diverted OR diversion OR caution* OR "caution plus" OR restorative OR "restorative caution" OR triage OR "final warning" OR reprimand OR "alternative* to custody" OR "pre-charge" OR "pre-caution" OR "pre-court" OR "pre-custody" OR "alternative program*" OR disposal OR disposition OR liaison) AND (police OR "law enforcement") AND (outcome OR evaluat* OR effect OR effectiv*eness OR experiment* OR quasi OR assessment OR RCT OR "random* control*") AND (recidivism OR arrest* OR rearrest* OR citation OR offend* OR reoffend* OR conviction OR reconviction OR adjudication OR adjudicated)

Date:
January 17, 2017 Search terms: "pre-court" OR "pre-custody" OR "alternative program*" OR disposal OR disposition OR liaison) AND (police OR "law enforcement") AND (outcome OR evaluat* OR effect OR effectiv* OR experiment* OR quasi OR assessment OR RCT OR "random* control*") AND (recidivism OR arrest* OR rearrest* OR citation OR offend* OR reoffend* OR conviction OR reconviction OR adjudication OR adjudicated)

Yield: 26
Database: youth OR child OR juvenile OR delinquent OR devian* OR student OR adolescent OR "young person" OR "young offender*" OR bully* OR "youthful offender" AND diverted OR diversion OR caution* OR "caution plus" OR restorative OR "restorative caution" OR triage OR "final warning" OR reprimand OR "alternative* to custody" OR "pre-charge" OR "pre-caution" OR "pre-court" OR "pre-custody" OR "alternative program*" OR disposal OR disposition OR liaison AND police OR "law enforcement" AND outcome OR evaluat* OR effect OR effectiv* OR experiment* OR quasi OR assessment OR RCT OR "random* control*" AND recidivism OR arrest* OR rearrest* OR citation OR offend* OR reoffend* OR conviction OR reconviction OR adjudication OR adjudicated   Notes: Direction of odds ratio coded so that values less than 1 reflect less recidivism for the diversion group relative to the comparison group. All analyses performed in Stata using the robumeta command that produces robust standard errors accounting for clustering (statistical dependencies across effect sizes). Selected Effect Sizes analyses had one effect size per treatment-comparison contrast, selecting the most general measure of official recidivism measured nearest 12-months post police encounter. There were 31 such contrasts nested within 19 independent studies, across 14 publications. All Effect Size analyses used all available effect sizes. There were 67 effect sizes available for analysis across the 31 contrasts.

RAND Documents
The traditional homogeneity test, Q, is not produced by robumeta. However, 2 provides information on heterogeneity.
When 2 is zero, the distribution is homogeneous. As 2 increases, it indicates increasing levels of heterogeneity. Running the above analyses using David Wilson's meanes macro for Stata suggests that for these data, any 2 above 0.05 reflects statistically significant heterogeneity.

Roles and responsibilities
• Content: David Wilson has extensive background knowledge on juvenile justice programs. Iain Brennan has experience with executing a research grant on a restorative justice program and has recently completed an evaluation of a police-led diversion scheme. Ajima Olaghere, along with David Wilson, are currently working on a meta-analysis focusing on trauma-informed interventions for at-risk and justiceinvolved youth.
• Systematic review methods: David Wilson has extensive expertise in systematic review methods. Ajima Olaghere has worked on meta-analyses with David Wilson and previously worked on systematic reviews. Iain Brennan has led a systematic review of interventions to reduce violence in licensed premises.
• Statistical analysis: David Wilson has developed tools that are in wide use for performing the statistical analyses related to meta-analysis. He also authored a book on these methods with Mark Lipsey.
• Information retrieval: David Wilson, Ian Brennan, and Ajima Olaghere all have experience performing systematic searches on various topics and retrieving studies and documents for review.

Sources of support
This systematic review is supported by funding from the Jacobs Foundation.