Contextualizing Depressive Contagion

Abstract

The purpose of this study is to examine microsocial and macrosocial contextual moderators of adolescent depressive contagion. Using data from the National Longitudinal Study of Adolescent Health (Add Health), the authors find evidence supporting the depressive contagion thesis. This effect is observed above and beyond key social relationship and sociodemographic controls. To examine the role of social context in moderating the effect of depressive contagion, the authors utilize a longitudinal mixed effects model using Wave 1 and Wave 2 of the Add Health survey. The results reveal that depressive contagion is more salient for adolescents who are embedded in dense peer networks and attend schools with high network density and mutuality. Furthermore, popular students, measured as the number of received friendship nominations, are more vulnerable to depressive contagion. Overall, the findings in this study demonstrate a differential vulnerability to depressive contagion dependent on microsocial and macrosocial context.

Keywords

adolescents depression networks childhood mental health

According to the Health for the World’s Adolescents report (World Health Organization 2014) conducted by the World Health Organization (WHO), depression is the number one cause of illness and disability among adolescents globally. Additionally, a National Institute of Mental Health survey (National Institute of Mental Health 2007) estimates that approximately 11 percent of adolescents in the United States have experienced depression by the age of 18 while the National Survey on Drug Use and Health in 2012 estimates that approximately 9.1 percent of adolescents 12 to 17 have experienced at least one major depressive disorder in the past year (Substance Abuse and Mental Health Services Administration 2013). Moreover, depression has significant effects on the adolescent’s self-esteem, academic performance, and interpersonal experiences, which in turn can increase depression within the adolescent (Joyner and Udry 2000).

An emerging literature on adolescent depression focuses on the explanations for the similarities in the level of depression between adolescents and their immediate friends. One explanation for this similarity is through the depressive contagion mechanism suggesting that adolescent depression is significantly influenced by friends’ level of depression. However, this contagion effect is often conflated with two other explanations: homophily and shared environments (Shalizi and Thomas 2011). Homophily assumes that depressive similarities in friendship networks are caused by individuals selecting friendships based on similar levels of depression. Shared environment effects point to spurious effects caused by shared contexts in which the friendship network is embedded, such as a classroom, neighborhood, or school. Studies aimed at disentangling contagion effects from homophily and shared environments use various statistical methods including longitudinal regression models (e.g., Conway et al. 2011) and stochastic actor-based models (e.g., Cheadle and Goosby 2012; Zalk et al. 2010a). However, few studies (e.g., Prinstein 2007) have investigated the role of environments in moderating the effect of depressive contagion.

This study aims to demonstrate the utility of a longitudinal multilevel mixed effects model in examining contextual moderators of depressive contagion. According to Prinstein (2007), contextual moderators “refer to aspects of the environment in which peer contagion potentially may occur.” Only one study to our knowledge examines contextual moderators of depressive contagion¹ (Conway et al. 2011). Using a measure of received friendship nominations, the authors find that depressive contagion is more salient in less popular students. Although this finding is a significant contribution to the literature, there is still much to be explained for other contextual moderators such as network structural characteristics.

Furthermore, much of the prior research on contextual moderators of contagion (e.g., Conway et al. 2011; Haynie 2001; Prinstein 2007) examines elements at purely the “microsocial,” or local peer group level (Dishion 2013). This tendency ignores higher level moderating effects where the adolescent is not only nested within a local peer group but also a classroom, a school, a community, and so forth. Dishion (2013) refers to these higher level social contexts as the “macrosocial.” As a result, we investigate several measures of social context at both the microsocial and macrosocial levels. The multilevel model used in this analysis presents a unique opportunity for this investigation. First, this model enables us to replicate prior findings of depressive contagion by statistically adjusting for (1) school-level effects that may bias parameter estimates of depressive contagion (Raudenbush and Bryk 2002) and (2) prior levels of depressive symptoms that in effect produces parameter estimates of change in ego depressive symptoms from Wave 1 to Wave 2 as a result of friends’ level of depressive symptoms during Wave 1 (Christakis and Fowler 2013).

Second, this model allows for the use of a large-scale, nationally representative sample of adolescents nested within schools from the first and second waves of the National Longitudinal Study of Adolescent Health (Add Health). This allows for a simultaneous test of both individual-level interactions between friends’ depressive symptoms and microsocial context and cross-level interactions between friends’ depressive symptoms and macrosocial context. Rather than viewing these social contexts as shared environments to be controlled, we examine their potential as social buffers or vulnerabilities to depressive contagion. In the following sections, we review the literature on the association between health and social relationships and depressive contagion.

Social Relationships and Health

The association between social relationships and health has long been established both empirically and theoretically (House, Landis, and Umberson 1988; Umberson and Montez 2010). Social relationships consist of potential features such as the level of social integration, the quality of the social relationship, and various social network characteristics (Umberson and Montez 2010). These features detail three significant components of the association: existence of social relationships, the nature of those relationships, and the context within which those relationships are embedded. First, those individuals who are less socially isolated and more socially integrated have better health outcomes (House et al. 1988). While social isolation simply means the lack of social relationships, social integration is defined as the “level of involvement” with informal and formal relationships (Umberson and Montez 2010). Informal relationships can include marriage or family structure while formal ones can include participation or membership in various religious or volunteer organizations.

Second, the nature of social relationships can produce both favorable and unfavorable health outcomes (Baller and Richardson 2009; Haynie 2001; House et al. 1988). They can be positive in that social and emotional support is offered through these social relationships that can help buffer stressful life events (Pearlin 1989; Pearlin et al. 1981; Wethington and Kessler 1986). Additionally, they can present opportunities for the social control, or regulation, of unhealthy behavior and the social learning of healthy ones (Umberson 1987). However, social relationships can also act as significant sources of role strain and role conflict. For instance, while marriage in general can provide social integration, a bad marriage reduces physical health by compromising immune and endocrine systems (Umberson et al. 2006). Additionally, in the same way that social relationships serve as pipelines for positive health resources, they can present opportunities for the social learning of unhealthy behavior. This aspect of negative influence has been researched extensively using explanations of differential association theory (Sutherland 1947) and social learning theory (Akers et al. 1979).

Finally, social relationships are embedded within social contexts or environments. These can refer to structural units such as the classroom or school within which students are nested (Entwisle et al. 2007), or it can be the web of social ties that ultimately form into a social network (Umberson and Montez 2010). The examination of contextual effects on individual phenomena has dramatically increased in the past several decades (Entwisle et al. 2007). For instance, researchers adopt census demographic measures at the neighborhood level to examine contextual effects on individual-level self-reported health while controlling for other individual-level covariates. Using multilevel models (Raudenbush and Bryk 2002), researchers are then interested in the effects of neighborhood above and beyond individual covariates (Blau 1960). Additionally, cross-level interactions may be examined to test whether variation in individual effects exists from one neighborhood to another. While this is an empirically rich area of sociological research, this same methodological investigation rarely exists when examining the social interactional aspect of social context (Entwisle et al. 2007) due in large part to the relative lack of large-scale social network data.

Social Contagion of Depression

More recently as a result of the increase in the availability of social network data, an emerging literature focuses on the network autocorrelation of depression (Cheadle and Goosby 2012; Conway et al. 2011; Hogue and Steinberg 1995; Prinstein 2007). This phenomenon can be traced to three social causes: homophily, contagion, and shared environment effects (Christakis and Fowler 2013). Homophily refers to the idiom of “birds of a feather flock together.” Individuals form ties based on similarities in their level of depression. The second cause of network autocorrelation involves the spread of phenomena among individuals within a social network, or what is often referred to as social contagion (Rosenquist, Fowler, and Christakis 2011). The final cause of network autocorrelation is shared environments, essentially spurious effects from the individuals attending the same school or residing in the same community.

Since these effects are often conflated with each other (Shalizi and Thomas 2011), many studies have utilized stochastic actor-based models, using the RSIENA package developed by Snijders, Van de Bunt, and Steglich (2010), to simultaneously examine the effects of homophily and contagion. A basic assumption of regression analysis is that actors in the analysis are independent from one another, resulting in uncorrelated errors. Actor-based models depart from this potentially “reductionist” assumption and directly analyze interactions between actors and behaviors. Moreover, whereas longitudinal regression models estimate the effects of network predictors at time t on the behavioral outcome at time t + 1, adjusting for the lagged effect of the behavioral outcome at time t, actor-based models simultaneously estimate network and behavioral change. Therefore, actor-based models are able to directly analyze the effects of network on behavior and potential feedback loops (El-Sayed et al. 2012). Despite the advantages, SIENA models require complete and longitudinal network data that are seldom available. For instance, although the Add Health survey collected complete network data for all 140 schools in the core probability survey in the Wave 1 in-school questionnaire, only 16 of these schools, referred to as the “saturated” oversample, have complete network data in subsequent waves. Additionally, only nine schools within the saturated oversample prove suitable for actor-based models (for further discussion, refer to Cheadle and Goosby 2012). As a result, it is apparent that longitudinal regression models remain useful in analyzing currently available, large-scale sample surveys.

Regardless of whether research utilizes actor-based models or longitudinal regression analysis, studies consistently show support for the effect of contagion in explaining similarities of depression in connected individuals (Cheadle and Goosby 2012; Conway et al. 2011; Hogue and Steinberg 1995; Prinstein 2007; Stevens and Prinstein 2005; Zalk et al. 2010a, 2010b). For instance, Cheadle and Goosby (2012) utilize stochastic actor-based models to analyze contagion and homophily in seven small schools. The authors find evidence supporting peer contagion effects of depression above and beyond social selection and social exclusion based on individual levels of depression. Conway et al. (2011) also find support for peer contagion effects using a multilevel model of more than 600 adolescents nested within peer groups.

Furthermore, researchers find that the effect of depressive contagion is moderated by several characteristics. Prinstein (2007) presents a theoretical typology of these potential moderators: target-oriented, prototype, relationship-oriented, and contextual. Target-oriented moderators refer to characteristics of the individual being socially influenced. For instance, Hogue and Steinberg (1995) find that peer contagion effects are stronger for male adolescents. Cheadle and Goosby (2012) echo this finding using actor-based models. On the other hand, prototype moderators refer to characteristics of the individual exerting influence on the target. Prinstein (2007) finds that male adolescents are more susceptible to peer influence in depression by friends who are perceived to be popular. Relationship-oriented moderators refer to characteristics of the relationship between the target and prototype such as strength of ties or directionality. In a study of almost 400 adolescents, Stevens and Prinstein (2005) find support for peer contagion of depression only between adolescents who were characterized as best friends.

The final type of moderator concerns the environment within which the contagion effect is embedded. Unfortunately, research on contextual moderators of depressive contagion is largely nonexistent. To our knowledge, only one study examines a contextual moderator of depressive contagion. In a study of more than 600 adolescents, Conway et al. (2011) find that popularity, measured as received nominations, is a significant moderator of depressive contagion. The authors find that popular adolescents were less susceptible to depressive contagion. According to Conway et al. (2011), this finding suggests that popular students are less likely to conform to group norms given their already high level of status. On the other hand, less popular students are more motivated to conform in order to achieve greater status. Although this explanation is conceivable, research on contagion of other phenomena suggests the opposite relationship (Aloise-Young, Graham, and Hansen 1994; Haynie 2001; Urberg et al. 2003). Popular students may instead be more susceptible to peer influence due to either a greater exposure to contagion or an increased motivation to maintain their status.

Consequently, it appears that more research on contextual moderators of depressive contagion is necessary. We suggest a potentially fruitful area of investigation involving structural characteristics of the microsocial context and the school-level macrosocial context. This suggestion is motivated by prior theoretical research asserting that social relationships are embedded within a web of other relationships that ultimately form a social interactional structure (Umberson and Montez 2010). As a result, we suggest that peer contagion of adolescent depression does not exist in a vacuum but rather within an environment with differential characteristics that can have significant influences on the saliency of peer contagion.

The Present Study

The following analysis examines contextual moderators of adolescent depressive contagion. First and foremost, we intend to replicate prior research on depressive contagion using a multilevel random effects model. Therefore, we expect that the average level of depressive symptoms among an adolescent’s peer network is significantly associated with the adolescent’s own level of depressive symptoms even after controlling for the potential effects of formal and informal social relationships and basic sociodemographic characteristics.

The next series of hypotheses concern the primary focus of this study: contextual moderators of depressive contagion. We operationalize context as two separate levels consisting of the microsocial, local ego-centric peer networks and the macrosocial, school-level structural characteristics (Dishion 2013). First, we test several hypotheses regarding interactions between the average level of depressive symptoms among friends and microsocial density, in-degree centrality (number of received friendship nominations), and out-degree centrality (number of sent friendship nominations). We expect that depressive contagion is more salient in dense peer networks as it allows for more indirect social ties, which in turn provides reinforcement of peer group norms (Conway et al. 2011). Moreover, we test the moderating effects of in-degree and out-degree centrality on depressive contagion. We do not hypothesize a specific direction for the association as prior research has shown varying results for in-degree centrality (Aloise-Young et al. 1994; Conway et al. 2011; Haynie 2001; Urberg et al. 2003) and no study to our knowledge has examined the effect of out-degree centrality.

Second, we test hypotheses regarding cross-level interactions between friends’ average level of depressive symptoms and school mean level of depressive symptoms, network size, density, and mutuality. Due to the lack of prior literature on school-level moderators of peer contagion, we do not suggest any one directional moderation over another. However, we suspect that depressive contagion would be more salient in schools characterized by small, dense networks with high levels of mutuality. Moreover, we suspect that the effect of depressive contagion is more notable in schools with higher average levels of depressive symptoms. These schools essentially provide reinforcement of peer group norms but at a higher level within which peer groups are nested.

Data

The analysis in this study uses data from the National Longitudinal Study of Adolescent Health (Add Health). This study follows a nationally representative cohort of adolescents beginning in grades 7 to 12. Adolescents were chosen based on a step-by-step stratified sampling process. First, high schools around the country were sampled based on region, urbanicity, size, public/private designation, and ethnicity. Qualifying schools were required to have an 11th grade and a school enrollment of 30 or more students. Subsequently, feeder schools for each high school were also identified. If high schools had multiple feeder schools, a feeder school was randomly selected based on a probability proportional to the feeder school’s enrollment contribution to the high school. The final sample consists of 80 high schools and 52 feeder schools nested in 80 communities.

The current analysis utilizes the first and second waves of the Add Health survey including a Wave 1 in-school questionnaire (n = 90,118), Wave 1 in-home interview (n = 20,745), both conducted between September 1994 and December 1995, and Wave 2 in-home interview (n = 14,738), conducted between April and August 1996. Social network variables in this analysis are constructed using friendship nominations gathered from Wave 1 in-school questionnaire. Students in each school were asked to list up to five male friends and five female friends in the in-school questionnaire.

The sample selection for this analysis is first restricted by valid data on the dependent variable measured in the Wave 2 in-home interview (n = 14,662), resulting in the elimination of 76 cases. The second restriction required cases to have valid Wave 1 network data including friendship nominations and school-level constructed network characteristics. In essence, this required students to be a part of a school that had at least a 50 percent completion rate for the Wave 1 in-school questionnaire. This restriction further limited our sample to 9,971 students nested within 121 schools. Due to the multilevel nature of this study, we further limited our sample to schools with at least 30 students. Finally, we utilize Wave 2 sample weights for both the individual and school levels to account for the complex sampling design of the Add Health survey (cases with missing sample weights were eliminated). Our final sample consists of 9,580 students nested within 112 schools with a range from 30 to 836 students per school and an average of 85.5. Missing data on independent variables are imputed using multiple imputation methods in Stata SE 13 (StataCorp 2013).

Analytical Strategy

Given the multilevel nature of our hypotheses, we use a multilevel mixed effects model (Raudenbush and Bryk 2002). Statistically, this allows us to investigate two potential effects of the environment: (1) the effect of school on the average level of adolescent depressive symptoms and (2) the effect of school on the relationship between friends’ depressive symptoms and the ego’s. The former refers to variation in the intercept of the second-level equation (random intercepts model), while the latter refers to variation in slopes of the second-level equation (random slopes model). Additionally, it allows us to test cross-level interactions between individual- and school-level variables in order to explain between-school variation in depressive contagion.

Second, we utilize longitudinal methods in predicting the level of depressive symptoms measured in Wave 2 using independent variables measured in the Wave 1 surveys while controlling for the level of depressive symptoms measured at Wave 1. Finally, we address the complex sampling design of the Add Health survey by including sample weights throughout the analyses at both the individual and school levels. This allows our sample to be representative of the overall U.S. population.

Measures

The outcome of interest in this analysis is respondents’ level of depressive symptoms at the time of the Wave 2 in-home interview. The predictor variable in question is network depression, measured by the average level of depressive symptoms among alters in the ego’s social network. We also control for prior levels of depressive symptoms, formal and informal social relationships, and basic demographic variables including age, sex, race, parent education, and parent income. Descriptive statistics of all variables included in the analyses can be seen in Table 1. The following sections describe the construction of each of these variables.

Table 1.

Descriptive Statistics of All Variables in the Analysis (N = 9,580 in 112 Schools).

Variables	Mean (Percentage)	SD	Minimum	Maximum
Wave 2 depressive symptoms^a	10.9	7.4	0	56
Wave 1 depressive symptoms	10.9	7.4	0	54
Friends’ depressive symptoms	11.0	5.8	0	50
School-level context
Mean depressive symptoms	11.0	1.4	6.3	14.8
Network size	1,001.4	580.5	44	2,559
Network density	.01	.02	.01	.13
Network mutuality	.38	.05	.23	.53
Ego-centric context
Density^b	.30	.15	.03	1
In-degree centrality^b	4.5	3.7	0	30
Out-degree centrality^b	4.4	3.0	0	10
Informal social relationships:
Single-parent household	.24		0	1
Other household	.05		0	1
Parent attachment	8.2	5.0	−16	16
School commitment	2.3	.67	0	3
Grade point average	2.8	.75	1	4
Formal social relationships
Religious youth groups	1.1	1.3	0	3
Extracurricular activities	.52		0	1
Sports team	.56		0	1
Individual-level controls
Age	15.2	1.5	11	20
Female	.50		0	1
Black	.18		0	1
Asian	.06		0	1
Hispanic	.14		0	1
Other race/ethnicity	.01		0	1
Parent income^a	46.5	.54	0	999
Parent education	13.3	.54	0	18
School-level controls
Private^c	.02		0	1
Catholic^c	.04		0	1
Rural^d	.12		0	1
Suburban^d	.59		0	1

Logged in the analysis.

Within-school standardized in the analysis.

Public school as reference category.

Urban as reference category.

Dependent Variable

During the Wave 2 interview, respondents were presented with 19 statements roughly corresponding to the CES-D scale used in order to measure depressive symptoms in the general population (Radloff 1977). The original CES-D scale consists of 20 statements, 17 of which correspond to statements included in the Add Health data set. Perreira et al. (2005) provide a more in-depth discussion on the differences between the two depression scales.

In response to the statements, the adolescents were then asked to gauge how often they agreed with the statements in the past seven days. Answer choices included four options: never or rarely, sometimes, a lot of the time, and most of the time or all of the time. We code these responses from 0 to 3, respectively, and a factor analysis indicates a single factor structure. The scores are thus aggregated to produce a final measure of depressive symptoms ranging from 0 to 56 with an alpha level of .859. Due to the skewed nature of the CES-D scale, we log-transform the dependent variable in order to satisfy normality assumptions in regression analysis.

Network Context of Depressive Contagion

Measures of the adolescent social network and network depressive symptoms are constructed using the friendship nomination data gathered during the Wave 1 in-school questionnaire. Adolescents were given the opportunity to nominate up to five male friends and five female friends. The nominations utilized in these analyses are limited to friendship nominations sent to and/or received from alters who attend the same school as the ego. First, we construct the measure of network depressive symptoms using the average of all CES-D indices among alters. This includes alters nominated by the ego and alters that nominated the ego. The classification of individuals as alters does not require reciprocation in nominations. Then, if ego has two alter friendships with CES-D scores of 13 and 17, the ego’s “friends’ depressive symptoms” would be coded as 15.

In order to contextualize depressive contagion, we utilize several social network measures at both the microsocial and macrosocial levels (Dishion 2013). We operationalize the microsocial level as the immediate network of alters that are directly connected to the ego and the macrosocial level as the school-level network in which the ego attends. At the microsocial level, we include measures of network density, in-degree centrality, and out-degree centrality. Network density is measured as the number of dyads that exist within an ego network divided by the total possible number of dyads. In-degree centrality is measured as the number of nominations an ego received from other individuals in the school. Out-degree centrality is measured as the number of nominations the ego sent out to other individuals in the school. Each of these ego-centric measures are standardized according to the means of each school. This enables our measures to capture ego-centric network characteristics in relation to each school’s network potential.

At the macrosocial level, we include three measures: network size, density, and mutuality. First, network size is measured as the total number of students within the school. Second, we measure density as the number of dyads that exists in the school friendship network divided by the total of potential dyads that could possibly exist in the network. This provides an indication of connectedness within a given school network. A network with high density can mean that there is relatively little clustering of social ties in the network. Low density networks can signify either of two possibilities: a high level of clustering or a high level of isolation of individuals. Finally, network mutuality is measured using a mutuality index developed by Katz and Powell (1955). This index measures the tendency for nodes within the network to reciprocate nominations. Additionally, we include a non-network contextual variable of average level of depressive symptoms measured at the school level.

Formal and Informal Social Relationships

In examining adolescent depressive symptoms, it is important to control for the effects of social relationships due to the potential support resources that can come from greater integration into the family, school, and community. We utilize three measures of formal social relationships: religious participation, extracurricular participation, and sports participation. We intend for these measures to account for the social resources these activities may provide in the same way that religious and volunteer organizations would provide social resources for adults.

First, we use a measure of athletic participation and non-athletic extracurricular participation in the school coded 0 for nonparticipating and 1 for participating in at least one activity. Second, we include a measure of religious participation to account for potential social resources coming from outside of the school. Students were asked how often they attended youth activities within a religious context such as youth groups, Bible classes, or choir. Answers were coded as 0 for never, 1 for less than a month, 2 for once a month or more, and 3 for once a week or more.

To account for more informal social relationships, we utilize measures of household structure, parental attachment, school commitment, and grade point average. Household structure is measured using self-reported answers of whether students lived with their mother, father, both, or neither. Variables of single-parent and other households are constructed using living with both parent as the reference category. To measure school commitment, we use the question asking students, “How hard do you try to do your school work well?” Answer choices were coded from 0 (never try at all) to 3 (try very hard).

Parental attachment is measured using an eight-item scale of questions asking students about their relationship with their parents. For example, students were asked whether they agreed with the statement, “Most of the time, your mother is warm and loving to you.” The students were given the option to strongly disagree, disagree, neither agree nor disagree, agree, and strongly agree. We coded these answer choices from −2 to +2, respectively. Finally, to calculate respondents’ GPA, self-reported grades for English, math, science, and history during the Wave 1 in-home interview were averaged together to generate a scale between 1 and 4 (the answer choices stopped at D or lower).

Demographic Variables

Age is computed using suggestions by the Add Health codebook. Respondents were only asked for their birth month and year, so therefore the 15th of each month was used to calculate the respondent’s age. For the variable sex, females were coded as 1, and males were coded as 0. For race, respondents were coded as either black, Asian, Hispanic, or other using white as the reference category. During the in-home interview, respondents were given the option to mark multiple racial and ethnic categories. Therefore, to construct the race dummy variable, we followed Add Health’s procedure documented in the codebook for the constructed network variables data set. To achieve mutual exclusivity, priority was given to Hispanic, Asian, black, white, and other in that order.

Furthermore, Add Health also conducted an interview with one of the parents or guardians during the Wave 1 in-home interview. Two measures, parent income and parent education, were drawn from these interviews for this analysis. Parent income is measured as annual income in thousands of dollars and logged in order to achieve normality. Highest level of parent education is originally measured as a categorical variable with responses such as high school graduate, some college, and college graduate. We recode high school graduate to 12, some college to 14, college graduate to 16, and so on. In addition to demographic variables at the individual level, we control for two measures at the school level indicating whether the school is public (reference), private, or Catholic and whether the school is located in an urban (reference), suburban, or rural location.

Results

Depressive Contagion

As a preliminary step, we investigate the effect of Wave 1 network depressive symptoms on Wave 2 ego depressive symptoms using a random intercepts model. Results in Model 1 of Table 2 show a contagion effect indicating that the average level of depressive symptoms among alters is significantly associated with ego depressive symptoms above and beyond the effects of control variables and social relationship variables. Models 2 through 4 present a step-wise introduction of macrosocial and microsocial context variables. Several findings are of note in these models. First, friends’ depressive symptoms remain a significant predictor of adolescent depressive symptoms even after adjusting for the main effects of microsocial and macrosocial variables. Second, none of the macrosocial variables are significant predictors of adolescent depressive symptoms. In fact, even after entering these school-level variables, there is no statistical reduction in the variance of the intercept (.007 in both Model 1 and 2). However, microsocial contextual variables of in-degree centrality and out-degree centrality are significant predictors of adolescent depressive symptoms.

Table 2.

Mixed Effects Regression Model: Influence of Friends’ Depressive Symptoms on Adolescent Ego Depressive Symptoms.

	Model 1		Model 2		Model 3		Model 4
	β	SE	β	SE	β	SE	β	SE
Intercept	1.936***	.128	2.095***	.171	1.957***	.127	2.053***	.181
Friends’ depressive symptoms	.004**	.002	.004**	.002	.004**	.002	.008*	.003
School-level context
Mean depressive symptoms			−.008	.010			−.007	.009
Network size			−.000	.003			.001	.003
Network density			−.677	.431			−.291	.333
Network mutuality			−.185	.229			−.178	.231
Ego-centric context
Density^a					−.001	.008	−.004	.009
In-degree centrality^a					.023*	.009	.022*	.010
Out-degree centrality^a					−.034***	.010	−.034***	.010
Informal social relationships
Single-parent household	.000	.022	.001	.022	−.001	.022	.002	.022
Other household	.027	.039	.026	.039	.024	.040	.023	.039
Parent attachment	−.013***	.002	−.013***	.002	−.013***	.002	−.014***	.002
School commitment	−.018	.013	−.018	.013	−.015	.013	−.015	.014
Grade point average	−.045***	.014	−.045***	.014	−.045***	.014	−.046***	.014
Formal social relationships
Religious youth groups	−.015	.009	−.015	.009	−.015	.009	−.015	.009
Extracurricular activities	−.020	.019	−.020	.019	−.016	.018	−.014	.019
Sports team	−.004	.018	−.004	.018	−.003	.018	−.004	.017
Individual-level controls
Wave 1 depressive symptoms	.047***	.002	.047***	.002	.047***	.002	.047***	.002
Age	.009	.006	.009	.007	.007	.007	.007	.007
Female	.079***	.019	.079***	.019	.082***	.019	.080***	.020
Black	.061*	.027	.059*	.028	.056*	.027	.056	.029
Asian	.166**	.056	.165**	.055	.163**	.055	.163**	.059
Hispanic	.121***	.030	.121***	.030	.118***	.031	.110***	.032
Other race/ethnicity	.160*	.073	.161*	.073	.166*	.073	.178*	.071
Parent income (logged)	−.029*	.012	−.029*	.012	−.029*	.013	−.027*	.012
Parent education	−.003	.004	−.003	.004	−.003	.004	−.003	.004
School-level controls
Private	.049	.048	.085	.053	.067	.047	.091	.053
Catholic	.022	.021	.051	.026	.022	.021	.029	.022
Rural	−.064	.032	−.035	.033	−.055	.032	−.031	.034
Suburban	−.023	.023	−.013	.023	−.026	.023	−.014	.026
Random effects
var(network depression)							.004**	.002
var(intercept)	.007***	.001	.007***	.001	.007***	.001	.050***	.020
var(residual)	.361***	.011	.361***	.011	.360***	.011	.360***	.011

Note. School level N = 112; individual level N = 9,580.

Within-school standardized.

p < .05. **p < .01. ***p < .001 (two-tailed).

In interpreting the coefficients in Table 2, recall that the dependent variable, adolescent depressive symptoms, in this analysis is log transformed. Therefore, according to Model 4, a one unit increase in friends’ level of depressive symptoms results in a .8 percent increase in adolescent level of depressive symptoms after adjusting for sociodemographic controls, social relationship, and contextual variables. Alternatively, with a one standard deviation increase in friends’ level of depressive symptoms (SD = 5.8), adolescent level of depressive symptoms increases by almost 5 percent. This finding suggests significant support for the presence of depressive contagion in adolescent peer networks.

Other notable significant effects include in-degree and out-degree centrality and socioedemographic variables. First, adolescents who sent more nominations are less depressed, but adolescents who received more nominations are more depressed. As received nominations is often used as a measure of popularity in adolescent networks, it is interesting that popular individuals are more depressed than unpopular individuals. This finding can potentially be explained using Falci and McNeely’s (2009) research suggesting that despite the common belief that social isolation causes emotional distress, having too many friends can also be harmful for adolescent mental health. Too many friends can result in role strain and role conflict. Additionally, results in Model 4 suggest large effect sizes for female, Asian, and Hispanic adolescents: roughly 8 percent increase, 16 percent increase, and 11 percent increase in adolescent level of depressive symptoms, respectively.

Contextual Moderators

The results in Table 3 concern the primary analysis in this study investigating whether the effect of friends’ level of depressive symptoms on adolescent level of depressive symptoms varies dependent on microsocial and macrosocial contexts. To examine this, we run a series of interaction models shown in Table 3. Models 5 through 7 enter microsocial density, in-degree centrality, and out-degree centrality one at a time to examine whether each of these measures significantly moderates the effect of depressive contagion. Model 8 examines the microsocial interactions simultaneously. Similarly, Models 9 through 11 enter macrosocial factors of school mean level of depressive symptoms, network size, density and mutuality, respectively, while Model 12 examines the macrosocial interactions simultaneously. Finally, Model 13 examines whether each of these interaction effects remains significant when combined into a single model.

Table 3.

Mixed Effects Regression Model: Examining Ego-centric and School-level Contextual Moderators of Depressive Contagion.

	Model 5	Model 6	Model 7	Model 8	Model 9	Model 10	Model 11	Model 12	Model 13	Model 14
Main effects
Friends’ depressive symptoms	.008* (.003)	.009** (.003)	.008* (.003)	.009** (.003)	.049** (.018)	.014* (.007)	−.003(.003)	−.047* (.019)	−.033(.024)	−.033(.024)
Microsocial context
Ego-centric density	−.032(.019)	−.003(.009)	−.004(.009)	−.053** (.019)	−.004(.009)	−.004(.009)	−.003(.009)	−.003(.009)	−.003(.009)	−.053** (.019)
In-degree centrality	.021* (.010)	−.016(.023)	.022* (.010)	−.032(.025)	.022* (.010)	.022* (.010)	.022* (.010)	.022* (.010)	.022* (.010)	−.032(.025)
Out-degree centrality	−.034*** (.010)	−.034*** (.010)	−.039* (.018)	−.042* (.020)	−.034*** (.010)	−.034*** (.010)	−.034*** (.010)	−.034*** (.010)	−.034*** (.010)	−.042* (.020)
Macrosocial context
Mean depression	−.006(.009)	−.008(.009)	−.007(.009)	−.007(.009)	.035(.020)	−.007(.009)	−.009(.010)	−.008(.009)	−.004(.021)	−.004(.021)
Network size	.000(.003)	.000(.003)	.000(.002)	.000(.003)	.000(.003)	.015(.009)	.001(.003)	.001(.003)	−.000(.000)	−.008(.007)
Network density	−.251(.314)	−.289(.324)	−.271(.319)	−.274(.319)	−.189(.355)	−.195(.312)	−5.26* (2.17)	−.179(.333)	−5.52* (2.55)	−5.49* (2.52)
Network mutuality	−.194(.231)	−.209(.232)	−.200(.230)	−.200(.234)	−.177(.232)	−.198(.232)	−.170(.240)	−1.75** (.651)	−1.01* (.480)	−1.05* (.480)
Interactions
Microsocial context
Ego-centric density	.003* (.001)			.005*** (.001)						.005*** (.001)
In-degree centrality		.004* (.002)		.005* (.002)						.005* (.002)
Out-degree centrality			.000(.001)	.001(.002)						.001(.002)
Macrosocial context
Mean depressive symptoms					−.004* (.002)				−.000(.002)	−.001(.002)
Network size						−.002* (.001)			.001(.001)	.001(.001)
Network density							.504** (.191)		.530* (.225)	.528* (.222)
Network mutuality								.144** (.055)	.078* (.038)	.082* (.038)

p < .05. **p < .01. ***p < .001 (two-tailed).

Findings in Table 3 suggest that of the three microsocial moderators tested, ego-centric density and in-degree centrality are significant moderators of depressive contagion. We illustrate these results in Figures 1 and 2 showing the effects of friends’ level of depressive symptoms on ego level of depressive symptoms at varying levels of ego-centric density and in-degree centrality. These effects are shown while holding all other covariates at their overall means. High levels of microsocial context are operationalized as 1.5 standard deviations above the mean and low levels as 1.5 standard deviations below the mean. Moreover, we set the x-axis in these to include only values within approximately two standard deviations of the average level of depressive symptoms among adolescents’ friends (mean = 11, SD = 5.8). In other words, approximately 95% of adolescents in the sample are represented within the figures.

Figure 1.

Ego-centric density by friends’ depressive symptoms.

Figure 2.

In-degree centrality by friends’ depressive symptoms.

As shown in Figure 1, the effect of network depressive symptoms is exacerbated when embedded in dense, cohesive peer networks. Although the statistical results indicate a moderation of microsocial density, Figure 1 shows that the difference in the effect of depressive contagion between an adolescent embedded in a high density peer group and a low density peer group is relatively modest. For instance, as friends’ level of depressive symptoms increases from 0 to 25, more than four standard deviations, adolescent depression in low density networks increases by .93 compared to adolescent level of depressive symptoms in high density networks, which increases by approximately 2.8. In relative terms, adolescents embedded in high density networks are three times more susceptible to depressive contagion. However, in response to a four standard deviation change in friends’ depression, the response for high density networks only represents less than half a standard deviation in adolescent’s own level of depressive symptoms.

In contrast to the modest moderating effects of microsocial density, the results in Model 6 show that in-degree centrality is a significant contextual moderator of depressive contagion. This finding is illustrated in Figure 2. According to the finding, students with high in-degree centrality are especially susceptible to depressive contagion. At low levels of friends’ level of depressive symptoms, there is essentially no difference in levels of depressive symptoms between adolescents with low and high in-degree centrality. However, as friends’ level of depressive symptoms increases, popular students experience higher levels of depressive symptoms while less popular students experience no real change. More specifically, the effect size of depressive contagion for popular adolescents is five times as large as that for unpopular adolescents.

The next analysis examines the variation of depressive contagion dependent on macrosocial contextual moderators operationalized as school-level characteristics. The first step is to test whether the effect of network level of depressive symptoms varies according to school. We do this by entering the variable network level of depressive symptoms as a random effect in Model 4 on Table 2. To examine the significance of the random effect, we perform two statistical tests. First, a Wald statistic is generated by dividing the variance estimate by its associated standard error. As shown in Model 4, this statistic is significant at the .01 level. Second, we conduct a likelihood ratio (LR) test of whether there is a statistically difference between Model 4 with random coefficients and Model 4 without random coefficients. The LR test indicates that the random coefficients model is statistically significant at the .001 level, LR χ²(2) = 100.97, p < .000. The next step is to explain this variation through a series of cross-level interactions between school-level network characteristics and network level of depressive symptoms. Models 9 through 12 show the results to these interaction effects.

Models 9 and 10 examine the roles of schools’ mean level of depressive symptoms and network size as moderating effects of depressive contagion. Although these models each reveal a statistically significant interaction term, both terms become null after entering all macrosocial moderators simultaneously in Model 13. The two macrosocial interaction coefficients that remain significant in Model 13 are network density and network mutuality. These effects are illustrated in Figures 3 and 4. As shown in Figure 3, friends’ level of depressive symptoms increases adolescent depression in high density schools. In average density schools, there does not appear to be any real effect of depressive contagion. However, most interestingly, friends’ level of depressive symptoms in low density schools actually decreases adolescents’ own level of depressive symptoms. This finding suggests important implications for future research as it demonstrates a significantly differing effect of depressive contagion dependent on a higher level contextual measure. For instance, the only other empirical study of contextual moderators of depressive contagion focuses only on peer group contexts. This finding in particular indicates the importance of examining contextual moderators at multiple nested levels.

Figure 3.

School network density by friends’ depressive symptoms.

Figure 4.

School network mutuality by friends’ depressive symptoms.

The final macrosocial moderator tested is network mutuality shown in Model 12. Results show that network mutuality exerts a significant moderating effect on depressive contagion. To better illustrate the effect, Figure 4 shows that depressive contagion is essentially nonexistent in schools with relatively low mutuality indices (holding all other variables at their means). However, the effect of depressive contagion is significantly greater and positive in high mutuality schools. Implications of these results are discussed in the next section.

Discussion

The literature on the association between social relationships and health has examined relational mechanisms that lead to both positive and negative health outcomes. Researchers in support of the former cite greater social integration and social control from risky behaviors as resources provided by having social relationships (House et al. 1988; Umberson and Montez 2010). Through these relationships, individuals are provided opportunities to socially learn (Akers et al. 1979) healthy behaviors. However, in this same process where healthy phenomena travel through the pipelines of social relationships, unhealthy phenomena can do the same. This hypothesized contagion effect where peer behaviors influence one’s own behavior has been historically studied in the branches of deviance (Akers et al. 1979; Sutherland 1947) and can even theoretically be traced back to Durkheim’s ([1897] 1951) discussion of imitation in his famous work Suicide.

In this study, we examine contextual moderators of adolescent depressive contagion. We operationalize contextual moderators using two separate levels: the microsocial and macrosocial. The results of this analysis demonstrate support for the effect of depressive contagion in adolescent networks. This finding demonstrates the robustness of the adolescent depressive contagion thesis suggested in other empirical studies (e.g., Cheadle and Goosby 2012; Conway et al. 2011; Hogue and Steinberg 1995; Prinstein 2007). Although our study utilizes a longitudinal design to control for prior level of depressive symptoms and a multilevel design to control for environmental main effects, evidence of depressive contagion does not imply the absence of a selection effect. In fact, multiple studies using various methodologies have indicated that depression is a significant quality when selecting or deselecting friends (e.g., Cheadle and Goosby 2012; Schaefer, Kornienko, and Fox 2011; Zalk et al. 2010a). However, many of these studies also indicate evidence of depressive contagion above and beyond homophilous effects (Shalizi and Thomas 2011). Moreover, the primary objective of this study is not to engage in whether contagion or homophily represents greater effect sizes for the network autocorrelation of depression. Rather, our study aims at shedding light on the significance of context and furthermore, context at multiple nested levels.

At the microsocial level, our results show that adolescents are more susceptible to depressive contagion when embedded in dense peer networks. Dense networks are characterized by not only the direct ties between ego and the ego’s friends but many indirect ties. These indirect ties can serve two primary purposes. First, the prevalence of indirect ties allow for more pathways through which depressive influence can travel. Second, indirect ties and consequently dense networks are more susceptible to the creation of small-group cultures where depressive behavior can be normalized (Conway et al. 2011). These norms, created through the interactions among ego and its alters, are subsequently internalized by the ego. It is no surprise then that this process is strengthened in conjunction with many social interactions occurring between not simply ego and alter but also alter and alters.

Furthermore, at the microsocial level, popularity plays an important role in moderating depressive contagion. Our findings suggest that popular students are more susceptible to the effects of depressive contagion. This finding may be indicative of the intrinsic personality traits of popular adolescents, or target-oriented factors (Prinstein 2007), rather than the network structure of individuals who receive more friendship nominations. Nonetheless, popular students may be inherently more vulnerable to emotional and behavioral changes of their peer environment. Moreover, these same students may be more likely to engage in and be influenced by peer pressure and adolescent social norms. Prior research shows differing results for the moderating role of popularity in depression. Particularly, Conway et al. (2011) find that less popular students are more susceptible to depressive contagion, a directly opposite result from our study. Conway et al. (2011) suggest that popular students may experience less pressure for conformity. However, other contagion research suggests that popular students are more likely to conform, potentially due to greater source exposures and status motivation (Aloise-Young et al. 1994; Urberg et al. 2003).

At the macrosocial level, our findings suggest a significant variation in the effect of depressive contagion contingent on school characteristics. To explain this variation, we examined four school-level measures: mean level of depressive symptoms, network size, density, and mutuality. We find that after considering all four measures simultaneously, school network density and network mutuality significantly moderate the effect of depressive contagion. This finding is conceivable for two reasons. Since network density is measured as the number of observed social ties divided by the total number of potential ties, a high density network should serve as a particularly vulnerable social environment for depressive contagion. In addition, mutuality accounts for the directionality of these ties. Networks with greater tendencies toward mutuality in friendship nominations should also serve as vulnerable contexts for depressive contagion.

Overall, our findings indicate the importance in examining contextual moderators of depressive contagion. Although this was theorized in Prinstein’s (2007) research almost a decade ago, empirical investigations of contextual moderators are limited (see Conway et al. 2011). Nonetheless, if we are to assume that behaviors or emotions are capable of spreading through social ties, then we must also understand that the structure of these social ties matter. In the same way that a telephone tree is considered an efficient way of diffusing information to a group of individuals, certain network structures are more susceptible to the social contagion of behaviors, emotions, and so on. Certain network structures may be particularly vulnerable to depressive contagion.

Implications

Our findings present several implications for the literature on depressive contagion and more broadly on contagion research in general. First, our findings demonstrate the robustness of the depressive contagion thesis. Despite the increasing use of actor-based models to simultaneously investigate homophily and contagion, future research must move further in order to examine factors that predict varying degrees in the effect of peer influence or selection. Second, our findings contribute to the contagion literature by providing evidence for the significance of macrosocial context. The results suggest that depressive contagion is moderated both at the peer network level and school level. Therefore, future research must view shared environments as a significant contextualizing factor for depressive contagion.

Limitations

Although the findings presented in this study provide significant contributions to the contagion literature, we acknowledge two limitations in our analyses. First, our study examines specifically peer contagion utilizing friendship nominations among adolescents. It is possible that influence comes from sources in addition to peers such as family, relatives, or so on. Second, the friendship nominations are limited to nominations to alters that attend the same school as the ego. As a result, the analysis does not capture the effects from alters that do not attend the same school. This boundary limitation is a common issue in social network research.

Despite these limitations, this study presents a valuable contribution to research on adolescent mental health and social contagion. Adolescence presents a unique stage in life where multiple biological, social, and psychological processes intersect. As individuals navigate through adolescence, sources of influence begin to shift from parental guidance and control to peer group influences due the growth in peer network sizes and increasing time spent within the school. This shift coincides with the salience of social and psychological identity formation and biological outcomes of puberty. The findings in this study suggest the importance in studying adolescent health in combination with social contextual factors.

Footnotes

Acknowledgements

This research uses data from Add Health, a program project directed by Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by grant P01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Special acknowledgment is due Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Information on how to obtain the Add Health data files is available on the Add Health website (). No direct support was received from grant P01-HD31921 for this analysis.

Notes

References

Akers

Ronald L.

Krohn

Marvin D.

Lanza-Kaduce

Lonn

Radosevich

Marcia

. 1979. “Social Learning and Deviant Behavior: A Specific Test of a General Theory.” American Sociological Review 44(4):636-55.

Aloise-Young

Patricia A.

Graham

John W.

Hansen

William B.

1994. “Peer Influence on Smoking Initiation during Early Adolescence: A Comparison of Group Members and Group Outsiders.” Journal of Applied Psychology 79(2):281-87.

Baller

Robert D.

Richardson

Kelly K.

2009. “The ‘Dark Side’ of the Strength of Weak Ties: The Diffusion of Suicidal Thoughts.” Journal of Health and Social Behavior 50(3):261-76.

Blau

Peter M.

1960. “Structural Effects.” American Sociological Review 25(2):178-93.

Cheadle

Jacob E.

Goosby

Bridget J.

2012. “The Small-school Friendship Dynamics of Adolescent Depressive Symptoms.” Society and Mental Health 2(2):99-119.

Christakis

Nicholas A.

Fowler

James H.

2013. “Social Contagion Theory: Examining Dynamic Social Networks and Human Behavior.” Statistics in Medicine 32(4):556-77.

Conway

Christopher C.

Rancourt

Diana

Adelman

Caroline B.

Burk

William J.

Prinstein

Mitchell J.

2011. “Depression Socialization within Friendship Groups at the Transition to Adolescence: The Roles of Gender and Group Centrality as Moderators of Peer Influence.” Journal of Abnormal Psychology 120(4):857-67.

Dishion

Thomas J.

2013. “Stochastic Agent-based Modeling of Influence and Selection in Adolescence: Current Status and Future Directions in Understanding the Dynamics of Peer Contagion.” Journal of Research on Adolescence 23(3):596-603.

Durkheim

Emile

. [1897] 1951. Suicide: A Study in Sociology. Glencoe, IL: Free Press.

10.

El-Sayed

Abdulrahman M.

Scarborough

Peter

Seemann

Lars

Galea

Sandro

. 2012. “Social Network Analysis and Agent-based Modeling in Social Epidemiology.” Epidemiologic Perspectives & Innovations 9(1):1-9.

11.

Entwisle

Barbara

Faust

Katherine

Rindfuss

Ronald R.

Kaneda

Toshiko

. 2007. “Networks and Contexts: Variation in the Structure of Social Ties.” American Journal of Sociology 112(5):1495-533.

12.

Falci

Christina

McNeely

Clea

. 2009. “Too Many Friends: Social Integration, Network Cohesion and Adolescent Depressive Symptoms.” Social Forces 87(4):2031-61.

13.

Haynie

Dana L.

2001. “Delinquent Peers Revisited: Does Network Structure Matter?” American Journal of Sociology 106(4):1013-57.

14.

Hogue

Aaron

Steinberg

Laurence

. 1995. “Homophily of Internalized Distress in Adolescent Peer Groups.” Developmental Psychology 31(6):897-906.

15.

House

James S.

Landis

Karl R.

Umberson

Debra

. 1988. “Social Relationships and Health.” Science 241(4865):540-45.

16.

Joyner

Kara

Udry

J. Richard

. 2000. “You Don’t Bring Me Anything but Down: Adolescent Romance and Depression.” Journal of Health and Social Behavior 41(4):369-91.

17.

Katz

Leo

Powell

James H.

1955. “Measurement of the Tendency toward Reciprocation of Choice.” Sociometry 18(4):403-09.

18.

National Institute of Mental Health. 2007. Children’s Mental Health Awareness: Depression in Children and Adolescents Fact Sheet. Bethesda: National Institute of Mental Health.

19.

Pearlin

Leonard I.

1989. “The Sociological Study of Stress.” Journal of Health and Social Behavior 30(3):241-56.

20.

Pearlin

Leonard I.

Menaghan

Elizabeth G.

Lieberman

Morton A.

Mullan

Joseph T.

1981. “The Stress Process.” Journal of Health and Social Behavior 22(4):337-56.

21.

Perreira

Krista M.

Deeb-Sossa

Natalia

Harris

Kathleen Mullan

Bollen

Kenneth

. 2005. “What Are We Measuring? An Evaluation of the CES-D across Race/Ethnicity and Immigrant Generation.” Social Forces 83(4):1567-601.

22.

Prinstein

Mitchell J.

2007. “Moderators of Peer Contagion: A Longitudinal Examination of Depression Socialization between Adolescents and Their Best Friends.” Journal of Clinical Child & Adolescent Psychology 36(2):159-70.

23.

Radloff

Lenore Sawyer

. 1977. “The CES-D Scale: A Self-report Depression Scale for Research in the General Population.” Applied Psychological Measurement 1(3):385-401.

24.

Rambaran

Ashwin J.

Dijkstra

Jan Kornelis

Stark

Tobias H.

2013. “Status-based Influence Processes: The Role of Norm Salience in Contagion of Adolescent Risk Attitudes.” Journal of Research on Adolescence 23(3):574-85.

25.

Raudenbush

Stephen W.

Bryk

Anthony S.

2002. Hierarchical Linear Models: Applications and Data Analysis Methods. Thousand Oaks, CA: Sage.

26.

Rosenquist

J. Niels.

Fowler

James H.

Christakis

Nicholas A.

2011. “Social Network Determinants of Depression.” Molecular Psychiatry 16(3):273-81.

27.

Schaefer

David R.

Kornienko

Olga

Fox

Andrew M.

2011. “Misery Does Not Love Company: Network Selection Mechanisms and Depression Homophily.” American Sociological Review 76(5):764-85.

28.

Shalizi

Cosma Rohilla

Thomas

Andrew C.

2011. “Homophily and Contagion Are Generically Confounded in Observational Social Network Studies.” Sociological Methods & Research 40(2):211-39.

29.

Snijders

Tom A. B.

Van de Bunt

Gerhard G.

Steglich

Christian E. G.

2010. “Introduction to Stochastic Actor-based Models for Network Dynamics.” Social Networks 32(1):44-60.

30.

StataCorp. 2013. Stata Multiple-imputation Reference Manual Release 13. College Station, TX: StataCorp LP.

31.

Stevens

Elizabeth A.

Prinstein

Mitchell J.

2005. “Peer Contagion of Depressogenic Attributional Styles among Adolescents: A Longitudinal Study.” Journal of Abnormal Child Psychology 33(1):25-37.

32.

Substance Abuse and Mental Health Services Administration. 2013. Results from the 2012 National Survey on Drug Use and Health: Mental Health Findings. Rockville, MD: Substance Abuse and Mental Health Services Administration.

33.

Sutherland

Edwin

. 1947. Principles of Criminology. 4th ed. Philadelphia: J. B. Lippincott.

34.

Umberson

Debra

. 1987. “Family Status and Health Behaviors: Social Control as a Dimension of Social Integration.” Journal of Health and Social Behavior 28(3):306-19.

35.

Umberson

Debra

Montez

Jennifer Karas

. 2010. “Social Relationships and Health: A Flashpoint for Health Policy.” Journal of Health and Social Behavior 51:S54-66.

36.

Umberson

Debra

Williams

Kristi

Powers

Daniel A.

Liu

Hui

Needham

Belinda

. 2006. “You Make Me Sick: Marital Quality and Health over the Life Course.” Journal of Health and Social Behavior 47(1):1-16.

37.

Urberg

Kathryn A.

Luo

Qing

Pilgrim

Colleen

Degirmencioglu

Serdar M.

2003. “A Two-stage Model of Peer Influence in Adolescent Substance Use: Individual and Relationship-specific Differences in Susceptibility to Influence.” Addictive Behaviors 28(7):1243-56.

38.

Wethington

Elaine

Kessler

Ronald C.

1986. “Perceived Support, Received Support, and Adjustment to Stressful Life Events.” Journal of Health and Social Behavior 27(1):78-89.

39.

World Health Organization. 2014. Health for the World’s Adolescents. A Second Chance in the Second Decade. Geneva: World Health Organization.

40.

Zalk

Maarten Herman Walter van

Kerr

Margaret

Branje

Susan J. T.

Stattin

Hakan

Meeus

Wim H. J.

2010a. “It Takes Three: Selection, Influence, and De-selection Processes of Depression in Adolescent Friendship Networks.” Developmental Psychology 46(4):927-38.

41.

Zalk

Maarten Herman Walter van

Kerr

Margaret

Branje

Susan J. T.

Stattin

Hakan

Meeus

Wim H. J.

2010b. “Peer Contagion and Adolescent Depression: The Role of Failure Anticipation.” Journal of Clinical Child & Adolescent Psychology 39(6):837-48.