Accepted Manuscript

Mental health and social networks in early adolescence: A dynamic study of

objectively-measured social interaction behaviors

Mark C. Pachucki, PhD Emily J. Ozer, PhD Alain Barrat, PhD Ciro Cattuto, PhD

PII: S0277-9536(14)00239-1
DOI: 10.1016/j.socscimed.2014.04.015
Reference: SSM 9417

To appear in: Social Science & Medicine

Received Date: 15 May 2013

Revised Date: 14 March 2014
Accepted Date: 13 April 2014

Please cite this article as: Pachucki, M.C., Ozer, E.J., Barrat, A., Cattuto, C., Mental health and social
networks in early adolescence: A dynamic study of objectively-measured social interaction behaviors,
Social Science & Medicine (2014), doi: 10.1016/j.socscimed.2014.04.015.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

[Title]
Mental health and social networks in early adolescence: A dynamic study of objectively-measured social
interaction behaviors

[Authors] Mark C. Pachucki, PhD 1 *, Emily J. Ozer, PhD2, Alain Barrat, PhD3, Ciro Cattuto, PhD4

[Present/Permanent Address]

PT
(1) Mongan Institute for Health Policy, Massachusetts General Hospital
General Academic Pediatrics, Massachusetts General Hospital for Children
Harvard Medical School

RI
50 Staniford Street, 9th floor
Boston, MA 02114
Phone: 617-643-9268

SC
Email: mpachucki@mgh.harvard.org (*Corresponding author)
PubMed Indexing: Pachucki, MC

(2) University of California, Berkeley School of Public Health

U
Department of Community Health and Human Development
PubMed Indexing: Ozer EJ
AN
(3) Aix Marseille Université, CNRS, CPT, UMR 7332, 13288 Marseille, France
Université de Toulon, CNRS, CPT, UMR 7332, 83957 La Garde, France
M
Data Science Laboratory, ISI Foundation, Torino, Italy
PubMed Indexing: Barrat A
D

(4) Data Science Laboratory, ISI Foundation, Torino, Italy

PubMed Indexing: Catutto C
TE

[Acknowledgements]
We would like to thank the students of Cal School for their participation in this project, and to the Cal
School faculty and staff for their assistance and facilitation. Nancy Adler, Ray Catalano, Wendi Gosliner,
EP

Bonnie Halpern-Felsher, Kris Madsen, Janet Tomiyama, Nicki Bush, Laura Gottlieb, Aric Prather, Emily
Jacobs, Chris Myer, and Shannon Iaffaldano provided valuable feedback on project design. We are grateful
for the expert research assistance of Vanessa Greenfield, the statistical modeling advice of Kevin Lewis,
C

and colleagues in the Behavior Change Research Network at University of California, Berkeley and the
Human Nature Lab at Harvard Medical School.
AC

Contributions: MCP and EJO designed the study, MCP conducted fieldwork, conducted all data analysis.
AB and CC provided the wearable RFID sensors infrastructure for monitoring the proximity networks.
MCP drafted the manuscript and EJO, AB, and CC contributed to manuscript revision.

Funding/conflicts of interest: This work was conducted while MCP was in the Robert Wood Johnson
Foundation (RWJF) Health & Society Scholars program, and the project benefited from a RWJF pilot
grant. There are no reported conflicts of interest. The Committee on Human Subjects at UC Berkeley
approved the original study protocol.
 ACCEPTED MANUSCRIPT

Title: Mental health and social networks in early adolescence: A dynamic study of
objectively-measured social interaction behaviors.

Abstract: How are social interaction dynamics associated with mental health during early
stages of adolescence? The goal of this study is to objectively measure social interactions
and evaluate the roles that multiple aspects of the social environment - such as physical
activity and food choice - may jointly play in shaping the structure of children's

PT
relationships and their mental health. The data in this study are drawn from a longitudinal
network-behavior study conducted in 2012 at a private K-8 school in an urban setting in
California. We recruited a highly complete network sample of sixth-graders (n=40, 91% of

RI
grade, mean age=12.3), and examine how two measures of distressed mental health (self-
esteem and depressive symptoms) are positionally distributed in an early adolescent
interaction network. We ascertain how distressed mental health shapes the structure of

SC
relationships over a three-month period, adjusting for relevant dimensions of the social
environment. Cross-sectional analyses of interaction networks revealed that self-esteem
and depressive symptoms are differentially stratified by gender. Specifically, girls with
more depressive symptoms have interactions consistent with social inhibition, while boys'

U
interactions suggest robustness to depressive symptoms. Girls higher in self-esteem tended
towards greater sociability. Longitudinal network behavior models indicate that gender
AN
similarity and perceived popularity are influential in the formation of social ties. Greater
school connectedness predicts the development of self-esteem, though social ties
contribute to more self-esteem improvement among students who identify as European-
M
American. Cross-sectional evidence shows associations between distressed mental health
and students’ network peers. However, there is no evidence that connected students’
mental health status becomes more similar in their over time because of their network
D

interactions. These findings suggest that mental health during early adolescence may be
less subject to mechanisms of social influence than network research in even slightly older
TE

adolescents currently indicates.

Keywords: social networks, self-esteem, depression, early adolescence, stochastic actor-

based modeling
EP

Word count: 8,000 (body text)

C
AC

1
 ACCEPTED MANUSCRIPT

1. Introduction

As children develop, they pass through a series of sensitive periods in which their

lived experiences can have disproportionate and lasting impacts on future health

(Hertzman & Boyce, 2010). The onset of adolescence – typically between 9 and 12 years of

PT
age – has been recognized as a critical inflection point for the emergence of depression and

RI
other symptoms of distress (Andersen & Teicher, 2008; Brooks-Gunn & Petersen, 1991;

Steinberg, 2005). During this time, hormonal transformations accompany a reorientation of

SC
motivation, decision-making, and risk-taking behaviors (Dahl, 2004; Eccles et al., 1993),

and children on the cusp of adolescence are increasingly responsive to peer interaction and

U
social influence (Forbes & Dahl, 2010; Steinberg & Monahan, 2007).
AN
Additional attention to these peer dynamics is warranted due to a growing body of
M
evidence connecting social relationships with social cognitive development. At the onset of

adolescence, changes in social interaction behaviors are occurring alongside complex

D

physiological transformations. Early adolescence is a period marked by heightened

TE

neurological plasticity (Choudhury et al., 2012), growth in social cognitive reasoning skills

(Blakemore & Choudhury, 2006), and increased neurological sensitivity to social influences
EP

(Grosbras et al., 2007). Changes in social-affective processing that begin around puberty
C

include refinements in how children interpret rewards and threats, self-perception, and
AC

seek inclusion in peer groups. Consideration of others, interpretation of their intentions,

and reciprocation of trust and judgments of fairness begin to mature during early

adolescence (Crone & Dahl, 2012). Socially, susceptibility to peer influence grows during

childhood and peaks in adolescence. At the same time, parental influence begins to wane as

children express a heightened desire for independence (Dishion & Tipsord, 2011; Gifford-

2
 ACCEPTED MANUSCRIPT

Smith et al., 2005; Steinberg & Monahan, 2007). Early adolescent relationships are dynamic

(Cairns et al., 1995; Chan & Poulin, 2007) as adolescents develop sophistication in their

interpersonal behaviors (Brown, 1990), and increase in sensitivity to others’ actual and

perceived appraisals of themselves (Hergovich et al., 2002). These complexities of

PT
interpersonal behavior may affect mental health through a variety of status and reputation

RI
processes associated with group membership.

Relative to other periods of the life course, there has been surprisingly little

SC
research that takes account of the dynamism of human interactions to investigate how

early adolescent relationships may contribute to mental health. While a large body of

U
developmental research documents how generalized peer acceptance and friendship
AN
quality contribute to disordered mental health in children (Parker et al., 2006), less is
M
understood about how specific patterns of social interaction, and changes in these

interactions may be implicated (Chan & Poulin, 2009). Network perspectives on early
D

adolescent development thus hold potential to clarify links between externalized social
TE

behaviors and internalized cognition. Moreover, advances in measurement technologies

enable us to identify behavioral social networks in more precise and unobtrusive ways
EP

than ever before.

C

These concerns motivate the present study of how social interaction dynamics are
AC

associated with mental health during the onset of adolescence. To do so, we partnered with

a school to examine changes in social interaction patterns and mental health indicators of a

nearly complete cohort of 6th-graders over a three-month period. One contribution of this

study is to use of a novel electronic technology to objectively gather network data among

this age cohort. A second contribution is to use these network data to examine how

3
 ACCEPTED MANUSCRIPT

multiple personal and peer characteristics may shape two indicators of early adolescent

mental health: self-esteem and depressive symptoms. While depression has been a subject

of some investigation in a social network framework, especially among older adolescents,

few network studies have examined relational determinants of self-esteem during any

PT
stage of adolescence. To the best of our knowledge, this is the first use of objectively-

RI
gathered micro-social interaction data in a study of adolescent mental health.

SC
2. Review: relevance of social networks for distressed mental health

A rapidly-expanding body of research attests to the fact that because humans are

U
connected, their health can also be connected (Smith & Christakis, 2008) in ways that differ
AN
across the life course (Umberson et al., 2010). Relational methods of social network
M
analysis have provided insight into interpersonal influences on mental health among

adolescents. Because depression and self-esteem during early adolescence are

D

understudied from a network perspective, this study focuses on these two constructs.
TE

Depression and social networks in adolescence

EP

Adolescent social relations are correlated with depressive symptoms, often due to
C

reputation and status mechanisms connected with peer acceptance and rejection. This
AC

phenomenon has been examined using several different approaches. Dyadic studies that

investigate psychometric properties of a sample of peers are more common among the

limited number of early adolescent studies of depression. Regression-based studies and

actor-based studies are more common among middle and later adolescents.

4
 ACCEPTED MANUSCRIPT

In the first approach, the dyadic relationship between a given individual and the

attributes of that individual’s single friend (or small group of friends) are typically

examined using a correlation or multiple regression analysis. For example, a study of 11th

graders found that girls’ social anxiety is associated with her immediate peer’s social

PT
anxiety (Prinstein, 2007). For boys, depressive symptoms tended to be linked with low

RI
friendship quality and having friends who are highly popular. Researchers examining

friendship changes among 6th-graders found that friendship instability is significantly

SC
related to subsequent depressive symptoms, and particularly sensitive to best friend

instability (Chan & Poulin, 2009). A study of non-suicidal self-injury (NSSI) among 6th to 8th

U
graders found that girls’ friends affect their propensity to enact non-suicidal self-injury
AN
behaviors, while there is no similar phenomenon among boys (Prinstein et al., 2010).
M
A second relational approach employs a regression-based framework to examine

health among a larger proportion of an individual’s social network than just a best friend or
D

close friends, sometimes taking into account the broader network structure in which
TE

individuals are embedded. In a study of suicide using data from the National Longitudinal

Study of Adolescent Health (Add Health), Bearman and Moody (2004) found that having a
EP

friend who commits suicide increased the likelihood of ideation and attempts. Another line
C

of investigation concerns the relationship between the size of the network and mental
AC

health. Using the same dataset, Ueno (2005) discovered that among late adolescents, higher

levels of social integration were associated with fewer depressive symptoms. Falci &

McNeeley (2009) extended this work by showing that the relationship between social

integration and depression is not linear, but convex. In other words, having either too many

or too few friends increases one’s risk of depression among teens, a phenomenon that

5
 ACCEPTED MANUSCRIPT

depends both on gender and the density of one’s personal network. Other research using

Add-Health data has shown that teenagers with depressive symptoms who report same-sex

and both-sex attraction are more socially isolated, have lower social status, and have fewer

network connections than opposite-sex attracted youths. These effects are especially

PT
pronounced in females (Hatzenbuehler et al., 2012).

RI
A third type of study uses a stochastic actor-based model (SABM) framework that

can evaluate changing social connections and a given individual-level health condition as

SC
inter-related outcomes (Snijders et al., 2010). This simulation-based approach capitalizes

on longitudinal sociocentric (whole-network) data, and allows investigators to evaluate the

U
contributions of common endogenous social mechanisms (i.e. reciprocity, transitivity,
AN
preferential attachment) to an outcome. A benefit of this approach is that it explicitly
M
models the interconnectedness and temporality inherent to social life, as opposed to

making post-hoc adjustments for the statistical dependence of observations (a limitation of

D

regression-based approaches). This framework has been useful for testing the reasons that
TE

depressed friends affiliate with one another through systematically evaluation of

alternative mechanisms for network formation and behavior change. For instance, in a
EP

study of approximately 850 Swedish 14-year olds, peer depressive symptoms were
C

predictive of adolescents’ depressive symptoms, after adjusting for network selection and
AC

de-selection effects over a four-year period (van Zalk et al., 2010). This is consistent with a

social contagion explanation for why depressed friends affiliate, and the authors also find

evidence that this process varies by gender.

In a study of American adolescents, Schaefer et al. (2011) extended earlier

regression-based research on the link between network size and depression with an actor-

6
 ACCEPTED MANUSCRIPT

based framework. They tested several competing explanations for depressive similarity

among 9th to 11th-grade friends, finding that connected adolescents are similar in

depression due to withdrawal from broader social networks. A study of nearly 1000

Finnish 16-year-olds revealed that depression tends to converge over time within friend

PT
groups (Kiuru et al., 2012). The authors evaluated the relative strength of social selection

RI
and social influence explanations, finding that depression convergence is not explained by

social contagion processes but because individuals select friends with similar depressive

SC
symptoms. Van Workum and colleagues (2013) examined network processes that explain

similarities in happiness among a group of approximately 400 socially-connected 15-year-

U
olds. The authors observed that while adolescents do not form friendships with similar-
AN
happiness peers, being similarly happy is an explanation for friendship maintenance, and
M
happiness differences are a reason for friendship termination. Additionally, they observed

that a given individual’s happiness becomes more similar to the average level of her
D

friends’ happiness.
TE

In sum, while the body of longitudinal studies of mid- and late-adolescent

depression that make use of sociocentric network information is relatively well-developed,

EP

the same level of attention has not yet been given to early adolescent depressive
C

symptoms. The present study aims to extend knowledge in this area by using longitudinal
AC

data to conduct sociocentric analyses of changing social ties and depressive symptoms of a

cohort of 6th-grade students. Moreover, because the network data is gathered objectively,

many of the biases associated with self-reported information can be avoided.

Self-esteem and social networks in adolescence

7
 ACCEPTED MANUSCRIPT

Self-esteem in early adolescence is linked to social relationships through feelings of

peer support, and self-perception of one’s place in peer groups and social status

hierarchies. Self-esteem is considered to be an important indicator of personal resilience,

defined as adaptability in the face of adversity, risk exposure, or disadvantage (Dumont &

PT
Provost, 1999; Fergus & Zimmerman, 2005). Thus, to the degree that social network

RI
structure is linked with self-esteem, it may provide insight into strategies for promoting

resilience during early adolescence and beyond.

SC
Social network research on self-esteem and peer relationships has been relatively

rare, though correlational studies of peer-derived psychological constructs are more

U
common. For example, early work on the transition from 6th to 7th grade documents strong
AN
associations between survey measures of peer support and self-esteem (Hirsch & Rapkin,
M
1987). In a study of the peer relationship quality and health of chronically ill 5th-graders,

McCarroll and colleagues (2009) found that self-reported high self-esteem is linked with
D

teacher-reported pro-social behavior and low aggression. A recent study of changes in

TE

mutual friendship patterns found that peer acceptance, friendship quality, and number of

friends are associated with loneliness (Kingery & Erdley, 2007). To the best of our
EP

knowledge, no studies of early adolescent self-esteem have made use of detailed

C

longitudinal network data (nor objectively-gathered social interaction data, used here) to
AC

conduct sociocentric analyses of changes in self-esteem.

Behavioral and social environment correlates of mental health in adolescence

Common correlates of mental health status are behaviors such as eating and

physical activity; see Salvy et al. (2012) for an extensive recent review. Bruening and

8
 ACCEPTED MANUSCRIPT

colleagues (2012) showed that best friends tended to have similar healthy choices (whole-

grain and dairy), and tended to eat breakfast. Recent actor-based network research on 8th-

grade low-nutrient energy-dense food consumption revealed that friends tend to emulate

each other’s junk food consumption in a way more “mindless” than deliberatively cognitive.

PT
In other words, adolescents were not thinking about their friends’ choices and then acting,

RI
but rather acting and then modifying behaviors (de la Haye et al., 2013). Network research

on physical activity in an afterschool program among a range of adolescents demonstrates

SC
peer influence on physical activity levels (Gesell et al., 2012).

Two important social environment correlates of mental health in adolescence

U
include social status position and level of school connectedness. Research by Goodman and
AN
colleagues (2001) showed that older adolescents’ perceived socioeconomic status is
M
related with depressive symptoms and self-esteem. Feelings of connectedness with one’s

school environment have been found to be associated with lower likelihood of risk
D

behaviors and distressed mental health (Resnick et al., 1997).

TE

Gaps in Research and Focus of Present Study

EP

The growing literature on teens’ social networks provides insight into how mental
C

health is related to different aspects of social connectedness. However, our review reveals
AC

several conceptual and methodological gaps. First, there is a relative shortage of early

adolescent social network research, which leaves knowledge of how relationships among

this age cohort shape mental health (and vice versa) relatively underdeveloped. For

instance, social networks of 6th graders tend to be more horizontally organized, while by 8th

grade (and even more strongly in high school years) networks ossify into clique-based,

9
 ACCEPTED MANUSCRIPT

mixed-gender, and more hierarchical structures (Steinberg & Morris, 2001). Thus, it may

be inappropriate to generalize network findings from later adolescent cohorts to those in

early adolescence, since interaction patterns change rapidly as children develop. Second,

much prior network research tends to rely solely on self-report of friendships, which can

PT
be highly sensitive to recall bias and context effects. Third, scholars have used the term

RI
“peer studies” in an inconsistent manner to mean anything from an amorphously defined

group of age-similar individuals, to a specific cohort of individuals considered friends, to

SC
studies involving a single peer (Fitzgerald et al., 2012). This problem of definition

undermines efforts to compare study settings and precisely identify causal effects.

U
Methodologically, many studies do not assess attributes or behaviors of the
AN
complete range of all of an individual’s alters, but rather only those of a close friend (i.e.
M
Prinstein, 2007), a small number of closest friends (i.e. Kiuru et al., 2012), or a delimited

number of alters (all research that uses Add-Health data, and other sophisticated designs,
D

i.e. van Workum et al., 2013; van Zalk et al., 2010). This type of missing data presents a
TE

challenge to causal inference when a large part of a child’s social network is not

represented in analyses. Last, with the exception of Chan and Poulin (2009), many of the
EP

studies that employ a longitudinal perspective rely upon long follow-ups, for example 12
C

months (Kiuru et al., 2012), or 18 months (Prinstein, 2007). This is problematic because it
AC

is well-known that during adolescence relationships can change very quickly, on the

timescales of days or weeks (Poulin & Chan, 2010). This variability is obscured by designs

with a lengthy follow-up horizon.

The present study seeks to address these concerns by investigating the following

questions: (1) How are network characteristics of young adolescents associated with

10
 ACCEPTED MANUSCRIPT

depressive symptoms and self-esteem? (2) How might network-based social selection and

social influence operate as mechanisms that contribute to change in self-esteem or

depressive symptoms? Face-to-face social interaction behaviors are electronically

measured between two individuals to generate social tie data. As others have argued,

PT
durable social relationships are built over time from repeated prior interactions, and

RI
understanding how patterns emerge over time is of crucial importance (Collins, 2004;

Hinde, 1976). Although we do not claim that interaction networks proxy the social

SC
relationship of friendship, examination of objective interaction measurements reveals a

high degree of correspondence with participants’ self-reports of friendship (see

U
Supplementary Data, Section 1).
AN
Another contribution of this design is to examine early adolescent network
M
structure in an educational setting at a finer temporal granularity than prior research.

Here, this temporal granularity was operationalized in terms of 20-second interaction

D

measurement intervals, within nine days of observation, over a three-month period.

TE

Although many others have conducted repeated measurements of social structure in

schools (e.g. Clack et al., 2005; McFarland, 2001), this is the first study to our knowledge to
EP

objectively examine the evolution of interaction networks in the context of early adolescent
C

mental health, and to examine self-esteem in a sociocentric network setting.

AC

Objective measurement of behaviors makes this study distinctive from prior

research on network processes in mental health, in that the research literature does not

offer strict guidance as to how behavioral (as opposed to self-report) measures of

children’s networks might change, nor how these changes might be related to their mental

health. Propositions formed on the basis of self-report data in adolescent mental health

11
 ACCEPTED MANUSCRIPT

research suggest that individuals who are positionally more peripheral in their school-

based social network may be higher in depressive symptoms, lower in self-esteem, lower in

school connectedness, and lower in social status. Research on peer socialization in later

adolescent cohorts suggests that socially-connected individuals are likely to have similar

PT
levels of depression. Because it has been demonstrated that both depressive symptoms and

RI
more sophisticated forms of social behaviors increase between late childhood and early

adolescence, it is hypothesized that there will be similarity in mental health status among

SC
socially-connected adolescents in this study population.

U
2. Methods
AN
M
Overview

The data in this study are drawn from the Pre-Adolescent Eating and Exercise
D

Network Study (preTEENS), conducted in 2012 by the authors at a private K-8 school in an
TE

urban setting in California (hereafter referred to as “Cal School”). preTEENS is best

characterized as a “network-behavior panel study” (Steglich et al., 2010) of early

EP

adolescent socialization, health behaviors, and key indicators of physical and mental health.
C

Based upon a review of the child development literature, a sixth-grade student sample was
AC

recruited because students in the 10-12 year-old range fall within a hypothesized

transition zone between decreasing parental influence and increasing peer influence

(Dishion & Tipsord, 2011).

School selection and partnership development

12
 ACCEPTED MANUSCRIPT

Criteria for study site selection were small grade size (less than 50 students), a

school administration with interest in our research questions about student wellness, and

stability in teaching and administrative staff so that the study could be sustained over the

yearlong study setup and longitudinal data collection period. Selecting for a stable and

PT
supportive school environment was prioritized in order to avoid challenges that can

RI
prevent collection of high-quality network data, such as absenteeism, severe behavior

problems, and lack of general interest by students and teachers. Because of the complexity

SC
of the methods being tested (i.e. collecting objective social interaction network and

physical activity data), we did not prioritize the selection of a research site to be

U
representative of urban middle schools, nor did we seek to conduct this phase of the

research at multiple sites.

AN
M
The approach to partnership with Cal School was collaborative and iterative.

Following several meetings with Cal School’s administrative leadership, school

D

psychologist, and 6th-grade teaching staff, a presentation about the research was made to
TE

the students to explain the purpose of the study and to answer their questions. Appropriate

participation incentives were provided commensurate with standards in the field. A follow-
EP

up debrief providing initial observations to children was conducted in small focus groups
C

at the completion of observations, and a pizza party was provided to the students as a
AC

further thank-you for their participation. The Committee for Human Subjects at

[REDACTED FOR REVIEW] approved the study protocol.

Design and Procedure

13
 ACCEPTED MANUSCRIPT

Students provided multiple types of data, including (i) self-report questionnaire

data concerning their health, eating, and exercise, (ii) objectively measured physical

activity using accelerometers; and (iii) objectively measured social interaction data with

wearable proximity sensors using active radio-frequency identification (RFID) technology,

PT
Parents of students were first required to provide written consent for their child’s

RI
participation as minors, and students completed an age-appropriate assent form when they

volunteered. Children were granted anonymity and informed their real names would not

SC
be used in presentations or publications, and that their responses to questionnaires and

identified social interaction data would not be shared with parents, teachers, or peers.

U
Following data-collection, responses associated with each child were coded to an electronic
AN
file using a unique random ID. Hard-copy questionnaires that listed students’ real names
M
were then destroyed.

Because study aims were to focus upon short-term changes in network structure,
D

health behaviors, and health status, social interaction network data were collected for three
TE

consecutive days at monthly intervals for three months during the second half of the 6th-

grade schoolyear. Nine days of measured interaction data were thus collected. To capture
EP

network socialization processes, an entire grade level of early adolescents was contacted to
C

participate in the study. By the commencement of the study, 90.9% (n=40) of the 44
AC

students in the sixth grade had volunteered. One of the consented students was absent

during the second observation period, and another declined to participate in the third

period owing to perceived response burden. Non-participating students were not

significantly different than participating students in terms of age or gender (using t-tests

14
 ACCEPTED MANUSCRIPT

and χ2 tests). Further consultation with teaching staff suggested no significant differences

in other social or behavioral dimensions.

The sixth grade at Cal School had a majority of girls (F: 62.5%, n=25, M: 37.5%,

PT
n=15), and the mean age at enrollment was 12.3 (SD: 0.29, Range=11.7-12.8). Self-report

from students indicates that at least half of students’ parents have advanced degrees,

RI
ascertained from five categories (neither parent went to college; one went to college; both

SC
went to college; one has an advanced degree; both have an advanced degree). The grade

contains more students who identify as European-American (72.5%, n=29) than not

U
(27.5%, n=11), ascertained from self-report categories (African-American, Asian, Latino or
AN
Hispanic, White, Other/Write-in). There were no statistically significant differences

between boys and girls in terms of age, parent education, or race/ethnic background.
M
D

Measures
TE

Social environment
EP

Two measures of children’s social status are included, the first based on subjective

self-report, and the second on peer nomination. Children’s subjective social status was
C

assessed using the youth ladder instrument adapted from the MacArthur Research
AC

Network on Socio-Economic Status & Health (Goodman et al., 2001). Students were shown

a picture of a 10-rung ladder, and indicated which rung they would place themselves in

their school environment. The stability of the measure from period 1 to period 2 was

relatively high (Spearman’s rho=0.85), and also between periods 2 and 3 (r=0.76). The

15
 ACCEPTED MANUSCRIPT

second measure was based on perceived popularity; participants were asked to indicate

the most popular students on a roster, and a summary measure was built. The stability of

the measure from period 1 to 2 (Spearman’s rho=0.96) and period 2 to 3 (r=0.93) was very

high.

PT
School connectedness was assessed with a measure of children’s social

RI
environment, using a scale created for the National Longitudinal Study of Adolescent

Health (Add Health)(Resnick et al., 1997). The instrument asks students the degree to

SC
which children feel close to people at school, are happy to be at this school, are a part of

this school, feel safe, and are treated fairly by teachers. Students rate statements using a

U
four-point Likert scale ranging from ‘strongly disagree’, ‘disagree a little’, ‘agree a little’, to
AN
‘strongly agree’. The connectedness scale has adequate reliability at all periods (Cronbach’s
M
alpha, P1 = 0.80, P2= 0.91, P3=0.76). Summary scores for each measure were constructed

using component questions and normalized between 0 and 1. The scale shows moderate
D

consistency from period 1 to 2 (Spearman’s rho=0.69) and from period 2 to 3 (r=0.66).

TE

Health behaviors
EP

Measures of exercise and food choice are included because low levels of exercise
C

and unhealthy food choices have been linked with children’s mental health. To obtain
AC

objective measurement of physical activity under free-living conditions, students were

provided with a consumer-grade accelerometer (Fitbit®) to wear during the entire school

day. While this device mainly captured incidental activity, such as travel between classes

and recess time, students also wore it during physical education class. The devices yielded

total steps per day and daily minutes of vigorous activity (averaged over the week of each

16
 ACCEPTED MANUSCRIPT

panel). Since this device is relatively new, it has not been validated for research purposes

against tools such as the Actigraph®, a more sensitive device considered the industry

standard in exercise physiology (Welk et al., 2012). However, the present priority is to

capture basic dimensions of physical activity, and because the same device was worn by

PT
each student, measurements are internally consistent within this population. As device

RI
measurement can vary depending upon where on the body it is worn (Vahdatpour et al.,

2011), all students were required to wear it on their waistband for consistency.

SC
The Healthy Eating Active Living (HEAL) questionnaire was used to measure food

choices. This instrument is publicly available for download from the website of the

U
Veronica Atkins Center for Weight and Health at UC Berkeley (Samuels et al., 2010). This
AN
age-appropriate questionnaire asks youths to indicate the consumption frequency of 18
M
common food and beverage items during the prior day. It has been recently used to assess

youth attitudes and behaviors surrounding school food (Gosliner et al., 2011) and
D

consumption changes following school nutrition reform (Woodward-Lopez et al., 2010). A

TE

typical question is: “Yesterday, did you eat candy of any kind? A: Did not eat, Ate at School,

Ate at home, Ate at Some other Place” (possible frequency: 0-3). Though these questions
EP

are too broad to evaluate nutrient content, the indicators allowed the creation of simple
C

summary measures of the consumption of healthier and unhealthier items. Healthier items
AC

included green salad, vegetables, fruit, baked chips, milk, water, fruit juice, and diet soda.

Unhealthier items included French fries, hot dogs/hamburgers, fried chicken, nachos,

candy, ice cream, sweets, chips, soda, and sweetened juice drinks.

Social interaction networks

17
 ACCEPTED MANUSCRIPT

Students’ changing interaction patterns were investigated during the three months

of the study. The lunchroom was chosen as the site for network observation because

students tend to sit with social intimates at lunchtime. Social interaction patterns were

recorded during lunchtime using an unobtrusive proximity-sensing active RFID tag worn in

PT
a neck lanyard. In conjunction with small wall-mounted receivers, these devices allow for

RI
the identification of who interacts with whom with a 20 seconds time resolution

throughout the approximately 25-minute long lunch period. Though 6th-graders ate lunch

SC
at the same time as 7th grade students, students self-segregated by grade, and 7th-graders

were not involved in this study. There was always a teacher present in the lunchroom, and

U
no aggressive or otherwise negative behavior incidents were observed during the study

term.
AN
M
Each tag recorded a digital signature of another tag only when two individuals were

having a face-to-face interaction within a range of approximately 1.5m, and the sensors
D

were tuned so that the proximity of two individuals wearing them could be assessed with a
TE

probability in excess of 99% over an interval of 20 seconds. Moreover, radio signals from

other proximate individuals were of such low power that they would be stopped by the
EP

water content of each person’s body. Thus, signals would not pass backwards through a
C

student’s chest and interact with a tag of another student sitting back-to-back with her at
AC

another table. This specific technology has been successfully used to measure person-to-

person interactions in settings as diverse as a hospital ward (Isella et al., 2011), elementary

school (Stehlé et al., 2011), and conferences (Cattuto et al., 2010). Similar wireless sensor

technologies have been used to measure interactions in a high school (Salathe et al., 2010)

18
 ACCEPTED MANUSCRIPT

and college (Eagle et al., 2009), and real-time experiences among adult park visitors

(Doherty et al., 2014).

The most basic unit of measurement is the occurrence of a digitally observed

interaction between two individuals, which then becomes coded as an undirected social tie.

PT
For each day, only those interactions between participants that had a minimum cumulated

RI
interaction length of 80 seconds were retained. First, it was important to reduce the

inclusion of transitory or accidental interactions measured at lower scan thresholds (i.e.

SC
20-40 seconds) as they can represent, at some level, practical and statistical noise and it is

doubtful that all of them capture a meaningful interaction. Second, sensitivity analyses of

U
self-report friendship vs. RFID-measured interactions over the three periods of this study
AN
suggest that longer-duration face-to-face interactions, especially above the 60-second
M
duration threshold, correspond to more meaningful social relations such as friendship

(discussed in Section 1 of Supplementary Data). Social ties from Days 1-3 of the first week
D

were then aggregated to constitute Period 1; days 4-6 comprised Period 2; and days 7-9
TE

comprise Period 3. It is hypothesized that there will be an association between these

interaction-derived social ties and mental health indicators, as social relationships have
EP

been previously implicated in mental health status in early adolescence.

C
AC

Dependent variables: Self-esteem and depressive symptoms

Two indicators of mental health were assessed as outcomes to be analyzed with

interaction-derived measures of social ties. The first is self-esteem, measured using eight

questions drawn from the DuBois Self-Esteem Questionnaire (DuBois et al., 1996). These

questions tap self-concept, children’s perceptions as people, and sense of contentment;

19
 ACCEPTED MANUSCRIPT

students rate statements using a four-point Likert scale. The self-esteem scale has good

reliability at all periods (Cronbach’s alpha, P1= 0.83, P2=0.86, P3=0.83). Summary scores

for each measure were constructed using component questions in order to evaluate the

stability of the construct over time. The stability of the summary scores is consistent from

PT
Period 1 to 2 (Spearman’s rho=0.74) and from Period 2 to 3 (r=0.72). At Period 1, 55%

RI
(n=21) of students are in the top two self-esteem quartiles; at Periods 2 and 3, 53% (n=20)

are in the upper half of the distribution.

SC
The second indicator is generalized depressive symptoms, assessed using the 10-

item short form of Kovacs’ Children’s Depression Inventory (CDI) (Sitarenios & Kovacs,

U
1999). It asks questions related to sadness, feeling alone, self-image, and social integration,
AN
and for each item participants choose one of three sentences that described their
M
symptoms over the past two weeks (0 = symptom absence; 2=the most severe symptom

expression). This measure has good within-scale reliability at all periods (Cronbach’s alpha,
D

P1=0.80, P2=0.82, P3=0.89). Over time, the scale is fairly stable from Period 1 to 2 (r=0.74),
TE

and a bit less stable from Period 2 to 3 (r=0.68). There are low rates of depressive

symptoms. At Period 1, two males and six females are above the 50th t-score percentile for
EP

depressive symptoms; the remaining students (80%) are below. During Period 2, one male
C

and three females are above; 93% are below. At Period 3, one male and five females are
AC

above; 85% are below.

Statistical Analysis

First, bivariate associations are calculated between time-invariant demographic

traits with time-varying social status, health behavior, and mental health dimensions at

20
 ACCEPTED MANUSCRIPT

each of the three observation periods. Correlation is used for continuous covariates, and X2

associations for categorical covariates. Next, various structural attributes of the student

interaction networks are described at each observation period, and associations are

assessed between the between socio-demographic traits and network attributes. The [sna]

PT
(Butts, 2013) and [igraph] (Csardi, 2013) packages for the R programming language were

RI
used for descriptive network analysis, and the [RSiena] package for R (Ripley et al., 2013)

was used for statistical social network analysis.

SC
Because the initial descriptive analysis reveals cross-sectional associations between

network characteristics and mental health, the dynamic co-evolution of mental health and

U
social interaction is analyzed as a joint dependent outcome. To do so, a stochastic actor-
AN
based modeling (SABM) framework is used to treat social ties and either depressive
M
symptoms or self-esteem as inter-related dependent variables. While detailed elaborations

of this approach are found elsewhere (Snijders et al., 2010; Steglich et al., 2010), a brief
D

treatment of procedures is appropriate. Relative to traditional regression-based

TE

approaches that assume individuals are statistically independent, actor-based models

instead treat individuals as interconnected. Models of interpersonal dependencies rely

EP

upon simulation methods that treat the observed network structure as a consequence of
C

micro-changes in actors’ behavior that proceed according to specified rules in a

AC

continuous-time Markov process.

One benefit of this approach is the ability to account for both stable and time-

varying properties of individuals, as well as changes in network structure that may result

from endogenous interpersonal mechanisms like preferential attachment (those with many

friends tend to attract more friends) and transitivity (friends of friends tend to become

21
 ACCEPTED MANUSCRIPT

friends). The particular processes of interest can then be parameterized in models. The

forward model selection procedure suggested by developers was used to choose a

parsimonious set of predictors for model inclusion. Important SABM considerations are

that continuously-coded variables need to be transformed into ordinal variables, and there

PT
must be adequate change in the outcome behaviors of interest. While there were relatively

RI
low rates of depressive symptoms and low self-esteem at baseline, diagnostics indicate

models are well-parameterized. Further information on model specification, variable

SC
transformations, and behavior changes are found in Supplementary Data (Section 3 and 4).

U
3. Results
AN
M
[Insert Table 1, “Participant characteristics, by gender”, about here]
D

Table 1 provides information on the baseline demographic composition of the

TE

cohort, as well as cross-sectional detail on participants’ social environment, health

behavior, and mental health characteristics, with bivariate tests for group differences by
EP

gender. There was no significant gender distinction in age, race/ethnicity, or parent

C

education. Across all observation periods, boys tend to be higher in subjective social status
AC

than girls, during all three periods (the last period only reaches a marginal level of

significance). There was no significant gender difference in peer-nominated popularity, or

school connectedness. In terms of health behaviors, boys are significantly more physically

active during the day, both in duration and intensity (p<0.001). Though boys report eating

fewer unhealthy food choices at the first and last period, neither healthy nor unhealthy

22
 ACCEPTED MANUSCRIPT

food varies significantly by gender. As expected in this age group, distributions are skewed

towards low depressive symptoms and high self-esteem, and there are few gender

differences. Only self-esteem at the first period reaches a marginal threshold for

significance (p=0.08), with boys reporting higher self-esteem than girls. Additional

PT
correlations between mental health indicators, socio-demographic, social environment, and

RI
health behavior covariates are reported in Supplementary Data (Section 2, Table A1 & A2).

SC
Network change and associations with mental health

Though measures of mental health and social status were obtained once per period,

U
three days of lunchtime interaction were observed per period, yielding nine days of
AN
network observation. Figure 1 shows the structure of the daily lunchroom networks, which
M
are then aggregated into three weekly networks to match the periodicity of the self-

reported covariates. As expected, gender was the most salient axis of socio-demographic
D

differentiation.
TE

There are several important details to note about daily lunchroom interaction

networks. First, with the exception of days 4 and 9, the majority of lunchtime periods
EP

contain several disconnected network components, reflecting subgroups that sat at

C

different tables. There appears to be a strong tendency towards gender homophily, though
AC

a strict polarization is not observed every day. Although there are a few students who are

unconnected to others, it is erroneous to think of these students as isolates. It is more

appropriate to state that these students did not meet the minimum 80-second interaction

criterion for tie creation. Additional information on network change (node- and graph-level

23
 ACCEPTED MANUSCRIPT

attributes) as well as change in dependent variables is found in Supplementary Data

(Section 3, Table A3).

[Insert Figure 1 – “Gender structure of 6th-grade interaction network during lunch, over

PT
nine days in a three-month period”]

RI
Table 2 reports on associations between individual-level network attributes and

SC
mental health status, and with few exceptions the direction of the association

(positive/negative) is described, rather than the strength of association (Pearson

U
correlation, r=0.0-1.0). Three node-level network statistics are presented: (a) size of
AN
personal networks, indicating the number of alters to whom a given individual is
M
connected; (b) transitivity (local clustering), an indication of how tightly grouped students

are (Wasserman & Faust, 1994); (c) closeness centrality, a measure of central or peripheral
D

structural location in one’s network (Freeman, 1979). Associations are reported by gender
TE

for each mental health outcome, with significant correlations indicated.

EP

[Insert Table 2 – Correlations between network indicators and mental health]

C
AC

At all periods, boys with more depressive symptoms tend to have strong positive

coefficients for network size, suggesting greater integration in social networks, though the

gender difference is only statistically significant at the second period. Girls are more

consistent in the direction of association between depressive symptoms and network

indicators at all periods. Girls with more depressive symptoms tend to be significantly less

24
 ACCEPTED MANUSCRIPT

integrated in their networks at two periods (-network size), are significantly more

peripheral at two periods (-closeness), and have significantly more transitive relationships

at one period (+transitivity), though even non-significant coefficients have the same

valence across periods. These patterns are consistent with a greater perceived social

PT
stigma related to depressive symptoms for girls, while boys’ depressive symptoms appear

RI
not to be associated with their propensity to interact. However, while Table 1 reports that

the difference in girls and boys’ depressive scores is not significant at any of the three

SC
periods, the boys’ level was consistently lower.

Examination of self-esteem reveals no significant or consistent relationship with

U
boys’ network attributes over time. Female students with higher self-esteem tend to have
AN
more interaction partners at all periods (+network size), though only one period was
M
statistically significant. In addition, girls with higher self-esteem tend to have greater

centrality (+closeness), though again only the last period was significant. These
D

associations suggest that girls with higher self-esteem tend towards greater sociability and
TE

social integration, while there is no consistent pattern among boys. This is consistent with

depressive symptoms as well.

EP

Taken together, cross-sectional evaluation of associations between indices of

C

network structure and mental health provides evidence of a great deal of variability even
AC

within a relatively narrow timespan (three months), indicating fluidity in children’s social

interactions and mental health status.

Co-evolution of social interaction behavior and mental health

25
 ACCEPTED MANUSCRIPT

Next, the extent to which individual attributes and aspects of network structure

predict the joint formation of ties and change in mental health status is assessed (first for

depressive symptoms, then for self-esteem). It should be noted here that not every

measured health behavior (physical activity and eating behaviors) was significant in the

PT
forward model-selection process for each mental health outcome. Results from alternative

RI
analyses that included all health behaviors did not differ significantly. See Supplementary

Data (Section 4) for additional detail on model specification.

SC
[Insert Table 3. “Social interaction and self-esteem co-evolution”, about here]

U
AN
Table 3 reports on the joint outcome of network formation and self-esteem change,
M
and indicates a highly significant rate of change in lunchtime interaction, with network

change from P2 to P3 greater than from P1 to P2. The degree parameter indicates that
D

students are selective in their choice of interaction partners, and a positive transitive triad
TE

parameter indicates a tendency towards children’s interaction partners being connected as

in three-person groups. Demographically, there is a negative tendency for girls to

EP

contribute to network tie formation, suggesting that boys are doing more tie initiation,
C

even after adjusting for their relatively fewer numbers in the gender distribution. We also
AC

observe a strong positive tendency towards gender homophily in tie formation, suggesting

that boys tend towards interactions with boys, and girls towards girls (Stehlé et al., 2013).

Last, there is a strong significant relationship between perceived popularity similarity and

subsequent tie formation, suggesting that students perceived as popular by others tend to

form ties with each another.

26
 ACCEPTED MANUSCRIPT

Changes in self-esteem were predicted by two actor attributes: student report of

school connectedness, and race/ethnicity. For the former, feeling connected to the school is

predictive of subsequent self-esteem. However, the latter term indicates a racial disparity,

insofar as European-American students shift towards higher self-esteem. Also of note is a

PT
significant negative quadratic shape effect. This type of negative feedback is known to

RI
indicate a self-correcting behavior and suggests that as individuals approach either very

high or very low self-esteem values, there is proportionally less change.

SC
Tie-formation findings for depressive symptoms/social interaction co-evolution

(Table 4) are similar to the self-esteem model, having the same significant parameters for

U
degree, transitivity, ego gender, same-gender, and popularity similarity. However, the
AN
behavior dynamics part of the model reveals that none of the actor attributes shape
M
depressive symptoms, and the lack of significance in average depression similarity suggests

that early adolescents do not shape one another’s depressive symptoms through a social
D

influence process. Additional sensitivity analyses used “average alter” and “total similarity”
TE

as alternative social influence effects, with no difference in findings.

EP

[Insert Table 4. “Social interaction and depressive symptom co-evolution”, about here]
C
AC

4. Discussion

Although prior research has established a relationship between depression and peer

group structure among older adolescents, the present study is novel in its objective

enumeration of dynamic social interaction networks in a middle school setting to examine

27
 ACCEPTED MANUSCRIPT

the development of both self-esteem and depression during early adolescence. While this

was a new application of this approach, we believe that reliance upon social interaction

data to construct network data, that reflects who students actually interact with, has great

potential to clarify how socialization processes unfold. In addition, this sociocentric focus

PT
on short-term changes in children’s networks reveals substantial social fluidity even within

RI
three months. While some prior work (Chan & Poulin, 2007) suggests that approximately

one-third of 6th-graders’ self-reported friendship ties are unstable from month to month,

SC
examination of who children actually interact with reveals that nearly two thirds of 6th-

graders’ ties change. This is similar to findings by Cairns and colleagues among 4th and 7th-

U
graders (Cairns et al., 1995).
AN
These longitudinal analyses of changes in mutually dependent tie-formation and
M
mental health confirm the important role that gender and popularity play in shaping social

interaction during early adolescent years (Stehlé et al., 2013). There is also an observed
D

propensity towards transitive closure at this age – friends of an individual tend to be

TE

friends as well. A further notable finding is that similarity in peer-perceived popularity has

a key role in determining interactions. The role of health behaviors such as eating and
EP

exercise are often emphasized as key determinants in physical and mental health
C

improvement. However, these health behaviors – even when physical activity is measured
AC

objectively rather than via self-report – do not seem to be as relevant to distressed mental

health among this age group.

Findings on the co-evolution of self-esteem and social interaction suggest that the

development of self-esteem in a network setting is socially moderated such that

individuals’ self-esteem neither becomes too inflated, nor dips too low. Rather, the

28
 ACCEPTED MANUSCRIPT

quadratic behavior shape parameter provides some evidence that self-esteem development

is self-correcting in an interconnected social system. Change in self-esteem is not

dependent upon just the individual, but upon those to whom the individual is socially

connected. The observation that increased sentiments of feeling connected to the school

PT
predict greater self-esteem suggests that fostering a welcoming environment for students

RI
may pay dividends in terms of their self-esteem and resilience. However, there is also

evidence that a racial disparity in self-esteem improvement has roots, in part, in a network

SC
mechanism. This highlights a phenomenon that may be proactively addressed through

attention to student interaction experiences, rather than simply attempting to address self-

U
esteem on an individual basis.
AN
Consistent with prior research, there are cross-sectional associations between
M
mental health status and school-based peer social network attributes. However, the

stringent causal methods employed here did not provide evidence of peer mental health
D

status similarity over time, either from social selection or from peer social influence. This
TE

stands in contrast to the small number of adolescent studies of peer mental health that find

peer influence or social selection mechanisms for depressive symptoms or positive affect in
EP

later adolescence (Kiuru et al., 2012; Prinstein, 2007; van Workum et al., 2013; van Zalk et
C

al., 2010). While present findings do not necessarily comport with prior studies, neither do
AC

they contradict them, as this research was conducted with adolescents at an earlier stage of

social development. The present objectively-assessed findings do harmonize in part with

those of Chan & Poulin (2009), who did not find that changes in self-reported network ties

affected subsequent depressive symptoms among 6th-graders.

29
 ACCEPTED MANUSCRIPT

There are several possible interpretations as to why study participants did not

demonstrate similarity in depressive symptoms nor self-esteem over time, despite

evidence in cross-sectional associations. First, children may not be fully aware of the

mental health status of others at so young an age. This is consistent with a developmental

PT
explanation that early in adolescence children are less influenced by peer sentiment, and

RI
that one’s perceptions of others’ opinions begin to play a stronger role in shaping mental

health beyond early adolescence. This explanation is also consistent with neurological

SC
development in that the youth in this study are at the cusp of shifts in social-affective

processing but not yet in the thick of these changes (Crone & Dahl, 2012). Second, it may be

U
that socialization shapes mental health status in ways unobservable from outward
AN
manifestations of social interactions. Third, features of the specific school studied may have
M
played a role in influencing the observed patterns. During the study collaboration, school

administrators and staff expressed that care for others was a key concern of the teaching
D

staff and a core value of the school culture. This care was clearly reflected in students’
TE

interactions with one another; students may thus have been socialized to interact with

others regardless of perceived mental health disposition. Despite this possibility, these data
EP

indicate that students nonetheless clearly discriminate in their social interaction patterns
C

and do make affiliation choices, largely on the basis of gender and popularity. Fourth, it
AC

may be that out-of-school interactions and other forms of school interaction not captured

by sensor network technology may also shape mental health.

Another possible explanation for a lack of social influence effects is that there was a

relatively low baseline level of distressed mental health in this population to begin with,

and less change over time in depressive symptoms than in self-esteem. However, model

30
 ACCEPTED MANUSCRIPT

specifications followed the process advocated by developers, and diagnostics suggested the

models were well-fit. Though we strove for complete participation, four of the 44 students

approached declined to participate. Sensitivity analyses were conducted that indicate non-

participating students are no different in terms of age or gender than their included peers.

PT
However, it may be that the connections those four students had with other 6th-grade peers

RI
were, in fact, relevant to the processes being examined here. These data do not allow us to

evaluate this possibility.

SC
It would be fruitful for the field to pursue objectively-based social network

enumeration approaches in community-based health research settings, not just among

U
children, but also in adult populations. RFID technologies are currently in their early stages
AN
of development, and more research is needed to cross-validate findings across social
M
settings, and types of devices. Additionally, the field would benefit from more research that

compares student self-reported networks with whom individuals actually socialize, and
D

that explores reasons for differences (Eagle et al., 2009; Smieszek et al., 2012). Given the
TE

rapid growth of computer-mediated lives, face-to-face encounters are but one form of

interaction that adolescents engage in on a daily basis. Future studies should examine the
EP

overlap between offline and online forms of interaction and their relevance for child health.
C

Children also likely have a variety of other social circles, including neighborhood and other
AC

extracurricular activity friendships that would benefit from scrutiny. As with other studies

that involve sensitive health information, detail about social interaction may represent a

special type of identifying information, and care should be taken that participant privacy

receives the highest priority.

31
 ACCEPTED MANUSCRIPT

We submit that the present study’s approach extends research on network

dynamics. For the questions investigated here involving the joint development of social

interaction and mental health, a SABM framework offers advantages over regression-based

techniques in terms of explicitly modeling the dependence of individuals. As a recent

PT
discussion of statistical social network methods makes clear (Christakis & Fowler, 2013),

RI
any choice of network modeling strategy makes its own assumptions, and this area is in its

infancy and enjoying a rapid burst of innovation. We have confidence that both regression

SC
and actor-based models will continue to develop with greater precision.

The present finding that one child’s mental health distress symptoms do not shape

U
another child’s subsequent symptoms may be of particular interest for parents and
AN
teachers of early adolescents. This study suggests that the onset of peer influence on
M
mental health is not especially strong in 6th grade, though signs of a social mechanism are

starting to emerge. While students are keenly aware of each other’s activities, fashions,
D

social status, and gossip of their everyday lives, early adolescent mental health does not yet
TE

appear to be affected by the mental health condition of socially-connected peers.

C EP
AC

32
 ACCEPTED MANUSCRIPT

5. References

Andersen, S.L., & Teicher, M.H. (2008). Stress, sensitive periods and maturational events in
adolescent depression. Trends Neurosci, 31, 183-191.
Bearman, P.S., & Moody, J. (2004). Suicide and friendships among American adolescents.
Am J Public Health, 94, 89-95.

PT
Blakemore, S.J., & Choudhury, S. (2006). Development of the adolescent brain: implications
for executive function and social cognition. J Child Psychol Psychiatry, 47, 296-312.
Brooks-Gunn, J., & Petersen, A.C. (1991). Studying the emergence of depression and

RI
depressive symptoms during adolescence. J Youth Adolesc, 20, 115-119.
Brown, B.B. (1990). Peer groups and peer cultures. In S.S. Feldman, & G.R. Elliott (Eds.), At
the threshold: The developing adolescent pp. 171-196): Harvard University Press.

SC
Bruening, M., Eisenberg, M., MacLehose, R., Nanney, M.S., Story, M., & Neumark-Sztainer, D.
(2012). Relationship between adolescents' and their friends' eating behaviors:
breakfast, fruit, vegetable, whole-grain, and dairy intake. J Acad Nutr Diet, 112,
1608-1613.

U
Butts, C. (2013). sna: Tools for Social Network Analysis, version 2.3-1.
Cairns, R.B., Leung, M.C., Buchanan, L., & Cairns, B.D. (1995). Friendships and social
AN
networks in childhood and adolescence: fluidity, reliability, and interrelations. Child
Development, 66, 1330-1345.
Cattuto, C., Van den Broeck, W., Barrat, A., Colizza, V., Pinton, J.F., & Vespignani, A. (2010).
M
Dynamics of Person-to-Person Interactions from Distributed RFID Sensor Networks.
PLoS One, 5, -.
Chan, A., & Poulin, F. (2007). Monthly changes in the composition of friendship networks in
D

early adolescence. Merrill-Palmer Quarterly-Journal of Developmental Psychology, 53,

578-602.
TE

Chan, A., & Poulin, F. (2009). Monthly Instability in Early Adolescent Friendship Networks
and Depressive Symptoms. Social Development, 18, 1-23.
Choudhury, S., McKinney, K.A., & Merten, M. (2012). Rebelling against the brain: public
engagement with the 'neurological adolescent'. Soc Sci Med, 74, 565-573.
EP

Christakis, N.A., & Fowler, J.H. (2013). Social contagion theory: examining dynamic social
networks and human behavior. Stat Med, 32, 556-577.
Clack, B., Dixon, J., & Tredoux, C. (2005). Eating together apart: Patterns of segregation in a
C

multi-ethnic cafeteria. Journal of Community & Applied Social Psychology, 15, 1-16.
Collins, R. (2004). Interaction ritual chains. Princeton, N.J.: Princeton University Press.
AC

Crone, E.A., & Dahl, R.E. (2012). Understanding adolescence as a period of social–affective
engagement and goal flexibility. Nature Reviews Neuroscience, 13, 636-650.
Csardi, G. (2013). igraph, version 0.6.5-2.
Dahl, R.E. (2004). Adolescent brain development: a period of vulnerabilities and
opportunities. Keynote address. Ann N Y Acad Sci, 1021, 1-22.
de la Haye, K., Robins, G., Mohr, P., & Wilson, C. (2013). Adolescents’ Intake of Junk Food:
Processes and Mechanisms Driving Consumption Similarities Among Friends.
Journal of Research on adolescence, 23, 524-536.
Dishion, T.J., & Tipsord, J.M. (2011). Peer contagion in child and adolescent social and
emotional development. Annu Rev Psychol, 62, 189-214.

33
 ACCEPTED MANUSCRIPT

Doherty, S.T., Lemieux, C.J., & Canally, C. (2014). Tracking Human Activity and Well-being in
Natural Environments Using Wearable Sensors and Experience Sampling. Social
Science & Medicine.
DuBois, D.L., Felner, R.D., Brand, S., & Phillips, R.S. (1996). Early adolescent self-esteem: A
developmental–ecological framework and assessment strategy. Journal of Research
on adolescence.
Dumont, M., & Provost, M.A. (1999). Resilience in adolescents: Protective role of social

PT
support, coping strategies, self-esteem, and social activities on experience of stress
and depression. J Youth Adolesc, 28, 343-363.
Eagle, N., Pentland, A.S., & Lazer, D. (2009). Inferring friendship network structure by using

RI
mobile phone data. Proc Natl Acad Sci U S A, 106, 15274-15278.
Eccles, J.S., Midgley, C., Wigfield, A., Buchanan, C.M., Reuman, D., Flanagan, C., et al. (1993).
Development during adolescence: the impact of stage-environment fit on young

SC
adolescents' experiences in schools and in families. American psychologist, 48, 90.
Falci, C., & McNeely, C. (2009). Too many friends: Social integration, network cohesion and
adolescent depressive symptoms. Social Forces, 87, 2031-2061.
Fergus, S., & Zimmerman, M.A. (2005). Adolescent resilience: a framework for

U
understanding healthy development in the face of risk. Annu Rev Public Health, 26,
399-419.
AN
Fitzgerald, A., Fitzgerald, N., & Aherne, C. (2012). Do peers matter? A review of peer and/or
friends’ influence on physical activity among American adolescents. Journal of
adolescence, 35, 941-958.
M
Forbes, E.E., & Dahl, R.E. (2010). Pubertal development and behavior: hormonal activation
of social and motivational tendencies. Brain and cognition, 72, 66-72.
Freeman, L.C. (1979). Centrality in social networks conceptual clarification. Social
D

Networks, 1, 215-239.
Gesell, S.B., Tesdahl, E., & Ruchman, E. (2012). The distribution of physical activity in an
TE

after-school friendship network. Pediatrics, 129, 1064-1071.

Gifford-Smith, M., Dodge, K.A., Dishion, T.J., & McCord, J. (2005). Peer influence in children
and adolescents: crossing the bridge from developmental to intervention science. J
Abnorm Child Psychol, 33, 255-265.
EP

Goodman, E., Adler, N.E., Kawachi, I., Frazier, A.L., Huang, B., & Colditz, G.A. (2001).
Adolescents' perceptions of social status: development and evaluation of a new
indicator. Pediatrics, 108, E31.
C

Gosliner, W., Madsen, K.A., Woodward‐Lopez, G., & Crawford, P.B. (2011). Would students
prefer to eat healthier foods at school? Journal of School Health, 81, 146-151.
AC

Grosbras, M.H., Jansen, M., Leonard, G., McIntosh, A., Osswald, K., Poulsen, C., et al. (2007).
Neural mechanisms of resistance to peer influence in early adolescence. J Neurosci,
27, 8040-8045.
Hatzenbuehler, M.L., McLaughlin, K.A., & Xuan, Z. (2012). Social networks and risk for
depressive symptoms in a national sample of sexual minority youth. Soc Sci Med, 75,
1184-1191.
Hergovich, A., Sirsch, U., & Felinger, M. (2002). Self-appraisals, actual appraisals and
reflected appraisals of preadolescent children. Social Behavior and Personality: an
international journal, 30, 603-611.

34
 ACCEPTED MANUSCRIPT

Hertzman, C., & Boyce, T. (2010). How experience gets under the skin to create gradients in
developmental health. Annu Rev Public Health, 31, 329-347 323p following 347.
Hinde, R.A. (1976). Interactions, relationships and social structure. Man, 1-17.
Hirsch, B.J., & Rapkin, B.D. (1987). The transition to junior high school: a longitudinal study
of self-esteem, psychological symptomatology, school life, and social support. Child
Development, 58, 1235-1243.
Isella, L., Romano, M., Barrat, A., Cattuto, C., Colizza, V., Van den Broeck, W., et al. (2011).

PT
Close Encounters in a Pediatric Ward: Measuring Face-to-Face Proximity and Mixing
Patterns with Wearable Sensors. PLoS One, 6.
Kingery, J.N., & Erdley, C.A. (2007). Peer experience as predictors of adjustment across the

RI
middle school transition. Education and Treatment of Children, 30, 73-88.
Kiuru, N., Burk, W.J., Laursen, B., Nurmi, J.E., & Salmela-Aro, K. (2012). Is depression
contagious? A test of alternative peer socialization mechanisms of depressive

SC
symptoms in adolescent peer networks. J Adolesc Health, 50, 250-255.
McCarroll, E.M., Lindsey, E.W., MacKinnon-Lewis, C., Chambers, J.C., & Frabutt, J.M. (2009).
Health status and peer relationships in early adolescence: The role of peer contact,
self-esteem, and social anxiety. Journal of Child and Family Studies, 18, 473-485.

U
McFarland, D.A. (2001). Student resistance: How the formal and informal organization of
classrooms facilitate everyday forms of student defiance1. American Journal of
Sociology, 107, 612-678.
AN
Parker, J.G., Rubin, K.H., Erath, S.A., Wojslawowicz, J.C., & Buskirk, A.A. (2006). Peer
relationships, child development, and adjustment: A developmental
M
psychopathology perspective. In D.J. Cohen (Ed.), Developmental psychopathology,
Theory and method pp. 419-493).
Poulin, F., & Chan, A. (2010). Friendship stability and change in childhood and adolescence.
D

Developmental Review, 30, 257-272.

Prinstein, M.J. (2007). Moderators of peer contagion: a longitudinal examination of
TE

depression socialization between adolescents and their best friends. J Clin Child
Adolesc Psychol, 36, 159-170.
Prinstein, M.J., Heilbron, N., Guerry, J.D., Franklin, J.C., Rancourt, D., Simon, V., et al. (2010).
Peer influence and nonsuicidal self injury: longitudinal results in community and
EP

clinically-referred adolescent samples. J Abnorm Child Psychol, 38, 669-682.

Resnick, M.D., Bearman, P.S., Blum, R.W., Bauman, K.E., Harris, K.M., Jones, J., et al. (1997).
Protecting adolescents from harm. Findings from the National Longitudinal Study on
C

Adolescent Health. JAMA, 278, 823-832.

Ripley, R., Boitmanis, K., & Snijders, T.A. (2013). RSiena: Siena - Simulation Investigation for
AC

Empirical Network Analysis, version. 1.1-241.

Salathe, M., Kazandjieva, M., Lee, J.W., Levis, P., Feldman, M.W., & Jones, J.H. (2010). A high-
resolution human contact network for infectious disease transmission. Proc Natl
Acad Sci U S A, 107, 22020-22025.
Salvy, S.J., de la Haye, K., Bowker, J.C., & Hermans, R.C. (2012). Influence of peers and
friends on children's and adolescents' eating and activity behaviors. Physiol Behav,
106, 369-378.
Samuels, S.E., Craypo, L., Boyle, M., Crawford, P.B., Yancey, A., & Flores, G. (2010). The
California Endowment's Healthy Eating, Active Communities program: a midpoint
review. Am J Public Health, 100, 2114-2123.

35
 ACCEPTED MANUSCRIPT

Schaefer, D.R., Kornienko, O., & Fox, A.M. (2011). Misery Does Not Love Company Network
Selection Mechanisms and Depression Homophily. American Sociological Review, 76,
764-785.
Sitarenios, G., & Kovacs, M. (1999). Use of the Children's Depression Inventory. The Use of
Psychological Testing for Treatment Planning and Outcome Assessment pp. 267-298).
Mahwah, NJ: Lawrence Erlbaum Associates Publishers.
Smieszek, T., Burri, E., Scherzinger, R., & Scholz, R.W. (2012). Collecting close-contact social

PT
mixing data with contact diaries: reporting errors and biases. Epidemiology and
infection, 140, 744-752.
Smith, K., & Christakis, N. (2008). Social networks and health. Annu. Rev. Sociol, 34, 405-

RI
429.
Snijders, T.A., Van de Bunt, G.G., & Steglich, C.E. (2010). Introduction to stochastic actor-
based models for network dynamics. Social Networks, 32, 44-60.

SC
Steglich, C., Snijders, T.A., & Pearson, M. (2010). Dynamic networks and behavior:
Separating selection from influence. Sociological Methodology, 40, 329-393.
Stehlé, J., Charbonnier, F., Picard, T., Cattuto, C., & Barrat, A. (2013). Gender homophily from
spatial behavior in a primary school: a sociometric study. Social Networks, 35, 604-

U
613.
Stehlé, J., Voirin, N., Barrat, A., Cattuto, C., Isella, L., Pinton, J.F., et al. (2011). High-Resolution
AN
Measurements of Face-to-Face Contact Patterns in a Primary School. PLoS One, 6.
Steinberg, L. (2005). Cognitive and affective development in adolescence. Trends Cogn Sci,
9, 69-74.
M
Steinberg, L., & Monahan, K.C. (2007). Age differences in resistance to peer influence.
Developmental Psychology, 43, 1531-1543.
Steinberg, L., & Morris, A.S. (2001). Adolescent development. Annu Rev Psychol, 52, 83-110.
D

Ueno, K. (2005). The effects of friendship networks on adolescent depressive symptoms.

Social Science Research, 34, 484-510.
TE

Umberson, D., Crosnoe, R., & Reczek, C. (2010). Social Relationships and Health Behavior
Across the Life Course. Annual Review of Sociology, Vol 36, 36, 139-157.
Vahdatpour, A., Amini, N., Xu, W., & Sarrafzadeh, M. (2011). Accelerometer-based on-body
sensor localization for health and medical monitoring applications. Pervasive Mob
EP

Comput, 7, 746-760.
van Workum, N., Scholte, R.H., Cillessen, A.H., Lodder, G.M., & Giletta, M. (2013). Selection,
Deselection, and Socialization Processes of Happiness in Adolescent Friendship
C

Networks. Journal of Research on adolescence, In press.

van Zalk, M.H., Kerr, M., Branje, S.J., Stattin, H., & Meeus, W.H. (2010). Peer contagion and
AC

adolescent depression: the role of failure anticipation. J Clin Child Adolesc Psychol,
39, 837-848.
Wasserman, S., & Faust, K. (1994). Social network analysis : methods and applications.
Cambridge ; New York: Cambridge University Press.
Welk, G.J., McClain, J., & Ainsworth, B.E. (2012). Protocols for evaluating equivalency of
accelerometry-based activity monitors. Med Sci Sports Exerc, 44, S39-49.
Woodward-Lopez, G., Gosliner, W., Samuels, S.E., Craypo, L., Kao, J., & Crawford, P.B. (2010).
Lessons learned from evaluations of California's statewide school nutrition
standards. Am J Public Health, 100, 2137-2145.

36
 ACCEPTED MANUSCRIPT

Table 1. Participant characteristics, by student gender

Period 1 Period 2 Period 3
Male (n=15) Female (n=25) Male (n=15) Female (n=25) Male (n=15) Female (n=25)
Med IQR Med IQR p-val Med IQR Med IQR p-val Med IQR Med IQR p-val
Demographic

PT
Age 12.2 (0.6) 12.3 (0.3) 0.65 - - - - - - - -
Race/ethnicity (Obs, %) 0.42
Not European-American 3 20% 8 32% - - - - - - - -

RI
European-American 12 80% 17 68% - - - - - - - -
Parent Education (Obs, %) 0.80
One or both went to college 6 40% 10 40% - - - - - - - -

SC
One or both have adv. Degrees 8 53% 12 48% - - - - - - - -
Don't know 1 7% 3 12% - - - - - - - -
Med IQR Med IQR p-val Med IQR Med IQR p-val Med IQR Med IQR p-val

U
Social Environment
School Social Status Ladder 9.0 (3.0) 8.0 (3.0) 0.012** 8.5 (1.5) 7.5 (1.0) 0.03** 8.5 (1) 8 (2) 0.09*

AN
Peer popularity nominations 2 (12) 1 (6) 0.97 2 (14) 1 (7) 0.81 2 (11) 2 (7) 0.56
School connectedness 19 (2) 19 (4) 0.31 20 (1) 20 (3) 0.45 20 (2) 20 (3) 0.65
Health behaviors
Steps (average freq./day) 7482 (1663) 5828 (778) <0.001*** 6473 (1514) 4864 (574) <0.001*** 7629 (2740) 6166 (755) 0.002***

M
Vigorous act. (avg. min/day) 57 (22) 40 (8) <0.001*** 37 (17) 25 (6) <0.001*** 40 (23) 29 (10) 0.003***
Healthier food 7 (3) 7 (3) 0.62 7 (3) 8 (3) 0.86 7.5 (4) 7 (4) 0.68
Unhealthier food 1 (2) 2 (2) 0.24 2 (1) 2 (2) 0.32 1.5 (2) 2 (2) 0.46

D
Mental health symptoms
Childhood depression inventory (1-20) 1 (2) 1 (3) 0.18
TE 0 (1) 0 (2) 0.22 0 0 0 (2) 0.12
Self-esteem (1-32) 29.0 (4.5) 27.5 (5.0) 0.08* 28.5 (4.0) 29.0 (6.8) 0.73 30.0 (2.0) 29.0 (6.0) 0.46

Note: p-value denotes significance level of bivariate association between gender and covariates. Chi-squared test used for categorical covariates (race/ethnicity,
EP

parent education). A nonparametric Mann-Whitney U test was used to assess group difference significance between gender and continuous variables with
C
AC
 ACCEPTED MANUSCRIPT

Table 2. Correlations between network indicators and mental health

CDI (P1) SE (P1)

M (n=15) F (n=25) M (n=15) F (n=25)
Network size (# of alters) 0.15 -0.21 0.06 0.37
Transitivity 0.17 0.03 -0.20 0.32
Closeness Centr. 0.18 -0.22 0.06 0.33

PT
CDI (P2) SE (P2)
M (n=14) F (n=25) M (n=14) F (n=24)
Network size (# of alters) 0.66* -0.47* -0.47 0.19

RI
Transitivity -0.33 0.65* 0.46 -0.31
Closeness Centr. 0.61 -0.56* -0.42 0.30

CDI (P3) SE (P3)

SC
M (n=14) F (n=25) M (n=14) F (n=25)
Network size (# of alters) 0.24 -0.56* -0.19 0.52*
Transitivity -0.04 0.22 0.25 -0.19
Closeness Centr. 0.30 -0.54* -0.35 0.51*

U
Note: CDI (Childhood Depression Inventory); SE (Self-esteem scale). Pairwise
correlation significance, * p<0.05.

AN
M
D
TE
C EP
AC
 ACCEPTED MANUSCRIPT

Table 3. Social interaction and self-esteem co-evolution

Self-esteem (final model)
PE SE p
Network dynamics
Rate of network change P1 to P2 15.33 (2.51) (na)
Rate of network change P2 to P3 19.78 (5.41) (na)
Structural effects
Degree (density) -1.09 (0.13) ***

PT
Transitive triads 0.13 (0.03) ***
Transitive ties 0.27 (0.14) *
Demographic & Social environment
Gender (ego's) -0.20 (0.09) **

RI
Gender (same) 0.49 (0.07) ***
Perceived popularity (ego's) -0.01 (0.05)
Perceived popularity similarity 0.24 (0.09) ***

SC
Self-esteem (ego's) 0.01 (0.05)
Self-esteem similarity -0.03 (0.21)
Behavior dynamics
Rate of behavior change P1 to P2 3.09 (1.30) (na)
Rate of behavior change P2 to P3 1.58 (0.58) (na)

U
Behavior linear shape -0.24 (0.15)
Behavior quadratic shape -0.30 (0.12) **
Attribute effects on behavior
Race/ethnicity
School connectedness
1.07
0.97
(0.46) **
(0.41) **
AN
M
Note: Parameter Estimate (PE), Standard Error (SE). Parameter estimates divided
by standard errors yield t-values for each parameter; t-values greater than 1.65
(p<0.1 *), greater than 1.96 (p<0.05 **), greater than 2.58 (p<0.01***)
D
TE
C EP
AC
 ACCEPTED MANUSCRIPT

Table 4. Social interaction and depressive symptom co-evolution

Dep. Symp. (final model)
PE SE p
Network dynamics
Rate of network change P1 to P2 15.28 (2.92) (na)
Rate of network change P2 to P3 19.48 (3.22) (na)
Structural effects

PT
Degree (density) -1.07 (0.13) ***
Transitive triads 0.13 (0.03) ***
Transitive ties 0.25 (0.14) *
Demographic & social environment

RI
Gender (ego's) -0.20 (0.09) **
Gender (same) 0.49 (0.08) ***
Perceived popularity (ego's) -0.01 (0.06)

SC
Perceived popularity similarity 0.23 (0.09) **
Depressive symptoms (ego's) 0.03 (0.06)
Depressive symptom similarity 0.23 (0.44)
Behavior dynamics

U
Rate of behavior change P1 to P2 3.51 (1.94) (na)
Rate of behavior change P2 to P3 4.57 (2.12) (na)
Behavior linear shape -2.80 (2.05)
Behavior quadratic shape
Average depressive similarity
-0.01
-32.29
(0.08)
(31.49)
AN
Attribute effects on behavior
Perceived popularity -0.41 (0.34)
M
Steps 0.21 (0.19)
Vigorous Activity 0.23 (0.19)
Note: Parameter Estimate (PE), Standard Error (SE). Parameter estimates divided
by standard errors yield t-values for each parameter; t-values greater than 1.65
D

(p<0.1 *), greater than 1.96 (p<0.05 **), greater than 2.58 (p<0.01***)
TE
C EP
AC
 ACCEPTED MANUSCRIPT

Gender
Female
Male

PT
RI
Day 1 Day 2 Day 3 Period 1

U SC
AN
M
D
Day 4
TE
Day 5 Day 6
EP
Period 2
C
AC

Day 7 Day 8 Day 9 Period 3

 ACCEPTED MANUSCRIPT

Mental health and social networks in early adolescence: A dynamic study of

objectively-measured social interaction behaviors

[Research highlights]

• Social interaction is associated with mental health status in early adolescence.

PT
• Girls high in self-esteem tend towards greater network social integration.

• Girls with depressive symptoms are more socially inhibited than boys with

RI
symptoms.

• Social influence does not shape self-esteem or depression at this age.

SC
• Electronic methods for network data-collection are useful in a school setting.

U
AN
M
D
TE
C EP
AC
 ACCEPTED MANUSCRIPT
Supplementary Data. “Mental health and social networks in early adolescence: A dynamic study of objectively-measured social
interaction behaviors”

1. Comparison of RFID-measured interactions with self-report interactions.

It is informative to assess the extent to which electronic measurement of social

interactions comports with student self-report of friendships. A comparison of RFID-measured
interactions at different thresholds (20 seconds, 40s, 60s, 80s, 100s, 120s) with student self-
report of friendship reveals a range of similarities in network relationship patterns. To obtain
self-report friendship data, students were asked to nominate up to five close friends. The

PT
question, which was accompanied by a grade-level roster of students, asked:

“Who are your five closest friends in your grade? (First & last names) Remember, your

RI
answers will not be seen by other students – this is just for researchers, no one in the
school or your family will see what you write.”

SC
The strongest similarities between self-report friendship and RFID-measured interactions appear
above the 60s interaction duration threshold. We decided to use the 80-second threshold for
RFID network enumeration because across all three observation periods, it appeared to have the
greatest overlap with student self-report on the basis of a number of metrics (% overlap of dyads

U
in each network, density, mean degree). We evaluated alternative depressive symptoms and self-
esteem models with 60s-minimum networks with no significant difference in results. As an
AN
example, Figure A1 depicts both self-report and RFID-measured networks; at this interaction
duration, 81% of dyads are identical in both networks. The density of self-report network (a) is
0.18 with an average of 7.2 alters, and the density of RFID network (b) is a very similar 0.17,
with individuals having an average of 6.6 alters.
M
D
TE
C EP
AC

Figure A1. Self-report versus RFID-measured interactions, Period 1

S1
 ACCEPTED MANUSCRIPT
Supplementary Data. “Mental health and social networks in early adolescence: A dynamic study of objectively-measured social
interaction behaviors”

2. Cross-sectional consistency over time of mental health across domains.

While the main body text described significant differences in mental health indicators by
gender, examination of the relationships between these indicators with other covariates relating
to social environment, health behaviors, and socio-demographic background is also informative.
Table A1 reports pairwise correlations between our two measures of mental health status
(depressive symptoms, self-esteem scale) with social environment and health behavior covariates

PT
in order to describe the distribution of mental health over time in the cohort. In terms of
depressive symptoms, the majority of the measures of children’s social environment are
negatively associated with the depression inventory at all three periods, suggesting that children

RI
with higher self-reported social status, peer-reported popularity, and self-reported connectedness
to school tend to be lower in depressive symptoms. Only school connectedness and subjective
social status are statistically significant. While other health behaviors (total steps,

SC
healthy/unhealthy food consumption) have mixed, and non-significant associations with
depressive symptoms, average duration of vigorous activity is negatively associated at all three
periods, suggesting that more physically active students are lower in depressive symptoms.
Self-esteem has relatively consistent and positive associations with measures of the social

U
environment, though associations do not attain statistical significance at all three periods.
Children who are higher in self-esteem tend to be higher in subjective social status, peer-reported
AN
popularity, and feelings of connectedness to the school. As with depressive symptoms,
associations between other health behaviors and self-esteem are relatively inconsistent. Strong
and statistically significant correlations between depression and self-esteem suggest that these
measures tap overlapping, yet not identical, dimensions of mental health. At all three periods,
M
there is an inverse association suggesting that children higher in self-esteem tend to be lower in
depression, a trend especially strong at the third period.
D

Table A1. Social environment, health behaviors, and mental health indicator correlations
TE

T1 T2 T3
CDI SE CDI SE CDI SE
Social Environment
Subj. social status -0.26 0.35 * -0.08 0.37 * -0.35 * 0.31
EP

Peer popularity 0.03 0.22 -0.01 0.32 * -0.16 0.25

Sch. connect. -0.32 * 0.55 * -0.15 0.25 -0.53 * 0.67 *
Health behaviors
C

Total steps/day -0.15 0.09 0.04 -0.03 -0.07 0.10

Vig. activity -0.14 0.18 -0.17 0.13 -0.14 0.10
AC

Healthier foods -0.18 0.01 0.24 -0.21 0.07 -0.09

Unhealthier foods 0.26 -0.17 0.10 -0.15 0.01 0.00
Mental Health
Depression - -0.66 * - -0.63 * - -0.86 *
Self-esteem -0.66 * - -0.63 * - -0.86 * -
Note: continuous associations given by Pearson correlation (rho=). CDI= childhood depression inventory, SE = self-esteem scale. Pairwise
correlation significance, * p<0.05.

S2
 ACCEPTED MANUSCRIPT
Supplementary Data. “Mental health and social networks in early adolescence: A dynamic study of objectively-measured social
interaction behaviors”

It is also informative to describe bivariate associations between race/ethnicity and parent

education, with depressive symptoms and self-esteem (Table A2). With respect to depressive
symptoms (CDI), though the median CDI is very low across both categories, non-whites are
higher in depressive symptoms at all three periods. There is no significant depressive symptom
difference by parent education. In terms of self-esteem, non-white students are again
consistently, and significantly lower on this mental health indicator. There are no significant self-
esteem differences by parent education.

PT
Table A2. Bivariate associations between race, parent education, & mental health indicators
CDI (P1) CDI (P2) CDI (P3)

RI
Med IQR p val Med IQR p val Med IQR p val
Race/ethnicity 0.01*** 0.004*** 0.01***
Not European-American 2 (6.0) 2 (4.0) 2 (6.0)

SC
European-American 1 (1.0) 0 (1.0) 0 (1.0)
Parent education 0.59 0.23 0.88
One/both went to college 1 (2.5) 0.5 (1.5) 0 (1.0)
One/both have adv. degrees 1 (2.5) 0 (1.5) 0 (2.0)

U
Don't know 0.5 (1.5) 0 0.0 0 (1.0)

SE (P1)
Med IQR
AN p val
SE (P2)
Med IQR p val
SE (P3)
Med IQR p val
Race/ethnicity 0.02** 0.04** 0.01***
M
Not European-American 25 (8.5) 25 (5.0) 26 (4.0)
European-American 28 (4.3) 29 (4.0) 30.3 (3.0)
Parent education 0.58 0.74 0.57
D

One/both went to college 28 (3.0) 29 (9.0) 29 (3.8)

One/both have adv. degrees 27.8 (4.5) 28 (5.0) 30 (6.0)
TE

Don't know 29.25 (3.8) 30 (4.8) 31 (4.5)

Note: CDI= childhood depression inventory, SE = self-esteem scale. The Mann-Whitney U test was used to assess significance of race/ethnicity
differences between depressive symptoms, self-esteem. The Kruskal-Wallis test was used for parent education.
C EP
AC

S3
 ACCEPTED MANUSCRIPT
Supplementary Data. “Mental health and social networks in early adolescence: A dynamic study of objectively-measured social
interaction behaviors”

3. Network and behavior change.

It is useful to examine how changes in additional aspects of network structure over time
are simultaneously occurring with changes in depressive symptoms and self-esteem. Table A3
describes descriptive network characteristics, and finds that networks are consistently sparse.
The average weekly degree statistic ranges from roughly 6.6 to 7.1 throughout the week,
suggesting that students interact with a number of different people over the course of a week

PT
beyond the same few adjacent tablemates. The Jaccard coefficients report network similarity
between the three periods, and confirms a large degree of change in student socialization; there is
roughly 30% overlap between the first two periods, and only 20% overlap in network structure

RI
between the last two periods. Edge change counts show that of participants who changed
interaction partners over time, students tended to make more ties than break them during the first
two periods, while more ties were broken during the last two periods. This suggests constant

SC
change rather than a consolidation of social cohesion during the second half of the 6th-grade year.
Changes in mental health outcome variables indicate a great deal of stability in
depressive symptoms, and that between the first two periods more children decreased their score,
though between the latter two periods, an equal amount decreased as did increase. Between both

U
periods, self-esteem increased at a steady rate, while there were fewer changes towards low self-
esteem in later periods. An important point to mention is that models require there to be
AN
sufficient change in the outcome behavior during the term of the study. Though there are no
strict cutoff thresholds for appropriate amount of behavior change, guidance from model
developers (Snijders et al., 2010) suggests that the amount of change in depressive symptoms
may be on the low side (31 changes), while a bit higher for self-esteem (40 changes). As they
M
write, “More changes will give more information and, thereby, allow more complicated models
to be fitted”(p.49). However, the low complexity of our models, excellent subsequent model
convergence (t-ratios all below 0.10 for all parameters), model goodness-of-fit statistics, and
D

careful forward covariate selection process to build a parsimonious set of predictors, all suggest
that these models are appropriately parameterized.
TE
C EP
AC

S4
 ACCEPTED MANUSCRIPT
Supplementary Data. “Mental health and social networks in early adolescence: A dynamic study of objectively-measured social
interaction behaviors”

Table A3. Change in 6th-grade networks and mental health

Structural attributes P1 P2 P3
Density 0.17 0.18 0.17
Avg. degree (weekly) 6.6 7.1 6.8
Number of ties 131 141 136
Similarity w/prior wave (Jaccard) - 0.27 0.20

PT
Edge changes P1/P2 P2/P3
0==>0 565 549
0==>1 84 90

RI
1==>0 74 95
1==>1 57 46

SC
Depressive symptom changes P1/P2 P2/P3
down 17 6
up 2 6
constant 21 27

U
missing 0 1

Self-esteem changes
down
AN P1/P2
15
P2/P3
7
up 9 9
constant 13 21
M
missing 3 3
D

4. SABM estimation considerations and covariate transformations.

TE

Several aspects of SABM estimation warrant additional detail, including covariate

selection, transformation, and tie definition.
EP

Covariate selection
We used a forward model selection procedure advocated by developers, in which
individual covariates, or groups of covariates, are added and their single or joint significance is
C

tested. Covariates that are statistically significant are retained for the next stage of model
evaluation. For the network formation component of the model, we tested a large range of
AC

covariates, including socio-demographic attributes (ego’s gender, parental education, race),

social environment attributes (ego’s school connectedness, ego’s subjective social status, ego’s
peer-reported popularity); and health behaviors (ego’s healthy eating, unhealthy eating, total
steps, vigorous activity). We also tested effects to account for attribute homophily between ego
and alter. In the case of categorical variables we tested same-status (gender, parent education,
race/ethnicity) attributes. For continuous predictors, we tested attribute similarity (school
connectedness similarity, subjective social status similarity, peer popularity similarity, healthy
eating similarity, unhealthy eating similarity, total step similarity, vigorous activity similarity).
Relevant network parameters evaluated relate to network structure and endogenous network
processes (degree, transitive triads, transitive ties, preferential attachment, cohesion).

S5
 ACCEPTED MANUSCRIPT
Supplementary Data. “Mental health and social networks in early adolescence: A dynamic study of objectively-measured social
interaction behaviors”

In the behavior outcome component of the models, we test several parameters. A key
predictor of each mental health outcome includes the likelihood that a child becomes more
similar over time to the prior average mental health status of connected peers (average
similarity); this parameter indicates evidence of peer influence. Other predictors tested include
mental health behavior evolution (linear shape, quadratic shape), effect of alter’s number of ties
(alter popularity), and the range of ego’s abovementioned demographic, social environment, and
health behavior effects on behavior. Additional analyses tested a range of social influence effects

PT
(i.e. average alter, total similarity).
Of this range of covariates evaluated for network formation and behavior change we
retained the covariates identified as significant (at p<0.05) during the forward model selection

RI
process. Though we largely follow developer guidance that score test significance should be
below p<0.05, others (e.g. De la Haye et al., 2013) have used lower thresholds (i.e. p<0.10).
While nearly all covariates were significant below 0.05, two covariates were significant at the

SC
0.10 threshold. Perceived popularity, a behavioral predictor in the self-esteem/friendship model
(Body text, Table 3) was significant at 0.08, and so we included it. Vigorous physical activity, a
behavioral predictor in the depression/friendship model (Body Text, Table 4), was significant at
0.06, and we also included it. Although they were not significant in score-type tests, we also

U
retained depressive symptoms and self-esteem as key constructs under investigation.

Covariate transformation AN
Due to the requirements of the modeling framework, mental health outcome variables
need to be specified as a limited set of ordinal integers. Because the summary scale of depressive
symptoms ranges from 5 to 11 (7 levels), it is appropriate to leave it untransformed. However,
M
because the distribution of the self-esteem summary scale ranges from 16 to 32 (16 levels), it
needs to be simplified to a more tractable set of ordinal integers (Snijders et al., 2010). For our
purposes, separating self-esteem into quintiles (5 levels) seemed appropriate because it offers a
D

trade-off between not oversimplifying (assuring an adequate number of changes over time for
model estimation) while preserving the character of the data. We also examined quartiles (4
TE

levels) and septiles (6 levels); results did not differ significantly from our primary results.
Though there is some debate as to the optimal number of levels of a given dependent variable,
guidance from developers (Snijders et al., 2010) and the online community (stocnet) suggest that
EP

less than 10 is acceptable, as long as estimation diagnostics show a well-fit model. With both
mental health and self-esteem models, goodness-of-fit tests showed an excellent fit in terms of
the standard degree, geodesics, and triad census.
It was also necessary to transform several of the continuous social environment and
C

health behavior covariates specified as predictors. To do so, we examined the distribution of each
within a given period and chose a partitioning scheme that would ensure adequate variation in
AC

categories. For physical activity variables, because boys and girls are so different in terms of
physical activity, we calculated the distribution of total steps and vigorous activity separately for
boys and girls, and calculated gender-specific PA quintiles so we could be sure that a significant
PA effect was not driven by gender. Since there were not appreciable gender differences in other
covariates, we transformed eating covariates (healthy food, unhealthy food) into quantiles,
school connectedness into a binary outcome (above/below the period-specific mean), and peer-
perceived popularity into tertiles. In the final model specifications guided by the forward model
selection process, only peer-perceived popularity and school connectedness were included, the

S6
 ACCEPTED MANUSCRIPT
Supplementary Data. “Mental health and social networks in early adolescence: A dynamic study of objectively-measured social
interaction behaviors”

latter in the depression model (Body text, Table 4) and the former in the self-esteem model
(Body text, Table 3).

Tie definition
In this model framework, adjacency matrices consisting of a binary “1” (for existence of
a non-directed tie between two individuals) or “0” (for absence of an non-directed tie) are
constructed for each panel of available data. The high-resolution longitudinal data during each

PT
day’s collection period (i.e. measured every 20 seconds) are assessed, and filtered to only retain
ties greater than or equal to a daily threshold of 80 seconds’ duration (for reasons specified above
in Section 1). A given day’s 80+ second ties are then aggregated with the next two days’ 80+

RI
second ties to obtain a cumulative 3-day weighted network, which represents a given Period’s
binary matrix. The outcome variables are the joint likelihood of forming a tie, with the given
mental health condition, and predictor variables include structural network properties as well as

SC
the range of individual-level covariates described above.
Current RFID technology did not permit us to observe directionality in social ties.
Although collecting social tie data every 20 seconds allows us to understand rapid changes in
social structure, advances in this technology are necessary to discern whether Student ‘A’

U
approached Student ‘B’, or vice versa. Hence, model estimations used the “unilateral initiative
and reciprocal confirmation” non-directed network option (i.e. modelType=3) available in
AN
RSiena (Ripley, Snijders, Preciado, RSiena Manual, Version August 27, 2013). Under this set of
assumptions, “one actor takes the initiative and proposes a new tie or dissolves an existing tie; if
the actor proposes a new tie, the other has to confirm, otherwise the tie is not created; for
dissolution, confirmation is not required”(p.75). Of the range of available options, this seemed to
M
most closely match the social situation of a school lunchroom. For example, a given student who
approaches a table to sit down with others will likely not do so if that student receives strong
indicators from already-seated students not to sit there. Thus, the act of joining a table is
D

equivalent to proposing a new tie, and those seated must confirm by consenting that the joiner
may be seated. Likewise, a student does not need to obtain approval to leave the table and
TE

dissolve the tie; others’ confirmation is not required.

C EP
AC

S7