ArticlePDF Available

Behavioral Heterogeneity in Adolescent Friendship Networks

Authors:

Abstract and Figures

Criminologists’ understanding of peer influences has been greatly advanced by social network methods; however, relatively scant attention has been paid to improving measurement. In particular, research has continued to measure peer influence by averaging the level of delinquency within a peer network, thereby neglecting the role of behavioral heterogeneity. The present study seeks to advance theory and research into peer influences on delinquency by explicitly modeling behavioral heterogeneity in peer networks measured as the variance. Drawing on social learning and opportunity theories, we argue that behavioral heterogeneity should attenuate the effect of average peer delinquency on individual offending. Models using social network data from the Add Health were estimated predicting involvement in two delinquent substance-use acts (cigarette smoking and getting drunk) as a function of peer influences. The results are consistent with our hypothesis, indicating that behavioral heterogeneity matters. Findings suggest that future research employing network models could incorporate peer behavioral heterogeneity to get a more accurate portrait of the processes of peer influence.
Content may be subject to copyright.
Behavioral Heterogeneity in Adolescent
Friendship Networks
Callie H. Burt and Carter Rees
Criminologistsunderstanding of peer influences has been greatly advanced
by social network methods; however, relatively scant attention has been
paid to improving measurement. In particular, research has continued to
measure peer influence by averaging the level of delinquency within a peer
network, thereby neglecting the role of behavioral heterogeneity. The
present study seeks to advance theory and research into peer influences on
delinquency by explicitly modeling behavioral heterogeneity in peer
networks measured as the variance. Drawing on social learning and opportu-
nity theories, we argue that behavioral heterogeneity should attenuate the
effect of average peer delinquency on individual offending. Models using
social network data from the Add Health were estimated predicting
involvement in two delinquent substance-use acts (cigarette smoking and
getting drunk) as a function of peer influences. The results are consistent
with our hypothesis, indicating that behavioral heterogeneity matters.
Findings suggest that future research employing network models could
incorporate peer behavioral heterogeneity to get a more accurate portrait
of the processes of peer influence.
Keywords delinquent peers; social networks; measurement; heterogeneity;
influence
Callie H. Burt is an Assistant Professor in the School of Criminology and Criminal Justice and a
Faculty Affiliate of the School of Social Transformation at Arizona State University. Her research
focuses on criminological theories, with particular emphasis on elucidating the social psychological
mechanisms through which social factors, such as racial discrimination, community crime, parent-
ing practices, peers, and life transitions, influence criminal offending across the life course. Her
work has recently appeared in the American Journal of Sociology, the American Sociological
Review,Criminology, and Justice Quarterly. Carter Rees is an Assistant Professor of Criminology
and Criminal Justice at Arizona State University. His work applies the theoretical and statistical
concepts of social network analysis to the longitudinal study of delinquency. He has recently
been published in the Journal of Quantitative Criminology and the Journal of Youth and
Adolescence. Correspondence to: Callie Burt, School of Criminology and Criminal Justice, Arizona
State University, 411 N. Central Ave, Ste. 600, Phoenix, AZ 85004. E-mail: Callie.Burt@asu.edu
!2014 Academy of Criminal Justice Sciences
Justice Quarterly, 2014
http://dx.doi.org/10.1080/07418825.2013.856932
Downloaded by [Arizona State University] at 23:15 26 March 2014
Peers matter. In criminology, research has shown the overwhelming tendency
for delinquents to have delinquent friends and for offending to occur in the
company of others (e.g. Akers, Krohn, Lanza-Kaduce, & Radosevich, 1979;
Elliott, Huizinga, & Ageton, 1985; Krohn, 1974; Matsueda & Heimer, 1987;
Warr, 1996). While debate continues as to why peers matter for youth develop-
ment, for example through attitudes (Sutherland, 1947), reinforcement (Akers,
1985), and/or opportunities (Jussim & Osgood, 1989), the evidence is clear
that peers influence behavior (Warr, 2002).
Peer influence is conditional. Researchers have long acknowledged that the
effects of delinquent peers are conditioned by specific features of social
relations (e.g. Krohn, 1986; Orcutt, 1983; Short & Nye, 1957; Voss, 1969), such
as friendship quality and social context (Bauman, Robert, Ennett, Hussong, &
Foshee, 2007; Zimmerman & Messner, 2011). In recent years, however,
scholarsaccess to a level of relational and behavioral detail and ability to
model these influences has been revolutionized by social network methodolo-
gies. The capacity to model the pattern of ties within these networks has
revealed that the structure of friendship network shapes the relationships
between peer influences and behavior (e.g. Haynie, 2001; Young, Barnes,
Meldrum, & Weerman, 2011). Social network methods have also renovated
work on measures of peer influence. No longer do scholars have to rely upon
respondentsreports of what they think their friends do and believe. Instead,
social network methods allow researchers to gain information directly from
individualsfriends themselves. Scholars have argued that indirect perceptual
measures were seriously flawed and may in some cases simply reflect assumed
similarity and/or pluralistic ignorance (e.g. Gottfredson & Hirschi, 1990;
Jussim & Osgood, 1989).
Criminologistsunderstanding of peer influences has thus advanced in a num-
ber of ways, and scholars continue to refine and apply new statistical modeling
techniques (e.g. Weerman, 2011). Equal attention, however, has not been paid
to measurement. With a few exceptions (e.g. Haynie, 2002; McGloin, 2009;
Weerman & Smeenk, 2005), scant attention has been paid to advancing mea-
surement of peer influence through peer behavior in social network models. In
particular, research has generally neglected variation in peer behaviors within
a friendship network. Although research demonstrates that considerable
variation in delinquency exists within peer groups (Elliott, Huizinga, & Menard,
1989; Elliott & Menard, 1996; Haynie, 2002), peer behavioral influence in net-
works is still frequently operationalized as the average frequency of delin-
quency across the network (e.g. Crosnoe, Muller, & Frank, 2004; Haynie &
Payne, 2006; Payne & Cornwell, 2007). As a result, we have increasingly
advanced statistical models relying upon a basic measure of peer influence.
In the present study, we seek to advance the measurement of peer influence
in social network models, drawing upon dominant theories of peer effects in
criminology. We argue that using an average level of delinquency across a peer
network collapses theoretically important behavioral heterogeneity. For
example, this average measure assigns a score of 5 for peer delinquency in all
2 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
of the following networks: five friends who each engage in five delinquent
acts; one friend who engages in five acts; five friends who engage in ten acts
and five friends committing zero acts; and two friends who engage in three
acts and two with seven acts. Drawing upon learning and opportunity theories
in criminology, we argue that there is good reason to believe that this peer
behavioral heterogeneity shapes the messages, reinforcers, and/or
opportunities for deviance that peers provide. To incorporate such behavioral
heterogeneity, we include a measure of the variance of peer delinquent behav-
ior in the network and argue that the variance conditions the effects of peer
average delinquency. Specifically, we propose that because the definitions,
messages, reinforcers, and/or opportunities for delinquency are more variable
and inconsistent among networks with more heterogeneity in delinquent
behavior, the influence of the average peer delinquency on respondent delin-
quency should be weaker in networks with more variance.
In addition, departing from much prior work, we focus on respondent and
peer involvement in specific offenses. Research indicates that considerable
variation exists in the types of delinquency engaged in by individuals and their
peers (e.g. Nieuwbeerta, Blokland, Piquero, & Sweeten, 2011). With a few
exceptions (e.g. Jackson, Tittle, & Burke, 1986; Rees & Pogarsky, 2011), stud-
ies have utilized omnibus measures of peer and individual delinquency. How-
ever, all offenses are not created equal, and acceptance, tolerance, and
endorsement of particular offenses do not necessarily transfer analogously to
other offenses (Warr, 2002). From the perspective of learning theory, having
friends who smoke marijuana, seemingly enjoy smoking marijuana, and profess
positive attitudes towards marijuana use should not have the same influence
on an individualslikelihood of shoplifting or starting a physical fight as it
would on that individuals marijuana use. In the same vein, opportunity factors
should also be more offense-specific; the opportunity factors conducive to
smoking marijuana are not the same as those for theft. For these reasons, we
examine peer and individual involvement in specific offenses.
In sum, this study seeks to extend prior research by focusing on an impor-
tant, yet relatively neglected, aspect of peer behavior: heterogeneity. Most
adolescents do not have friends who engage in equivalent levels of offending,
but instead are affiliated with friends with varying levels of offending (Elliott
et al., 1989; Elliott & Menard, 1996; Haynie, 2002; Weerman & Smeenk, 2005).
We argue that measures of the average or absolute levels of delinquency in a
network are conditioned by the amount of behavioral heterogeneity in the net-
work as measured by the variance. To test our hypothesis, we use detailed
adolescent friendship network data from the National Longitudinal Study of
Adolescent Health and social network methods. Controlling for structural prop-
erties of friendship networks (density, ego centrality, and ego popularity), we
examine whether peer variance in a given delinquent behavior attenuates the
effect of average peer involvement in the behavior for two delinquent
substance-use outcomes: smoking cigarettes and drunkenness. In doing so, we
seek to contribute to our understanding of how peers matter for delinquency.
BEHAVIORAL HETEROGENEITY 3
Downloaded by [Arizona State University] at 23:15 26 March 2014
Behavioral Heterogeneity and Peer Influence: Implications from Theory
Although it is certainly the case that self-selection processes account for a
portion of the peer-delinquency associationprinciple of homophily at dyad
level; homogeneity at group level (e.g. Cairns & Cairns, 1994; Lazarsfeld &
Merton, 1954; McPherson, Smith-Lovin, & Cook, 2001)–research has clearly
demonstrated that peers shape individual behavior (e.g. Thornberry, Lizotte,
Krohn, Farnworth, & Jang, 1994; Warr, 2002). Two theoretical explanations for
this peer effect predominate in criminology: differential association/social
learning theories (Akers, 1985; Sutherland, 1947) and social ecological
approaches or opportunity theories (Cohen & Felson, 1979; Haynie & Osgood,
2005; Osgood et al., 1996). In this section, we discuss these two explanations
and their implications for understanding the impact of behavioral heterogene-
ity among peers.
Learning Perspective
According to Sutherlands(1947) differential association theory, techniques of
committing crime and definitions (motives, drives, and attitudes) are learned
in intimate social networks. The central proposition of the theory is that an
individual becomes delinquent because of an excess of definitions favorable
to the violation of law over definitions unfavorable to the violation of law
(Sutherland, 1947, p. 7). Thus, in friendship networks delinquency is transmit-
ted through attitudes about the appropriateness of delinquent behavior and, to
a lesser extent, techniques for committing the offense. Importantly, Suther-
land posited that it is the balance or ratio between definitions encouraging or
excusing and discouraging law violation that determines whether an individual
becomes delinquent (see also Haynie, 2002).
From this perspective, behavioral heterogeneity in peer delinquency is
manifestly influential, as it is not simply the number of delinquent peers one is
exposed to or the average amount of peer offending that shapes individuals
definitions, but the ratio of delinquent to non-delinquent definitions. This ratio
is not captured in a measure of average peer delinquency across a network.
Therefore, we argue that Sutherlands(
1947) theoretical notion of the ratio of
definitions implies a more nuanced measure than peer average delinquency.
The idea inherent in the notion of ratio of definitions is that delinquent and
non-delinquent messages offset each other. Thus, all else equal, the more
consistent pro-delinquent messages of an individual whose peers all engage in
moderate levels of delinquency should result in more pro-delinquent
definitions, on balance, than an individual with the same average level but has
friends who engage in high levels of delinquency and those that eschew delin-
quency. This implies that the amount of heterogeneity around the average
level of peer delinquency in a network should attenuate the effect of the aver-
age peer delinquency.
4 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
Our interpretation of Akerss(1985) extension of differential association the-
ory also suggests that variation in delinquency among peers in a network should
affect peer influence. By more clearly specifying the process of learning
through operant conditioning, Akerss(
1985) social learning theory emphasizes
the role of reinforcement. Reinforcement refers to the anticipated or actual
consequences of given behaviors and can be social or non-social. Behaviors are
differentially reinforced, and this differential reinforcement determines indi-
vidual behavior: individuals engage in behaviors that they believe will result in
more rewards or help them avoid punishments. These rewards and punishments
can be social (e.g., gaining status or avoiding ridicule; e.g. Warr, 2002) or non-
social (physiological pleasure or pain), but the theory asserts that the principle
behavioral effects come from the interaction in or under the influence of those
groups which control individualsmajor sources of reinforcement and punish-
ment and expose them to behavioral models and normative definitions(Akers
et al., 1979, p. 638; emphasis omitted). Peers can influence this process in two
major ways: by reinforcing behavior directly and/or by providing behavioral
models. We posit that variation in delinquency by peers should be expected to
influence both of these processes and expound upon this idea below.
Social learning theory proposes that the probability and frequency of
delinquency is higher when there is greater exposure to delinquent vs.
non-delinquent models, when there is greater association with delinquent vs.
non-delinquent friends, when delinquency is differentially reinforced over
non-delinquency, and when there are more positive or neutral than negative or
discouraging definitions towards delinquency (Akers, 1985; Akers et al., 1979).
As with differential association, this implies a ratio of delinquent vs. non-
delinquent definitions, which is not captured in a measure of the number of
delinquent friends (as this does not incorporate number of non-delinquent
friends or account for differences in peerslevels of delinquency) or the
average amount of delinquency among friends.
To use an example, Warr (2002) has noted that ridicule, as negative
reinforcement, is a powerful inducement to conformity among peers.
Individuals may choose to go along with the delinquent acts of their peer group
because not doing so would entail ridicule. Insofar as a peer group varies in
the behavior, for example half smoke and half do not, ridicule becomes less of
a potent conforming mechanism. Similarly, according to social learning theory,
if one-third of an individuals peers shoplift frequently and two-thirds do not,
all else equal, the balance of definitions and models of shoplifting would be
against the behavior. This ratio of definitions is distinct from a network with
the same average where all individuals engage in low to moderate levels of
shoplifting. While these examples are certainly simplifications of what is ulti-
mately a complex process for illustrative purposes, our argument is that, from
the perspective of learning theories, behavioral heterogeneity among peers
should shape peer influence. Specifically, we argue that based on the idea of
ratios of definitions, the amount of variation around the mean levels of peer
delinquent behavior should weaken the effects of this average level, such that
BEHAVIORAL HETEROGENEITY 5
Downloaded by [Arizona State University] at 23:15 26 March 2014
the less behavioral heterogeneity in a peer network, the more the more consis-
tent the definitions and reinforcements and, therefore, the stronger the influ-
ence of mean peer offending.
Opportunity Perspective
Opportunity theory provides another theoretical explanation for peer influence
and reason to believe that peer behavioral heterogeneity matters. Whereas
learning theories emphasize the normative influence of social relationships,
the opportunity perspective emphasizes the role of peer networks in structur-
ing activities and providing resources and fields for behaviors–legal and illegal
(Haynie & Osgood 2005; Osgood, Wilson, OMalley, Bachman, & Johnston
1996). Osgood and colleagues (1996) extended the routine activities perspec-
tive (Cohen & Felson, 1979) to focus on individual offending, highlighting the
role of peers in shaping situational opportunities for deviance. Because of the
different routine activities in which they engage and thus the opportunities
they provide, delinquent and non-delinquent peers should differentially influ-
ence opportunities for delinquency and therefore the likelihood of offending.
1
For parsimonys sake, opportunities for delinquency can be broadly divided into
two categories: lack of supervision (or absence of external controls) and
resources needed for the act (goods, substances, etc.). There is good reason
to expect that behavioral heterogeneity in delinquency in peer networks can
influence both facets of opportunity.
Regarding the supervision or absence of guardians facet of delinquent oppor-
tunity, scholars have argued that individuals with high criminal propensity are
more likely to select themselves into situations (activity fields) that facili-
tate doing what they like to do (including offending; e.g. Wikstro
¨m et al.,
2010; Wikstro
¨m, Oberwittler, Treiber, & Hardie, 2012). Thus, more delinquent
individuals are more likely to gravitate towards settings that provide more
opportunities for offending (abandoned houses, friendshouses whose parents
are out of town, bars that allow underage patrons, etc.). It is reasonable to
expect that behavioral heterogeneity in delinquency among peers would affect
the consistency of or amount of time spent in settings void of guardians that
would interfere with delinquency. If all friends in a network like to party and
drink heavily, we would expect that this group would more consistently gravi-
tate towards activity fields conducive to underage drinking than one in which
some friends drink heavily and others do not. Thus, through its influence on
1. This is confounded by Osgood and colleagues (1996) prediction that peer relations are not con-
nected to delinquency by the type of peers one chooses, but rather by the amount of time one
spends with peers in unstructured socialization. In their test of this idea using the Add Health data,
Haynie and Osgood (2005) found support for the idea that unstructured socializing increased delin-
quency net of peersaverage delinquency; however, peersdelinquency continued to have a signifi-
cant effect on respondentsdelinquency.
6 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
the consistency of selection into settings without capable guardians, we expect
that peer behavioral heterogeneity attenuates the effects of average peer
delinquency on adolescent offending.
In addition, as routine activities theory (Cohen & Felson, 1979) emphasizes,
most delinquent acts not only require the absence of controls, but also
resources. This includes such things as spray paint for tagging, guns for shoot-
ing, substances for imbibing, and the like. As with the supervision aspect of
opportunity, we argue that there is reason to expect that variation in
delinquent behaviors among peers should influence the availability of resources
for specific offenses. Smoking cigarettes, for example, not only requires the
absence of guardians and motivation but also the presence of cigarettes. In an
individuals friendship network in which all friends smoke, we would expect a
more consistent presence of cigarettes when that individual hangs out with
friends than in the case of an individual with no peers who smoke. Insofar as
there is variation in peer smoking in a friendship network, we would expect
the presence of cigarettes to be less consistent than in more homogenous
networks.
2
Behavioral Heterogeneity in Friendship Networks
From the perspectives of learning and opportunity theories, we argue that
there is good reason to believe that behavioral heterogeneity–which is not
captured in measures of average peer delinquency or the number of delinquent
peers–in a peer network matters. Two studies, in particular, have incorpo-
rated some aspect of peer behavioral heterogeneity and provide support for its
importance.
Building on research demonstrating that friendship networks are heteroge-
neous in delinquency as well as theoretical insights from differential associa-
tion and control theories, Haynie (2002) explored whether the proportion of a
network that engages in delinquency has more influence than the absolute
amount of delinquency among the peers in an individualsnetwork. Haynie
(2002, p. 106) argued: Assuming that delinquent friends provide favorable
definitions and modeling of delinquent behavior and non-delinquent friends
provide unfavorable definitions and modeling of prosocial behavior, the
proportion of delinquent friends may be more important than the frequency of
delinquent acts committed by friends.The results using the Add Health data
2. Notably, however, the relevance of the opportunity perspective for explaining behavioral heter-
ogeneity on resource opportunity is likely greater for certain offenses, such as substance use,
which necessarily rely on the availability of substances. Certain behaviors, including those Warr
(2002) highlights as less groupy–especially violent offenses, and to a lesser extent shoplifting
and other forms of stealing–rely less on the presence of a resource that is difficult to procure and
more on the spontaneity that often characterizes much of youth delinquency such as various forms
of vandalism or senseless destruction (see the excellent example that opens Warrs(
2002) book).
BEHAVIORAL HETEROGENEITY 7
Downloaded by [Arizona State University] at 23:15 26 March 2014
support Haynies(2002) hypothesis, showing that when examined together, the
proportion of delinquent friends significantly predicts an individuals
delinquency while the absolute level does not.
Although the proportional measure represents an advance over average peer
measures in its recognition of behavioral heterogeneity among peers by more
explicitly modeling the ratio of delinquent peers, it does so at the expense of
capturing variation in within-peer levels of offending. The proportional mea-
sure captures the ratio of delinquent and non-delinquent peers, thereby mod-
eling between-peer variation, by collapsing peer offending into a dichotomy
(delinquent or non-delinquent). As such, in this proportional measure, individu-
als who engage in all delinquent acts at a high level or one delinquent act on a
single occasion–and all those in between–are coded as delinquent.The
measurement strategy we propose allows for the incorporation of both
between and within levels of offending among peers, based on our theoretical
argument that both forms of variation matter.
Additionally, we depart from Haynies(
2002) general delinquency measures
by examining peer and ego involvement in specific offenses. Although, like
Haynie (2002) most prior research has utilized an omnibus measure of delin-
quency, we argue that learning theories, and to a lesser extent opportunity
theories, imply that the influence of peers on behavior is more offense-specific
than general (e.g. Jackson et al., 1986; Matsueda, 1982), and this may be
especially true for the status use substance offenses that are more common
and more groupyamong adolescents.
3
Thus, having friends who smoke
should have only a tenuous relationship to the likelihood that the individual
should start a physical fight, net of those friendspropensities to smoke. Warr
(2002, pp. 134–135) articulately notes:
[I]nvestigators who employ broad composite indices of offending seem to be
relying on a theory of peer influence that emphasizes some form of attitude
transference with very generalized consequences. Having friends who commit
any kind of delinquency, in other words, supposedly opens the door for adoles-
cents to commit any other kind of delinquency. Available evidence, however,
consistently fails to support the notion of [general] attitude transference when
it comes to peer influence and delinquency (Warr & Stafford, 1991)
In a second study that supports the relevance of behavioral heterogeneity,
Rees and Pogarsky (2011) examined the relative influence of a best friends
delinquency compared to the remaining social network on an individuals
offending. Consistent with their expectations, they found that a best friends
3. To be sure, it is not our argument that general delinquency or unspecialized pro-delinquent def-
initions have no influence on adolescent offending. Instead, we argue that when predicting more
groupyand common(often viewed as less serious) forms of offending among samples of the
general adolescent population, peer behavioral influences should be greater for offense-specific vs.
general delinquency scales for the reasons we have described (e.g. Jackson et al., 1986; Matsueda,
1982; Warr & Stafford, 1991).
8 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
influence on respondent drunkenness and fighting was weakened when the
remaining social network was different from the best friend. This provides
support for the arguments presented here that peers present competing behav-
ioral models and that peer behavioral heterogeneity conditions peer influence.
We build upon these excellent extensions of peer influence focusing on the
variation in peer behaviors.
Current Study
As we have noted, average measures of peer delinquency capture only the cen-
tral tendency of the network, and thereby ignore variation around the mean.
Peer influences are much more complex than what is captured in the average
measure (e.g. Sampson, 1999). We argue that incorporating behavioral
heterogeneity qua variance adds some needed context to the mean peer scores.
Specifically, we hypothesize that the variance in peer delinquent behavior in a
friendship network attenuates the effect of the average amount of peer delin-
quency, net of individual and network controls (viz., density, ego centrality, and
ego popularity). In the present study, we examine the effects of behavioral het-
erogeneity on adolescent involvement in two substance-use offenses–smoking
cigarettes and getting drunk–and peer involvement (and variation) in these
offenses.
4
These two offenses were selected for several reasons. First, unlike
the general delinquency measure in the wave 1 in-home survey, which is mea-
sured as a variety count, the measurement of these outcomes in the Add Health
allowed us to capture variation in levels both within and between peers. In addi-
tion, as we have argued, given features of substance-use offenses in adolescence
(their group nature, reliance on opportunity), we believe these two outcomes
are apposite for investigating the relevance of behavioral heterogeneity on peer
influences in adolescence. Smoking and getting drunk represent illegal behaviors
that are both groupyto a great extent, strongly influenced by peers, risky, and
potentially harmful (e.g. Warr, 2002; Zimmerman & Vasquez, 2011).
Methods
Data
We assess our hypotheses with data from the National Longitudinal Study of Ado-
lescent Health (Add Health), a nationally representative sample of adolescents
in grades 7 through 12 at the start of the study (Udry, 2003). Multistage stratified
sampling techniques (schools were stratified by school type, region, urbaniza-
tion, racial-ethnic composition, and size) were employed to identify a nationally
representative sample of 80 high schools and 52 middle schools. All students
4. Unfortunately, measures of other substances (e.g. marijuana use) were not included in the in-
school survey; thus, we could not incorporate a lagged measure to control for selection.
BEHAVIORAL HETEROGENEITY 9
Downloaded by [Arizona State University] at 23:15 26 March 2014
attending the schools were administered a self-report survey which asked about
their peer relationships, the dynamics of their family, and various behaviors.
More than 90,000 students, most of whom were between the ages of 12 to 18,
completed the in-school survey administered during the 1994–1995 school year.
Within this school-based sample, a stratified random sample of approximately
20,000 students (stratified by grade and sex) was selected for in-depth follow-up
interviews in their homes between April and December of 1995. A total of 20,745
adolescents participated in the Wave 1 in-home survey (Harris et al., 2003).
The core sample for the ensuing analyses consist of respondents who com-
pleted the in-school and wave 1 in-home interviews, have valid data on the
dependent variables at both of these time points, have at least one nominated
friend with valid delinquency data on the in-school questionnaire, and have
non-missing sample weights.
5
Because data on the dependent variables vary by
outcome, the sample size differs for the smoking (n= 7,394 persons) and get-
ting drunk (n= 7,379) models.
6
To model the appropriate time order of vari-
ables, the outcomes are measured at the wave 1 in-home interview and
predictor variables are measured at the in-school interview, with the exception
of receipt of public assistance which was not included in the in-school survey.
Each model also includes a lagged outcome to examine the change in delin-
quency as a result of peer influence (the regressor variable method; Allison,
1990). All analyses corrected for the design of the Add Health survey data by
using the PSU, Stratum, and Wave 1 grand sample weights (see Chantala &
Tabor, 1999). As a result of these corrections, our analyses produce unbiased
point estimates of regression parameters along with corrected estimates of
variances, standard errors, and confidence intervals.
7
Dependent Variables
We examine involvement in two substance-use offenses: cigarette smoking and
getting drunk. Smoking was measured with respondentsanswers to the question
on how many days out of the past thirty have you smoked. This variable is thus
measured as the number of days the respondents smoked, ranging from 0 to 30.
8
5. Slightly more than 30% of the original wave 1 in-home sample was lost due to missing sample
weights. In order to properly analyze the in-home sample, respondents must have a valid probabil-
ity weight (see Chantala & Tabor, 1999).
6. We also estimated the models with multivariate imputation of missing values under missing-at-
random assumptions. Multiple imputation using chained equation techniques in StataMP version 12
(StataCorp, 2011) was used to produce parameter estimates and standard errors based on the com-
bination of models from 20 imputed data-sets. The results using the multiply imputed data are
analogous to that presented here and are available upon request.
7. The guidelines for correcting analyses for design effects and unequal probability of selection
in the Add Health can be found here: http://www.cpc.unc.edu/projects/addhealth/data/guides/
wt-guidelines.pdf.
8. The smoking item we incorporate asks about the frequency of smoking over the past month
rather than the past year. This difference in measurement is due to data limitations. In the wave 1
in-home survey, a measure of frequency of smoking in the past year was not included.
10 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
Heavy alcohol consumption (drunkenness) in the past 12 months was measured
by asking on how many days the respondent had gotten drunk or very, very
highon alcohol. This variable was also coded from 0 neverto 6 every day or
almost every day.Approximately 25 and 18% reported engaging in some level of
smoking and getting drunk, respectively. See Appendices Aand Bfor descriptive
statistics of all study variables by outcome.
Lagged dependent variables for the outcomes were included to estimate the
change in offending. Respondents were asked identical questions for getting
drunk, and responses were coded the same way as above. For smoking, the ques-
tion asked about their cigarette use during the past 12 months. For this measure,
responses were coded 0 neverto 6 every day or almost every day.
Peer Influence Variables: Respondent Nominated Network
Network Average Delinquency Outcome
The same substance-use questions asked of respondents were presented to
respondent-identified friends. Average peer smoking and getting drunk are
measured as the mean of responses of all respondent-identified friends to
these questions asked during the in-school interviews.
Peer Behavioral Heterogeneity
We have argued that the amount of variation around the mean level of
delinquency should attenuate the effects of this average level on respondent
delinquency. We propose a simple measure of behavioral heterogeneity:
variance around the mean. In other words, we capture behavioral heterogene-
ity by incorporating the variance for each of the average network offenses
using the formula:
peer behavioral heterogenity ¼Pðxith friend delinq #!
xnetwork mean delinqÞ2
ntotal friends in network
Higher values on this variable reflect greater variance around the average
delinquency level of the network. Network variance is at a minimum (0) when
all members of the network have the same level of delinquency and at a maxi-
mum when all members are different from all other members.
9
9. Of course, network variance for respondents who identify one friend is zero. We include respon-
dents who identify one friend in the analyses given our arguments about consistency of definitions
and opportunities, which we believe apply to a network of 1 or 10. However, we also estimated
the models excluding respondents who only nominate 1 friend and therefore cannot have a vari-
ance different than 0. The pattern of results from these models is equivalent to that presented
here; these models are available upon request.
BEHAVIORAL HETEROGENEITY 11
Downloaded by [Arizona State University] at 23:15 26 March 2014
Individual-level Control Variables
We include the following demographic variables in the models: sex (1 = Male),
race/ethnicity (White, Black, and Other), and age (at wave 1). Consistent
with prior work using the Add Health data, we also control for a number of
family- and school-level variables that competing theories identify as impor-
tant to delinquency and substance use (e.g. Hirschi, 1969).
10
Friend attach-
ment is the average response of the network to the question how much do
you feel your friends care about you?Friend involvement captures how much
time the respondent spent (or in communication) with each nominated friend
during the past week. Respondents were asked if they had gone to a friends
house, met with a friend to hang out, spent time with a friend on the prior
weekend, and talked on the phone with the friend (1 = yes, 0 = no). These
responses were summed for each friend and then summed across all friends.
Family-level variables measured include: parental attachment, family struc-
ture, and public assistance. Parental attachment (Hirschi, 1969) is the average
of respondentsresponses (on a scale from 1 not at allto 5 very much) to
two questions measuring how much their mothers and fathers care about them
(How much do you think he/she cares about you?). Family structure is a
dichotomous variable representing the presence of two parental figures in the
home vs. one parental figure in the home. Public assistance measures whether
the household received at least one form of public assistance during the month
preceding the interview, including: social security or railroad retirement, sup-
plemental security income, Aid the Families with Dependent Children (AFDC),
food stamps, unemployment or workers compensation, or some form of hous-
ing subsidy.
Individual school-related variables include school attachment and GPA.
School attachment is measured as the average response to three questions on
a five-point scale from 1 (not at all) to 5 (very much): I feel close to
people at this school,”“I feel like I am part of this school,and I am happy
to be at this school.GPA, calculated using the average of a respondents Eng-
lish/Language Arts, Mathematics, Social Science/History, and Science grades
from the most recent grading period at the time of the wave 1 in-school inter-
view, ranges on a continuous and roughly normally distributed scale from 1 D
or lowerto 4 A.We also control for three network structural properties,
including: density,centrality, and popularity, which prior research has shown
influences on individuals delinquency (e.g. Haynie, 2001; Kreager, Rulison, &
Moody, 2011).
10. Notably, our results are not dependent on the inclusion of these controls, which we chose out
of the many possible ones given their theoretical and empirical relationships with peer relations,
influences, and substance use. The pattern of results is the same whether or not these competing
theoretical constructs are included or excluded.
12 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
Analytic Strategy
We test our hypothesis with four models, two for each outcome. Negative
binomial models were estimated because the distributions of these variables
are highly skewed, truncated at zero, and discrete, and therefore best mod-
eled as event counts (e.g. Long, 1997; Long & Freese, 2003).
11
In the first
model for each outcome, we link our analyses to prior works by replicating the
association between respondent offending and his or her network average
offending, along with other network and other controls. In the second model,
we incorporate the measure of behavioral heterogeneity–the network vari-
ance–as well as its interaction with the average level of network delinquency
to test our hypothesis that the effect of peer average offending is attenuated
by behavioral heterogeneity. We are interested in how the effect of average
network delinquency changes in magnitude and significance over values of
network variance. The interaction was created following the usual product
term protocol (average peer level ×variance; Aiken & West, 1985).
Notably, the interpretation of interaction effects in non-linear models is
not as straightforward as that from linear ones. Although it is still informa-
tive to graphically depict interaction effects from non-linear models by
calculating marginal effects (Buis, 2010; Greene, 2010), the interpretation
of interaction terms using marginal effects in non-linear models requires
additional considerations (Karaca-Mandic, Norton, & Dowd, 2012). In linear
models, the marginal effect of the interaction can be accurately quantified
by the partial effect coefficient of X ×Z holding all other covariates
constant. This does not hold for non-linear models because, in contrast to
linear models, the values of the other variables in the model influence the
partial effects (Ai & Norton, 2003; Norton, Wang, & Ai, 2004). Following
the suggestion of Norton and colleagues (2004), we therefore calculate
average marginal effects (AME), computed by predicting the outcome for
each level of the focal variable and other levels of the covariates and then
averaging these predictions within each level. In addition, to further test
the robustness of our interpretation, we also calculate marginal effects at
the means (MEM). These effects are calculated while holding all other
covariates not involved in the interaction term at their respective means. In
analyses not shown, we also conducted sensitivity analyses holding covari-
ates at their minimum, maximum, and IQR values each of which reported
equivalent substantive interpretations.
12
11. In each case, Poisson regression was inappropriate because the overdispersion parameter was
significant.
12. These results are available from the authors upon request.
BEHAVIORAL HETEROGENEITY 13
Downloaded by [Arizona State University] at 23:15 26 March 2014
Results
Before discussing the results of our multivariate models, a few zero-order
correlations deserve mention. First is the relationship between behavioral het-
erogeneity and network density. We would expect that more dense networks,
those in which more of an individuals friends are also friends with each other,
would have less behavioral heterogeneity; correlations are consistent with this
expectation, with values of r=.19 for drunkenness and r=.25 for smoking
(p< .01). Thus, while more dense networks have less behavioral heterogeneity,
the relationship is only moderate, indicating that behavioral heterogeneity
clearly is not merely a proxy or outcome of network density. The correlations
between network variance and respondent popularity and centrality are
significant for each outcome but small in size. Popularity correlations with the
variance are r= .07 for drunkenness and r= .08 for smoking, whereas central-
ity correlates at r= .10 for smoking and drunkenness (p< .01 for all). Thus,
not surprisingly, individuals who are more popular (received more nominations
from others) and more central have networks with more behavioral heteroge-
neity in their networks. Lastly, the correlation between the peer average mea-
sure and the variance is approximately .58 for smoking and .59 for
drunkenness (p< .001).
13
A scatterplot of these two variables for each out-
come is presented in Figures 1(a) and 1(b). In short, these correlations and the
scatterplot indicate that behavioral heterogeneity qua variance is associated
with other network characteristics in expected ways but is not reducible to
them.
The results of our multivariate models are presented in Table 1. In the A
models of Table 1, we present the baseline model for the outcome containing
controls derived from theory and past research. These results are consistent
with prior work. For example, for both offenses, a respondents past offending
strongly predicts future offending, with the effect being strongest for smoking.
Sex is not significantly associated with these outcomes, net of the controls.
Being black is associated with less smoking and drunkenness, especially for
smoking. The heterogeneous other racecategory did not differ significantly
from whites in their expected rate of substance use. The A models also show
that respondents are more likely to get drunk when they do not live with two
parents, but family structure does not significantly influence the likelihood of
smoking. In line with control theories, both parental and school attachments
as well as GPA are negatively associated with both substance-use outcomes.
Moving further down the A models, we can see that more popular respon-
dents (those who received more friendship nominations) and those who occupy
13. This moderate correlation is to be expected given the inherent relationship between the peer
mean and variance at certain levels. For individuals whose peers all eschew substance use (25% for
smoking and 32% for drunkenness), the variance is zero (a perfect correlation). Even so, the corre-
lation is far from unity. Moreover, sensitivity analyses confirm that multicollinearity is not a prob-
lem in any of the models.
14 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
Figure 1(a) Average network smoking by network smoke variance.
Figure 1(b) Average network drunkeness by network drunk variance.
BEHAVIORAL HETEROGENEITY 15
Downloaded by [Arizona State University] at 23:15 26 March 2014
a more central location in their network get drunk more frequently, net of
other factors in the model, but neither were significantly associated with
smoking. Individuals with larger networks (network size measured by number
of friends respondents selected) had lower expected counts of smoking and
getting drunk. Last among the controls, respondents in more dense networks
had lower rates of smoking than those in less dense networks, net of other
factors; network density was not significantly related to drunkenness.
Consistent with prior research using peer average measures, the A models
reveal that average peer smoking and drinking have a strong influence on
respondent smoking and drinking, net of other variables in the models. For
example, for a standard deviation increase in peer average smoking, the
expected count (number of days smoked in the past month) increases by a fac-
Table 1 Negative binomial models examining the influence of peer delinquency on
respondentsdelinquency
Outcomes Smoking Getting drunk
Model A Model B Model A Model B
Independent variables Exp^βExp^βExp^βExp^β
Lagged outcome 3.20
***
3.26
***
1.80
***
1.81
***
Sex (1 = Male) 1.01 1.02 .99 .99
Age 1.09 1.09 1.33
***
1.33
***
Age squared .93 .92 .91
**
.92
*
Respondent race: Black .68
***
.70
***
.83
***
.82
***
Respondent race: Other .91 .90 .96 .96
Public assistance 1.01 .99 .99 .99
Two parents at home .93 .91 .90
**
.90
**
Parental attachment 1.10 1.11 .93
*
.93
*
Friend attachment 1.15
*
1.16
**
.94 .94
Friend involvement 1.24
**
1.24
**
1.20
***
1.20
***
School attachment .87
*
.86
*
.91
**
.91
**
GPA .70
***
.70
***
.84
***
.84
**
Respondent popularity 1.08 1.08 1.18
***
1.17
***
Respondent centrality 1.01 1.07 1.16
*
1.16
*
Network density .85
*
.88 .97 .98
Network size .73
*
.66
***
.77
***
.76
***
Network average delinquent outcome 1.32
***
1.26
***
1.23
***
1.27
***
Network variance around average peer delinq. 1.71
***
1.26
**
Network average ×Variance .68
**
.78
**
N7,394 7,379
Note. Exp^βindicates the factor change in the expected count of the outcome for a standard
deviation increase in the predictor, holding all other variables constant. Time between interviews
is controlled for but not shown.
***
p< .001;
**
p< .01;
*
p< .05 (two-tailed tests).
16 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
tor of 1.32, holding all other variables constant. Similarly, a standard deviation
increase in peer average drunkenness increases the expected counts of respon-
dentsdrunkenness by a factor of 1.23. Thus, consistent with prior work and
theory, average peer substance use significantly increases respondent sub-
stance use.
We have argued, however, that measures of peer influence through peer
average offending collapse meaningful variation. Thus, we next test the
hypothesis that the variance in peer substance use conditions the significant,
positive effect of average peer substance use on the outcomes. These results
are displayed in the B Models of Table 1and provide support for our hypothe-
sis.
14
Looking first at smoking, we can see that the interaction term is signifi-
cant and the standardized event rate ratio is less than one (Exp^β= .68),
indicating a negative effect. Thus, as predicted, the positive effect of average
peer smoking on respondent smoking is attenuated by smoking variance in the
peer network. The relationship between average network smoking and network
variance on respondent cigarette use are presented as marginal effects in
Figure 2along with associated 95% confidence intervals. This figure shows that
network variance weakens the effect of average peer smoking, such that at
high levels of variance, average peer smoking does not significantly influence
the likelihood of respondent smoking. Specifically, a one unit increase in
average network smoking when there is no variation in peer behavior leads to
an average of 0.96 additional number of days smoked by a respondent. This
effect however is decreasing across increasing values of network smoking
variance with the final significant effect being just over an average of half a
day (.53) at a variance level of 1.5. The average network smoking effect is no
longer significant when peer smoking heterogeneity levels are above 1.5.
The regressions predicting getting drunk, displayed in the last column of
Table 1, also provide support for our variance hypothesis. There is a negative
and significant interaction between friendsaverage frequency of getting drunk
and the variation. This implies that the effect of average peer drunkenness on
the expected frequency of respondent drunkenness decreases as network
variation increases. The implications of the relationship between average
network smoking and network variance on respondent heavy alcohol consump-
tion are presented in Figure 3and again indicate that peer networks that are
more homogenous in their involvement in heavy drinking have a stronger
influence on individual drunkenness than those with more behavioral
heterogeneity.
Illustratively, a unit increase in average network drunkenness when all peers
engage in the same level of drunkenness (i.e. variance = 0) leads to an average
.187 unit increase in respondent drunkenness. This effect however is
decreasing across increasing values of network smoking variance with the final
14. Appendix C displays the marginal effects of the network average offending across different lev-
els of network variance based on the B models in Table 1.
BEHAVIORAL HETEROGENEITY 17
Downloaded by [Arizona State University] at 23:15 26 March 2014
significant effect being just over an average of .073 at a variance level of 2.
The average network drunkenness effect is no longer significant when the vari-
ance of peer drunkenness is 2.5 or greater.
Our findings are consistent with prior research in that average friendship
network smoking and getting drunk are a strong and significant predictors of
respondent use of these substances. Most focally for the present study, we also
find that variation in the behavior of the friends in a respondents friendship
network conditions this relationship. Generally, these findings suggest that at
least for these two substance-use outcomes, heterogeneity in peer behavior
measured as variance around the mean moderates the influence of average
peer offending.
Discussion and Conclusion
In recent years, scholarsunderstanding of peer influences on delinquency has
been greatly advanced by network methods, which have allowed for the mod-
eling of the structure of peer relations and direct measures of peer character-
istics and behaviors. In this paper, we have argued that these advances have
not been matched in the measurement realm, such that with a few exceptions
(e.g. Haynie, 2002; McGloin, 2009; Rees & Pogarsky, 2011), measures of peer
influence have not been improved. Despite the great advances in social
network analyses, peer influence is still invariably measured by simply averag-
ing the level of offending across the observed network. In the present paper,
we have argued that this average measure of peer delinquency within a friend-
ship network collapses variation that is theorized to be relevant to peer influ-
ence. In particular, learning theories posit that it is the ratio of exposure to
delinquent and non-delinquent behaviors and attitudes that shapes individual
Figure 2 Average marginal effects of average network smoking (95% Cls).
18 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
attitudes and behavior (e.g. Akers, 1985; Sutherland, 1947). Similarly, opportu-
nity theories imply that variation in offending among the peers in an individ-
uals network should affect peer influence through their direct influence on the
availability of resources and situational factors conducive to or necessary for
offending. This balance of definitions, reinforcements, and/or opportunities is
not captured in average measures, as these neglect between peer variation,
nor proportional measures (i.e. Haynie, 2002), as these neglect within-peer
variation in levels of offending.
Drawing upon theory and prior research, we sought to advance research into
peer influences on delinquent behaviors by recognizing and explicitly modeling
behavioral heterogeneity in peer networks. We incorporated a novel measure
of peer behavioral heterogeneity to capture the competing behavioral models
and implied ratio of definitions that individuals are exposed to in their peer
groups and examined whether this variation shapes peer influence. Specifically,
we incorporated a simple measure of the variance around the mean level of
peer delinquency and hypothesized that the variance attenuates the effect of
average peer delinquency through the processes specified in learning and
opportunity theories.
Consistent with our expectations, the results indicated that peer behavioral
heterogeneity weakens the effect of average peer levels for both smoking and
heavy alcohol consumption. At moderate to high levels of variance across the
network for both peer drunkenness and smoking, the average peer measure did
not significantly increase respondentslikelihood of smoking or drunkenness.
Conversely, at low levels of peer variance in smoking and drunkenness, the
effects of average peer levels were particularly pronounced. This finding sug-
gests that through its effects on the consistency of opportunities and/or the
ratio of definitions to which an adolescent is exposed in his or her peer net-
work, behavioral heterogeneity conditions peer influence.
Figure 3 Average marginal effects of average network drunkeness (95% Cls).
BEHAVIORAL HETEROGENEITY 19
Downloaded by [Arizona State University] at 23:15 26 March 2014
In sum, consistent with learning theories and implications from the oppor-
tunity perspective, these results suggest that variation in involvement in
delinquency among the peers in an adolescentsfriendship network influ-
ences peer smoking and drunkenness in ways that are not captured by sim-
ply averaging or summing the levels of delinquency among peers. Although
further research is needed to unpack variation in peer attitudes, reinforce-
ments, and opportunities, these findings suggest that future research on
peer influences employing network methods could benefit from considering
variation in behavior across a network to get a more accurate portrait of
the processes of peer influence.
This research, combined with that of others (Haynie, 2002; Rees & Pogar-
sky, 2011), has practical implications as well. These findings underscore the
idea that non-delinquent peers can counterbalance the influence of delin-
quent peers. Thus, adding pro-social or at least non-delinquent peers to a
youths network can counteract some of the influence of delinquent peers.
This further underscores work suggesting that our many policies that entail
grouping adjudicated delinquents or troubled youth together or segregating
them from non-delinquent youth may only serve to further contribute to
their delinquency and later crime (Dishion, McCord, & Poulin, 1999; Dishion,
Poulin, & Burraston, 2001; Gifford-Smith, Dodge, Dishion, & McCord, 2005).
These findings also imply that given the struggle or fears that many caregiv-
ers have about their children hanging around with troubled friends, one
avenue for mitigating potentially badinfluences is exposing their children
to a range of pro-social individuals, institutions, and networks. Of course,
as Haynie (2002, p. 126) warns, we should not naively engage in these
practices as, [b]uilding bridges connecting delinquent adolescents to non-
delinquent adolescents may be helpful to the delinquent adolescents but
potentially damaging to the non-delinquent youth.One potential solution
for particularly vigilant parents would be to expose children to a range of
pro-social individuals and networks prior to their becoming delinquent.
Future research that explores the developmental nature of peer influences
and the changing nature of networks across youth and adolescence are
needed to explore these ideas.
Although we believe the results of this study make an important contribu-
tion to the growing body of work exploring peer influences, it is not without
limitations. First, we examined only two substance-use (status) offenses.
Although we believe these offenses are particularly relevant for understanding
peer influences in adolescence given that they fall under what Warr (2002)
deems the groupyoffense category and, thus, are less likely to be engaged
in surreptitiously or solo, it is the case that the dynamics of network influences
may differ for different types of offenses (e.g. theft, vandalism) or more seri-
ous offenses (e.g. robbery, drug dealing, homicide). We hope future research
explores the effect of behavioral heterogeneity on these offense types. More
focused network data, such as among gangs, are needed to explore the dynam-
20 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
ics of serious offenses, as the prevalence and thus the behavioral heterogene-
ity across nationally representative student samples are quite low. Notably,
this limitation is partially vitiated by the reality that most existing studies that
employ the average measure of peer delinquency are invariably dominated by
involvement in less serious offenses (e.g. Haynie, 2001; Payne & Cornwell
2007; Schreck & Fischer, 2004). Nonetheless, this study represents a first look
at the relevance of behavioral heterogeneity qua variance, and it is important
that future work examine a variety of different outcomes. In addition, while
we examined offense-specific models, future research should also investigate
the effects of behavioral heterogeneity on individual and peer involvement in
general offending.
Another caveat is related to our measure of behavioral heterogeneity. The
formula for the variance is based upon the property of least squares, which
involves squaring the difference between each friends delinquency score and
the mean delinquency of the network. This squaring, of course, does not line-
arly increase differences, but rather makes them larger in a non-linear fashion
(an exponential increase). This raises the question of whether the observed
effects of the variance are being driven by larger differences magnified by the
squaring. To test this possibility, we conducted additional analyses where we
censored the variance measure at 2 standard deviations above the mean (e.g.
a variance of 7 for smoking). For the smoking model, we recoded the 347 cases
(4.7%) with values above the 2 standard deviation threshold to 7 and
re-estimated the model. The results from this model, as well as one predicting
drunkenness, are analogous to that presented in the paper.
15
Thus, while it is
hoped that future research might explore other means of capturing behavioral
heterogeneity, such as the median absolute deviation, it does not appear that
our results are an artifact of the squaring of the differences in the variance
calculation.
16
Another limitation worth noting concerns the operationalization of friend-
ship networks in the Add Health data. Respondents can nominate up to ten
individuals as friends, which may appear as only a minor limitation given
Haynies(
2002) finding that the average number of nominations is roughly 6
peers. However, respondents were only able to nominate five same-sex
peers as well as five opposite-sex peers. It is quite possible that this nomi-
nation constraint distorts the true composition of friendship networks in
ways that could influence estimates of peer influence and heterogeneity. Of
particular relevance for the present study, this nomination constraint might
underestimate the number of same-sex peers and overestimate the rele-
vance of opposite-sex peers, thereby misrepresenting actual patterns of
15. For the drunkenness model, we also censored at two standard deviations above the mean,
which was approximately equal to a variance of 4. Ninety-five percent of the scores were less than
or equal to 4 and thus were not recoded (censored). These results are available upon request.
16. Credit to an anonymous reviewer for suggesting the median absolute deviation.
BEHAVIORAL HETEROGENEITY 21
Downloaded by [Arizona State University] at 23:15 26 March 2014
friendship and association. We might expect that had youth not been con-
strained to five same-sex peers and five opposite-sex peers, we might have
seen a higher number of same-sex nominations, fewer opposite-sex nomina-
tions, and more total nominations. There is some evidence to suggest that
this restricting of same-sex peer nominations to five might be distorting the
composition of peer networks. For example, the youth in Poulon and Peder-
sens(
2007) study were allowed to list up to ten friends with no sex restric-
tions. In this study, the average number of same-sex peers was greater than
6 for both males and females, and the average number of opposite-sex
peers was less than 2 for males and less than 3 for females across all waves
(grades 7–10). Future research surveying networks without specifying the
sex of friendship nominations or without size constraints is needed to
explore the potential effects of this nomination constraint on peer influence
and heterogeneity. In addition, given evidence that same-sex and opposite-
sex peers have different degrees of influence, and this differs by the sex of
the respondent, future research might also explore the influence of the sex
of the respondent and his or her peers as it relates to influence and hetero-
geneity (e.g. Gaughan, 2006).
Notwithstanding these limitations, we believe that these findings represent a
useful contribution to the body of empirical literature on the effect of peers in
the etiology of delinquency. Although birds of a feather may flock together,
friendship networks are heterogeneous in delinquency (Elliott & Menard 1991;
Haynie, 2002). Consistent with past research, this study indicates that peers
matter in the explanation of adolescent substance use, but the amount of
variation in substance-use behavior among an individuals friends influences this
effect. This should be given consideration in an effort to best represent the
social and empirical realities of adolescent friendships. Clearly, more research
on peer behavioral heterogeneity is needed. In addition to replicating these
findings and extending them to other delinquent outcomes and general delin-
quency scales, future work might take a longitudinal perspective in regards to
variation in peer network behavior, recognizing that the peer networks are
dynamic (Berndt, 1982; Brown, Classen, & Eicher, 1986; Cairns & Cairns, 1994).
Such research would also provide an opportunity to study how peer behavioral
heterogeneity affects the creation and dissolution of ties within a network (see
Cohen, 1977). Network data-sets like Add Health and sophisticated modeling
techniques such as longitudinal network models (SIENA, Snijders, 2005) make
answering these and other time-dependent network questions possible. Given
the existence and potential relevance of behavioral heterogeneity within peer
networks, further exploring these and other questions is worthwhile.
Acknowledgements
An earlier version of this paper was presented at the 2012 Annual Meeting of
the American Society of Criminology. The authors thank Scott Decker, Alyssa
22 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
Chamberlain, Jacob Young, Greg Zimmerman, three anonymous reviewers, and
Cassia Spohn for insightful comments on earlier drafts of this paper. This
research uses data from Add Health, which is a program project directed by
Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, and
Kathleen Mullan Harris, and it is funded by grant P01-HD31921 from the Eunice
Kennedy Shriver National Institute of Child Health and Human Development,
with cooperative funding from 23 other federal agencies and foundations. Spe-
cial acknowledgment is due to Ronald R. Rindfuss and Barbara Entwisle for
assistance in the original design. Information on how to obtain the Add Health
data files is available on the Add Health website (http://www.cpc.unc.edu/
addhealth). No direct support was received from grant P01-HD31921 for this
analysis.
References
Ai, C., & Norton, E. C. (2003). Interaction terms in logit and probit models. Economics
Letters,80, 123-129.
Aiken, L. S., & West, S. G. (1991). Multiple regression: Testing and interpreting inter-
actions. Thousand Oaks, CA: Sage.
Akers, R. L. (1985). Deviant behavior: A social learning approach. Belmont, CA:
Wadsworth.
Akers, R. L., Krohn, M. D., Lanza-Kaduce, L., & Radosevich, M. (1979). Social learning
and deviant behavior: A specific test of a general theory. American Sociological
Review,44, 636-655.
Allison, P. (1990). Change scores as dependent variables in regression analysis. Sociolog-
ical Methodology,29, 93-114.
Bauman, K. E., Robert, F., Ennett, S. T., Hussong, A., & Foshee, V. A. (2007). Adding
valued data to social network measures: Does it add to associations with adolescent
substance use? Social Networks,29, 1-10.
Berndt, T. J. (1982). The features and effects of friendship in early adolescence. Child
Development,53, 1447-1460.
Brown, B. B., Classen, D. R., & Eicher, S. A. (1986). Perceptions of peer pressure, peer
conformity dispositions, and self-reported behavior among adolescents. Developmen-
tal Psychology,22, 521-530.
Buis, M. L. (2010). Stata tip 87: Interpretation of interactions in nonlinear models. Stata
Journal,10, 305-308.
Cairns, R. B., & Cairns, D. B. (1994). Lifelines and risks: Pathways of youth in our time.
New York, NY: Cambridge University Press.
Chantala, K., & Tabor, J. (1999). Strategies to perform a design-based analysis using
the Add Health data. Chapel Hill, NC: Carolina Population Center, University of North
Carolina.
Cohen, J. M. (1977). Sources of peer group homogeneity. Sociology of Education,50,
227-241.
Cohen, L. E., & Felson, M. (1979). Social change and crime rate trends: A routine
activity approach. American Sociological Review,44, 588-605.
Crosnoe, R., Muller, C., & Frank, K. (2004). Peer context and the consequences of
adolescent drinking. Social Problems,51, 288-304.
Dishion, T. J., Poulin, F., & Burraston, B. (2001). Peer group dynamics associated with
iatrogenic effects in group interventions with high-risk young adolescents. In C.
BEHAVIORAL HETEROGENEITY 23
Downloaded by [Arizona State University] at 23:15 26 March 2014
Erdley & D. W. Nangle (Eds.), Damon’s new directions in child development: The
role of friendship in psychological adjustment (pp. 70-92). San Francisco, CA: Jos-
sey-Bass.
Dishion, T. J., McCord, J., & Poulin, F. (1999). When interventions harm: Peer groups
and problem behavior. American Psychologist,54, 755-764.
Elliott, D. S., Huizinga, D., & Ageton, S. S. (1985). Explaining delinquency and drug
use. Beverly Hills, CA: Sage.
Elliott, D. S., Huizinga, D., & Menard, S. (1989). Multiple problem youth: Delinquency,
drugs and mental health problems. New York, NY: Springer.
Elliott, D. S., & Menard, S. (1996). Delinquent friends and delinquent behavior: tempo-
ral and developmental patterns. In D. J. Hawkins (Ed.), Delinquency and crime cur-
rent theories (pp. 28-67). New York: Cambridge University Press.
Gaughan, M. (2006). The gender structure of adolescent peer influence on drinking.
Journal of Health and Social Behavior,47, 47-61.
Gifford-Smith, M., Dodge, K. A., Dishion, T. J., & McCord, J. (2005). Peer influence in
children and adolescents: Crossing the bridge from development to intervention
science. Journal of Abnormal Child Psychology,33, 255-265.
Gottfredson, M., & Hirschi, T. (1990). A general theory of crime. Stanford, CA: Stanford
University Press.
Greene, W. (2010). Testing hypotheses about interaction terms in nonlinear models.
Economics Letters,107, 291-296.
Harris, K. M., Florey, F., Tabor, J., Bearman, P. S., Jones, J., & Udry, J. R. (2003). The
national longitudinal study of adolescent health: research design. Retrieved from
http://www.cpc.unc.edu/projects/addhealth/design
Haynie, D. L. (2001). Delinquent peers revisited: Does network structure matter? Ameri-
can Journal of Sociology,106, 1013-1057.
Haynie, D. L. (2002). Friendship networks and delinquency: The relative nature of peer
delinquency. Journal of Quantitative Criminology,18, 99-134.
Haynie, D. L., & Osgood, D. W. (2005). Reconsidering peers and delinquency: How do
peers matter? Social Forces,84, 1109-1130.
Haynie, D. L., & Payne, D. C. (2006). Race, friendship networks, and violent
delinquency. Criminology,44, 775-805.
Hirschi, T. (1969). Causes of delinquency. Berkeley: University of California Press.
Jackson, E. F., Tittle, C. R., & Burke, M. J. (1986). Offense-specific models of the
differential association process. Social Problems,33, 335-356.
Jussim, L., & Osgood, D. W. (1989). Influence and similarity among friends: An integra-
tive model applied to incarcerated adolescents. Social Psychology Quarterly,52,
98-112.
Karaca-Mandic, P., Norton, E. C., & Dowd, B. (2012). Interaction terms in nonlinear
models. Health Services Research,47, 255-274.
Kreager, D., Rulison, K., & Moody, J. (2011). Delinquency and the structure of
adolescent peer groups. Criminology,49, 95-127.
Krohn, M. D. (1974). An investigation of the effect of parental and peer associations on
marijuana use: an empirical test of differential association theory. In M. Reidel & T.
Thornberry (Eds.), Crime and delinquency: Dimensions of deviance (pp. 75-89). New
York: Praeger.
Krohn, M. D. (1986). The web of conformity: A network approach to the explanation of
delinquent behavior. Social Problems,33, 81-93.
Lazarsfeld, P. F., & Merton, R. K. (1954). Friendship as a social process: A substantive
and methodological analysis. In M. Berger, T. Abel & C. H. Page (Eds.), Freedom and
Control in Modern Society (pp. 18-66). New York, NY: Van Nostrand.
Long, J. S. (1997). Regression models for categorical and limited dependent variables.
Thousand Oaks, CA: Sage.
24 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
Long, J. S., & Freese, J. (2003). Regression models for categorical dependent variables
using Stata, rev ed. College Station, TX: Stata.
Matsueda, R. L. (1982). Testing control theory and differential association: A causal
modeling approach. American Sociological Review,47, 489-504.
Matsueda, R. L., & Heimer, K. (1987). Race, family structure, and delinquency: A test
of differential association and social control theories. American Sociological Review,
52, 826-840.
McGloin, J. M. (2009). Delinquency balance: Revisiting peer influence. Criminology,47,
439-477.
McGloin, J. M., & Shermer, L. O. (2009). Self-control and deviant peer network
structure. Journal of Research in Crime and Delinquency,46, 35-72.
McPherson, M., Smith-Lovin, L., & Cook, J. M. (2001). Birds of a feather: Homophily in
social networks. Annual Review of Sociology,27, 415-444.
Nieuwbeerta, P., Blokland, A., Piquero, A. J., & Sweeten, G. (2011). A Life-course anal-
ysis of offense specialization across age: Introducing a new method for studying indi-
vidual specialization over the life course. Crime and Delinquency,57, 3-28.
Norton, E. C., Wang, H., & Ai, C. (2004). Computing interaction effects and standard
errors in logit and probit models. Stata Journal,4, 154-167.
Orcutt, J. D. (1983). Analyzing deviance. Homewood, IL: Dorsey Press.
Osgood, D. W., Wilson, J. K., OMalley, P. M., Bachman, J. G., & Johnston, L. D.
(1996). Routine activities and individual deviant behavior. American Sociological
Review,61, 635-655.
Payne, D. C., & Cornwell, B. (2007). Reconsidering peer influences on delinquency: Do
less proximate contacts matter? Journal of Quantitative Criminology,23, 127-149.
Poulon, F., & Pedersen, S. (2007). Developmental changes in gender composition of
friendship networks in adolescent girls and boys. Developmental Psychology,43,
1484-1496.
Rees, C., & Pogarsky, G. (2011). One bad apple may not spoil the whole bunch: Best
friends and adolescent delinquency. Journal of Quantitative Criminology,27, 197-
223.
Sampson, R. J. (1999). Techniques of research neutralization. Theoretical Criminology,
3, 438–451.
Schreck, C., & Fischer, B. (2004). Specifying the influence of family and peers on vio-
lent victimization: Extending routine activities and lifestyles theories. Journal of
Interpersonal Violence,19, 1021-1041.
Short, J. F., & Nye, F. I. (1957). Reported behavior as a criterion of deviant behavior.
Social Problems,5, 207-213.
Snijders, T. A. B. (2005). Models for longitudinal network data. In P. J. Carrington, J.
Scott & S. Wasserman (Eds.), Models and methods in social network analysis
(pp. 215-247). New York: Cambridge University Press.
Sutherland, E. H. (1947). Criminology. JB: Lippincott.
Thornberry, T. P., Lizotte, A. J., Krohn, M. D., Farnworth, M., & Jang, S. J. (1994).
Delinquent peers, beliefs, and delinquent behavior: A longitudinal test of
interactional theory. Criminology,32, 47-83.
Udry, J. R. (2003). The National Longitudinal Study of Adolescent Health (Add Health),
Waves 1 and II, 1994-1996 [Machine-readable data file and documentation]. Chapel
Hill: Carolina Population Center, University of North Carolina.
Voss, H. (1969). Differential association and containment theory: Theoretical
convergence.Social Forces,47, 381-391.
Warr, M. (1996). Organization and instigation in delinquent groups. Criminology,34, 11-37.
Warr, M. (2002). Companions in crime: The social aspects of criminal conduct. New
York: Cambridge University Press.
BEHAVIORAL HETEROGENEITY 25
Downloaded by [Arizona State University] at 23:15 26 March 2014
Warr, M., & Stafford, M. (1991). The influence of delinquent peers: What they think or
what they do? Criminology,29, 851-866.
Weerman, F. M. (2011). Delinquent peers in context: A longitudinal network analysis of
selection and influence effects. Criminology,49, 253-286.
Weerman, F. M., & Smeenk, W. H. (2005). Peer similarity in delinquency for different
types of friends: A comparison using two measurement methods. Criminology,43,
499-524.
Wikstro
¨m, P. H., Ceccato, V., Hardie, B., & Trieber, K. (2010). Activity fields and the
dynamics of crime. Journal of Quantitative Criminology,26, 55-87.
Wikstro
¨m, P. H., Oberwittler, D., Treiber, K., & Hardie, B. (2012). Breaking rules: The
social and situational dynamics of young people’s urban crime. Oxford: Oxford
University Press.
Young, J. T. N., Barnes, J. C., Meldrum, R. C., & Weerman, F. M. (2011). Assessing and
explaining misperceptions of peer delinquency. Criminology,49, 599-630.
Zimmerman, G. M., & Messner, S. F. (2011). Neighborhood context and nonlinear peer
effects on adolescent violent crime. Criminology,49, 873-903.
Zimmerman, G. M., & Vasquez, B. E. (2011). Decomposing the peer effect on
adolescent substance use: Mediation, nonlinearity, and differential nonlinearity.
Criminology,49, 1235-1273.
26 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
Appendix A
Appendix B
Descriptives for drunk model variables NMSD Range
Dependent variable
Respondent drunk wave 1 7,379 0.58 1.14 0–6
Independent variables
Lagged outcome (drunk) 7,379 0.61 1.19 0–6
Network average drunk 7,379 0.71 0.85 0–6
Network variance around average drunk 7,379 0.94 1.37 0–9
Sex (1 = Male) (%) 7,379 0.44 0.5 0–1
Respondent race: Black (%) 7,379 0.19 0.39 0–1
Respondent race: other (%) 7,379 0.08 0.27 0–1
Age 7,379 0 1 1.86 to 1.89
Age squared 7,379 1 1.07 0–3.56
Two parents in home (%) 7,379 0.78 0.41 0–1
Public assistance (%) 7,379 0.06 0.24 0–1
Grade point average 7,379 2.87 0.78 1–4
School attachment 7,379 3.64 0.94 1–5
Parental attachment 7,379 4.73 0.59 1–5
Friend attachment 7,379 4.31 0.74 1–5
Friend involvement 7,379 71.4 5.73 0–40
Network density 7,379 0.4 0.2 0.10–1
Respondent centrality 7,379 0.99 0.59 0.04–4.29
Network size 7,379 5.18 2.52 1–10
Respondent popularity 7,379 4.99 3.79 0–30
Time between surveys (days) 7,379 228.6 47.2 60–405
Descriptives for smoking model variables NMSD Range
Dependent variable
Respondent smoking wave 1 7,394 3.78 8.95 0–30
Independent Variables
Lagged outcome (smoking) 7,394 1.11 1.95 0–6
Network average smoking 7,394 1.16 1.37 0–6
Network variance around average smoking 7,394 2.08 2.51 0–9
Sex (1 = Male) (%) 7,394 0.44 0.5 0–1
Respondent race: Black (%) 7,394 0.19 0.39 0–1
Respondent race: other (%) 7,394 0.08 0.27 0–1
Age 7,394 0 1 1.85 to 1.89
Age squared 7,394 1 1.07 0–3.65
Two parents in home (%) 7,394 0.78 0.41 0–1
Public assistance (%) 7,394 0.06 0.24 0–1
(Continued)
BEHAVIORAL HETEROGENEITY 27
Downloaded by [Arizona State University] at 23:15 26 March 2014
Appendix C
Appendix B (Continued)
Descriptives for smoking model variables NMSD Range
Grade point average 7,394 2.87 0.78 1–4
School attachment 7,394 3.64 0.94 1–5
Parental attachment 7,394 4.73 0.59 1–5
Friend attachment 7,394 4.31 0.74 1–5
Friend involvement 7,394 7.14 5.74 0–40
Network density 7,394 0.4 0.2 0.10–1
Respondent centrality 7,394 0.99 0.59 0.04–4.29
Network size 7,394 5.18 2.52 1–10
Respondent popularity 7,394 4.98 3.79 0–30
Time between surveys (days) 7,394 228.55 47.25 60–405
Value of
network
variance
Change in respondent
smoking per unit increase
in network average
Change in respondent
drunkenness per unit increase
in network average
@y=@x@y=@x
0 0.959
**
0.187
**
0.5 0.821
**
0.160
**
1 0.679
**
0.132
**
1.5 0.530
*
0.103
**
2 0.373 0.073
*
2.5 0.207 0.042
3 0.031 0.009
3.5 0.158 0.025
40.361 0.062
4.5 0.581 0.101
50.820 0.143
5.5 1.081 0.188
61.366 0.237
6.5 1.679 0.290
72.023 0.346
7.5 2.403 0.408
82.823 0.475
8.5 3.289 0.548
93.805 0.627
N 7,394 7,379
*
p< 0.05;
**
p< 0.01 (two tailed tests)
28 BURT AND REES
Downloaded by [Arizona State University] at 23:15 26 March 2014
... Empirical studies examining the relationship between peer delinquency and individual delinquency and substance are based primarily on adolescent samples (Agnew, 1991;Alexander et al., 2001;Ary et al., 1999;Bahr et al., 2005;Barnes et al., 2007;Boman et al., 2014;Brendgen et al., 2000;Brezina & Piquero, 2007;Brook et al., 1998;Burt & Rees, 2015;Chapple et al., 2014;Church et al., 2012;De Kemp et al., 2006;De Vries et al., 2003;Dishion & Owen, 2002;Dishion et al., 1996;Eklund et al., 2010;Ennett et al., 2008;Gallupe & Bouchard, 2013, 2015Gardner et al., 2009;Garnier & Stein, 2002;Gaughan, 2006;Gault-Sherman, 2013;Haynie, 2002;Haynie & Payne, 2006;Haynie et al., , 2014Henry et al., 2001;Hochstetler et al., 2002;Hwang & Akers, 2006;Jang, 1999;Jewell et al., 2015;Keijsers et al., 2012;Knecht et al., 2010;Kreager & Haynie, 2011;Lakon et al., 2015;Lonardo et al., 2009;Maggs & Hurrelmann, 1998;Matsueda, 1982;Matsueda & Anderson, 1998;Matthews & Agnew, 2008;Maxwell, 2002;McGloin, 2009;Miller et al., 2009;Neppl et al., 2015;Osgood et al., 2014;Patterson et al., 2000;Payne & Cornwell, 2007;Pierce et al., 2015;Piquero, Gover, MacDonald, & Piquero, 2005;Posick, 2013;Ragan, 2014;Rebellon, 2006;Rees & Pogarsky, 2011;Reitz et al., 2006;Sanchagrin et al., 2014;Schaefer et al., 2012;Siennick et al., 2015;Trucco et al., 2011;Vitaro et al., 2005;Vitulano et al., 2010;Warr, 1993b;Warr & Stafford, 1991;Weerman, 2011;Weerman et al., 2015a;Wong, 1998;Worthen, 2012;Wright et al., 2008;Zhang & Messner, 2000). There is less research, however, assessing the association between peer delinquency and individual delinquency using young adult samples (e.g., Boman & Gibson, 2011;Fergusson et al., 2002;Gardner & Steinberg, 2005;Overbeek et al., 2011). ...
... From a methodological standpoint, these associations have been established using cross sectional data (Ary, Duncan, Duncan, & Hops, 1999;Bahr et al., 2005;Barnes et al., 2007;Boman, Miller, Stogner, Agnich, & Krohn, 2014;Brendgen et al., 2000;Brezina & Piquero, 2007;Brook et al., 1998;Burt & Rees, 2015;Chapple et al., 2014;Church et al., 2012;De Kemp et al., 2006;De Vries et al., 2003;Dishion & Owen, 2002;Dishion, Spracklen, Andrews, & Patterson, 1996;Eklund et al., 2010;Ennett et al., 2008;Fergusson et al., 2002;Gallupe & Bouchard, 2013, 2015Gardner et al., 2009;Gaughan, 2006;Gault-Sherman, 2013;Haynie & Payne, 2006;Haynie et al., , 2014Hochstetler et al., 2002;Hwang & Akers, 2006;Jang, 1999;Jewell, Brown, & Perry, 2015; Keijsers et al., 2012;Knecht et al., 2010;Kreager & Haynie, 2011;Lakon et al., 2015;Lonardo et al., 2009;Maggs & Hurrelmann, 1998;Matsueda & Anderson, 1998;Matthews & Agnew, 2008;McGloin, 2009;Osgood et al., 2014;Payne & Cornwell, 2007;Pierce et al., 2015;Posick, 2013;Rebellon, 2006;Rees & Pogarsky, 2011;Reitz et al., 2006;Sanchagrin et al., 2014;Trucco et al., 2011;Vitaro et al., 2005;Vitulano et al., 2010;Weerman, 2011;Weerman et al., 2015a;Wong, 1998;Worthen, 2012;Wright et al., 2008;Zhang & Messner, 2000) and longitudinal data (Alexander et al., 2001;Ary et al., 1999;Bahr et al., 2005;Brendgen et al., 2000;Brezina & Piquero, 2007;Brook et al., 1998;Burt & Rees, 2015;Chapple et al., 2014;Church et al., 2012;De Kemp et al., 2006;De Vries et al., 2003;Dishion et al., 1996;Dishion & Owen, 2002;Eklund et al., 2010;Ennett et al., 2008;Fergusson et al., 2002;Gallupe & Bouchard, 2013, 2015Gardner et al., 2009;Garnier & Stein, 2002;Gaughan, 2006;Gault-Sherman, 2013;Haynie, 2002;Haynie & Payne, 2006;Henry et al., 2001;Hochstetler et al., 2002;Hwang & Akers, 2006;Jang, 1999;Jewell et al., 2015; Keijsers et al., 2012;Knecht et al., 2010;Kreager & Haynie, 2011;Lakon et al., 2015;Lonardo et al., 2009;Maggs & Hurrelmann, 1998;Matsueda & Anderson, 1998;Matthews & Agnew, 2008;Maxwell, 2002;McGloin, 2009;Miller et al., 2009;Neppl, Dhalewadikar, & Lohman, 2015;Osgood et al., 2014;Patterson, Dishion, & Yoerger, 2000;Payne & Cornwell, 2007;Pierce et al., 2015;Ragan, 2014;Rebellon, 2006;Rees & Pogarsky, 2011;Reitz et al., 2006;Sanchagrin et al., 2014;Schaefer et al., 2012;Siennick, Widdowson, Woessner, & Feinberg, 2015;Trucco et al., 2011;Vitaro et al., 2005;Vitulano et al., 2010;Weerman, 2011;Wong, 1998;Worthen, 2012;Wright et al., 2008;Zhang & Messner, 2000). ...
... From a methodological standpoint, these associations have been established using cross sectional data (Ary, Duncan, Duncan, & Hops, 1999;Bahr et al., 2005;Barnes et al., 2007;Boman, Miller, Stogner, Agnich, & Krohn, 2014;Brendgen et al., 2000;Brezina & Piquero, 2007;Brook et al., 1998;Burt & Rees, 2015;Chapple et al., 2014;Church et al., 2012;De Kemp et al., 2006;De Vries et al., 2003;Dishion & Owen, 2002;Dishion, Spracklen, Andrews, & Patterson, 1996;Eklund et al., 2010;Ennett et al., 2008;Fergusson et al., 2002;Gallupe & Bouchard, 2013, 2015Gardner et al., 2009;Gaughan, 2006;Gault-Sherman, 2013;Haynie & Payne, 2006;Haynie et al., , 2014Hochstetler et al., 2002;Hwang & Akers, 2006;Jang, 1999;Jewell, Brown, & Perry, 2015; Keijsers et al., 2012;Knecht et al., 2010;Kreager & Haynie, 2011;Lakon et al., 2015;Lonardo et al., 2009;Maggs & Hurrelmann, 1998;Matsueda & Anderson, 1998;Matthews & Agnew, 2008;McGloin, 2009;Osgood et al., 2014;Payne & Cornwell, 2007;Pierce et al., 2015;Posick, 2013;Rebellon, 2006;Rees & Pogarsky, 2011;Reitz et al., 2006;Sanchagrin et al., 2014;Trucco et al., 2011;Vitaro et al., 2005;Vitulano et al., 2010;Weerman, 2011;Weerman et al., 2015a;Wong, 1998;Worthen, 2012;Wright et al., 2008;Zhang & Messner, 2000) and longitudinal data (Alexander et al., 2001;Ary et al., 1999;Bahr et al., 2005;Brendgen et al., 2000;Brezina & Piquero, 2007;Brook et al., 1998;Burt & Rees, 2015;Chapple et al., 2014;Church et al., 2012;De Kemp et al., 2006;De Vries et al., 2003;Dishion et al., 1996;Dishion & Owen, 2002;Eklund et al., 2010;Ennett et al., 2008;Fergusson et al., 2002;Gallupe & Bouchard, 2013, 2015Gardner et al., 2009;Garnier & Stein, 2002;Gaughan, 2006;Gault-Sherman, 2013;Haynie, 2002;Haynie & Payne, 2006;Henry et al., 2001;Hochstetler et al., 2002;Hwang & Akers, 2006;Jang, 1999;Jewell et al., 2015; Keijsers et al., 2012;Knecht et al., 2010;Kreager & Haynie, 2011;Lakon et al., 2015;Lonardo et al., 2009;Maggs & Hurrelmann, 1998;Matsueda & Anderson, 1998;Matthews & Agnew, 2008;Maxwell, 2002;McGloin, 2009;Miller et al., 2009;Neppl, Dhalewadikar, & Lohman, 2015;Osgood et al., 2014;Patterson, Dishion, & Yoerger, 2000;Payne & Cornwell, 2007;Pierce et al., 2015;Ragan, 2014;Rebellon, 2006;Rees & Pogarsky, 2011;Reitz et al., 2006;Sanchagrin et al., 2014;Schaefer et al., 2012;Siennick, Widdowson, Woessner, & Feinberg, 2015;Trucco et al., 2011;Vitaro et al., 2005;Vitulano et al., 2010;Weerman, 2011;Wong, 1998;Worthen, 2012;Wright et al., 2008;Zhang & Messner, 2000). ...
Article
Purpose. Peer delinquency and unstructured socializing have been identified as important correlates of delinquency and substance use. This state-of-the-art review explicates research into these associations to identify important trends in the literature and directions for future research. Methods. A search of the criminological literature and literatures of allied disciplines was executed to identify studies that have examined the potential influence of peer delinquency and unstructured socializing on delinquency and substance use. Results. The review highlights the theoretical underpinnings of the two constructs, issues of measurement quality, the generality of effects on delinquency and substance use, advances in the respective literatures, and important remaining gaps for future research to fill. Conclusions. While considerable attention has been given to studying the potential influence of peer delinquency and unstructured socializing on delinquency and substance use, there remain a number of ways in which these literatures can be advanced to provide a more complete understanding of the relevance of these constructs for the etiology of delinquency and substance use. Note: All authors contributed equally to the preparation of this publication. Author order is listed alphabetically.
... That is, illicit hard and soft drugs were combined with alcohol and tobacco and treated as one single outcome. This is particularly concerning given that primary research has revealed varying degrees of support for concepts of SLT for different types of substance use (Burt & Rees, 2015;Gallupe & Bouchard, 2013;Gray, Durkin, Call, & Evans, 2015;HeavyRunner-Rioux & Hollist, 2010;Kim, Akers, & Yun, 2013;Kim, Kwak, & Yun, 2010;Miller, Jennings, Alvarez-Rivera, & Miller, 2008;Miller et al., 2011;Monroe, 2004;Musher-Eizenman, Holub, & Arnett, 2003;Onica-Chipea, Saveanu, & Buhas, 2014;Peralta & Steele, 2010;Schaefer, Vito, Marcum, Higgins, & Ricketts, 2015a, 2015bWhaley, Smith, & Hayes-Smith, 2011;Yun & Kim, 2015). ...
... More recently, SLT concepts have been used to explain a variety of soft-drug use including alcohol (Gallupe & Bouchard, 2013;Kim et al., 2013, Musher-Eizenman et al., 2003, tobacco (Burt & Rees, 2015;Monroe, 2004;Onica-Chipea et al., 2014), marijuana (Gray et al., 2015;HeavyRunner-Rioux & Hollist, 2010;Miller et al., 2008) and saliva divinorum (Kim et al., 2010;Miller et al., 2011). Although less frequent, SLT concepts have also been applied to hard-drug use such as amphetamines (Whaley et al., 2011;Yun & Kim, 2015), cocaine (Schaefer et al., 2015b), ecstasy (Norman & Ford, 2015;Whaley et al., 2011), nonprescription pain pills (Higgins, Mahoney, & Ricketts, 2009;Peralta, Stewart, Steele, & Wagner, 2016;Peralta & Steele, 2010;Schroeder & Ford, 2012), and heroin (Schaefer et al., 2015a). ...
Article
Background: Despite ample empirical research testing components of Akers’ Social Learning Theory (SLT) on substance use, no research to date has attempted to synthesize the empirical evidence. Objectives: The purpose of this article is to synthesize prior research that has examined the utility of SLT for predicting specific types of substance use, both legal and illegal. Methods: Using a systematic review and meta-analysis, the current study estimated the effect size results from 83 primary studies published between 1974 and 2018 that had empirically tested concepts of Akers’ SLT regarding substance use. In addition, moderator analyses examined variations in effect sizes across measurement constructs and among specific types of substance use. Results: Results indicated medium-sized weighted mean effect size estimates for SLT in relation to substance use. Regarding conceptualization of SLT, measures of Differential Association produced the strongest effect size estimates. Moderator analyses also revealed that mean effect size estimates were largest for soft drugs, for studies conducted in the context of the United States, and for adult samples. Conclusions: The authors conclude that SLT constructs may be better suited for explaining soft drug use than hard drug use. Given the relatively sparse primary research that has controlled for temporal ordering, collected data from multiple differential associates, or considered opportunity effects, caution in the interpretation of synthesis results is warranted.
... Studies of adolescent alcohol use have tended to capture the social norms of the peer drinking behavior as the number of friends using, the proportion of friends using ( Windle et al., 2008), or the average level of alcohol use amongst friends ( Ali & Dwyer, 2010;Crosnoe et al., 2004;Rees & Pogarsky, 2011). These measures implicitly ignore the behavioral heterogeneity found within friendship networks and assume equal influence and norm consensus ( Burt & Rees, 2014). ). ...
... We opted not to impute data on the dependent variables or the peer alcohol use measure, but we included these variables as well as the Add Health survey design variables in our imputation models as auxiliary variables ( see Enders, 2010;Reiter, Raghunathan, & Kinney, 2006;Von Hippel, 2007). 3 See Burt and Rees (2014) for similar arguments. ...
Article
Full-text available
Adolescent peer groups with pro-drinking group norms are a well-established source of influence for alcohol initiation and use. However, classic experimental studies of social influence, namely 'minority influence', clearly indicate social situations in which an individual can resist conforming to the group norm. Using the National Longitudinal Study of Adolescent Health ("Add Health"), a nationally representative sample of adolescents, we find evidence that being a non-drinking adolescent does not unilaterally put youth at risk for drinking onset when faced with a friendship network where the majority of friends drink. Our results also show that a non-drinking adolescent with a majority of drinking friends is significantly less likely to initiate alcohol abuse if he or she has a minority of non-drinking friend(s). Furthermore, a drinking adolescent with a majority of friends who drink has a decreased probability of continuing to drink and has overall lower levels of consumption if he or she has a minority of friends who do not drink. Our findings recognize that adolescent in-group friendships are a mix of behavioral profiles and can perhaps help adolescents continue or begin to abstain alcohol use even when in a friendship group supportive of alcohol use.
... Considering the role of antisocial peers, several studies have confirmed a strong influence on delinquent behaviour (see 2, 5, 8, 9, 15, 19, 20, 21, 22, 24, 26, 28, 29, 42, 43, 44, 51, 52, 53, 55, 61, 62, 64, 67, 69, 71, 75, 77, 78, 80, 82, and 84). Antisocial peers increased the probability of substance abuse, such as alcohol consume (Burt & Rees, 2015;Miller, 2013), violent behaviour (Hoeben & Weerman, 2016;Hughes & Short, 2014;Maimon & Browning, 2010;Tanner et al., 2015), vandalism, theft, or general delinquency (Hoeben & Weerman, 2016). Hoeben and Weerman (2016) suggest that antisocial peers act as situational motivators. ...
Article
This paper aims to summarise and integrate the accumulated knowledge on recent literature analysing the situational factors of juvenile delinquency and draw future avenues for research. A review of recent empirical papers published on two relevant databases allowed us to gather a sample of 88 papers published from 2010 to 2017 that analyse the role of situational factors of delinquent behaviour using quantitative methods and applying one of the situational theories. The results highlight the robustness of some situational and environmental variables for a deeper understanding of juvenile antisocial behaviour by putting it into context. There is a considerable amount of evidence to corroborate the impact of unstructured leisure activities on antisocial behaviour, or the role of home location in establishing a geographic area of action. However, there is ambiguous evidence on other aspects: i.e. guardianship needs to be understood and measure in a more complex way, and the role of physical design of the places where juvenile delinquency happens deserves further analysis. Future research lines from this perspective are needed and will offer relevant improvements for better understanding juvenile delinquency and designing more effective preventive measures.
... Esbensen, Peterson, Taylor, & Osgood, 2012) may be enough to dissuade a portion of students from joining. Prior studies have shown the structure of, and position in, school peer networks can influence involvement in delinquency (Burt & Rees, 2015;Gallupe, 2017;Haynie, 2001;McGloin & Shermer, 2009;McGloin, Sullivan, & Thomas, 2014). Related to the fact that schools provide a venue for students to connect with prosocial influences, an important focus for future research would be to examine social network influences on gang desistance. ...
Article
Full-text available
Schools are venues in which gang and non-gang involved youth converge. It is therefore a likely venue for gang recruitment. The extent to which this occurs depends upon the ability of gang members to connect with non-gang members. In this study, we compare the social network positions of high social status gang members who are well integrated into school networks with low status members who are not. Using network data from the Add Health study (n = 1,822), we find that not only are high status gang members strongly embedded within school networks, but that this status is driven by their ability to connect with non-gang members rather than other gang members (indicated by the high number of friendship nominations they receive from non-gang members). These gang members are potentially in optimal positions to influence others to join gangs. The implications of these results for school-based gang prevention programs are discussed.
... In detail, peer violent delinquency was an important factor in distinguishing chronic group membership from other groups, while peer non-violent delinquency was an important factor in distinguishing membership in the sharp-decreasing group from non-involved group. These results are suggesting the possibility of causal mechanism of delinquent peer influence is mostly rooted on the hypotheses of similar levels of delinquency resemblance (Burt & Rees, 2015;Young, Rebellon, Barnes, & Weerman, 2014). The results of the current study are also supportive of the taxonomic separation between adolescent-limited and chronic offenders in regard to seriousness of deviancy (Moffitt, 1993(Moffitt, , 2003. ...
Article
Cyberbullying has been subject to a debate about whether it is a subtype of traditional bullying or a distinct deviant behavior from traditional bullying. Applying a longitudinal South Korean youth sample and latent group-based trajectory modeling, the current study examines: 1) an overlap of developmental trajectories between cyberbullying and traditional bullying, and 2) effects of predictors on developmental trajectory groups for both cyberbullying and traditional bullying. It is concluded that cyberbullying is close to a variation of bullying rather than a distinct deviant behavior and reported an overlap of developmental trajectories between cyberbullying and traditional bullying and strong associations between both forms of bullying and peer-related predictors. Implications of the research findings and suggestions for future research are also discussed.
... In a subsequent study using the same data, Zimmerman and Vásquez (2011) reported a similar nonlinear relationship between substance using peers and individual substance use. Research by Burt and Rees (2014) comports with these findings, suggesting the association between individual substance use and exposure to peer substance use was nonlinear. Moderate to high levels of peer behavior did not increase respondents' likelihood of engaging in these behaviors; however, those exposed to low levels of peer smoking and alcohol use were strongly influenced by their peers' behaviors. ...
Article
Research on youth delinquency has been essential for gaining a deeper understanding of the etiology of delinquent behavior. Studies considering the environmental perspective have increased during the last decade, but relatively little attention has been paid to temporal patterns and weather conditions. The present study explores the seasonality of youth delinquency as well as the association between violent and non-violent youth offenses and temperature, rainfall, level of darkness, type of day, type of place, and companionship, using data gathered by the police along with data obtained from official weather agencies. To this end, we conducted ANOVA and contingency table analyses. Seasonality was found for non-violent crimes. Companionship, semi-public, and public places were all associated with a higher likelihood of non-violent crime, while darkness and public holidays raise the odds of violent crime to happen. No direct association was found between temperature and type of crime.
Article
We explore the implications of multiple reference groups for symbolic interaction explanations of delinquency. Specifically, we test the idea that dissimilarity in reflected appraisals across reference groups weakens the effect of the appraisals of a single group on behavior. We also assess whether the relative influence of different reflected appraisals is moderated by differences in the amount of time spent in and attachment to the roles taken among specific reference groups. Using data from the National Youth Survey, our findings mostly support symbolic interactionist predictions: Dissimilarity in reference groups' reflected appraisals weakens the influence of a single group's appraisals on behavior. We also found some evidence that, relatively, the more time an individual spends and the more attached they are to a role in a particular reference group, the more strongly correlated the respective appraisals are with delinquent behavior. Contrary to our predictions, we found that relatively strong attachments to one's role with their parents weakens the influence of parental reflected appraisals as a rule violator.
Article
Background: Despite ample empirical research testing components of Akers’ Social Learning Theory (SLT) on substance use, no research to date has attempted to synthesize the empirical evidence. Objectives: The purpose of this article is to synthesize prior research that has examined the utility of SLT for predicting specific types of substance use, both legal and illegal. Methods: Using a systematic review and meta-analysis, the current study estimated the effect size results from 83 primary studies published between 1974 and 2018 that had empirically tested concepts of Akers’ SLT regarding substance use. In addition, moderator analyses examined variations in effect sizes across measurement constructs and among specific types of substance use. Results: Results indicated medium-sized weighted mean effect size estimates for SLT in relation to substance use. Regarding the conceptualization of SLT, measures of Differential Association produced the strongest effect size estimates. Moderator analyses also revealed that mean effect size estimates were largest for soft drugs, for studies conducted in the context of the United States, and for adult samples. Conclusions: The authors conclude that SLT constructs may be better suited for explaining soft drug use than hard drug use. Given the relatively sparse primary research that has controlled for temporal ordering, collected data from multiple differential associates, or considered opportunity effects, caution in the interpretation of synthesis results is warranted.
Book
Why do certain people commit acts of crime? Why does crime happen in certain places? Presenting an ambitious new study designed to test a pioneering new theory of the causes of crime, Breaking Rules: The Social and Situational Dynamics of Young People's Urban Crime demonstrates that these questions can only go so far in explaining why crime happens - and, therefore, in preventing it. Based on the work of the Peterborough Adolescent and Young Adult Development Study (PADS+), Breaking Rules presents an analysis of the urban structure of Peterborough and its relation to young people's social life. Contemporary sciences state that behaviour is the outcome of an interaction between people and the environments to which they are exposed, and it is precisely that interaction and its relation to young people's crime involvement that PADS+ explores. Driven by a ground-breaking theory of crime, Situational Action Theory, which aims to explain why people break rules, it implements innovative methods of measuring social environments and people's exposure to them, involving a cohort of 700 young people growing up in the UK city of Peterborough. It focuses on the important adolescent time window, ages 12 to 17, during which young people's crime involvement is at its peak, using unique space-time budget data to explore young people's time use, movement patterns, and the spatio-temporal characteristics of their crime involvement. Presenting the first study of this kind, both in breadth and detail, with significant implications for policy and prevention, Breaking Rules should not only be of great interest to academic readers, but also to policy-makers and practitioners, interested in issues of urban environments, crime within urban environments, and the role of social environments in crime causation.
Article
The abstract for this document is available on CSA Illumina.To view the Abstract, click the Abstract button above the document title.
Article
It is argued that the work of Reckless and Dinitz concerning the socialization and self-concepts of “good” boys and “potential” delinquents offers the basis for a test and extension of the theory of differential association. This position is based on recognition that these investigators, like Sutherland, emphasized socialization processes; to explain the non-delinquent Reckless and Dinitz rely on the same mechanisms Sutherland employed to explain the delinquent. In the analysis of data obtained in Honolulu it was found that the joint effects of the measures of differential association and socialization proposed, respectively, by Short and Reckless and Dinitz, account for delinquent behavior more fully than does the separate effect of either measure. While they overlap considerably, the use of both measures accounts for delinquent behavior, defined according to official and unofficial criteria, more fully than does either measure treated singly.
Article
Our current understanding of the role of the social environment in crime causation is at best rudimentary. Guided by the theoretical framework of Situational Action Theory, and using data from the ESRC financed Peterborough Adolescent and Young Adult Development Study (PADS+), this paper aims to propose how we can better theorise and study the role of the social environment, particularly the person and place interaction, in crime causation. We will introduce, and illustrate the usefulness of, a space–time budget methodology as a means of capturing people’s exposure to settings and describing their activity fields. We will suggest and demonstrate that, combined with a small area community survey and psychometric measures of individual characteristics, a space–time budget is a powerful tool for advancing our knowledge about the role of the social environment, and its interaction with people’s crime propensity, in crime causation. Our unique data allows us to study the convergence in time and space of crime propensity, criminogenic exposure and crime events. As far as we are aware, such an analysis has never before been carried out. The findings show that there are (a) clear associations between young people’s activity fields and their exposure to criminogenic settings, (b) clear associations between their exposure to criminogenic settings and their crime involvement, and, crucially, (c) that the influence of criminogenic exposure depends on a person’s crime propensity. Having a crime-averse morality and strong ability to exercise self-control appears to make young people practically situationally immune to the influences of criminogenic settings, while having a crime-prone morality and poor ability to exercise self-control appears to make young people situationally vulnerable to the influences of criminogenic settings.
Article
Identifying interpersonal influence involves distinguishing it from two other sources of similarity among friends: selective attraction and effects of the shared environment. We specify several conceptual and methodological requirements for accomplishing this goal and we develop a causal model that meets those requirements. Our approach is illustrated through a study of delinquent values among a sample of 72 pairs of incarcerated adolescents. Results showed little evidence of actual interpersonal influence, primarily because respondents did not perceive their friends' attitudes accurately. There was strong evidence of subjective influence, however: respondents were influenced by their perceptions of their friends' attitudes. In addition, respndents greatly overestimated their similarity to their friends. This study illustrates the value of our conceptual framework for understanding sources of similarity among friends. The findings have important implications for understanding social influence among incarcerated adolescents.
Article
A number of strong theoretical statements have been based on analyses of delinquency data from the Richmond Youth Project. Hirschi (1969) and Jensen (1972), in particular, found that Hirschi's control theory was empirically supported over Sutherland's theory of differential association. This paper reanalyzes these data and reassesses this negative evidence pertaining to differential association theory. It is shown that the ratio of learned behavior patterns favorable and unfavorable to violation of legal codes, the critical variable in Sutherland's theory, can be operationalized by explicitly modeling its measurement error structure. In turn, this allows the testing of specific hypotheses derived from the theory. The analysis based on this strategy finds differential association theory supported over control theory. Specifically, the unobservable construct representing the ratio of learned behavior patterns successfully mediates the effects on delinquency of the model's other variables.