PROTOCOL: Organised crime groups: A systematic review of individual-level risk factors related to recruitment

3.1.1 Study design

This systematic review aims at identifying and evaluating existing knowledge on the risk factors leading to recruitment to OCGs. Because recruitment into OC cannot be the object of experimental interventions, experimental and quasi-experimental studies are not relevant to the aim of this systematic review. This review will examine empirical evidence resulting only from studies using an observational research design.

To be included, studies must report on recruitment into OCGs as one of the main objectives of the analysis, and provide details on the sampling strategy, data collection and the type of analysis conducted, that is, the relation between a risk factor and recruitment into OCG. This review will exclude literature reviews, theoretical and conceptual contributions and editorial pieces. Based on the recommendations of the anonymous reviewers and of our understanding of the field, this systematic review will retrieve and screen both quantitative and qualitative studies. Quantitative studies will undergo the selection process described in the Statistical Procedures subsection. Qualitative studies will be systematically retrieved, screened for inclusion and coded. In accordance with current Campbell Collaboration policy on systematic reviews, they will be used to inform and contextualise the evidence and findings of the quantitative studies.

For quantitative synthesis, we will rely on studies with variability in recruitment into OC, measuring and comparing at least two groups (e.g., OC prisoners and non-OC prisoners). The review will include studies based on longitudinal and cross-sectional designs. To be included in a meta-analysis, each study must report at least an effect size, or allow calculation of an effect size based on the information provided.

We will not exclude studies based on their geographical scope or year of publication. In addition, we will not exclude studies based on their quality. We will evaluate the risk of bias resulting from study quality using a risk-of-bias tool adapted from Higginson et al. (2018) recent Campbell systematic review and PROBAST tool for prediction studies (see below, Quality assessment subsection).

3.1.2 Types of OCGs

As discussed in Section 1, the definition of OC has generated a long-lasting debate in the literature. To favour inclusion of the largest number of possible studies, this systematic review will rely on the definition provided by Article 2 of the United Nations Convention against Transnational Organized Crime (United Nations, 2000):

“Organized criminal” group shall mean a structured group of three or more persons, existing for a period of time and acting in concert with the aim of committing one or more serious crimes or offences established in accordance with this Convention, in order to obtain, directly or indirectly, a financial or other material benefit.

This definition includes a variety of OCGs, ranging from traditional mafias to DTOs and adult gangs. Given the important share of adult offenders in OC and the relevance of the ties to the legitimate world, the systematic review will exclude youth (street) gangs, prison gangs and terrorist groups. The literature generally considers youth street gangs as different from OCGs (Decker & Pyrooz, 2014). Furthermore, recent systematic reviews have already assessed the factors leading to youth gang membership (Higginson et al., 2018; Klein & Maxson, 2006). As for prison gangs, while some are extension of criminal organisations active outside the prison, others exist and establish themselves in the isolation of the prison setting. For this reason, this study does not consider prison gangs, as they occur in a specific and institutionalised settings, and therefore individuals’ recruitment is influenced by different contextual factors (Blevins, Johnson Listwan, Cullen, & Lero Jonson, 2010; Wood, Alleyne, Mozova, & James, 2014). Furthermore, while there is a relevant literature on prison gangs, this field is mostly separate from the literature on OC, which emphasises the social embeddedness into the legitimate world. The exclusion of terrorist groups is due to the ideological/political motivation of such organisations. Furthermore, a Campbell systematic review on the factors leading to radicalisation and recruitment into terrorism is currently ongoing (Litmanovitz et al., 2017).

3.1.3 Types of risk factors

This systematic review aims at identifying the risk factors associated to recruitment to OCGs. With regards to the measurement of the risk factors, we will only include measures taken at the individual level. Among the types of factors identified by our review, we expect to include demographic, social, economic, psychological and criminal history factors.

To consider a variable as a risk factor, the variable must occur prior to the outcome (Murray, Farrington, & Eisner, 2009). The risk factor therefore must precede the outcome, that is, OCG membership. Some factors, however, may be considered as preceding the outcome even if included in cross-sectional studies, as they do not vary over the life course (e.g., sex and race). Some scholars argue that such time-invariant factors cannot be considered as risk factors due to their fixed nature (see Murray et al., 2009). However, this systematic review will consider as risk factors for OCG membership not only those predictors resulting from longitudinal studies—measuring the factors preceding the occurrence of the outcome—but also time-invariant factors estimated from cross-sectional studies. Self-reported retrospective data assessing risk factors preceding the outcome will also be considered, though they present some biases as they are based on individual's recall of past events (Murray et al., 2009). This choice is driven by the goal to include as many studies as possible given the lack of any systematic review on the recruitment into OC. Due the difficulties of collecting longitudinal data on OCGs, we expect to find few longitudinal studies on OCG membership (see Bruinsma, 2015).

We recognise the difficulty of establishing causation for risk factors deriving from observational designs. We acknowledge that the option may cause some factors to be measured only after the recruitment into OCGs has already occurred (e.g., unemployment, low education). In line with previous systematic reviews (Higginson et al., 2018; Klein & Maxson, 2006), this systematic review will attempt to classify as predictors the risk factors measuring conditions preceding the recruitment into OCGs and as correlates the risk factors measuring conditions occurring simultaneously or after the recruitment. Effects for predictors and correlates will be reported separately.

3.1.4 Types of outcome measures

The outcome of interest in this systematic review is the recruitment into OCGs. As discussed in Section 1, recruitment refers to the different processes leading individuals to the stable involvement into OCGs. We will not differentiate among different forms of recruitment to OCGs. Therefore, we will include studies considering recruitment, affiliation and other forms of stable involvement. If relevant, the impact the risk factors on different forms of recruitment will be analysed through a moderator analysis.

The systematic review includes only studies that measure recruitment into OCGs at the individual level, measured with either a dichotomous or a categorical variable. In the case of a categorical variable (e.g., OCG membership, former membership, nonmembership, etc.), separate meta-analyses will be carried out for each paired OCG–non-OCG for which effect sizes can be extracted (e.g., OCG membership vs. former membership), with the outcomes being compared and discussed in the review.

The review will include self-reported, peer-reported, practitioner-reported and police-reported measures of individual OCG membership. If applicable, we will assess heterogeneity due to measurement methods with moderator analysis.

3.2.1 Search terms

This review relies on a threefold query structure that ensures systematic, thorough and efficient results. The queries incorporate all aspects that are relevant to the risk factors relating to the recruitment into different types of OCGs. The search terms from each of the three main categories (i.e., OCGs, factors and recruitment) combined formed the queries (Figure 1). The Boolean Operator “OR” connected keywords pertaining to the same category, while the Boolean Operator “AND” connected keywords from different categories (see Table A1 in Appendix A). This query structure ensured to retrieve all the studies containing at least one term from each word category.

OPEN IN VIEWER

1

Query structure

Given the transdisciplinary approach of this systematic review, the search for relevant studies relies on 12 databases relating to different research disciplines. ² The suitable studies encompass academic and grey literature written in English, French, German, Italian and Spanish, and pertaining to social, psychological and economic disciplines. ³ No limitations apply as to their year of publication or geographic origin. Both academically published and grey literature is being considered. Table 1 reports the list of databases indicating in which language the search was conducted and which search technique was applied. When available, the preferred technique was to search title, abstract and keywords.

To validate the search terms and queries, the research team attended two meetings with a librarian to ensure the inclusion of all databases relevant to this systematic review. Table A3 in Appendix A shows the list of databases and the related queries used to perform the research (Table A4).

Apart from identifying relevant literature through scientific databases, researchers will also contact experts to receive suggestions on relevant studies that may not have been included in the systematic review yet. First, several renown authors in the field of OCGs will be contacted: Jay Albanese (Virginia Commonwealth University, USA), Paolo Campana (University of Cambridge, UK), Scott Decker (Arizona State University, USA), Edward Kleemans (Vrije University of Amsterdam, NL), Klaus Von Lampe (John Jay College of Criminal Justice, USA), Carlo Morselli (University of Montreal, CA), Arthur Lurigio (Loyola University Chicago), Letizia Paoli (Katholieke Universiteit Leuven, BE), David Pyrooz (University of Colorado Boulder, USA), Sonja Wolf (Centro de Investigación y Docencia Económicas, MEX). Second, this list of experts will be implemented on the basis of the screening of the literature done by this systematic review. More precisely, the authors of the literature included after the full text screening will be also contacted.

The research team will further identify relevant literature from the bibliographies of the studies that will be selected for full-text screening. As for the selection of studies from the database searches, these additional studies will be assessed for full-text eligibility.

3.3.1 Preparatory activities

The review process will incorporate all the studies retrieved through database search. Metadata for each study will be imported to the Covidence online platform, which provides an environment to manage and conduct systematic reviews. ⁴

After the removal of duplicate entries, the research team will be trained for the screening of relevant studies. The training will include a comprehensive briefing on the purpose and scope of the systematic review, followed by a tentative screening phase during which each reviewer will independently conduct the title-and-abstract screening of a set of 100 studies. The results will then be discussed among all researchers to reveal divergent interpretations and other issues, and maintain common criteria for the inclusion of studies in the systematic review.

To ensure reliability, throughout the screening process two reviewers will screen each document. A third researcher will settle divergent screening decisions, where necessary in consultation with the full review team.

3.3.2 Eligibility screening criteria

As a first step, the screening will be based on the information reported in titles and abstracts. If the document is relevant in light of the aim of this systematic review, that is, investigates recruitment into OCGs as main aim of the study, it will be filtered in. If the document is irrelevant, it will be filtered out. If the information report in the title and abstract do not allow to include/exclude the document, the study will be kept for full-text screening. In other words, we will keep every study that cannot be dropped, rather than the other way around.

As a second step, the screening will be based on the information reported in the full-text. ⁵ Each document will have to meet all the eligibility criteria listed in the “Eligibility screening form” (see Appendix B). The “Eligibility screening form” will guide the selection process by including only empirical documents that are focused on OGCs as defined in the paragraph “Types of organised crime groups”, examining clearly defined factors leading to recruitment into OCGs at an individual level. If the document meets all the eligibility criteria, it will be filtered in. If none of the eligibility criteria can be definitively answered in the positive based on the full-text screening, the study will be filtered out. While in the previous phase we have favoured inclusivity, in this phase every criterion needs to be conclusively met, on penalty of study exclusion.

3.3.3 Study coding

The quantitative, mixed-method and qualitative studies that met all full-text screening criteria will be independently coded by two reviewers based on a detailed coding guide (see Appendix B). Mixed-method studies will be coded two times, one each for their empirical qualitative and quantitative sections. Item-based questionnaire-style coding documents have been used in previous reviews (e.g., Higginson et al., 2018). Types of OCGs will be initially coded into different categories, that is, mafias, drug trafficking groups, adult gangs and outlaw motorcycle gangs, and a residual category of other criminal groups (see items 12 and 13 of Table A4 in Appendix C). Such categories may be redefined based on the types of OCGs addressed by included studies and will also serve to conduct moderator analysis. The results will be compared and any coding conflict will be resolved through exchanges with a third reviewer.

3.5.1 Effect size metric and calculations

To perform the formal meta-analysis, the different statistical measures reported in the quantitative and mixed-method studies must be transformed into comparable effect size measures. If effect sizes are not directly included in the studies, we will extrapolate them based on reported statistics. Our coding document contains a subsection to help with this process (see items 35–57 in Appendix C). If the studies do not contain the necessary data for effect size extrapolation we will contact the authors of the studies.

We will code all effect sizes from our screened studies. Thanks to the coding guide, we will be able to group them based on several dimensions relevant for synthesis and interpretation. In particular, each effect size will be coded based on its document of origin, the nature of the two groups the effect was assessed on (e.g., OCG affiliates for the OCG and general criminals for the non-OCG), and the risk factor it refers (items 1–4, 18–19 and 35 of our coding guide, respectively). We will carry out the statistical synthesis for all the comparable effect sizes between similar pairs of groups. Risk factors will also be classified based on their focus domain (sociodemographic, psychological, etc.) for easier comparation, synthesis and presentation (see item 36 in our coding guide).

Effect sizes can be calculated using three categories of statistics: Group means, for continuous variables; risk-based association measures between two binary variables; and correlation measures between two either continuous, ordinal or categorial variables. We expect studies in our review to report their results using mainly group means differences and standard deviations for continuous variables, odds ratios for binary variables, and correlation measures, such as Pearson's correlation or regression coefficients. These three different forms of data will be transformed into effect sizes in the form of log odds ratios in order to perform meta-analysis.

The logic of using log odds ratios as a common statistic is twofold. First, both odds ratios and log odds ratios are symmetrical across the two variables they reference. We expect the studies in our review to often consider OCG recruitment as an independent variable and what we would call a risk factor as dependent variable, in particular when reporting the difference in a continuous variable between an OCG group and a non-OCG. Once the same statistical information is transformed into an odds ratio, the issue of directionality disappears: We can interpret the resulting effect size as the likelihood of OCG recruitment between groups with and without the risk factor, as intended for review.

Second, log odds ratios have the property of symmetry around their null value. While odds ratios are defined between 0 and positive infinity with a null value of 1 and asymmetrical standard errors, log odds ratios “normalize” the null value to 0 and are defined between negative infinity and positive infinity, with symmetrical standard errors regardless of sign (see Borenstein, Hedges, Higgins, & Rothstein, 2009, p. 35). This makes it easier to use them for analysis.

The conversion to log odds ratios entails, respectively:

1.

For continuous variables for which group means and variance are reported, calculating first

Cohen's d:

d = x ̅ OC − x ̅ NOC ( ( n OC − 1 ) s OC 2 + ( n NOC − 1 ) s NOC 2 n OC + n NOC − 2 )

d 's standard error SE _d :

SE d = ( n OC + n NOC n OC n NOC + d 2 2 ( n OC + n NOC ) )

Where

OPEN IN VIEWER

x ̅ OC	Mean value of the variable of interest in the OC sample
x ̅ NOC	Mean value of the variable of interest in the non-OC sample
n OC	OC sample size
n NOC	non-OC sample size
s OC 2	Variance of the variable of interest in the OC sample
s NOC 2	Variance of the variable of interest in the non-OC sample

These measures will then be used to calculate:

Log odds ratio:

log OR = π d 3

Log OR standard error:

SE log OR = π SE d 3

2.

For binary variables for which contingency tables or odds ratios are reported, calculating:

Log odds ratio:

log OR = ln OR = ln ( n 1 OC n 0 NOC n 0 OC n 1 NOC )

Log OR standard error:

SE log OR = 1 n 1 OC + 1 n 0 OC + 1 n 1 NOC + 1 n 0 NOC

Where

OPEN IN VIEWER

n 1 OC	Number of OC individuals with the variable of interest
n 0 OC	Number of OC individuals without the variable of interest
n 1 NOC	Number of non-OC individuals with the variable of interest
n 0 NOC	Number of non-OC individuals without the variable of interest

3.

For continuous variables for which only Pearson's correlation is reported, calculating:

r's standard error SE _r :

SE r = 1 − r 2 n − 1

Then, using r and SE _r , calculating

Cohen's d :

d = 2 r 1 − r 2

d's standard error SE _d :

SE d = 2 SE r ( 1 − r 2 ) 2

And finally, using d and its SE _d to calculate the log OR and its SE _{log OR} using the same formulas used for continuous variables for which group means were reported.

Another source of effect sizes for review are coefficients from regression models reported in our studies. As regression coefficients are sensitive to the set of covariates the models use, they need to be adjusted before analysis in order to remove covariate effect. For OLS regression models, this entails calculating the semi-partial correlation r _sp , which we will do following procedures suggested by Aloe and Thompson (2013). We will then calculate the log odds ratios following the same procedure used for “regular” product–moment correlations. In logistical regression models, regression coefficients are already presented as log odds ratios. In this case we will simply code them as they are, together with their standard error, to be directly used in meta-analysis.

Figure 2 synthetically represents how effect size extraction and conversion will be carried out.

OPEN IN VIEWER

2

Effect sizes extraction by type of relation. [Color figure can be viewed at wileyonlinelibrary.com]

3.5.2 Method of synthesis

If at least two studies provide effect sizes for the same predictor or correlate, we will conduct a random-effects meta-analysis on that factor using inverse variance weighting. This way we will calculate the overall weighted mean effect estimate of each separate factor on OCG recruitment. The result will be presented in a forest plot with 95% confidence intervals. In keeping with previous reviews (Hawkins et al., 2000; Higginson et al., 2018), we will carry out meta-analysis using log odds ratios, then convert the results into odds ratios for presentation. As each effect size will refer to the pairing of an OCG and a non-OCG (e.g., involved in an OCG vs. general criminals), we will only carry out meta-analyses among effect sizes that measure the same factor for the same group pairing. On the other hand, meta-analyses that refer to the same factor across different OCG and non-OCG pairings will be presented in the same forest plot but not further synthesized.

If the paucity of the studies analysing a factor prevents us from completing a formal meta-analysis for that factor, we will present forest plots and confidence intervals for each factor without attempting a statistical synthesis. This may be the case of studies reporting on factors similar in nature but for which a meta-analytic synthesis would be hardly meaningful (e.g., unemployment and low socioeconomic conditions).

3.5.3 Assessment and investigation of heterogeneity

While the main scope of a meta-analysis is to assess the global effect of a factor on a given phenomenon with a degree of precision superior to that of any single study, the study of heterogeneity can provide indications on how to interpret that effect (while quantitatively describing the efficacy of the calculated global effect size; Borenstein et al., 2009). For instance, a small degree of heterogeneity emerging when comparing otherwise similar studies made on different populations tells us that population differences do not appear to play a large part on that factor-outcome association—a finding which may otherwise have flown under the radar. For this reason, it is useful both to attempt meta-analyses between studies we suspect to have some degree of methodological incomparability, and to give a statistical value to that incomparability. To this end, we will assess heterogeneity between studies with the Q, I ², τ and χ ² statistics.

3.5.4 Sensitivity and subgroup analysis

If enough studies are available, we are going to conduct sensitivity analyses to further assess the effect of study heterogeneity and risk level on the results of the review. In particular, the relevant subgroups of studies are going to be selected based on bias risk (as assessed in questions 58–85 of our coding document, see Appendix C) and geographic scope of the study. We may also choose to conduct complete subgroup analyses on the same or different subgroups to formally explore how the study variables the group division was based on impact global effect sizes.

3.5.5 Assessment of publication bias

To assess potential publication bias in each subgroup we will use funnel plots, a specialised form of scatter plots used in meta-analysis to visually identify publication and other bias (Sterne, Becker, & Egger, 2006) Furthermore, publication bias will be adjusted with trim and fill analysis, aiming to “both identify and correct for funnel plot asymmetry arising from publication bias” (Higgins & Green, 2011). These steps will follow the methodology suggested by Rothstein, Sutton, and Borenstein (2005).