Skip to main content
SearchLoginLogin or Signup

Selection and influence: A meta-analysis of the association between peer and personal offending

Gallupe, O., McLevey, J., & Brown, S. (2019). Selection and influence: A meta-analysis of the association between peer and personal offending. Journal of Quantitative Criminology, 35(2), 313-335. http://doi.org/10.1007/s10940-018-9384-y

Published onMay 23, 2018
Selection and influence: A meta-analysis of the association between peer and personal offending
·

Abstract

Objectives: Whether people are affected by the criminal behavior of peers (the “influence” perspective) or simply prefer to associate with others who are similar in their offending (the “selection” perspective) is a long-standing criminological debate. The relatively recent development of stochastic actor-oriented models (SAOMs – also called SIENA models) for longitudinal social network data has allowed for the examination of selection and influence effects in more comprehensive ways than was previously possible. This article reports the results of a systematic review and meta-analysis of studies that use SAOMs to test for peer selection and influence effects. Methods: A systematic review and 3-level random effects meta-analysis of studies that have used SAOMs to test selection and influence dynamics for offending behavior. Results: There is support for both influence (mean log odds ratio=1.23, p<.01, 21 effects, pooled n=21,193) and selection dynamics (mean log odds ratio=0.31, p<.01, 28 effects, pooled n=21,269). Type of behavior, country, and the year of the first wave of data collection are found to moderate the influence effect; no significant moderation effects are found for peer selection on offending. Conclusions: People are both influenced by the offending of their peers and select into friendships based on similarity in offending.

Acknowledgements

The authors would like to thank Nate Doogan for his valuable comments on an earlier version of this article.

This is a post-peer-review, pre-copyedit version of an article published in the Journal of Quantitative Criminology. The final authenticated version is available online at: http://doi.org/10.1007/s10940-018-9384-y

Keywords: Selection; influence; crime; meta-analysis; stochastic actor-oriented models

Whether the association between individual and peer offending is due to the influence of peers (as argued by social learning theorists such as Akers, 2009; Sutherland, 1947) or is simply a product of people preferring to associate with others who are similar to themselves (as argued by control theorists such as Gottfredson and Hirschi, 1990; Hirschi, 1969) has long been debated and evidence has been presented for both sides. For example, the finding by the Gluecks (1950, p. 163) that 98% of delinquents in their sample had friends who were also delinquent (compared to only 7% for non-delinquents) has been interpreted as supporting the idea that delinquent similarity is due to selectivity in friendship choices. Supporting the peer influence perspective, the meta-analysis by Pratt et al. (2010) found significant mean effects across 133 studies for the various components of social learning theory. In this article, we present a meta-analysis of research that explicitly tests the selection and influence hypotheses using relatively recent methodological advances (stochastic actor-oriented models) that were designed for this purpose. In addition to summarizing the evidence for these behavioral dynamics, this study has implications for empirical research in terms of whether peer selection on offending, influence, or both dynamics should be included in statistical models of offending behavior.

1. TRADITIONAL TESTS OF PEER INFLUENCE

The effect of peers on individual behavior has generally been tested by asking survey respondents to indicate the extent to which their peers engage in offending behavior and the extent to which they are personally involved in similar behaviors. Correlational tests are then conducted to measure the relationship between the two. Typically, peer influence researchers interpret statistically significant positive relationships as evidence supporting differential association theory (Sutherland, 1947) and/or social learning theory (Akers, 2009). This approach has garnered extensive empirical support (see, for example, Akers and Jensen, 2006; Pratt et al., 2010). However, critics suggest that “projection bias” – the tendency for people to overestimate the similarity between their behaviors and those of their peers – undermines these interpretations. As Haynie (2001, p. 1032) notes, “when asked to report their peers’ delinquent behavior, adolescents show a proclivity to report their own delinquent behavior. If this is the case, it casts considerable doubt on the frequently observed relationship between peers’ delinquency and adolescents’ own reports of their delinquency.” Gottfredson and Hirschi (1990, p. 157) take a more extreme view by stating that measures of peer delinquency are no more than “another measure of self-reported delinquency” (see also Gottfredson and Hirschi, 1987; Hirschi, 1969).

Some research has attempted to address selection effects by including measures of prior offending on current offending. The idea is that, by controlling for time 1 offending in models predicting time 2 offending, unmeasured factors such as friendship selection dynamics are accounted for. If the effect of peer behavior on individual behavior is significant even in statistical models that include prior offending, the influence perspective is thought to be supported. But in these models, peer effects (and usually all other effects) are typically quite small, which suggests that peer effects are not nearly as strong when accounting for selection processes. For example, Weerman and Hoeve (2012) found that the effect of peer offending was reduced from 0.31 to 0.25 for girls and from 0.18 to 0.05 when prior offending was added to the models. Similarly, Haynie and Osgood (2005) found that the standardized effect of peer delinquency on personal delinquency dropped from 0.14 to 0.05 with the inclusion of prior delinquency. However, this approach suffers from methodological issues of its own. As Haynie and Osgood (2005, p. 1119) note, lagged measures of delinquency tend to be “too strong a control for selection factors, thereby underestimating the contribution of peer relations to delinquency.” This ‘washout’ effect associated with lagged dependent measures is not limited to delinquency (Achen, 2000), but it does suggest that the common approach to accounting for selection effects in criminology is less than ideal.

Furthermore, lagged measures of offending are an imprecise proxy for peer selection. They are thought to capture all unexplained variability in the dependent variable of which similarity in offending is only one of many potential factors. So, a significant lagged measure of offending can only tell us that previous offending is related to current offending; it cannot tell us whether people tend to become more similar to their friends over time in terms of their level of offending or whether similarity in offending behavior is driven by the tendency for friendships to form over time among people that already offend at similar levels. In essence, the standard approach to longitudinal modeling leaves the specific selection and influence processes implicit. Finally, regression/structural equation modeling approaches to testing peer dynamics, even when employing full network data and therefore not subject to projection bias, violate independence assumptions since “the individual respondent, whose behavior figures centrally in one observation, will also appear among the peers for other observations” (Steglich et al., 2010, p. 342). In other words, the same individual can be part of the reference group upon which measures of peer offending are based for multiple others in the data set. Longitudinal network models, outlined in the following section, have the capacity to both isolate selection and influence dynamics and appropriately handle dependence among actors.

2. STOCHASTIC ACTOR-ORIENTED MODELS TO TEST SELECTION AND INFLUENCE EFFECTS

Stochastic actor-oriented models (SAOMs – also called SIENA models) were developed by Snijders and colleagues and implemented in the SIENA package (Simulation Investigation for Empirical Network Analysis - https://www.stats.ox.ac.uk/~snijders/siena/) for R. These models allow the researcher to test patterns and predictors of tie evolution using simulation methods with longitudinal sociometric data. They are ideally suited for examining selection and influence effects in ways that get around the aforementioned problems of regression-based approaches. They are not subject to projection bias as every individual in the network (e.g., a school or classroom) is asked to self-report their own offending; self-reported offending by members of the peer group can then be linked back to the individual via friendship nominations. As these are measured at multiple time points, patterns of network ties and their relationship to personal characteristics such as delinquency can be tested longitudinally. As with all social network methods, the interdependence among cases does not violate model assumptions but is expected and explicitly modeled.

At the most basic level, SAOMs allow for the simultaneous testing of the effects of predictors on tie formation/dissolution and behavioral congruence among peers while controlling for the other effects. They start from the assumption that behaviors and tie formation are influenced by the structure of the network as well as characteristics of the individuals within the network. The continuous-time transition of the observed network state at one discrete point in time to a subsequent point in time is modeled via a large number of simulations in which changes to the network are broken into individual “micro steps” (Steglich et al., 2006, p. 49) in which one actor has the opportunity to change one tie or behavior. The first observation of the network serves as the starting point for the simulations. SAOMs can be considered Markov chain processes (Steglich et al., 2010, p. 355). For much more extensive descriptions of the method, see Ripley et al., 2018, Snijders et al., 2010, Steglich et al., 2006, 2010.

There are numerous effects that can be included in SAOMs to explicitly test selection and influence (see Ripley et al., 2018 for further elaboration and formulae). Typically, studies use the ‘average similarity on X’ parameter to examine influence dynamics and ‘peer similarity on X’ parameter to examine selection dynamics. The interpretation of this type of influence effect parameter relates to the tendency for respondents to become more or less like their peers over time. For example, a significant positive average similarity parameter for violent crime would indicate that network members tend to adjust their levels of violence to more closely match that of their peers (interpretations of other influence parameters are very similar). Selection effect parameters are interpreted as the tendency for people to become friends with others who are already like themselves. For example, a significant positive similarity (selection) parameter for violent crime would indicate that network members tend to choose friends who report similar levels of violence. Essentially, influence measures the extent to which people become more like their friends over time while selection measures the extent to which people are attracted to others who already engage in similar levels of offending.

3. CURRENT STUDY

The goal of this study is to assess the state of the research on peer influence and selection effects related to offending behavior by conducting a systematic review and meta-analysis. The focus is strictly on studies that use SAOMs as they provide the most sophisticated analytic approach to testing these dynamics. SAOMs are ideally suited to testing selection effects as they have the capacity to track the likelihood of friendship ties being made in response to similarity in behavioral characteristics, however, it can be argued that they are less well suited to testing the specific perceptual peer influence dynamics outlined by Sutherland (1947) and Akers (2009). Akers (2009) and others (see, for example, McGloin and Thomas, 2016) have argued that perceptions of peer behavior have important consequences for individual behavior even if these perceptions are inaccurate. From this perspective, it is inconsequential if one does not know what level of offending their peers are actually involved in; the important thing is thinking that friends are engaged in a particular level of offending. That is, people are influenced by the amount of offending they perceive their friends to commit, not the amount they actually commit. The argument from proponents of social network studies that directly asking peers about their behavior constitutes a more ‘objective’ measure of peer offending does not undermine the importance of perceptions; they simply place more emphasis on the actual behavior of peers than perceptual studies. But despite research showing that people do not accurately perceive the offending behavior of friends (e.g., Rebellon and Modecki, 2013; Boman et al., 2012), ‘objective’ peer offending measures do tend to overlap with traditional perceptual measures of peer offending (e.g., Boman and Ward, 2014, p. 570 found that the correlation between a perceptual measure of peer offending and the self-reported delinquency of peers was 0.471). Despite differing opinions on which method is better suited to testing peer influence effects, the fact remains that only one, the perceptual method, has been meta-analyzed (see Pratt et al., 2010). It would be difficult to continue to take social learning theories seriously if the dynamics they describe only held up when examined using one particular approach. Therefore, this study can be viewed as a complement to the meta-analysis by Pratt et al. that offered support for social learning theory by aggregating the results of mostly perceptual peer influence research. If our study displays similar results despite focusing on studies that employ a different methodological approach and that control for factors that are not well captured in perceptual studies (i.e., selection), it will highlight the robustness of the criminogenic peer influence effect. However, discrepant findings would suggest that the significant results of previous peer influence research using perceptual methods may simply be a methodological artefact.

4. METHODS

In this section, we outline inclusion/exclusion criteria for the systematic review and meta-analyses, the search and screening strategy, moderating variables, and analytic approach. The study methods followed a pre-established protocol.

4.1. Inclusion Criteria

There were a number of specific criteria that previous research had to meet to be included in the systematic review:

  1. The study had to employ SAOMs. Inherent within this requirement are a number of characteristics that must be met in order to use these models.

    1. The sample had to be drawn from a network (or group of networks) with defined boundaries such as a school, grade level, or classroom. No limitations were placed on the type of network (school, community).

    2. The study had to use sociometric data, meaning that ties between the members of a network were explicitly measured. Ties can be of any form (e.g., co-offending, sexual partners, event attendance), but in this study are limited to friendships. This is because the selection versus influence debate centers around whether people are influenced by their friends or whether they select friends based on particular characteristics.

    3. The study had to use longitudinal data. The number of waves is not critical as long as it is at least two.

  2. The study had to examine offending behavior. To avoid confusion over what this includes, we limited the study to behaviors that were clearly illegal in all countries represented in our analysis. Establishing this conceptual boundary meant excluding such behaviors as substance use, where legal status varies across national contexts. We feel that the constraints imposed on the included acts were the result of a clearly defined research question while at the same time being sufficiently broad that it does not fall into the category of “correct but noninformative” (Ioannidis, 2016, p. 501). But for this reason, this meta-analysis should be interpreted with the caveat that other behaviors cannot necessarily be interpreted in the same way and we encourage other researchers to conduct similar meta-analyses focusing on other behaviors.1

  3. The study manuscript must have been available by May 2017 when our updated database searches were conducted.

While not part of the inclusion criteria outlined in the protocol, an additional constraint was placed on the influence analysis. Only average similarity effects were used in the meta-analysis of influence dynamics. Average similarity is one of numerous effects that can be used to assess peer influence on offending. Others include total similarity and average alter. They all approach the same issue (influence) in slightly different ways (see sections 5.7, 12.2.1, and 13 of Ripley et al., 2018 for a discussion of the various effects). However, the different influence effects are not directly comparable; therefore, only the most common one (average similarity) is included. Since similarity effects are not influenced by the scaling of the behavioral variable, the raw form of the effect as reported in the individual articles is sufficient (section 13.2 of Ripley et al., 2018).

4.2. Search Strategy

Searches for relevant materials, including peer-reviewed journals, theses/dissertations, conference proceedings, and book chapters, were conducted in June 2015 and updated in April/May 2017. The following academic databases were searched: Scopus, PsycINFO (PsycNET), Web of Science, H.W. Wilson Social Sciences (EBSCO), Sociological Abstracts (ProQuest), National Criminal Justice Reference Service (NCJRS), and ProQuest Dissertations and Theses Global (formerly Dissertations Abstracts). Searches conducted in each database were limited to English2 materials published since January 1, 2001 when Snijders (2001) published the article outlining the SAOM technique. Search strategies were developed by a social science librarian (SB) in consultation with the principal researcher (OG), and consisted of keywords and database specific subject headings for three main concepts: delinquency/crime/aggression, peers/friendship networks, and SAOM/SIENA/selection/influence. Full search strategies are available in Appendix 1.

As relevant studies may only discuss the specific analytic techniques in the methods section of the paper rather than the abstract, full text databases were included in the search. Multiple searches were conducted in Google Scholar and JSTOR, a full text journal database. Due to the high number of results retrieved, results were sorted by relevance, the first 300 results from each search in JSTOR and the first 200 from each search in Google Scholar were included in the screening. To minimize the possibility of the results being influenced by publication bias (the possibility that the research literature can distort reality by selectively publishing stronger effects), search results included theses and dissertations whenever possible. The use of Google Scholar and ProQuest Dissertations and Theses Global is beneficial in terms of picking up studies not published in peer reviewed journals. Four graduate theses/dissertations were included (De Cuyper, 2008; De La Rue, 2015; Knecht, 2008; Lee, 2011). The fact that positive and negative coefficients were found as well as statistically significant and non-significant effects for both influence and selection suggests that the results were not entirely driven by preferential publication of certain types of results.

In addition, a cited reference search was conducted for articles citing Snijders (2001) within Scopus and Web of Science. Snijder’s article was selected for a cited reference search as it was determined by the lead author to be a key early paper for SAOMs. This search was to identify any additional studies that may not have been found through the searches mentioned above. See Appendix 2 for the study flow diagram that displays the number of articles found at each step of the literature search. Finally, the list of studies was presented to a prominent author who has published in this area to confirm the completeness of the list. This did not result in other studies being added that were not captured through other means.

Title and abstract screening was performed independently by two authors (JM and OG) and resulted in a set of 120 articles to be full text screened. From the full text screening, 26 articles were selected as meeting the inclusion criteria (Baerveldt et al., 2008; Burk et al., 2007, 2008; Dahl and Van Zalk, 2014; De Cuyper, 2008; De La Rue, 2015; Dijkstra et al., 2011; Haynie et al., 2014; Jose et al., 2016; Kerr et al., 2012; Knecht et al., 2010; Lee, 2011; Light and Dishion, 2007; Logis et al., 2013; Molano et al., 2013; Osgood et al., 2015; Rulison et al., 2013; Shin, 2017; Sijtsema et al., 2010; Snijders and Baerveldt, 2003; Snijders et al., 2010; Svensson et al., 2012; Turanovic and Young, 2016; Van Zalk and Van Zalk, 2015; Weerman, 2011; Weerman et al., 2017). See the supplementary online material (table S1) for more extensive details on these articles and effects.

Of the 26 articles that met the inclusion criteria, one was excluded from the meta-analysis (Sijtsema et al., 2010) as the selection and influence effects they reported were interactions with reciprocity, which differs conceptually from the included studies. Supplementary analyses including this study showed that the decision to exclude this study did not alter the results (available on request). One other article (Light and Dishion, 2007) was excluded as it did not include a selection effect and the influence effect (average alter) was not comparable to the majority of other included influence effects (average similarity). Effects from a total of 24 separate articles were therefore included in either the selection or influence meta-analysis.

Each study was coded separately by two authors (JM and OG). We compared similarity across the two coders by examining Spearman correlations of their independently recorded effect sizes and standard errors. These correlations were all between .90 and .97. All discrepancies were resolved by discussing the rationale for the coded value.

4.3. Moderators

A number of pre-specified potential effect size moderators were tested (see table 1 for descriptive statistics):

Table 1. Summary of study characteristics.

Mean

SD

Minimum

Maximum

Age

13.84

2.30

8.62

17.27

Lag

0.80

0.30

0.04

1.00

Start yeara

7.93

6.25

0.00

20.00

N

%

Country

USA

12

42.86

Netherlands

7

25.00

Sweden

7

25.00

Other

2

7.14

Type of sample

General sample

19

67.86

High risk

6

21.43

Other

3

10.71

Type of behavior

General offending

17

60.71

Aggression

9

32.14

Other

2

7.14

Neffects=28.

aNumber of years since 1994.

  1. Age was the average age of the sample at the first wave analyzed in the article (mean=14).

  2. Lag time between waves (in years). Most studies had a lag time of approximately one year between data collection waves (61%); the smallest lag was two weeks (0.04 years) (mean=0.80).

  3. Start year. It is possible that methodological approaches to the collection of social network data or methods of communication between friends changed over time in ways that might affect patterns of ties. For this reason, we tested whether the time period of the study (as indicated by the year of the first wave of data collection) influenced effects. This variable was scaled as the number of years from the time the first coded study began (1994) (mean=8). E.g., a study that started in 2002 would be coded 8 since it started 8 years after 1994.

  4. Country where the study was conducted (USA, Netherlands, Sweden, Other). Due to estimation issues, the one effect each from Chile and South Korea were combined into an ‘other’ category. The largest number of effects came from American data (43%) with substantial proportions from the Netherlands and Sweden (25% each).

  5. Type of sample (general, high risk, other). We took an inductive approach to coding this moderator. A study was coded as ‘high risk’ if the authors indicated that the sample was subject to some sort of broadly defined social disadvantage (not strictly socioeconomic in nature) that may be related to a greater likelihood of involvement in offending behavior (e.g., “students in the lower educational strata of a major Dutch city with inner-city problems were overrepresented” - Weerman, 2011, p. 260). Where there was no suggestion that the study differed from the broader population, it was coded ‘general’. The majority of the effects were from general samples (68%). Separate categories for ‘high poverty’ and ‘high immigrant’ networks were initially coded, however estimation issues in the influence analysis owing to the small numbers of effects in these groups resulted in combining them into an ‘other’ category.

  6. Type of behavior (general offending, aggression, other). General offending scales combined numerous types of offending (e.g., shoplifting, fighting). Aggression involved physical aggression though it could be in combination with relational aggression. However, if it had only included relational aggression (e.g., spreading gossip), it would not have been considered offending behavior and therefore would not have met the inclusion criteria. The “other” category combined property crime and illegal political behavior out of necessity as there were estimation issues when leaving them as distinct categories. The majority of effects were for measures of general offending (61%).3

4.4. Analysis

Some articles provided more than one effect (between 1 and 3 effects per article). This occurred when separate results were reported for different groups within a particular data set. In these cases, there was no overlap in the samples upon which separate estimates within the same article were based. However, multiple articles included in the meta-analysis were based on the same data set (between 1 and 3 articles per data set; e.g., 3 articles used the Add Health data) which introduces more complex dependencies between effect sizes. To account for this nesting structure, 3-level multilevel models were estimated (using the metafor package in R - Viechtbauer, 2010). Given the unknown covariance among these dependent effect sizes,4 we used a sandwich-type estimator with a small-sample adjustment to produce cluster robust standard errors which have been shown to perform well under these conditions (see Hedges et al., 2010). As a sensitivity test, we also estimated models in which dependence among effect sizes was eliminated by selecting a single effect from each data set (the selected effect was from the model with the largest number of covariates) (contact the corresponding author for these results)[supp3]. Aside from minor variation in mean effect sizes, the results were highly congruent with the multilevel estimates reported in the tables below. One study reported an influence effect that was an outlier compared to the others in our meta-analysis (b=17.25 – De La Rue, 2015). Estimates excluding this effect did not change the results [supp6].

Results of SAOMs are presented as log odds ratios and therefore require no further conversion to use in a meta-analysis. Since they are a product of multivariate analyses, however, it is important that the models from which effects are drawn are similar enough to compare across studies. While there is variation across studies, SAOMs contain similar network effects (e.g., out-degree, reciprocity, rate functions) regardless of the focus and in that way are quite similar. Individual effect sizes were inverse-variance weighted based on standard errors reported in the original articles.

Regarding the moderation analysis, some rules of thumb exist as to the number of effects required for use in a meta-regression analysis. For example, Borenstein et al. (2009, p. 188) suggest that there should be a minimum of ten studies per covariate. Since there are not enough studies to include all moderators in the same model, we estimated separate models for each moderator (see Pratt et al., 2014 for a similar approach). It should be noted, however, that the value of moderation analysis with a small number of effect sizes has been questioned (Higgins and Green, 2011, section 9.5.3). For this reason, the moderation analysis should be considered descriptive.

5. RESULTS

Table 2 presents the mean effect size and moderation analysis for the 21 peer influence effects nested within 19 articles and 13 data sets (pooled n=21,193) (also see figure 15). The overall effect was positive and statistically significant (mean log odds ratio=1.23, p<.01), suggesting that individual offending is influenced by the offending of friends. Of the 21 individual effects, 13 (62%) were significantly different from zero. The fact that over a third of all effects were non-significant and yet the mean effect size was significant highlights the importance of taking stock of the entire body of literature. In order to interpret the results in terms of a one unit change in the offending scale, we divided each coefficient and standard error by the range of the behavioral variable minus one and estimated a model with these converted values (see, for example, Dijkstra et al., 2011, 2012; Haynie et al., 2014; Osgood et al., 2015).6 Results revealed a significant, positive mean effect size that support the main results (mean log odds ratio=0.19, odds ratio=1.21, p<.01). This can be interpreted to mean that the odds of a person adjusting their level of offending to be one unit closer to that of their friends is 21% higher than not changing their level of offending.

Table 2. Effect size estimates for peer influence.

Log odds ratio

95% CI

Qa

Intercept

Fb

Lower

Upper

Overall

1.23**

0.74

1.71

129.12**

Coefficient

Moderators

Age

-0.18

-0.38

0.02

115.43**

3.60*

4.06

Lag

-0.21

-2.58

2.16

118.89**

1.41

0.04

Start year

0.08*

0.02

0.13

126.26**

0.60**

9.08*

Country (USA omitted)

59.24**

1.21*

49.88**

Netherlands

-0.73

-1.64

0.18

Sweden

0.64

-0.24

1.52

Other

0.32

-0.95

1.59

Type of sample (general omitted)

98.25**

1.35**

2.62

High risk

-0.57

-1.41

0.27

Other

0.61

-1.02

2.24

Type of behavior (general offending omitted)

105.20**

0.83**

51.11**

Aggression

1.46*

0.11

2.80

Other

1.61**

1.25

1.96

* p<.05, ** p<.01

aHeterogeneity test

bModeration test

Figure 1. Forest plot of influence effects - b(95% CI).

To give a sense for the strength of the effect in units more commonly seen in meta-analyses, we converted the log odds ratio to Cohen’s d (using equation 7.17 in Borenstein et al., 2009). For a log odds ratio of 1.23, the d value of 0.68 places it in the ‘medium’ to ‘large’ range using the convention proposed by Cohen (1988). This is somewhat higher than the differential association effect size reported by Pratt et al. (2010, p.782) of d=0.45 (using equations 6.5 and 7.58 of Borenstein et al., 2009 to convert from z=0.225 to d=0.45). This is somewhat unexpected since Pratt et al. (2010) mostly examined studies that employed perceptual measures of peer offending which are thought to inflate the peer/personal offending relationship. However, caution is always recommended when applying standardized conventions, especially in light of Borenstein et al.’s (2009, chapter 7) warning that “observational studies that report correlations may be substantially different from observational studies that report odds ratios.” Further, if we use the rescaled estimates that allow for an interpretation in terms of a one unit change in the behavioral variable (arguably a better comparison with the standardized regression coefficients in Pratt et al., 2010, pp. 775--776), the log odds ratio of 0.19 translates to d=0.10, in the ‘small’ range according to Cohen’s rule of thumb and substantially smaller than the differential association effect size in Pratt et al. (2010). Given this, the most sensible interpretation is that, when combining the results of the available literature, peer influence is found to be related to offending regardless of whether a) perceptual measures of peer offending with effects generally based on regression methods are considered (as shown by Pratt et al.); or b) whether more direct measures of peer offending collected via longitudinal network methods and analyzed using SAOMs are considered (the present study).

The Q statistic indicated significant heterogeneity in influence effect sizes (Q=129.12, p<.01). Some of this heterogeneity can be attributed to the moderator variables. In particular, the year of the first wave of data collection, country the sample was drawn from, and type of behavior all had significant moderation effects. Larger influence effects were associated with more recently collected data (b=0.08) and aggression (b=1.46) and ‘other’ behaviors (b=1.61) (property crime, illegal political behavior) relative to general offending. For the country moderator, the F score was significant indicating that the country from which the sample was taken moderated the influence effect, though none of the individual coefficients were significant at the .05 level. Significant heterogeneity remained even after accounting for moderators. The age of the sample, lag time between waves, and type of sample were not found to moderate the effect of peer influence.

Table 3 presents the mean effect size and moderation analysis for the 28 selection effects nested within 24 articles and 15 data sets (pooled n=21,269) (also see figure 2). Like peer influence, the overall selection effect was positive and statistically significant (mean log odds ratio=0.31, p<.01) which indicates that people with similar levels of offending tend to become friends over time. A minority of the individual selection effects were statistically significant (13 out of 28 – 46%). Rescaling the parameters so that the mean effect size could be interpreted in terms of a one unit change in offending did not alter the interpretation that similarity in offending is a determinant of friendship formation (mean log odds ratio=0.05, odds ratio=1.05, p<.01). More specifically, this can be interpreted to mean that a person has a 5% higher odds of forming a friendship tie with someone who has the same score on the offending scale than someone 1 unit away.

Table 3. Effect size estimates for peer selection.

Log odds ratio

95% CI

Qa

Intercept

Fb

Lower

Upper

Overall

0.31**

0.01

0.52

121.68**

Coefficient

Moderators

Age

0.02

-0.05

0.10

117.48**

-0.02

0.54

Lag

0.44

-0.13

1.02

120.09**

-0.02

2.74

Start year

0.01

-0.01

0.03

111.79**

0.23

1.11

Country (USA omitted)

90.08**

0.23**

2.71

Netherlands

-0.11

-0.28

0.06

Sweden

0.58

-0.19

1.35

Other

0.06

-0.17

0.28

Type of sample (general omitted)

112.63**

0.31

0.00

High risk

0.01

-0.65

0.67

Other

-0.00

-0.40

0.40

Type of behavior (general offending omitted)

121.36**

0.34**

0.16

Aggression

-0.08

-0.36

0.21

Other

0.04

-0.72

0.80

* p<.05, ** p<.01

aHeterogeneity test

bModeration test

Figure 2. Forest plot of selection effects - b(95% CI).

Unlike for the influence effect where we could offer tentative comparisons with the effect sizes presented by Pratt et al. (2010), there are no known relevant comparisons for peer selection on similarity in offending characteristics. However, we provide the d value in hopes that it can serve as a benchmark for future work in this area. The d of 0.17 (corresponding to log odds ratio=0.31 presented in table 3) as well as a d of 0.03 for the rescaled estimates (corresponding to log odds ratio=0.05) fall into the ‘small’ range on Cohen’s rule of thumb.

As with influence, there was significant heterogeneity in selection effects (Q=121.68, p<.01). However, where the peer influence effect was significantly moderated by a number of variables, no moderation effects emerged for selection on offending.

6. DISCUSSION

The existence of a relationship between peer and personal offending is widely acknowledged, but the mechanism underlying it is subject to debate. The peer influence perspective argues that offending behavior spreads between people through a process of contagion as they are affected by the behaviors of their friends. The selection perspective argues that the relationship is driven by homophily, where people who offend at similar levels are more likely than others to become friends over time. The development of SAOMs (Snijders, 2001; Snijders et al., 2010), a relatively recent statistical advance, has allowed for an examination of peer selection versus influence in a way that avoids the criticism that (a) the peer influence effect is driven by projection bias, (b) that lagged dependent variables are not ideal measures of selection dynamics, and (c) that dependence among cases violates model assumptions (as in the case of regression modeling with network measures). By measuring longitudinal changes in friendship patterns and incorporating offending only based on self-reports, SAOMs circumvent those criticisms while also controlling for the effect of selection on influence and vice versa. Sufficient numbers of studies employing SAOMs to examine peer selection and influence on offending have been published that aggregating the results via meta-analysis is now a fruitful endeavor.

The results suggest that the relationship between peer offending and individual behavior is related both to people choosing to befriend others with similar offending profiles (selection) as well as adjusting their offending behavior to more closely match that of their friends (influence). Influence effects appear to be somewhat dependent on study characteristics such as the type of offending behavior measured and when the sample was collected. But overall, these results support the basic tenet of social learning theory (Akers, 2009) and differential association theory (Sutherland, 1947) that the offending behavior of one person increases the likelihood of offending among others. The results presented here complement the meta-analytic results of Pratt et al. (2010) who found similar support for differential association predominantly among studies that use indirect (non-network) measures of peer offending. This speaks to the robustness of the peer effect to methodological specification. Proponents of Akers’ social learning theory will argue that perceptions of peer behaviors are what matter; Pratt et al. (2010) show that these perceptions are important determinants of individual offending behavior. But those who are concerned more with actual peer behaviors raise the issue of projection bias within perceptual measures of offending (e.g., Haynie, 2001). Social network studies get around this issue by asking each participant about their own offending. Our study provides evidence supporting the peer influence effect when taking this approach.

However, the counterpoint, often stemming from control theorists, is that homophily on offending underlies the peer/personal offending relationship (Gottfredson and Hirschi, 1990; Hirschi, 1969). This concern is shown here to be on solid ground since friendship ties are found to be at least partly a product of pre-existing similarity in offending. Theoretical claims that the peer offending/personal offending relationship is driven solely by friendship selection is not supported. Given the differences in the meaning and calculation of the selection and influence parameters, we hesitate to state that influence is in some way more important than selection even though the magnitude of the mean effect size for influence is larger. However, we do feel justified in stating that evidence supporting both dynamics is robust.

The major implication of this study for researchers who model individual offending behavior is that they should account for both influence and selection. SAOMs are particularly well suited to this task (especially if projection bias is a concern) and we encourage greater use of this approach. And there are longitudinal sociometric data available that can be used to address criminological interests (e.g., the National Longitudinal Study of Adolescent to Adult Health - Add Health). However, there are limits to the scope of the publicly available data which means that original data must often be collected. Longitudinal social network data is difficult and often expensive to collect, but that does not negate the importance of being able to account for both selection and influence when examining individual offending. Our results therefore extend beyond recommending SAOMs. Even research that does not employ a social network approach should incorporate measures of both dynamics. Peer influence is suitably measured by taking the traditional non-network approach of asking respondents to estimate the level of offending of their peers assuming the researcher accepts Akers’ (2009) argument that perceptions of peer offending are what truly matter (i.e., that peer behaviors are unlikely to affect personal behavior if the individual does not know what their friends are actually up to). But accounting for selection effects is trickier without network data. Lagged dependent variables are imprecise and subject to statistical issues, but there are no well-established alternatives. A more direct approach to measuring friendship selection processes that can be incorporated into standard survey research would be to ask respondents if they have a preference for associating with people who engage in offending behavior (e.g., “Given the choice, I would rather hang out with people who physically fight”; response categories on a Likert-type scale). While this would not contribute to modeling the tie formation process in the way that SAOMs do, it is likely to capture preferences for friends with particular characteristics that may underlie tie formation. Therefore, it is likely to be a much more precise proxy for peer selection than lagged dependent variables.

6.1. Limitations

There are a number of limitations to this study. The first is that there were a small number of effect sizes combined in the meta-analysis. While there is some evidence that peer influence in particular varies by characteristics of the study, we recommend that these variations not be considered conclusive without further research confirmation.

A second limitation in terms of drawing conclusions on the generality of selection and influence effects is that the studies in this analysis were based exclusively on youth samples. This is the current state of the literature, but it means that we cannot be sure that these dynamics will hold among adults who tend to place less importance on peers than adolescents (Warr, 2002). This situation is understandable given the availability of well-bounded child and adolescent school-based networks. Replicating these analyses among adult samples could be done, for example, using work-based networks for examinations of theft or prison-based networks for examinations of violence. There has been some progress on the latter (Kreager et al., 2016) so there is hope that such analyses will soon be produced.

The third limitation is that missing data was reported too inconsistently to be used as an effect moderator. Defining what constitutes ‘missing’ is a challenge in the context of longitudinal social network studies given the various forms that are potentially present. Some network members may enter or leave the network at different waves, ties may be missed if there are limits placed on the number of nominations that can be made, and respondents may choose not to answer questions. Complicating matters, covariate non-response is commonly imputed, but missing tie data is not.9 It is unclear whether missingness would alter effects in any systematic manner (other than widening confidence intervals given reduced sample sizes). It would be helpful in the future for authors of articles that employ SAOMs to include information on all forms of missing data at each wave and overall.

Despite these limitations, we find evidence that adolescents are both influenced by the delinquency of their peers and select friends based on similarity in levels of offending. This suggests that rather than debating selection versus influence, criminologists should be discussing selection and influence.

7. NOTES

1 Articles where the behavioral measure was a composite of substance use and delinquency were excluded to ensure that the focus on delinquency was clear. For example, Franken et al. (2016) was excluded as they combined involvement in delinquency, alcohol use, and tobacco use in their measure of externalizing behavior. We did, however, test supplementary models that included weapon carrying (Dijkstra et al., 2010, 2012), a behavior that does not meet the “clearly illegal” criterion, and found that doing so did not alter the results in any meaningful way[supp1].

2 Four potentially relevant non-English language articles were found since they had English abstracts. We were able to find the English language master's thesis that one was drawn from (De Cuyper et al., 2009), the author of two others sent us a copy of an article that was soon to be released online that used the same data (Shin, 2017), and the author of the final potentially relevant non-English article confirmed that SIENA models were not used.

3 We conducted supplementary analyses in which the nature of the scale (multi-item average, variety scale, dichotomized, etc.) was tested as a moderator. The mean effect sizes were not found to depend on the form of the scaling.

4 The exception here is Haynie et al. (2014). They reported separate results for violent and non-violent crime, but since the correlation between these outcomes is known (not reported in the article, but we used the same data to generate this correlation – r=0.562), a suitable synthetic estimate can be produced that is not based on an assumed correlation but on the observed data. We combined these results into a single effect by averaging the individual effects and calculating the variance using the equation recommended by Borenstein et al. (2009, p. 227):

VY=14(VY1+VY2+2rVY1VY2)V_{\overline{Y}} = \frac{1}{4}\left( V_{Y_{1}} + V_{Y_{2}} + 2r\sqrt{V_{Y_{1}}}\sqrt{V_{Y_{2}}} \right)

VY1V_{Y_{1}}= variance of effect 1

VY2V_{Y_{2}}= variance of effect 2

We also tested supplementary models in which the violent and non-violent crime estimates were included as separate effects. The difference in results was negligible (available on request)[supp5].

5 An outlier (De La Rue, 2015) was excluded from this figure as it altered the horizontal axis scaling such that it obscured the other effects that had narrower confidence intervals. The forest plot including this effect can be accessed in the online supplementary material.

6 For example, Osgood et al. (2015) used a 4-category delinquency scale; the reported influence effect was 1.108 (SE=0.084). This was rescaled as follows:

Log odds ratio=1.108/(4-1)=0.369

SE=0.084/(4-1)=0.028

7 Equation 7.1: Converting log odds ratio to d.

d=Log odds ratio x 3πd = Log\ odds\ ratio\ x\ \frac{\sqrt{3}}{\pi}

8 Equation 6.5: Converting z to r.

r= e2z1e2z+1r = \ \frac{e^{2z} - 1}{e^{2z} + 1}

Equation 7.5: Converting r to d.

d= 2r1 r2d = \ \frac{2r}{\sqrt{1 - \ r^{2}}}

9 Imputation methods exist for network ties (Wang et al., 2016), but they are not yet widely implemented.

8. References

Achen, C. H. (2000). Why lagged dependent variables can suppress the explanatory power of other independent variables. Paper presented at the Annual Meeting of the Political Methodology Section of the American Political Science Association, Los Angeles, CA.

Akers, R. L. (2009). Social Learning and Social Structure: A General Theory of Crime and Deviance, Transaction Publishers, New Brunswick, NJ.

Akers, R. L., and Jensen, G. F. (2006). The empirical status of social learning theory of crime and deviance: The past, present, and future. In Cullen, F. T., Wright, J. P. and Blevins, K. R. (eds.), Taking Stock: The Status of Criminological Theory, Transaction Publishers, New Brunswick, NJ, pp. 37--76.

Baerveldt, C., Volker, B., and Van Rossem, R. (2008). Revisiting selection and influence: An inquiry into the friendship networks of high school students and their association with delinquency. Can J Criminol Crim, 50(5): 559--587.

Boman, J. H., Stogner, J. M., Miller, B. L., Griffin, O. H., and Krohn, M. D. (2012). On the operational validity of perceptual peer delinquency: Exploring projection and elements contained in perceptions. J Res Crime Delinq, 49(4): 601--621.

Boman, J. H., and Ward, J. T. (2014). Beyond projection: Specifying the types of peer delinquency misperception at the item and scale levels. Deviant Behav, 35(7): 555--580.

Borenstein, M., Hedges, L. V., Higgins, J., and Rothstein, H. R. (2009). Introduction to Meta-Analysis, John Wiley and Sons, Ltd, Chichester, UK.

Burk, W. J., Kerr, M., and Stattin, H. (2008). The co-evolution of early adolescent friendship networks, school involvement, and delinquent behaviors. Rev Fr Sociol, 49(3): 499--522.

Burk, W. J., Steglich, C. E. G., and Snijders, T. A. B. (2007). Beyond dyadic interdependence: Actor-oriented models for co-evolving social networks and individual behaviors. Int J Behav Dev, 31(4): 397--404.

Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.), Lawrence Erlbaum Associates, Hillsdale, NJ.

Dahl, V., and Van Zalk, M. (2014). Peer networks and the development of illegal political behavior among adolescents. J Res Adolescence, 24(2): 399--409.

De Cuyper, R. (2008). Good Friends, Bad Behavior?! The Co‐Evolution of Friendship Relations and Delinquency Involvement Among Dutch Adolescents. (Master's thesis), Radboud University Nijmegen.

De Cuyper, R., Weerman, F., and Ruiter, S. (2009). De co-evolutie van vriendschapsrelaties en delinquent gedrag onder Nederlandse jongeren. Mens en Maatschappij, 84(3): 300--328.

De La Rue, L. (2015). The Influence of Family and Friends on Girls’ Delinquency: A Social Network Analysis. (PhD dissertation), University of Illinois at Urbana-Champaign.

Dijkstra, J. K., Berger, C., and Lindenberg, S. (2011). Do physical and relational aggression explain adolescents' friendship selection? The competing roles of network characteristics, gender, and social status. Aggressive Behav, 37(5): 417--429.

Dijkstra, J. K., Gest, S. D., Lindenberg, S., Veenstra, R., and Cillessen, A. H. N. (2012). Testing three explanations of the emergence of weapon carrying in peer context: The roles of aggression, victimization, and the social network. J Adolescent Health, 50(4): 371--376.

Dijkstra, J. K., Lindenberg, S., Veenstra, R., Steglich, C., Isaacs, J., Card, N. A., and Hodges, E. V. E. (2010). Influence and selection processes in weapon carrying during adolescence: The roles of status, aggression, and vulnerability. Criminology, 48(1): 187--220.

Franken, A., Moffitt, T. E., Steglich, C. E. G., Dijkstra, J. K., Harakeh, Z., and Vollebergh, W. A. M. (2016). The role of self-control and early adolescents’ friendships in the development of externalizing behavior: The SNARE study. J Youth Adolescence, 45(9): 1800--1811.

Glueck, S., and Glueck, E. (1950). Unraveling Juvenile Delinquency, Harvard University Press, Cambridge, MA.

Gottfredson, M., and Hirschi, T. (1987). The methodological adequacy of longitudinal research on crime. Criminology, 25(3): 581--614.

Gottfredson, M., and Hirschi, T. (1990). A General Theory of Crime, Stanford University Press, Stanford, CA.

Haynie, D. L. (2001). Delinquent peers revisited: Does network structure matter? Am J Sociol, 106(4): 1013--1057.

Haynie, D. L., Doogan, N. J., and Soller, B. (2014). Gender, friendship networks, and delinquency: A dynamic network approach. Criminology, 52(4): 688--722.

Haynie, D. L., and Osgood, D. W. (2005). Reconsidering peers and delinquency: How do peers matter? Soc Forces, 84(2): 1109--1130.

Hedges, L. V., Tipton, E., and Johnson, M. C. (2010). Robust variance estimation in meta-regression with dependent effect size estimates. Res Synth Methods, 1(1): 39--65.

Higgins, J. P. T., and Green, S. (2011). Cochrane Handbook for Systematic Reviews of Interventions: The Cochrane Collaboration.

Hirschi, T. (1969). Causes of Delinquency. University of California Press, Berkeley, CA.

Ioannidis, J. P. A. (2016). The mass production of redundant, misleading, and conflicted systematic reviews and meta-analyses. Milbank Q, 94(3): 485--514.

Jose, R., Hipp, J. R., Butts, C. T., Wang, C., and Lakon, C. M. (2016). Network structure, influence, selection, and adolescent delinquent behavior: Unpacking a dynamic process. Crim Justice Behav, 43(2): 264--284.

Kerr, M., Van Zalk, M., and Stattin, H. (2012). Psychopathic traits moderate peer influence on adolescent delinquency. J Child Psychol Psyc, 53(8): 826--835.

Knecht, A. (2008). Friendship Selection and Friends’ Influence: Dynamics of Networks and Actor Attributes in Early Adolescence. (PhD dissertation), Utrecht University, Netherlands.

Knecht, A., Snijders, T. A. B., Baerveldt, C., Steglich, C. E. G., and Raub, W. (2010). Friendship and delinquency: Selection and influence processes in early adolescence. Soc Dev, 19(3): 494--514.

Kreager, D. A., Schaefer, D. R., Bouchard, M., Haynie, D. L., Wakefield, S., Young, J., and Zajac, G. (2016). Toward a criminology of inmate networks. Justice Q, 33(6): 1000--1028.

Lee, S. T. (2011). An Actor-Oriented Approach to Friendship Networks: Re-Analysis and Further Model Specifications of the Published Model by Baerveldt, Volker & Van Rossem (2008). (Master’s thesis), University of Oxford.

Light, J. M., and Dishion, T. J. (2007). Early adolescent antisocial behavior and peer rejection: A dynamic test of a developmental process. New Dir Child Adolesc Dev, 118: 77--89.

Logis, H. A., Rodkin, P. C., Gest, S. D., and Ahn, H.-J. (2013). Popularity as an organizing factor of preadolescent friendship networks: Beyond prosocial and aggressive behavior. J Res Adolescence, 23(3): 413--423.

McGloin, J. M., and Thomas, K. J. (2016). Considering the elements that inform perceived peer deviance. J Res Crime Delinq, 53(5): 597--627.

Molano, A., Jones, S. M., Brown, J. L., and Aber, J. L. (2013). Selection and socialization of aggressive and prosocial behavior: The moderating role of social‐cognitive processes. J Res Adolescence, 23(3): 424--436.

Osgood, D. W., Feinberg, M. E., and Ragan, D. T. (2015). Social networks and the diffusion of adolescent problem behavior: Reliable estimates of selection and influence from sixth through ninth grades. Prev Sci, 16(6): 832--843.

Pratt, T. C., Cullen, F. T., Sellers, C. S., Winfree, L. T., Jr., Madensen, T. D., Daigle, L. E., Fearn, N. E., and Gau, J. M. (2010). The empirical status of social learning theory: A meta-analysis. Justice Q, 27(6): 765--802.

Pratt, T. C., Turanovic, J. J., Fox, K. A., and Wright, K. A. (2014). Self-control and victimization: A meta-analysis. Criminology, 52(1): 87--116.

Rebellon, C. J., and Modecki, K. L. (2013). Accounting for projection bias in models of delinquent peer influence: The utility and limits of latent variable approaches. J Quant Criminol, 30(2): 163--186.

Ripley, R. M., Snijders, T. A. B., Boda, Z., Voros, A., and Preciado, P. (2018). Manual for RSiena (version April 17, 2018), University of Oxford, Department of Statistics; Nuffield College, Oxford, UK.

Rulison, K. L., Gest, S. D., and Loken, E. (2013). Dynamic social networks and physical aggression: The moderating role of gender and social status among peers. J Res Adolescence, 23(3): 437--449.

Shin, H. (2017). Friendship dynamics of adolescent aggression, prosocial behavior, and social status: The moderating role of gender. J Youth Adolescence, 46(11): 2305--2320.

Sijtsema, J. J., Ojanen, T., Veenstra, R., Lindenberg, S., Hawley, P. H., and Little, T. D. (2010). Forms and functions of aggression in adolescent friendship selection and influence: A longitudinal social network analysis. Soc Dev, 19(3): 515--534.

Snijders, T. A. B. (2001). The statistical evaluation of social network dynamics. Sociol Methodol, 31(1): 361--395.

Snijders, T. A. B., and Baerveldt, C. (2003). A multi-level network study of the effects of delinquent behavior on friendship evolution. J Math Sociol, 27(2-3): 123--151.

Snijders, T. A. B., Van de Bunt, G. G., and Steglich, C. E. G. (2010). Introduction to stochastic actor-based models for network dynamics. Soc Networks, 32(1): 44--60.

Steglich, C., Snijders, T. A. B., and Pearson, M. (2010). Dynamic networks and behavior: Separating selection from influence. Sociol Methodol, 40(1): 329--393.

Steglich, C., Snijders, T. A. B., and West, P. (2006). Applying SIENA: An illustrative analysis of the coevolution of adolescents’ friendship networks, taste in music, and alcohol consumption. Methodology, 2(1): 48--56.

Sutherland, E. H. (1947). Principles of Criminology (Vol. 4), J. B. Lippincott, Philadelphia, PA.

Svensson, Y., Burk, W. J., Stattin, H., and Kerr, M. (2012). Peer selection and influence of delinquent behavior of immigrant and nonimmigrant youths: Does context matter? Int J Behav Dev, 36(3): 178--185.

Turanovic, J. J., and Young, J. T. N. (2016). Violent offending and victimization in adolescence: Social network mechanisms and homophily. Criminology, 54(3): 487--519.

Van Zalk, M. H. W., and Van Zalk, N. (2015). Violent peer influence: The roles of self-esteem and psychopathic traits. Dev Psychopathol, 27(4pt1): 1077--1088.

Viechtbauer, W. (2010). Conducting meta-analyses in R with the metafor package. J Stat Softw, 36(3): 1--48.

Wang, C., Butts, C. T., Hipp, J. R., Jose, R., and Lakon, C. M. (2016). Multiple imputation for missing edge data: A predictive evaluation method with application to Add Health. Soc Networks, 45: 89--98.

Warr, M. (2002). Companions in Crime, Cambridge University Press, New York.

Weerman, F. M. (2011). Delinquent peers in context: A longitudinal network analysis of selection and influence effects. Criminology, 49(1): 253--286.

Weerman, F. M., and Hoeve, M. (2012). Peers and delinquency among girls and boys: Are sex differences in delinquency explained by peer factors? Eur J Criminol, 9(3): 228--244.

Weerman, F. M., Wilcox, P., and Sullivan, C. J. (2017). The short-term dynamics of peers and delinquent behavior: An analysis of bi-weekly changes within a high school student network. J Quant Criminol, 1--33.

Appendix 1. Database search terms.

All searches conducted by Sarah Brown, Liaison Librarian, University of Waterloo.

Scopus Search

  • Performed June 3, 2015; updated on April 22, 2017

  • Database: Scopus

  • Retrieved 754 results and 958 results respectively

((TITLE-ABS-KEY(peer* OR friend* OR gang* OR social OR "collective behavior*" OR "collective behavior*") AND PUBYEAR > 2000) AND ((TITLE-ABS-KEY(delinquen* OR firearm* OR weapon* OR aggressi* OR devian* OR criminal* OR crime OR assault OR vandal* OR perpetrator* OR arson OR "driving under the influence" OR "drunk driving" OR "drunk driver*" OR "human traffick*" OR kidnapping OR "sex* offen*") OR TITLE-ABS-KEY(theft OR burglar* OR fraud OR "illegal drug distribution" OR "drug traffick*" OR violen* OR stalking OR stalker* OR terroris*))) AND ((TITLE-ABS-KEY(influenc* AND select*) OR TITLE-ABS-KEY(stochastic AND actor) OR TITLE-ABS-KEY(siena) OR TITLE-ABS-KEY("simulation investigation") OR TITLE-ABS-KEY(stochastic AND model*)))) AND ( LIMIT-TO(LANGUAGE,"English" ) )

Cited Reference search: Snijders, T. A. B. (2001). The statistical evaluation of social network dynamics. Sociological Methodology, 31(1), 361–395.

  • Run on June 23, 2015; updated on May 2, 2017

  • Retrieved 331 results and 443 results respectively

PsycInfo Search

  • Performed June 3, 2015; updated on April 22, 2017

  • Database: PsycNET

  • Retrieved 442 results and 562 results respectively

(Any Field:(delinquen* OR firearm* OR weapon* OR aggressi* OR devian* OR criminal* OR crime OR assault OR vandal* OR perpetrator* OR arson OR "driving under the influence" OR "drunk driving" OR "drunk driver*" OR "human traffick*" OR kidnapping OR "sex* offen*" OR theft OR burglar* OR fraud OR "illegal drug distribution" OR "drug traffick*" OR violen* OR stalking OR stalker* OR terroris*) AND Year:[2001 TO 2018]) AND (Any Field:(influenc* AND select*) OR Any Field:(stochastic AND actor) OR Any Field:(siena) OR Any Field:("simulation investigation") OR Any Field:(stochastic AND model*)) AND (Subject:(peer* OR gang OR gangs OR friend* OR social* OR "collective behavior*" OR "collective behavior*") OR Abstract:(peer* OR friend* OR gang OR gangs OR "collective behavior*" OR "collective behavior*") OR Abstract:("social network*" OR "social support*" OR "social influence*"))  

Wilson Social Sciences Full Text

  • Performed June 3, 2015; not updated in 2017 (library subscription expired).

  • Database: EBSCO

  • Retrieved 116 results

(TX delinquen* OR firearm* OR weapon* OR aggressi* OR devian* OR criminal* OR crime OR vandal* OR perpetrator* OR arson OR "driving under the influence" OR "drunk driving" OR "drunk driver" OR "drunk drivers" OR "human trafficking" OR "human trafficker" OR "human traffickers" OR kidnapping OR "sex offenses" OR "sex offender" OR "sex offenders" OR "sexual offenses" OR "sex offences" OR "sexual offences" OR assault OR theft OR burglar* OR fraud OR "illegal drug distribution" OR "drug trafficker" OR "drug trafficking" OR "drug traffickers" OR violen* OR stalking OR stalker* OR terroris*)  AND ( (DE "Friendship" OR DE "Peer relations" OR DE "Friendship -- Sociological aspects" OR DE "Social networks" OR DE "Social support" OR DE "Social interaction" OR DE "Social pressure" OR DE "Peer pressure" OR DE "Social influence" OR DE "Gangs") ) OR SU ( social* or peer* or friend* OR gang* ) OR TI ( social* or peer* or friend* OR gang* ) OR AB ( social* or peer* or friend* OR gang* )  AND ( influenc* AND select* ) OR ( stochastic AND actor ) OR siena OR "simulation investigation" OR ( stochastic AND model* )

Web of Science

  • Performed June 3, 2015; updated April 24, 2017

  • Database: Web of Science

  • Retrieved 821 results and 1,338 results respectively

#5

#3 AND #2 AND #1

Refined by: LANGUAGES: (ENGLISH)

DocType=All document types; Language=All languages;

#4

#3 AND #2 AND #1

DocType=All document types; Language=All languages;

#3

TOPIC: (delinquen* OR firearm* OR weapon* OR aggressi* OR devian* OR criminal* OR crime OR vandal* OR perpetrator* OR arson OR "driving under the influence" OR "drunk driving" OR "drunk driver" OR "drunk drivers" OR "human trafficking" OR "human trafficker" OR "human traffickers" OR kidnapping OR "sex offenses" OR "sex offender" OR "sex offenders" OR "sexual offenses" OR "sex offences" OR "sexual offences" OR assault OR theft OR burglar* OR fraud OR "illegal drug distribution" OR "drug trafficker" OR "drug trafficking" OR "drug traffickers" OR violen* OR stalking OR stalker* OR terroris*)

DocType=All document types; Language=All languages;

#2

TOPIC: (social* OR peer* OR friend* OR gang*)

DocType=All document types; Language=All languages;

#1

TOPIC: (influenc* AND select*) OR TOPIC: (stochastic AND actor) OR TOPIC: (siena) OR TOPIC: (stochastic AND model) OR TOPIC: ("simulation investigation")

DocType=All document types; Language=All languages;

Cited Reference search: Snijders, T. A. B. (2001). The statistical evaluation of social network dynamics. Sociological Methodology, 31(1), 361–395.

  • Performed June 23, 2015; updated May 2, 2017

  • Retrieved 172 results and 241 results respectively

Sociological Abstracts

  • Performed April 24, 2017

  • Database: ProQuest

  • Retrieved 902 results

(delinquen* OR firearm* OR weapon* OR aggressi* OR devian* OR criminal* OR crime OR vandal* OR perpetrator* OR arson OR "driving under the influence" OR "drunk driving" OR "drunk driver" OR "drunk drivers" OR "human trafficking" OR "human trafficker" OR "human traffickers" OR kidnapping OR "sex offenses" OR "sex offender" OR "sex offenders" OR "sexual offenses" OR "sex offences" OR "sexual offences" OR assault OR theft OR burglar* OR fraud OR "illegal drug distribution" OR "drug trafficker" OR "drug trafficking" OR "drug traffickers" OR violen* OR stalking OR stalker* OR terroris*) AND ((influenc* AND select*) OR (stochastic AND model*) OR (stochastic AND actor) OR SIENA OR "simulation investigation") AND (ab(peer* OR friend* OR social* OR gang*) OR ti(peer* OR friend* OR social* OR gang*) OR su(peer* OR friend* OR social* OR gang*)) AND pd(20010101-20171231)

[Note: Limited to English language only]

ProQuest Dissertations and Theses Global

  • Performed April 24, 2017

  • Database: ProQuest

  • Retrieved 691 results

all(delinquen* OR firearm* OR weapon* OR aggressi* OR devian* OR criminal* OR crime OR vandal* OR perpetrator* OR arson OR "driving under the influence" OR "drunk driving" OR "drunk driver" OR "drunk drivers" OR "human trafficking" OR "human trafficker" OR "human traffickers" OR kidnapping OR "sex offenses" OR "sex offender" OR "sex offenders" OR "sexual offenses" OR "sex offences" OR "sexual offences" OR assault OR theft OR burglar* OR fraud OR "illegal drug distribution" OR "drug trafficker" OR "drug trafficking" OR "drug traffickers" OR violen* OR stalking OR stalker* OR terroris*) AND (all(influenc* AND select*) OR all(stochastic AND model*) OR all(stochastic AND actor) OR all(SIENA) OR all("simulation investigation")) AND all(peer* OR friend* OR social* OR gang*)

[Note: All terms searched “Anywhere – Except Full-text”. Limiters included: date range 2001 – 2017 and English only.]

National Crime Justice Reference Service (NCJRS) Abstracts

  • Performed April 24, 2017

  • Database: ProQuest

  • Retrieved 294 results

(delinquen* OR firearm* OR weapon* OR aggressi* OR devian* OR criminal* OR crime OR vandal* OR perpetrator* OR arson OR "driving under the influence" OR "drunk driving" OR "drunk driver" OR "drunk drivers" OR "human trafficking" OR "human trafficker" OR "human traffickers" OR kidnapping OR "sex offenses" OR "sex offender" OR "sex offenders" OR "sexual offenses" OR "sex offences" OR "sexual offences" OR assault OR theft OR burglar* OR fraud OR "illegal drug distribution" OR "drug trafficker" OR "drug trafficking" OR "drug traffickers" OR violen* OR stalking OR stalker* OR terroris*) AND ((influenc* AND select*) OR (stochastic AND model*) OR (stochastic AND actor) OR SIENA OR "simulation investigation") AND (ab(peer* OR friend* OR social* OR gang*) OR ti(peer* OR friend* OR social* OR gang*) OR su(peer* OR friend* OR social* OR gang*))

[Note: Limiters: date range 2001 – 2017 and English only.]

Full Text Searches

JSTOR (3 searches)

(((peer* OR friend*) AND (delinquen* OR crim* OR aggressi*)) AND ("stochastic model"~5)) AND la:(eng OR en)

  • Performed June 5, 2015; updated April 25, 2017

  • Date Range: 01/01/2001 – current date; 01/01/2001 – 31/12/2017

  • Results = 242 and 246 respectively

(((peer* OR friend*) AND (delinquen* OR crim* OR aggressi*)) AND ("stochastic actor"~5)) AND la:(eng OR en)

  • Performed June 5, 2015; updated April 25, 2017

  • Date Range: 01/01/2001 – current date; 01/01/2001 – 31/12/2017

  • Results = 18 [same results on update]

(((peer* OR friend*) AND (delinquen* OR crim* OR aggressi*)) AND ("selection influence"~10)) AND la:(eng OR en)

  • Performed June 5, 2015; updated April 26, 2017

  • Date Range: 01/01/2001 – current date; 2015 - 2017

  • Results = 1071 (First 300 included); 65

Google Scholar (2 searches)

(peer OR friend OR friendship) (crime OR delinquent OR delinquency OR aggression OR aggressive) (SIENA)

  • Performed June 5, 2015; updated April 26, 2017

  • Date Range: 2001 – current date; 2015 – current date

  • Results = 7760; 2020 (first 200 included for both)

(peer OR friendship OR friend) (crime OR delinquent OR delinquency OR aggression OR aggressive) (stochastic actor)

  • Performed June 7, 2015; updated April 26, 2017

  • Date Range: 01/01/2001 – current date; 2015 – current date

  • Results = 1280; 2540 (first 200 included for both)

Appendix 2. Study flow diagram.

Caption: Adapted from Higgins and Green (2011, figure 11.2.a).

Table S1. Details of included studies.

Study

# of effect sizes (range)

Country

Type of behavior

Data set (sample size range)

Influence

Selection

Burk et al., 2007

1

(2.44)

1

(2.54)

Sweden

Other

10-to-18 project

(260)

Burk et al., 2008

1

(1.44)

1

(1.55)

Sweden

General offending

10-to-18 project

(445)

Kerr et al, 2012

1

(1.61)

1

(1.53)

Sweden

General offending

10-to-18 project

(847)

Molano et al., 2013

1

(2.91)

1

(0.18)

USA

Aggression

4Rs program

(900)

Svensson et al., 2012

2

(1.39 – 1.99)

2

(0.53 – 0.90)

Sweden

General offending

7 schools project

(352 – 817)

Van Zalk and Van Zalk, 2015

-

1

(0.77)

Sweden

Aggression

7 schools project

(1772)

Haynie et al., 2014

1

(1.57)

1

(0.95)

USA

General offending

Add Health

(1857)

Jose et al., 2016

2

(0.29 – 1.16)

2

(-0.07 – 0.14)

USA

General offending

Add Health

(976 – 1284)

Turanovic & Young, 2016

-

3

(0.23 – 0.43)

USA

Aggression

Add Health

(361 – 1022)

Dijkstra et al., 2011

1

(0.96)

1

(0.18)

Other

Aggression

Chilean Adolescent Study

(274)

Logis et al., 2013

1

(3.53)

1

(0.10)

USA

Aggression

Classroom Peer Ecologies Project

(613)

Baerveldt et al., 2008

1

(0.49)

1

(0.85)

Netherlands

General offending

Dutch Social Behavior Study

(859)

Lee, 2011

1

(1.23)

1

(0.35)

Netherlands

General offending

Dutch Social Behavior Study

(914)

Snijders and Baerveldt, 2003

-

1

(-0.49)

Netherlands

General offending

Dutch Social Behavior Study

(990)

Knecht et al., 2010

1

(0.03)

1

(0.16)

Netherlands

General offending

Dutch study

(544)

Snijders et al., 2010

1

(6.08)

1

(3.22)

Netherlands

General offending

Dutch study

(26)

De La Rue, 2015

1

(17.25)

1

(0.52)

USA

General offending

Illinois study

(401)

Rulison et al., 2013

-

1

(0.11)

USA

Aggression

Longitudinal Study of Early Adolescent Peer Networks & School Adjustment project

(480)

De Cuyper, 2008

1

(1.73)

1

(0.11)

Netherlands

General offending

NSCR School Study

Weerman, 2011

1

(0.17)

1

(0.06)

Netherlands

General offending

NSCR School Study

(1156)

Osgood et al., 2015

1

(1.11)

1

(0.22)

USA

General offending

PROSPER

(9135)

Dahl & Van Zalk, 2014

-

1

(0.08)

Sweden

Other

Political Socialization Program

(1006)

Shin, 2017

1

(1.94)

1

(0.37)

Other

Aggression

South Korea study

(736)

Weerman et al., 2017

1

(-0.56)

1

(-0.18)

USA

General offending

TEENS

(155)

Figure S1. Forest plot of influence effects including De La Rue - b(95% CI).

Comments
0
comment
No comments here
Why not start the discussion?