The Small School Friendship Dynamics of Adolescent Depressive Symptoms

Abstract

Adolescence is a time when depressive symptoms and friendships both intensify. We ask whether friendships change in response to depressive symptoms, whether individual distress is influenced by friends’ distress, and whether these processes vary by gender. To answer these questions we use longitudinal Simulation Investigation for Empirical Network Analysis (SIENA) models to study how changes in friendships and depressive symptoms intertwine with each other among all adolescents, boy-only, and female-only networks in seven smaller K-12th grade Add Health schools. Findings indicate that distressed youth are more likely to be socially excluded, though depressive symptoms are also a basis for friendship formation. Moreover, friends influence each other’s mood levels. These processes differ for boys and girls, however, such that distressed girls are more likely to face exclusion and distressed boys are more likely to befriend and subsequently influence each other. Differences in these processes have implications for intervention efforts since the joint selectivity-influence mechanisms may undercut intervention efforts.

Adolescence is an emotionally tumultuous time when depressive symptoms increase (Wade et al. 2002), which correspondingly elevates risks for more severe forms of distress like suicide ideation and psychological disorders (Wilson and Deane 2010; Pine et al. 1999). Approximately 35% of adolescents experience depressive symptoms (Compas 1993) and findings by gender indicate that rates of depression are 2 or 3 times higher for girls than boys (Nolen-Hoeksema and Girgus 1994). Adolescence is also a very social period when interpersonal connections to peers and friends become central to young peoples’ lives (Crosnoe 2011). Given the co-emergence of emotional distress and increased socializing over this period of the early life course, we ask whether peer interactions amplify or mollify emotional distress. Studies of adolescent mental health have focused on family factors for years (e.g., Avison and McAlpine 1992), but less attention is given both to how emotional distress is embedded in friendships and how those mechanisms are gendered (Rubin, Bukowski, Parker 2006).

The way that distress is socially intertwined with the interpersonal lives of adolescents has important consequences for ameliorative efforts depending upon the particular social mechanisms at play. For example, distressed adolescents may withdraw from others or be viewed more negatively by them, with the result that others withdraw from them (Schaefer et al. 2011; Rose et al. 2007). Additionally, distress may diffuse among friends, or be a basis for friendships (e.g., Kandel 1978; Weerman 2011). Consequently, interventions addressing emotional distress will benefit from different strategies depending upon the specific social mechanisms at work (see Bauman and Ennett 1996). Because youth cultures are typically gendered based upon broader societal norms and subsequent socialization efforts by parents, adults (Rhodes and Lowe 2009; West and Zimmerman 1987), and educational institutions (Eder and Parker 1987), preventing distress may require special attention to gender-specific friendship dynamics.

In order to address these issues we interlink individual changes in depressive symptoms to the longitudinal friendship networks in seven small (N<300) Add Health high schools. Our focus on these small schools reflects the circumscribed nature of the social networks in these settings. Size places constraints on knowledge about other peers, and this structural constraint increases quickly with school size to the point where most students do not know each other and have little knowledge about each other’s affective states and tendencies. In addition, prior studies suggest that emotional distress processes differ among boys and girls (Rudolph 2002), so we assess both the social dynamics of distress among all adolescents within schools, and among boys and girls separately. We therefore evaluate multiple mechanisms connecting depressive symptoms and social relations to each other during adolescence using new longitudinal social network models (Snijders et al.’s [2007] SIENA model). This approach allows us to study both friend selection and influence mechanisms directly while simultaneously accounting for structural network and other background factors.

Literature Review

Most research on adolescent social relationships and distress focuses on social support (e.g., Cornwell 2003; Turner and Turner 1999) and utilizes biased adolescent perceptions rather than actual social network data and direct friend reports (e.g., Beam et al. 2002). Though much attention has been given to the social dynamics of externalizing behaviors (e.g., Mercken et al. 2010; Pearson et al. 2006), there is a growing, albeit relatively small, social network-based sociological literature studying adolescent psychological outcomes. This research largely reflects the sociological legacy of research on social integration (Durkheim 1966) and the emphasis on network structures in the social network tradition (Wellman 1988).

Recently, Falci and McNeely (2009) reported that Add Health adolescents with either very small or very large personal networks have more depressive symptoms. Moreover, their research indicates important gender differences. Specifically, the ill effects of over-integration only occurred at low levels of network cohesion for girls, while the consequences of over-integration only occurred at high levels of network cohesion for boys. Their study builds off of Ueno (2005) and Hansell’s (1985) earlier examinations of social integration, which together showed that a linear operationalization of integration had only a weak association with depressive symptoms.

We focus here on two limitations of this prior work. First, the network measures used (i.e., Ueno 2005; Falci and McNeely 2009) did not reflect the actual depressive symptoms of friends and consequently did not shed light on social influence processes. Psychologists, however, argue that having distressed friends increases individual risk for distress (Hogue and Steinburg 1995; Stevens and Prinstein 2005). Second, the issue of friend selection has not been addressed directly in studies of social influence. Selectivity is a generally acknowledged challenge for peer influence studies (e.g., Haynie 2001) since the outcomes researchers and others care about can be sources of friendship rather than merely consequences of them (Cohen 1977). Yet despite concerns that social selection processes bias peer effect estimates (e.g., Billy and Udry 1985), and thus expectations for interventions utilizing them (Ali and Dwyer 2010), no studies of adolescent distress – or social integration –directly incorporate friend selection into models estimating the mutual influences friends have on one another. However, our purpose is not merely to isolate social influence effects by controlling for selection, but also to begin studying the role of distress in friendship processes directly (e.g., House et al. 1988:308; see also Hartwell and Benson 2007; Schaefer et al. 2011).

Social Influence

We use social influence to refer to the mechanism by which friends influence each other’s emotional distress interpersonally (Friedkin 1998; see also Umberson, Crosnoe, and Reczek 2010). Interactional views of depression describe a social influence process through which depressed affect spreads via social interaction (Coyne 1976a,b; Prinstein 2007). Heider (1958), for example, suggests that individuals try to find emotional and other forms of balance with their social relations (see also Prinstein et al. 2005). Moreover, friends influence and perpetuate loneliness (Cacioppo et al. 2009) and both positive and negative mood in one another (Fowler and Christakis 2008; Hogue and Steinberg 1995; Stevens and Prinstein 2005), suggesting that assimilation in distress is similar to that for some externalizing behaviors (Pearson, Steglich, and Snijders 2006). We thus hypothesize that adolescent depressive symptoms will change over time to become more similar among friends.

Evidence, however, is mixed regarding sex differences in social influence (Johnson 1991; Hogue and Steinberg 1995; see also Giordano 2003). Hogue and Steinburg (1995), for example, suggest that boys’ distress is more susceptible to friend influences, while Stevens and Prinstein (2005; see also Prinstein et al. 2005) report the opposite. More generally, some studies show that boys are more vulnerable to peer influences than girls for certain types of antisocial behaviors (Brown et al. 1986) and other internalizing outcomes (Prinstein 2007). To the extent that relationships among girls are more intensive and require more emotion work to provide empathy (e.g., Hochschild 1979), distress may more strongly diffuse among them. Alternatively, Hochschild’s (1990) notion of emotional suppression suggests that influence may be greater among boys if girls are able to provide support for each other (Vaughan et al. 2010) while boys are unable or unwilling to because they tend to suppress and hide their distress for fear of social exclusion (Johnson 1991). Though the evidence on gender differences in social influence are mixed, we hypothesize that social influence will be stronger among boys than girls given the greater supportiveness and expressiveness of female relationships.

Friend Selection

Fine (1980) argues that friendships cannot form and be maintained without three factors: (1) structural constraints conducive to ongoing interactions, such as the social contexts provided by schools; (2) individual inclinations that allow friendships to form; and (3) the existence of satisfying interactions. For example, youth who share characteristics are more likely to become friends (Goodreau et al. 2009; Shrum et al. 1988) and same sex friendship groups are more common than mixed-sex groups in adolescence (Cairns et al. 1998; Shrum et al. 1988). The degree to which students form relationships with each other thus reflects a variety of social and interpersonal mechanisms drawing youth together and influencing the quality of their relationships. We focus here on three specific ways that emotional distress can influence friend selection in general, and among boys and girls differentially. These mechanisms include (1) homophilous selection, (2) social exclusion, and (3) social withdrawal (c.f. Brown and Larson 2009).

Homophilous selection is a mechanism reflecting the idea that “birds of a feather flock together” (i.e. McPherson et al. 2001), with the associated hypothesis that adolescents with similar levels of depressive symptoms are more likely to be friends (e.g., Billy, Rogers and Udry 1984; Hallinan and Kubitschek 1990). There is evidence that psychological predisposition plays an important role in social network formation and structure (Kalish and Robbins 2006) and this is true both among adolescents (Hogue and Steinberg 1995) and adults (Wenzlaff and Prohaska 1989; Rosenblatt and Greenberg 1988; Merikangas 1984). Moreover, adolescents appear to select friends with similar substance use and aggression (e.g., Cairns et al. 1988). Thus, emotional distress may be a source of interpersonal bonding, though one recent study raises doubts about the importance of this mechanism (Schaefer et al. 2011).

Two additional friendship mechanisms have been hypothesized. The first, social exclusion, captures the idea that distressed individuals are pushed out to the periphery of the friend network and end up socially marginalized (Bendgren 2002; Coyne and Downey 1991; Cacioppo et al. 2006; Link et al. 1989). This can happen because distressed youth express more aversive behavior and lower reciprocity, thereby requiring more maintenance effort and increasing the likelihood that they are social excluded by others as a result (Coyne 1976a,b; Bendgren et al. 2002; Youngren and Lewinsohn 1980; Coyne and Bolger 1990). The other mechanism, social withdrawal, may also lead to social isolation or marginalization (Crosnoe et al. 2008). While healthier individuals have larger networks and are more likely to actively seek support from friends, depressed youth may withdraw socially to cope with depressive symptoms or because they perceive themselves to be socially excluded (Schaefer et al. 2011; Rose et al. 2007). These two mechanisms lead us to hypothesize that more distressed adolescents will be less popular and that more distressed adolescents will consider fewer peers to be their close friends. Determining the differential contributions between these and the homophilous selection mechanism is one of the central goals of this study.

Research on gender differences in interpersonal social processes (Deaux and Major 1987) and the influence of friendship networks on distress also suggests that homophilous selection processes may differ between boys and girls (Falci and McNeely 2009; Hogue and Steinburg 1995; Stevens and Prinstein 2005). Girls express more intimacy and self-disclosure and attribute greater importance to loyalty and closeness within relationships (Rudolph 2002). Relative to girls, boys have greater uncertainty about social support and acceptance, which raises the stakes of securing accepting and supportive friends for them. One way they might socially cope is to attempt to identify and befriend others with similar needs, so we hypothesize that the role of distress-based homophilous selection is stronger among boys than girls.

Additionally, boys have more dispersed networks and may be more likely to be socially sanctioned because role expectations preclude expressions of emotional distress (Johnson 1991). Thus, we hypothesize that boys with higher depressive symptom levels are more likely to be excluded from social networks than is the case among girls. Because the greater value placed on intimacy, self-disclosure, empathic understanding, and emotional support among girls (Hall 2011) may keep distressed girls socially connected, we hypothesize that emotional distress will lead to greater social withdrawal for boys. Thus, we suggest that the overall impact of distress on friendship networks will be larger among boys than girls because of different socialized patterns of emotional acceptance and supportiveness (Ridgeway and Smith-Lovin 1999).

Network and Background Factors

In addition to the core selection and influence mechanisms, we also account for additional factors in both the selection and influence models. For example, youth with elevated depressive symptoms are more likely to be female (Cyranowski et al. 2000), have less educated parents (Goodman et al. 2003), have single parents and less support (Carlson 2006), and lower self-esteem (Dumont and Provost 1999). There is also evidence of homophily among youth based on gender, grade level, race, and SES (Goodreau et al. 2009; Moody 2001). Furthermore, social exclusion by family and friends due to pre-existing dispositional characteristics may lead troubled youth to associate with each other (Giordano et al. 1986). Because these factors are related both to friend selection and depressive symptoms, they are included in both the selection and influence models.

In addition, Steglich et al. (2010) discusses the fact that network processes are themselves sources of influence and change. Along with reciprocity, network closure captures a set of higher-order structural processes driving friendship change, (Snijders et al. 2010). By network closure we specifically refer to triadic processes capturing the overlapping sets of friends that dyads are embedded in. For example, transitive closure captures the processes where a focal adolescent forms a new friendship with another friend’s friend. In this way, networks ‘close,’ not because of depressive symptoms or other factors, but because of the way that the structure of the friendship network creates opportunities for new friendships. In addition, as the work of Falci and McNeely shows (2009; see also Keiley et al. 2000), popularity or social integration can likewise be an important structural feature that influences change in depressive symptoms.

Data & Methods

Data come from waves 1 and 2 of the in-home components of Add Health. Add Health is a stratified longitudinal study of 7-12^th grade youth begun in 1994 with in-school questionnaires administered to approximately 90,000 students in 140 schools. A nationally representative sample of over 20,000 students was drawn from the in-school study and data were collected in-home in 1995 and again approximately one year later at wave 2. This longitudinal sample consists of a core probability sample and special oversamples (racial/ethnic, disabled, genetic) including 16 “saturated” school-settings where efforts were made to collect data on all attending 7-12th grade students so that a network sample could be maintained over time. Of these 16 schools, two were large (N≈1,000; 2,100) and 14 were much smaller (N<300).

We used seven of the saturated settings, all K-12th grade schools that are relatively racially and ethnically homogeneous, to construct our sample. Our decision to use these schools was based on several criteria (this is also discussed further in the Discussion section). First, because our analysis requires longitudinal measures of friendship networks, we were limited to the saturated schools. Second, one of the schools was a special education school and another six were 6^th-8^th grade. We chose not to use the latter schools because the 8^th graders moved into high schools for which full network data is not available. Third, larger schools capture different macro-settings than the small schools as indicated by the enormous size differences with the grade cohorts of the big schools being larger than the entire 7-12^th grade cohorts of the smaller schools. The result is that we focus on the social dynamics in a collection of smaller, more homogeneous settings. The joint sample size of the small schools comprising this study is 798 mostly white 7-12^th grade students. The largest school contributed 163 students to the analysis and the smallest contributed 61. Three were public rural schools (N=363), the remaining four were private (N=435), three of which were urban (N=374). Network data was present for 70-89% of the students on the school-provided roster and these rates have previously been shown to be acceptable for social network analysis (Huisman 2009; Kossinets 2006).

Measures

Dependent & Focal Independent Variables

The first variable, the friendship network matrix, is used to analyze friend affiliation over two waves. These two matrices at waves 1 and 2 map the interconnections between individuals. The adolescent friendship networks at both waves are constructed from two sets of variables requesting nominations of up to five male and five female friends from the school roster. The total sample makes use of all available nominations and the sex-specificity in the questions allowed us to construct male- and female-only networks separately for the sex-specific analyses. The psychological variable used for the analysis is based on 19 ordinal items from the Center for Epidemiologic Studies Depression Scale (CES-D; Radloff 1977). The scale is based on a series of questions inquiring how often during the past week the respondent felt depressed, enjoyed life, etc, with response categories ranging from never or rarely (0) to most or all of the time (3). Because the models require ordinal dependent actor variables, we recoded the CES-D scale into deciles.¹ Cronbach’s α=.85 at both waves.

Control variables

For controls we include whether the respondent is female (=1), grade (range: 7-12th), whether the youth is white (=1), and whether the parent is single (=1). Parent education of the responding parent is included as a five-value variable with categories ranging from (1) did not graduate from high school to (5) received postgraduate training. Scales for parent support and respondent self-esteem were created from wave 1 in-home items. The standardized parent support scale (α =.86) is based on an 8-item ordinal (range: 1-5) “closeness to parents” scale (see Cornwell 2003). The standardized self-esteem scale (α =.81) is based on five ordinal items (range: 1-5) inquiring whether the adolescent feels like they are doing everything just about right, has a lot of good qualities, has a lot to be proud of, likes themselves just as they are, and whether they feel socially accepted.

Finally, we include the number of off list nominations provided by the adolescent during the network portion of the survey. Although the majority of nominations in Add Health are to friends at school, close to 30% are not (Falci and McNeely 2009). In addition, we also include an indicator for whether the respondent was in the restricted nomination sample because some adolescents were allowed to nominate only one male and female friend due to a survey implementation error. The result of this error is that the full friendship network was not captured at the wave 1 in-home survey for 40% of the youth in the sample. We carried the wave 1 in-school nominations forward for these youth² (note that the present study relies on the subsequent wave 1 and 2 in-home surveys) to preserve the full network so that we could conduct the longitudinal social network analysis. There will thus be greater change in the networks for the restricted nomination than regular sample, so we have constructed this indicator to reflect the fact that overall change in friendships will be greater for these adolescents.

2.2. The Model

The analysis employs the new class of Simulation Investigation for Empirical Network Analysis models (SIENA) developed by Snijders (1996; 2001) and colleagues (e.g., Snijders et al. 2007). The model has two components: a network model and a behavioral model that together comprise a system of interdependent equations. The models decompose the influences of selection and influence (see Steglich et al. 2010) by conditioning on wave 1 and then modeling subsequent changes in friendships and CES-D from that point. Thus, wave 1 is a starting point for the estimation routine and is not modeled directly. Coefficients are calculated using a method of moments estimator capturing aggregate changes in social networks and CES-D between observations. The parameter estimates are refined with an agent-based simulation model that is used to calculate their uncertainties and guides their interpretation. The simulation model decomposes changes in the network into a series of the smallest possible changes in either one tie or a 1-decile change in CES-D at a time for a randomly chosen adolescent. In this way, very complicated patterns of change are modeled as the accumulation of many small changes across micro-steps in a way consistent with the total aggregate observed pattern of network and CES-D change.

The model constitutes a continuous time Markov process such that each actor’s decision on whether or not to change one tie or behavior is determined by the current state of his or her network-behavioral configuration. A rate parameter governs how often actors have an opportunity to make changes so that only one actor can act at a time, actors cannot coordinate with each other, or optimize beyond their current state and the next state. The selection model is thus concerned with tie changes in the friendship network, x, where ties between ego (i; rows) and alter (j; columns) are denoted as x_ij = 1, and the lack of a tie is x_ij = 0. The network evaluation function, $f_i^{\mathit{net}}\left(x\right)$, for actor i is defined as

$$f_i^{\mathit{net}}(x,z,v)=\sum _k{\beta }_k^{\mathit{net}}{s}_{ik}^{\mathit{net}}(x,z,v)$$

where the ${\beta }_k^{\mathit{net}}$ are the parameters and $s_{ik}^{\mathit{net}}(x,z,v)$ are the effect paramaterizations defined in Table 1, and which include CES-D (z) and additional control variables (v). In each micro-step the configuration with the most positive $f_i^{\mathit{net}}$ value plus a small amount of randomness determines how the network is modified. There is either no change, a new friend is nominated, or an existing friendship is terminated. The model thus determines actor i’s modification of their network by choosing their optimal tie configuration across all other actors (j).

Table 1.

Description of the main model parameters for the network selection model

Parameter	$\text{Formula}=s_{ki}^{\mathit{net}}$	Description
Selection Parameters & Covariate Effects
Alter (potential friend)	${\sum }_j{x}_{ij}{v}_j$	Main effect of potential friend’s CES-D or varname on the selection of friends (−β=exclusion)
Ego (adolescent)	$v{\sum }_j{x}_{ij}$	Main effect of adolescent’s CES-D or varname on the selection of friends (−β=withdrawal)
Similarity potential friend & adolescenta	${\sum }_j{x}_{ij}({\mathit{sim}}_{ij}-\overline{\mathit{sim}})$	Tendency to choose a friend based upon CES-D or varname similarity (+β=homophilous selection)
Same potential friend & adolescentb	${\sum }_j{x}_{ij}I(v_i=v_j)$	Tendency to choose friends with exactly the same varname
Structural Network Effects
Outdegree	${\sum }_j{x}_{ij}$	General tendency to choose a friend
Reciprocity	${\sum }_j{x}_{ij}{x}_{ji}$	Tendency to have reciprocal friendships
Transitive triplets	${\sum }_{j,h}{x}_{ih}{x}_{ij}{x}_{jh}$	Tendency to become the friend of a friend’s friend
3-Cycles*	${\sum }_{j,h}{x}_{ij}{x}_{jh}{x}_{hi}$	Tendency for a friend’s friend to chose the focal adolescent as a friend
Number distance=2**,c	#(j|xij = 0, G = 2)	Tendency to be indirectly connected through one intermediary
Rate Parameters
Basic rate parameter	---	Determines the average number of change opportunities in each period
varname	---	Differences in the rate of change by varname

^*Notes for structural network effects: A positive effect implies generalized reciprocity while a negative effect with a positive transitive triplet effect suggests local hierarchies (Ripley and Snijders 2011).

^**This effect is an inverse effect of network closure so effects tend to be negative, suggesting that indirect connections tend to close through, e.g., the formation of transitive triplets, or else the indirect connections are lost.

Notes for equations and notation: All covariates are centered at the grand mean and v generically refers to CES-D and additional covariates (varname).

^asim_ij = 1 – |v_i – v_j|/max_ij |v_i – v_j| and $\overline{\mathit{sim}}$ is the average similarity.

^bI(v_i = v_j) is a function indicating whether v_i = v_j (=1) or v_i ≠ v_j (=0).

^cThis function is the number (#) of actors at geodesic distance=2 (G=2) to which the actor is not directly tied (x_ij = 0).

The specific parameters in Table 1 capture the ways that covariates are operationalized in this study to influence changes in ties. Positive values on these effects contribute to the evaluation function and thus express preferences for ties, while negative values indicate the opposite. The alter, ego, and similarity effects are the key selection mechanisms central to this process. The alter effect is the sum of the covariate values for each alter (j) that ego (i) is tied to, and thus captures the influence of CES-D_j (or another v_j) on the likelihood of being viewed as a friend. A negative parameter estimate is therefore an indicator of exclusion. The ego effect reflects the ego’s covariate value on z_i or v_i and the count of their nominations. For CES-D, this indicates the extent to which distress is related to the extensiveness of their activity in the network, with withdrawal indicated by a negative ${\beta }_k^{\mathit{net}}$. Similarity is a dyadic effect expressing homophilious selection that is based on how similar ego (z_i) and alter (z_j) are to each other. This is captured as their absolute difference on v relative to the observed range of v (0=maximally dissimilar, 1=maximally similar). A positive ${\beta }_k^{\mathit{net}}$ thus indicates a preference for ties among those with similar CES-D or v levels, as predicted by the homophilous selection hypothesis.

Additional parameters and textual descriptions are provided in Table 1, including structural parameters for reciprocity and network closure. Accounting for network closure processes is of substantial importance because they reflect alternative confounding mechanisms driving changes in friendships (Steglich et al. 2010). For example, when the friends of friends become friends, not because of depressive symptoms, but because of the opportunities afforded from socializing together due to shared friendships (transitive triplets). Closure is also captured with the distance=2 effect, which is expected to be negative as an indication that adolescents prefer to be directly rather than indirectly connected to one another. The 3-cycles effect reflects cycles of friendships such that i → j → k → i, and because a posititve parameter expresses generalized reciprocity suggests relative equality in the popularity off all friends locally within the network, is expected to be negative.

In the models we present, each randomly chosen actor’s decision is actually composed of two parts: they can either change their network, $f_i^{\mathit{net}}$, or their CES-D value can adjust up or down 1-unit (but only one-unit and for only one outcome in a single micro-step). In the behavioral evaluation function

$$f_i^{\mathit{beh}}(x,z,v)=\sum _k{\beta }_k^{\mathit{beh}}{s}_{ik}^{\mathit{beh}}(x,z,v)$$

the statistics $s_{ik}^{\mathit{beh}}(x,z,v)$ predict changes in CES-D as a function of current CES-D values (z), the state of the network (x), and other variables (v) such that +${\beta }_k^{\mathit{beh}}$ indicates increases, and −${\beta }_k^{\mathit{beh}}$ indicates decreases in CES-D levels much like an ordinal logit model. Average friend similarity is the focal parameter capturing the influence of friends’ CES-D, and for which the friend influence hypothesis predicts a positive ${\beta }_k^{\mathit{beh}}$. This parameter is defined as the average of the egos (i) CES-D similarity to each of their friends (j; thus the summation over j and division by the count over j denoted by $x_{i+}^{-1}$). Additional covariates (v) enter the equation as interactions with CES-D (z_i) but are interpreted as main effects. The linear and quadratic shape coefficients are two additional parameters that we do not view substantively in this application. They express the shape of the distribution and thus contribute to the evaluation function, $f_i^{\mathit{beh}}(x,z,v)$, by indicating movement towards globally optimal values given the current value of z_i.³

Analysis

The analysis uses the SIENA software (Ripley et al. 2011) to model friendship and depressive symptom changes in the joint combined social network of the schools. Because youth in different schools are unable to select each other as friends, out-of-school elements in the sociomatrices are fixed (see Ripley et al. 2011 for a discussion of this and other approaches⁴). All respondents were included in the analysis and were allowed to enter the study later or leave early (i.e., those who graduated from high school) using the composition change method of Huisman and Snijders (2003). Missing attributes and CES-D data were treated as non-informative following the method described by Huisman and Steglich (2008) so that missing values are imputed within the model, but only observed values contribute to the estimated change statistics in the estimation algorithm. Additional parameters that were not included in the analysis are also presented at the bottom of Table 2. Whether these parameters should be included were assessed using score tests to determine if they improved the model performance against a baseline model including the network structure effects and CES-D influence and selection parameters (Schweinberger 2011). Score-tests determining the improvement in the model fit were also used to simplify the model structure with respect to the control variables so that not all ego, alter, and similarity parameters are included for each covariate.

Table 2.

Description of the main model parameters for the network selection model

Parameter	$\text{Formula}=s_{ik}^{\mathit{beh}}(x,z)$	Description
Focal Influence Parameter
Average friend similaritya	$x_{i+}^{-1}{\sum }_j{x}_{ij}({\mathit{sim}}_{ij}-\overline{\mathit{sim}})$	Main effect of friends’ average CES-D similarity on individual CES-D (+β=friend influence)
Varname b	zivi	Main effect of varname on CES-D
Shape Parameters
Linear shape parameter	zi	The overall shape of the CES-D distribution modeled as a quadratic function
Quadratic shape parameter	$z_i^2$
Rate Parameters
Basic rate parameter	---	Determines the average number of change opportunities in each period
Score Tests: Extra CES-D Effects Tested
Incoming friendships	$z_i{\sum }_j{x}_{ji}$	Test of main effect of adolescents number of received nominations
Outgoing friendships	$z_i{\sum }_j{x}_{ij}$	Test of main effect of adolescents’ number of nominated friends
CES-D average friend similarity X reciprocityc	$x_{i\left(r\right)}^{-1}{\sum }_j{x}_{ij}{x}_{ji}({\mathit{sim}}_{ij}-\overline{\mathit{sim}})$	Test of whether the effect of average friend CES-D similarity with differs among reciprocal and non-reciprocal friends

Notes: All covariates are centered at the grand mean.

^asim_ij = 1 – |v_i – v_j/max_ij|v_i – v_j|. Note too that x_i+ is the total number of friends nominated by i, and $\overline{\mathit{sim}}$ is the average similarity.

^bvarname = v, CES-D = z.

^c$x_{i\left(r\right)}^{-1}$ is the inverse of the count of the egos reciprocated friendships.

Additionally, the contribution of the different processes to the autocorrelation between the friendship network and CES-D is decomposed by the simulation method described in Steglich et al. (2010; see also Mercken et al. 2010). The spatial network-CES-D autocorrelation is calculated using Moran’s I (Moran 1950) across a special model series disaggregating the contributions of the different mechanisms to this correlation. In this way, depressive symptom similarity is decomposed into the proportionate contributions of selection (by type), influence, alternative mechanisms from the other covariates and structural network effects (i.e. controls), and general trend effects indicating state dependencies in friendships and individual distress.

Results

Descriptive statistics for the total, boy, and girl samples are presented in Table 3. On average, approximately 3.5 in-school and nearly 2.5 additional (unmatched within school off list) friends were nominated. Of the youth in the sample over the entire study period who did not graduate, dropout, or change schools, CES-D scores were consistent for 25% while the rest, evenly split, increased or decreased by at least one decile. Girls had slightly more depressive symptoms than boys, and whereas girls’ scores increased slightly, boys’ decreased. The overall percentages moving up and down, however, were similar. With respect to the focal parameters, the social exclusion and withdrawal counts were widely dispersed and indicated that lower CES-D youth were more likely to be nominated and less likely to nominate others. There were slight gender differences in exclusion with girls above the mean on CES-D slightly more likely to be selected as friends. Average friend similarity on depressive symptoms, relevant for the homophilous selection and assimilation hypotheses, was .66, indicating that most friends are relatively similar to each other. The boys were slightly more similar to each other than the girls were (.68 to .63).

Table 3.

Descriptive statistics for the total, boy, and girl samples

	Total Sample					Male		Female
Variable	Obsa	Mean	SD	Min	Max	Mean	SD	Mean	SD
Dependent Variable: CES-D
Wave 1 in home	796	4.19	(2.82)	1	10	4.04	(2.67)	4.36	(2.96)
Wave 2 in home	627	4.17	(2.83)	1	10	3.88	(2.71)	4.45	(2.92)
Prop. decreased	625	0.37		0	1	0.37		0.36
Prop. stayed the same	625	0.27		0	1	0.26		0.27
Prop. increased	625	0.37		0	1	0.37		0.36
Network Characteristics
Off list nomination count	798	2.39	(2.20)	0	10	1.12	(1.36)	1.19	(1.21)
Restricted nom. sample	798	0.40		0	1	0.41		0.39
Alter CES-D (exclusion)	798	−0.61	(5.18)	−25.0	20.4	−0.53	(3.61)	0.18	(4.12)
Ego CES-D (withdrawal)	796	−0.50	(10.90)	−27.2	40.2	−0.19	(5.88)	−0.07	(6.91)
Average friend similarity (homophily)	688	0.66	(0.19)	0	1	0.68	(0.22)	0.63	(0.22)
Received nominations	798	3.28	(2.93)	0	19	1.89	(1.72)	1.92	(1.73)
Out nominations	779	3.36	(2.27)	0	10	1.94	(1.41)	1.99	(1.37)
Reciprocal ties count	779	1.20	(1.31)	0	7	0.85	(0.99)	0.93	(1.06)
Transitive triplets count	779	3.37	(4.83)	0	29	1.16	(1.70)	1.36	(2.22)
Covariates
Female	798	0.50		0	1
Grade	795	9.47	(1.69)	7	12	9.47	(1.72)	9.47	(1.66)
White	798	0.97		0	1	0.98		0.96
Parent education	717	2.63	(1.04)	1	5	2.61	(1.03)	2.66	(1.04)
Single parent	721	0.22		0	1	0.22		0.22
Parent support	610	0.27	(0.93)	−4.12	1.31	0.31	(0.82)	0.24	(1.03)
Self-esteem	798	0.14	(0.94)	−3.52	1.56	0.29	(0.85)	−0.01	(1.00)

^aNotes: Observations for CES-D change between waves 1 and 2 largely because of high school graduation.

Dynamics in the Total Sample

The model series for the full sample is presented in Table 4. We reiterate here that we used score-tests to build the model parsimoniously so that not all possible effects were estimated. Coefficients are in logit metrics for both the network change model and depressive symptoms. In the first model, A0, network and distress dimensions evolve independently from each other so that the network dynamics can be considered in isolation from those for CES-D. The outdegree density parameter is large and negative because adolescents were only able to nominate (or only nominated) a few friends out of the total pool and the large positive reciprocity indicates the tendency for friendships to be reciprocated. That is, the odds of a friendship are (exp[2.03]=) 7.6 times larger if the tie is reciprocal, all else equal. The triadic effects, namely the positive transitive triplets effect and the proscription against distance=2 connections, show that friendship changes reflect opportunities available by virtue of existing connections and thus adolescents become friends with their friends’ friends (triadic closure). Moreover, when combined with this finding, the negative 3-cycles coefficient implies that there is tendency for some youth to be locally more popular than others (Ripley et al. 2011).

Table 4.

Parameter estimates (in logits) and standard errors for models of joint network and CES-D evolution over two waves (N=798)

	A0	-	A1	-	A2	-	A3	-	A4
Parameter	b/se		b/se		b/se		b/se		b/se
Network: CES-D
	Alter (potential friend; exclusion)		−0.028	*	−0.032	*	−0.028	*	−0.026	*
			[0.013]		[0.013]		[0.014]		[0.013]
Ego (adolescent; withdrawal)			−0.018		−0.003		−0.001		−0.01
			[0.013]		[0.015]		[0.013]		[0.015]
Similarity (potential friend & adolescent)			0.525	*	0.406	+	0.47	*	0.505	+
			[0.219]		[0.216]		[0.220]		[0.259]
Network: Structural Parameters
Outdegree density	−2.047	*	−2.486	*	−2.075	*	−2.726	*	−2.943	*
	[0.085]		[0.043]		[0.108]		[0.147]		[0.152]
Reciprocity	2.026	*	1.977	*	2.018	*	1.721	*	1.722	*
	[0.082]		[0.065]		[0.088]		[0.075]		[0.082]
Transitive Triplets	0.426	*			0.421	*	0.384	*	0.396	*
	[0.035]				[0.040]		[0.038]		[0.036]
3-cycles	−0.476	*			−0.48	*	−0.48	*	−0.492	*
	[0.063]				[0.074]		[0.064]		[0.067]
Number distance==2	−0.333	*			−0.331	*	−0.245	*	−0.226	*
	[0.039]				[0.046]		[0.035]		[0.037]
Network: Control Variables
Off list nominations, alter	−0.083	*	−0.08	*	−0.08	*	−0.071	*	−0.069	*
	[0.013]		[0.012]		[0.014]		[0.013]		[0.013]
Restricted nominations, same sample	0.196	*	0.156	*	0.196	*	0.217	*	0.231	*
	[0.043]		[0.043]		[0.044]		[0.044]		[0.043]
Same sex							0.267	*	0.276	*
							[0.043]		[0.043]
Same grade							0.704	*	0.712	*
							[0.045]		[0.049]
Same race/ethnicity (white/nonwhite)							0.066		0.217	+
							[0.121]		[0.125]
Parent education similarity							0.300	*	0.305	*
							[0.115]		[0.107]
Single parent, ego							−0.105		−0.095
							[0.068]		[0.066]
Single parent, same							0.102	*	0.111	*
							[0.051]		[0.049]
Parent support (z), ego									0.023
									[0.033]
Self esteem (z), ego									−0.053	+
CES-D: Assimilation & Covariates
CES-D average friend similarity			2.420	*	2.727	*	2.592	*	2.128	+
			[0.882]		[1.376]		[0.980]		[1.124]
Private school			−0.074	*	−0.067	*	−0.022		−0.023
			[0.035]		[0.032]		[0.034]		[0.033]
Parent education							−0.061	*	−0.062	*
							[0.017]		[0.017]
Parent support (z)									−0.056	*
									[0.021]
Self-esteem (z)									−0.048	*
									[0.020]
Rate & Behavioral Shape Parameters
Network: basic rate parameter	12.842	*	10.043	*	12.876	*	14.382	*	14.401	*
	[0.593]		[0.426]		[0.620]		[0.795]		[0.770]
Network: restricted nomination sample	0.455	*	0.428	*	0.452	*	0.566	*	0.571	*
	[0.110]		[0.092]		[0.121]		[0.136]		[0.176]
CES-D: basic rate parameter	14.548	*	14.14	*	14.1	*	13.938	*	15.456	*
	[1.040]		[1.606]		[1.678]		[1.094]		[1.402]
Linear shape parameter	−0.034	*	−0.012		−0.009		−0.012		−0.015
	[0.015]		[0.018]		[0.021]		[0.018]		[0.018]
Quadratic shape parameter	−0.006	+	0.014	+	0.015		0.013		0.003
	[0.003]		[0.007]		[0.012]		[0.008]		[0.008]

⁺p<.1,

^*p<.05

Turning to the control variables in model A0, those with more off list nominations report fewer friends in school, as expected. Score tests revealed that off list nominations were not related to receipt of nominations or homophilous selection (similarity). The positive effect for being in the restricted nomination sample suggests that these youth were more likely to become friends over time, reflecting the unequal probability of being in the restricted nomination sample across schools. The remaining rate and CES-D shape parameters indicate that adolescents, on average, were given 13 opportunities (microsteps) to change their friendship ties in the simulation portion of the model and that those in the restricted nomination sample received approximately .5 more opportunities. Youth were also provided 14.5 opportunities to adjust their CES-D values. These rates merely express opportunities, not actual change, since those youths with optimal values for their evaluation function do not change values. Finally, the shape parameters describe the shape of the CES-D as mostly flat and slightly decreasing, reflecting the decile coding and slight decline in depressive symptoms over time.

The remaining models explicitly address the substantive hypotheses. Model A1 removes the triadic network effects and includes the effects of CES-D on friendship dynamics, while the network dynamics enter the CES-D equations through the average friend similarity effect capturing influence. First, this model provides support for both the exclusion and homophilous selection mechanisms. The negative exclusion coefficient (b=−.028) indicates that youth with more depressive symptoms are less likely to receive friendship nominations while the positive similarity (homophilous selection) parameter (b=.53) documents a tendency for friends to have similar CES-D scores such that two adolescents perfectly similar have odds of being friends 1.7 times larger than those who are perfectly dissimilar.

Second, the positive similarity coefficient in the CES-D model (b=2.4) further shows social influence effects – changes in CES-D scores among friends move towards each other or become more similar. In other words, there is evidence that though more depressed youth are excluded in the network, they tend to find each other, and that over time their mood adjusts to their friends’. Regarding the odds of a decile CES-D change, a .1 increase in similarity is associated with a (exp[2.4*.1]=1.3) 30% increase in the odds. Finally, the negative private school coefficient indicates greater decreases in depressive symptoms relative to those in public schools.

Model A2 addresses the possible biasing role of network processes by adding the network closure effects. Overall, the results change little from those reported previously, although the homophilous selection coefficient decreases to .41 (p<.1) and is no longer statistically significant at p<.05. In fact, the general results pattern continues into model A3 when a subset of the control variables is added. Both the homophilous selection and influence parameters are marginally significant (p<.1) in model A4 when the parent support and self-esteem parameters are included. Both parameters remain relatively large, however. Overall, these findings provide evidence of the role of social exclusion and homophilous selection in friendship networks and that friends influence each other’s emotional distress.

The remaining parameters in A4 correspond with general homophilous trends as well –friends are more likely to be same sex, same grade, same race, and have parents with similar education backgrounds and marital status. The self-esteem ego coefficient is unexpectedly negative, indicating a social withdrawal effect conditional on the other parameters included in the model. In the behavioral model, parent education, support, and self-esteem are all negatively related to depressive symptoms. Finally, the private school indicator effect is the result of different parent education levels between children attending the different school types. These small school results thus point to the importance of larger social structural factors arranging relationships locally.

In order to provide a better sense of effect magnitudes, Table 5 presents the estimated network-CES-D autocorrelation along with a % decomposition of this autocorrelation into trend effects (i.e., state dependence in friendships and CES-D). Contributions of control variables and structural network effects, exclusion, withdrawal, homophilous selection, social influence, and the residual correlation are also included. Approximately 11% of the network-CES-D autocorrelation is due to general stability or trend effects, while another 17% reflects the control variables. Notably, the most salient process is social influence at 37% while the second most important process is social selection at 25%. Overall, neither the withdrawal or exclusion mechanisms play a very strong role. In summary, these findings suggest that adolescents prefer friends with levels of distress similar to their own, but also that their own levels of distress even more strongly react to that of their friends’.

Table 5.

Percentage decomposition of the estimated network-CES-D autocorrelation by mechanism for the total, male, and female friendship networks

Parameter	Total	Male	Female
Trend	11	16	22
Control variables	17	3	33
Alter (potential friend; exclusion)	1	1	2
Ego (adolescent; withdrawal)	0	0	3
Selection (similarity)	25	27	3
Influence (average friend similarity)	37	47	24
Residual autocorrelation	8	7	14
Estimated Network Autocorrelation	0.14	0.20	0.11

We also explored a number of additional structural effects on CES-D changes that were not included in the presented models. Adolescent depressive symptoms are unrelated to friend nominations received (χ²=.01, df=1, p=.9), suggesting that social exclusion does not increase CES-D scores. The same is true for the number of nominations sent (χ²=.1, df=1, p=.35), indicating that social withdrawal is not related to changes in depressive symptoms. Additionally, distress is not exacerbated within reciprocated friendships (χ²=.1, df=1, p=.75). In other words, the influence of social relationships on depressive symptoms appears to largely operate through the depressive symptoms of those one views as friends, not through other structural network effects (Ueno 2005), reciprocity (Stevens and Prinstein 2005), or a lack of social integration (e.g., Falci and McNeely 2009) in smaller schools.

Dynamics among Boys & Girls

Next, we turn to the gender-specific analysis assessing whether or not friendship and CES-D processes differ among boys and girls in gender-specific networks. The effect sizes captured as autocorrelation decompositions are presented in Table 5 and the coefficient estimates and standard errors are reported in Table 6. The final column of Table 6 contains z-values and significance levels for a comparison of the boy-girl coefficient differences. First, as shown in Table 5, the autocorrelation is two times larger among boys than girls, primarily reflecting the combination of influence (47 vs. 24%) and homophilous selection (27 vs. 3%). The coefficients capturing these two processes indicate that neither selection nor influence is significant among girls and that the differences relative to boys are marginally significant (p<.1). Overall, then, homophilous selection and social influence in the total network are likely to be driven by the males. Globally, the findings presented in Table 6 are consistent with the hypothesis that the social dynamics of depressive symptoms are different among girls and boys (Rudolph 2002).

Table 6.

Parameter estimates (in logits) and standard errors for models of joint network and CES-D evolution over two waves for boys and girls separately

	Males			Females			Compare z-value
Parameter	b		se	b		se
Network: CES-D
Alter (potential friend; exclusion)	0.01		(0.03)	−0.07	*	(0.03)	1.78	+
Ego (adolescent; withdrawal)	0.07	+	(0.04)	−0.05		(0.04)	2.26	*
Similarity (potential friend & adolescent)	2.30	*	(0.96)		0.43	(0.57)	1.67	+
Network: Structural Parameters
Outdegree density	−3.19	*	(0.40)	−3.66	*	(0.32)	0.92
Reciprocity	1.91	*	(0.17)	2.56	*	(0.19)	2.55	*
Transitive Triplets	0.74	*	(0.10)	0.76	*	(0.11)	0.09
3-cycles	−0.64	*	(0.16)	−1.02	*	(0.20)	1.49
Number distance==2	−0.17	*	(0.08)	−0.40	*	(0.10)	1.87	+
Network: Control Variables
Off list nominations, alter	−0.09	*	(0.03)	−0.02		(0.03)	1.39
Off list nominations, similarity	−0.15		(0.31)	0.96	*	(0.32)	2.50	*
Restricted nominations, same sample	0.27	*	(0.11)	0.37	*	(0.10)	0.74
Same grade	0.66	*	(0.10)	0.90	*	(0.10)	1.65	+
Same race/ethnicity (white/nonwhite)	−0.25		(0.28)	0.32		(0.24)	1.52
Parent education similarity	0.70	*	(0.26)	0.29		(0.24)	1.18
Single parent, ego	−0.27	+	(0.16)	−0.06		(0.16)	0.93
Single parent, same	0.01		(0.11)	0.24	*	(0.12)	1.42
Parent support (z), similarity	0.18	*	(0.09)	−0.05		(0.07)	2.10	*
Self esteem (z), similarity	−0.04		(0.08)	−0.06		(0.07)	0.18
CES-D: Assimilation & Covariates
CES-D average friend similarity	6.71	*	(2.67)	0.82		(1.42)	1.95	+
Private school	−0.04		(0.04)	−0.07	*	(0.03)	0.78
Parent education	−0.03	(0.04)		−0.06	*	(0.03)	0.66
Parent support (z)	−0.09	*	(0.03)	−0.04	+	(0.02)	1.21
Self-esteem (z)	−0.01		(0.06)	−0.07		(0.05)	0.86
Rate & Behavioral Shape Parameters
Network: basic rate parameter	5.64	*	(0.46)	6.91	*	(0.74)	1.45
Network: restricted nomination sample	0.28		(0.18)	0.54	*	(0.25)	0.86
CES-D: basic rate parameter	10.09	*	(1.22)	14.87	*	(1.89)	2.13	*
Linear shape parameter	0.02		(0.03)	0.03	+	(0.02)	0.44
Quadratic shape parameter	−0.01		(0.01)	−0.03	*	(0.01)	1.11

⁺p<.1,

^*p<.05

The evidence also points to girls being more likely to exclude those with higher CES-D scores, contrary to our expectations, and suggesting that processes among girls are more consistent with prior mental health research indicating that depressive symptoms lead to social isolation (e.g., Kawachi and Berkman 2001; Barnett and Gotlib 1988). Boys, however, are more consistent with the adolescent externalizing behavior literature that suggests boys are more likely to assimilate to the behavior of others in their friend networks (Brown et al. 1986). Additional findings show that reciprocity is higher among girls and that there is a marginally stronger tendency towards triadic closure (number distance=2) so that that friend groups are tighter for girls than for boys (Cairns 1998). Friendships among girls are also more likely to be withingrade. Though the differences are not statistically significant between boys and girls, the following results may be substantively meaningful because the estimated coefficient is different from zero for one group and not the other: Boys are more likely to be friends with those from similar socioeconomic backgrounds, with similar levels of parent support, and show a social withdrawal effect related to living with a single parent. Girls, on the other hand, are more likely to be friends with those residing in the same family structure.

Depression

We also assessed whether homophilous social selection is driven by more distressed youth. First, we estimated an alternative homophilous selection parameter (an “ego × alter” interaction) assessing whether selectivity increases with higher CES-D scores. A further extension (not shown) dichotomized the CES-D scores at cutoffs of 16 (17% of the sample) and 20 (12%) to approximate the dynamics of clinical depression (Roberts et al. 1990).⁵ The results indicated that the social dynamics of depression operate across the distribution of symptoms and are not driven only by those with the most symptoms, though there is a small increase in selectivity for boys with higher CES-D scores. In addition, Fowler and Christakis (2008) reported stronger and more consistent effects using the happiness subscale of the CES-D, but we were not able to replicate those results in this study. In other words, distress, not happiness, appears to be more important among adolescents.

Discussion

The analysis we presented contributes to the sociological literature on adolescent distress in a number of ways. First, the sociological literature itself is not large and with the exception of only a few studies (Falci and McNeely 2009; Ueno 2005; Schaefer et al. 2011), most are either small samples or not nationally representative (Hogue and Steinburg 1995; Hansell 1985; Stevens and Prinstein 2005; Prinstein et al. 2005). This is a long-standing issue in social network research. Though our study, which focuses on adolescents who attended seven small (N<300) Add Health high schools where longitudinal social network data was collected, is neither large nor nationally representative, it contributes to this growing literature by utilizing a new set of schools and innovative new analytic approaches to further develop an emerging picture of the social dynamics of adolescent distress across different school settings.

Second, we modeled friend influence and selection jointly, controlling each for the other (Steglich et al. 2006), which other studies have either failed to do or have had to use ad hoc methods to address (see Steglich et al. 2010). This allows us to contribute both to the growing literature on how social networks influence depressive symptoms and to the new discussion about how depressive symptoms influence friend selection as a new wave of scholarship turns the social network into a dependent variable (Schaefer et al. 2011; see House et al. 1988). Finally, we hypothesized gender differences with the findings we present contributing to the growing body of research indicating that some social processes among young men and women differ in important ways (Eder, Evans, and Parker 1995).

Having used complete networks and individual data together, the results suggest, consistent with our first two hypotheses, that the social dynamics of distress involve a complicated joint process with depressive symptoms influencing and being influenced by friendship. That is, friendships form among those with similar distress levels, even while friendships with more distressed youth can be harmful (though those with healthier peers can be beneficial). Moreover, the findings by gender illustrate that friendship processes vary for different groups even within the same network as a result of broader social structural processes. Girl networks are tighter and more cohesive and distress increases the risk for social exclusion compared to boys, contrary to our expectations. This latter effect is not present for boys. Instead, and consistent with our hypotheses, homophilous selection and social influence are important among boys but not girls (Hogue and Steinberg 1995).

The fact that the network processes differ among boys and girls points to what qualitative researchers have argued for some time – that social dynamics within school settings are nuanced, variant, and gendered (see Eder, Evans, and Parker 1995). Eder and Hallinan (1978) argued that different socialization and structured interactional emphases between girls and boys lead to different social skill sets (see also Giordano 2003; Rubin et al. 2006). These skill sets arise from the accumulating histories of patterned interactions such that boys’ social skills come to reflect group decision making, leadership, and other group processes, while girls’ focus on dyadic interaction leads to more intimate and intense relationships. The focus on dyadic relations and self-disclosure leads to greater emotional supportiveness in the maintenance of those relations too (Cyranowski et al. 2000; Kort-Butler 2009), but possibly also burdensome reassurance-seeking behavior that increases the risk that more distressed girls are marginalized in the network (Schaefer et al. 2011).

To the extent that distressed girls are more likely to be marginalized, the tendency towards social exclusion we found among them may reflect the fact that nurturing distressed friends is costly (Prinstein et al. 2005), even if those costs do not appear to lead to a diffusion of distress through friend influence processes. In fact, friend distress was unrelated to subsequent changes in distress, which contributed to the finding that the network-CES-D autocorrelation was 50% smaller for girls than for boys. Consistent with the contention that girls experience more risk factors for depression (Nolen-Hoeksema and Girgus 1994), the network-CES-D autocorrelation was also more strongly related to background factors for the female sample. This indicates that adolescent emotional distress among girls in these small school settings responds primarily to other factors and also suggests that despite greater intimacy, girls are better able to protect themselves from the distress of their intimates than are boys.

Consistent with our hypotheses, the results for boys show both homophilous selection and friend influences. Not only do boys select similarly distressed male friends, they also influence each other over time. In general, male relationships tend to be less close and cohesive, and many dimensions of male friendships are guided by the complex intersection of the desire to fit into larger peer groups, male social competitiveness, and emotional needs (Rudolph 2002; Prinstein 2007). However, our findings suggest that this story needs qualification as it applies to emotional distress. In this case, it appears that boys show a preference for others with similar distress levels and that they may even increase their participation in the social network in order to find those associations. For distressed young males, being friends with peers sharing their mood may provide protection from potential rejection within the network, but at the cost of being negatively influenced by the mood of their friends. This finding is consistent with those in the externalizing literature suggesting that youth whose peer networks consist of high concentrations of delinquent peers are more likely to engage in subsequent delinquent behaviors (Haynie 2002; Heimer 1996).

Our results raise important questions about the specification of network effects and the role of gender in moderating social processes (see also Gaughan 2006). As we have shown here, the social mechanisms of distress found in the total network were a mixture of separate processes happening in girl and boy networks. Clearly, understanding gendered social processes in school settings is an important avenue for future adolescent health research on both theoretical and applied levels since distinctions in these processes have implications for intervention designs that seek to either incorporate, address, or redress the social dimensions of adolescent mental health.

There are limitations to our analysis and the results we present are tentative. First, the weight of the statistical evidence in the final model for the total sample, and the boy-girl differences in the homophilous selection and influence parameters, are marginal (p<.1, two-tailed⁶). Second, this study covers only a very short one-year period over adolescence. While studies among adults suggest similar social dynamics of mood across the life-course (e.g., Fowler and Christakis 2008; Cacioppo et al. 2009), more studies during adolescence, when the prevalence of depressive symptoms increase, are needed. Third, there are two additional processes that we were not able to incorporate into this study that reflect cross-gender selection processes (e.g., girl→boy). Fourth, we have focused on only friendship processes and social relationships are much more varied. Negative relationships networks of bullying and social aggression may be even more critical for understanding the mental health outcomes of many youth (e.g., Faris and Felmlee 2011).

Finally, the schools in this study are small in size, which limits generalizability across settings. The longitudinal network component of Add Health is restricted to a small subset of schools, so we have chosen to focus on smaller, more homogenous settings. With respect to friend selection, depressive affect may be a relatively subtle signal that is more visible in small schools where everyone knows one another. If so, this would explain why Schaefer et al.’s (2011) selection findings differ from ours. They used a larger set of Add Health schools than we did and did not find evidence of homophilous selection. Indeed, the smaller Add Health school settings are substantially different from the larger settings (N<300 for K-12th versus N>800 for 10-12th). Adolescents are more connected to each other in smaller schools (McNeely, Nonnemaker, and Blum 2002), more attached to their schools (Crosnoe, Johnson, and Elder 2004), and individuals in the school are more likely to know each other (Leithwood and Jantzi 2009). Network processes also vary across schools (Mouw and Entwisle 2006), and school size influences the structure of the curriculum (Leithwood and Jantzi 2009), thereby constraining friendship opportunities in larger schools (Kubitschek and Hallinan 1998). For these reasons, we have chosen to limit the heterogeneity in our sample, exchanging a broader set of schools for one more narrowly circumscribed but also more specifically targeting certain types of social environments where the influence of distress on social processes is likely to be the most evident.

This study has a number of strengths as well. For example, we were able to control for a wider range of structural network processes in our models than others directly studying selectivity have been able to (i.e., Crosnoe et al. 2008). Although we did not focus on how structural processes influence depressive symptoms (i.e., Falci and McNeely 2009), our findings suggest that immediate friend influences are more central, at least for boys. In addition, we used full network data that is not biased by the social cognitions of more depressed youth who tend to view friendships negatively, even when others view them favorably (Rose et al. 2007). Finally, the models we used are designed to deal with the inherent dependencies in network data and so inferences are not undercut by the limited statistical assumptions inherent when traditional statistical approaches are applied to network data (Steglich et al. 2010).

Conclusion

Though the vast majority of adolescents in the United States attend much larger schools than those studied here, nearly 800k students are currently enrolled in K-12th grade schools (NCES 2011). This is a small proportion of the total number of students, but a non-trivial number of people all the same. Youth in these schools interact in a relatively circumscribed educational environment and so have known each other for years. As the current analysis indicates, distress-based homophilous selection and social influence processes are strongest for boys. Specifically, emotional distress influences the meso-level school friendship network, even while this pattern of friendships, and the distribution of distress within it, is related to individual-level changes in depressive symptoms. Interventions and programs targeting internalizing problems and disorders in small school settings should be aware that processes between boys and girls differ in these ways and, that at least among boys, depressive symptoms have roots that extend beyond individuals and out into the interpersonal social worlds they participate in.