Similarity of Friends Versus Nonfriends in Adolescence: Developmental Patterns and Ecological Influences

Abstract

This study examines developmental change across adolescence in the similarity of friends versus nonfriends. This differential in similarity is a key aspect of the organization of the peer context of development: The stronger the correlation between friends for an attribute, the more the attribute delineates clustering and divisions of friendships. We investigated change in the correlation between friends across twelve attributes covering demographics, orientations to key institutions (family, school, religion), and problem behavior, and we expected that the link between similarity and friendship would increase during adolescence for most attributes other than gender. We also predicted that the social ecological factors of school size and attribute variability would be associated with stronger correlations between friends and partially mediate developmental change. Data are from two grade cohorts of 27 small school districts, followed from 6^th through 11^th grades (N = 454 time-specific networks and over 65,000 person/waves of data; 84.2% White, 6.8% Hispanic/Latino, 3.2% African-American, 1.3% Asian, .5% Native American, 3.9% other or multiple). The data analysis takes the form of a three-level random effects meta-analysis of network level correlations between friends (Moran’s I). As expected, declining dominance of gender was offset by the emergence of moderate correlations across a broader profile of attributes. The ecological opportunity factors of grade cohort size and attribute variability significantly mediated these increases in correlations between friends, accounting for 23% to 73% of age-related change for 10 of the 11 attributes other than gender.

Through their friendship choices and daily interactions, children and adolescents create a dynamic social world that serves as the peer context of development. The present study examines the progression of that social world across adolescence in terms of the shifting profile of differences in similarity between friends and nonfriends. The interplay between individual development and social contexts is a cornerstone of prominent developmental perspectives. Bronfenbrenner’s ecological view of development, for example, highlights not only interactions between individuals and their settings, but also how these interactions change with age (1979; Bronfenbrenner & Morris, 1998). Our study moves from the usual developmental emphasis on individual change to focus instead on age-related change in a social context, specifically the social organization of peers. Notably, this context is itself defined by the aggregate behavior of the developing individuals. Our results show that the organization of peer relationships, specifically similarity between friends versus nonfriends, systematically changes across adolescence. Furthermore, we demonstrate that this change is a function both of school size, which is an externally imposed aspect of adolescent life, and of developmental change in distributions of individual attributes.

The ecology of adolescent development comprises multiple, interconnected settings at several levels of aggregation (Bronfenbrenner 1979; Bronfenbrenner & Morris, 1998). Our interest is the setting of peers, and abundant research has demonstrated the developmental significance of peers through processes such as social influence and acceptance (e.g., Brown & Larson, 2009). A large share of this research concerns dyads of two people who are connected in a relationship and thereby constitute a simple and immediate context for each other. The set of dyadic relationships within a population, in turn, defines a social network. Accordingly, the concepts and methods of social network analysis are well-suited to studying the emergent properties of the larger peer setting. As Wellman (1988) noted, social networks provide an avenue for studying social relationships as connections within a structure, as opposed to mere averages of independent attributes. Our interest lies in understanding the larger pattern of peer relations that characterize who is friends with whom, which in turn constitutes the input to the proximal developmental processes that occur between individual peers.

The current study examines the organization of adolescents’ friendships within their grade cohorts at school. In Western, industrialized society, school is the dominant setting of public life from middle childhood through adolescence, and it is the primary source of friendships beyond family associates and the neighborhood (McPherson et al., 2001; Shum et al., 1988). Not only do schools provide education and instill societal values, but as scholars such as Coleman (1961), Brown (1990), and Crosnoe (2011) have shown, school contexts also play a prominent role in adolescents’ social lives. The students who constitute a grade cohort at a school spend hundreds of hours in a shared setting each year, defining a community of peers that serves as a pool of potential friends and an unavoidable audience for one another’s behavior and reputations. Thus, the school becomes the primary locus of organization in adolescents’ social lives, hosting the peers and norms that become increasingly important to adolescents’ self-identity.

Most developmental peer research concerns friendships, and our work follows this tradition. Friendships are children’s primary voluntary associations, and they have profound implications for the development of adolescents’ identities (Erikson, 1950; Meeus et al., 1999), the potential for social influence (Brechwald & Prinstein, 2011), and the clustering and divisions that emerge within larger groups (McFarland et al., 2014). Drawing on rich social network data collected from students in small towns in Iowa and Pennsylvania, our research focuses on one prominent characteristic of adolescent friendship networks, namely, that adolescents who are friends tends to be more similar than those who are not.

Differences in Similarity for Friends Versus Nonfriends: A Varying Feature of Adolescents’ Social Structure

The tendency to have friendships with others like oneself is among the strongest and most robust findings in the study of social relationships (McPherson et al., 2001). The greater similarity of friends than nonfriends, also known as homophily, has been documented for a wide range of personal attributes, including race (Shrum et al., 1988), religion (Louch, 2000), education (Louch, 2000), and delinquency (Warr, 2002). Indeed, this consistent pattern led Bukowski and colleagues (2009) to list similarity as a key component of friendship. Homophily results from several prominent social processes, including social influence, selection preferences, relational patterns (such as reciprocating friendships and becoming friends of mutual friends), proximity or propinquity (such as being in the same class or after-school activity), and shared environmental influences (Brechwald & Prinstein, 2011; Feld, 1982; McPherson et al., 2001; Rivera et al., 2010). These processes all play important roles in linking similarity with friendship, and research continues to advance our understanding of them.

In contrast to most developmental research, our interest is in the difference in similarity between friends and nonfriends in its own right rather than in the processes that produce it. We believe this differential similarity merits attention because it directly reflects the organization of adolescents’ social worlds, indicating the degree to which an attribute is a basis of clustering and division. For instance, if friends were far more similar than nonfriends for school performance, then more and less successful students would occupy largely separate social worlds. The changing magnitude of friendship similarity across a set of attributes would provide a picture of a setting’s dynamic social organization as a map of the domains for which there is greater mixing versus division.

Patterns of differences in the similarity of friends versus nonfriends are informative about the peer experiences that will be typical in a population, shedding light on the form and direction of these inputs to development. Lower similarity differentials for race, ethnicity, social class, and immigrant status, for example, would correspond to greater proportions of adolescents having friendships with students who are different from themselves in these ways. Moody’s (2001) study of friends’ similarity for race is an especially strong example. He showed that features of school organization, such as integration of extracurricular activities and academic tracks, were associated with higher rates of cross-race friendships. Decades of research on contact theory indicate that such cross-race friendships are key to reducing intergroup prejudice (Pettigrew & Tropp, 2006).

Similarity between friends versus nonfriends also has strong implications for peer influence processes because the direction of the influence depends on who an adolescent’s friends are. An strong link between similarity and friendship for antisocial behavior, for instance, implies that most deviant youth will have deviant friends who influence them toward maintaining or increasing their deviant behavior, while conforming youth usually will have conforming friends who deter deviance. In contrast, low similarity for antisocial behavior implies that, regardless of their own behavior, most adolescents have friends whose behavior is on average close to the mean (as shown by Haynie, 2002), in which case peer influence should dampen both rigid conformity and extreme deviance.

Similarity for Which Attributes?

A few early studies reported levels of correlations between friends on many attributes, giving some sense of which types of attributes tend to be stronger and weaker axes of grouping among adolescents. The most extensive study, by Kandel (1978), found that similarity was greatest for the demographic attributes of age, gender, and race, relatively strong for several forms of substance use, moderate for a variety of other attitudes, and weak for family relationships. The findings of Eiser et al. (1991) and Tolson and Urberg (1993) were broadly consistent with this pattern.

Our research will focus on three types of attributes, available in our dataset, that research suggests are important in adolescent life and friendships. The first type is demographic variables, which are a perennial emphasis of research on friendships and similarity (McPherson et al., 2001) and the focus of societal concern about inequality. Our work includes indicators of four demographic variables historically associated with inequality in the United States, namely, gender, race/ethnicity, social class, and family composition (a topic of special interest for children and adolescents). Next, we will investigate indicators of adolescents’ orientations to three of the major institutions for this life stage: family, school, and religion. Family is the dominant institution of early life in most cultures, including the United States, and developmental research indicates that parents have considerable impact on their offspring’s friendships (Cui et al., 2002; Rubin et al., 2004). Despite Kandel’s (1978) early findings, recent research also indicates at least some tendency for adolescents to choose friends whose relationships with their parents are similar to their own (Ragan et al., 2014). For some students, the school is a setting of success and satisfaction, while for others this prominent institution brings frustration and a sense of rejection. Accordingly, previous research indicates that friends tend to be more similar than nonfriends for school performance and interest in school (Eiser, 1991; Kandel, 1978; Tolson & Urberg, 1993). Religion plays a more variable role in the lives of adolescents and their families in the United States, central for many but absent for others. McPherson and colleagues (2001) concluded that similarity of religious orientation played a meaningful role in friendships, but the relationship was variable across settings or groups, and it also appeared to be declining from the 1970s onward. Finally, problem behavior has been a major focus of research on adolescents’ similarity with friends, and many studies demonstrate that both selection and influence contribute to this correlation between friends (e.g., Gallupe et al., 2019; Osgood et al., 2015; Tucker et al., 2014). We will investigate four forms of problem behavior prominent for adolescents in the United States: delinquency, drinking alcohol, smoking cigarettes, and using marijuana.

Inbreeding Versus Baseline Homophily

McPherson and colleagues (2001) distinguished two ways that friendship could be associated with similarity: inbreeding homophily and baseline homophily. Our interests focus only on inbreeding homophily because it reflects the organization of friendships within an interacting population, such as students who attend school together (Hafen et al., 2011). In contrast, baseline homophily is differential similarity arising not from friendships within settings, but from population differences between settings. A byproduct of such population differences is greater average similarity of people in the same setting, whether or not they are friends, compared to their similarity with people in other settings. For instance, in a single-gender school, baseline homophily guarantees all friends match on gender, while the same uniformity precludes inbreeding homophily and means that gender is irrelevant to friendship groupings and divisions in that school. For a school with equal numbers of boys and girls, in contrast, a preponderance of same-gender friendships could only arise from differential friendship pairings that correspond to inbreeding homophily for gender. Although only inbreeding homophily is relevant to the social organization generated by friendship choices, most earlier studies of similarity between friends have, unfortunately, conflated the two types of homophily by failing to differentiate data from multiple schools (e.g., Kandel, 1978; Eiser et al., 1991; Tolson & Urberg, 1993).

We limit our attention to inbreeding homophily because it maintains a focus on the organization of friendship within schools, where students have at least the potential to be aware of one another and form friendships. Note that inbreeding homophily includes all friendship patterns within schools, whether they stem from individual preferences or the setting’s organization and structure. For instance, homophily for school performance might arise because students like being with others who are similarly successful or unsuccessful, but inbreeding homophily is not limited to that process. This friendship pattern would also be inbreeding homophily if it occurred because the school tracked students into different classrooms based on their grades, after which spending time together led to the friendships.

Change Across Adolescence in Similarity between Friends

A useful starting point for considering change across adolescence in the greater similarity of friends than nonfriends is that, during middle childhood, gender dominates who is friends with whom. Friendship groups within elementary school classrooms typically are fully segregated by gender (Maccoby, 1998), but by late adolescence gender separation diminishes and mixed-gender friendship groups emerge (Poulin & Pedersen, 2007; Shrum et al., 1988). We will investigate a) how much the connection between gender and friendship declines across adolescence, b) if that decline is accompanied by increasing connections of friendship with similarity on other attributes, and c) which attributes become more prominent bases of division versus integration in adolescent friendship networks.

We expect that, in industrialized Western cultures like the United States, similarity between friends versus nonfriends will increase during adolescence for most attributes due to several key changes in daily life. Adolescents gain increasing autonomy, such as being allowed to spend longer periods away from adult supervision and to be farther away from home (Osgood et al., 2005). One element of this increased independence is that time spent with the family declines and is replaced by time spent with peers (Felson & Gottfredson, 1984; Larson & Richards, 1991; Larson et al., 1996). Thus, peers become increasingly important to adolescents’ personal and social development, while parents exert decreasing control over their offspring’s friendship choices and social activities. Accordingly, adolescents have the potential to exercise greater choice in selecting friends (Kandel, 1978) and greater opportunity to influence one another (Richmond et al., 2019), both of which should contribute to increased similarity of friends versus nonfriends. Furthermore, adolescents’ capabilities, interests, and activities expand and become more differentiated, in concert with gains in cognitive and social skills. This dynamic would further promote similarity of friends through propinquity and shared environmental influences because adolescents spend time in a growing variety of specialized classes and activities where they will encounter potential friends who share their interests, whether in sports, student government, or video games.

A pair of early studies compared the extent to which friends were more similar than nonfriends at different ages: Eiser and colleagues (1991) measured student similarities and friendships in eighth and eleventh grades and Tolson and Urberg (1993) did so at ages eleven through sixteen. Both reported modest, mixed evidence of an increasing differential on lists of attributes dominated by substance use. More recent studies of alcohol use by Burk et al. (2012) and of delinquency by Richmond et al. (2019) reported clearer patterns of increased homophily from early or pre-adolescence through mid-adolescence, followed by leveling or declines in later adolescence. Thus, prior work suggests similarity may increasingly characterize friendship through much, if not all, of adolescence, for at least some attributes. Our study will examine this more closely through a larger data set, a broader set of attributes, and analytic methods that isolate inbreeding homophily, while also assessing the potential mediating role of two ecological factors.

Two Ecological Predictors of Similarity Between Friends

We test two potential ecological sources of increased similarity between friends relative to nonfriends across adolescence. We expect that increasing size of school grade cohorts and rising attribute variability will provide adolescents with more options for highly differentiated friendships as they are progressing from elementary to secondary schools, gaining additional freedom, and developing more individualized interests.

Following themes from Barker and Gump’s (1964) social ecological analysis of setting size, Bahns et al. (2012) reasoned that students in larger schools should have more opportunities for forming friendships with others who are distinctly similar to themselves (and different from many others) simply because there is a larger pool of people at every level of any attribute. Kinney (1993) offered a similar argument in qualitative form. Consider two schools of very different sizes, but with the same distributions of student attributes. At the smaller school, students with less common characteristics (perhaps being a minority group member or preferring goth music over sports) will have few choices of friends like themselves on that attribute, compared to students at the larger school. Further, students at the small school would be less likely to find friends who closely match them in several ways at once, dampening the overall association of friendship with similarity across attributes. Bahns and colleagues (2012) emphasized the role of selective attraction, but the ecological argument applies to other similarity processes as well. For instance, large schools are likely to offer a more finely differentiated curriculum and a greater range of sports, activities, and interest groups, all of which enhance propinquity-based friendships with students distinctly similar to oneself. Further, the new friends may become even more similar through shared experiences in the activities and classes where they came to know one another.

In the United States, grade cohorts often become larger with age through students from multiple elementary schools or middle schools transitioning to single middle schools or high schools. These school mergers bring considerable change in friendship patterns (Felmlee et al., 2018; Temkin et al., 2018). We predict that the association between similarity and friendship will be stronger in larger grade cohorts, and that this relationship will account for part of the increase in this association across adolescence.

Attribute variability also should affect the differential in similarity between friends and nonfriends because it further impacts the range of options for forming friendships (Bahns et al., 2012; Kinney, 1993). For any given school size, greater variability on an attribute leads to increased opportunities for individuals across the full attribute range to encounter potential friends like themselves and more clearly different from others. As variation diminishes, in contrast, stronger and finer distinctions are required for the attribute to play an equal role in differentiating friends and nonfriends. Otherwise, the restriction of variance will dampen the association between friendship and similarity (Meade, 2010). Moody (2001) showed that similarity between friends versus nonfriends on race was strongly tied to the variance of race at a school, with the tendency to have friends of the same race/ethnicity growing markedly as groups became more equal in size. During adolescence in the United States, delinquency and substance use increase dramatically, school performance is graded more competitively, and many students become disaffected with schooling. These developmental trends constitute increased variability (reflected in larger standard deviations), which translates to additional opportunities for individuals to sort themselves on these attributes.

Bahns and colleagues (2012) believed that attribute variability would account for a stronger association between similarity and friendship in larger schools, but they could not test this well with their comparison between 55 dyads at a large university and 79 dyads pooled from four small colleges. We anticipate that school size and attribute variability will each contribute to the association, above and beyond the other, because they provide distinct routes to greater opportunity for more highly differentiated friendships.

Current Study

The present study examines developmental change in the tendency for friends to be more similar than nonfriends, from early- through mid-adolescence. This change provides a window on the evolving organization of the peer context of adolescent development. Specifically, we examine the changing importance of demographic variables, orientations to prominent institutions, and problem behaviors for the organization of adolescent friendships. We go beyond prior research by assessing the extent to which changes in the strength of the correlation between friends are potentially explained by two ecological factors, school-level measures of network size and attribute variability. We also improve upon prior work by using stronger data and applying methods that isolate the similarity differential within school grade-cohorts and eliminate contributions from differences between schools.

Cultural Context

The setting of our study is a sample of small school districts that serve rural and small-town communities in the Midwestern state of Iowa and the mid-Atlantic state of Pennsylvania. Their enrollments were predominately White and English-speaking, and at least 15% of the families in each district were eligible for free or reduced-cost school lunches. A serious cultural analysis of this setting is beyond the scope of the present study, but we will make note of some features especially relevant to our topic.

With regard to the size and organization of the schools, many districts had multiple schools in the earlier grades, but each district had a single school per grade for 9^th through 12^th grades. Accordingly, a strong majority of same-age adolescents in the geographic area will attend the same school (Bielick & Chapman, 2003), and we expect that the grade cohort friendship networks we study will largely encompass neighborhood networks as well. This would not be true in communities where school selection is largely independent of residential location and residents in any area may attend a variety of schools, as is more common in large metropolitan areas of the United States and standard in many countries.

The typical pattern of within-grade classroom organization in the United States shifts from fixed classroom groupings in elementary school to different groupings throughout the day by high school. Thus, our students are exposed to a relatively large number of peers, although that exposure may be selective, based on course interests and tracking by performance. The schools offer a variety of extra-curricular activities that also provide opportunities for interest-based mixing among students.

The demographic composition of these communities also sets the context for our study. All schools have roughly equal proportions of boys and girls, providing ample opportunity for sorting on this basis. Students who identify as white, non-Hispanic predominate in all the sampled communities, so our study does not represent minority groups well. The proportion of non-white students varies from less than 5% to 35%, however, which means there is considerably more potential for friendships to be divided by race/ethnicity in some schools than in others. About 60% of students also live with both biological parents, and this varies relatively little across communities. There is thus ample opportunity for adolescents to group themselves by family structure, if they find this important, but variability of this attribute has limited potential to account for differences among communities or over time. Income levels were relatively low in all these communities, but also varied considerably. The proportion of students reporting that they received free or reduced-price school lunch in the sixth grade ranged from 20% to 50%. Thus, our research is not informative about either high-income professional communities or the most heavily disadvantaged communities, but it does permit us to investigate a sizable range of within-community variability for family income.

Methods

Sample

Data for the current paper are drawn from the PROSPER study, which gathered friendship choices from students in a large sample of schools, following two grade cohorts across much of adolescence. The PROSPER study was designed as a community-level program evaluation for testing a community-university partnership model for implementing evidence-based interventions. The intervention sought to promote prosocial attitudes and behaviors and to reduce adolescent delinquency and substance use. PROSPER randomly assigned 14 school districts to receive a set of family- and school-focused interventions and 14 districts to serve as a comparison group (Spoth et al., 2007; Spoth et al., 2011).

Students in the first cohort were enrolled in the 6^th grade during the 2002–2003 school year, and students in the second cohort were enrolled in the 6^th grade during the following year. In-school data collection occurred in the Fall of the 6^th grade and every Spring thereafter through the 12^th grade, resulting in eight waves of data for each of the two cohorts. The race/ethnicity composition of the sample was 84.2% White, 6.8% Hispanic/Latino, 3.2% African-American, 1.3% Asian, .5% Native American, 3.9% other or multiple identifications.

Enrollment into the PROSPER study was open at each wave, allowing new students to be added to the sample, and students were not retained in the study if they left the school district. This sample definition is appropriate to our focus on the schools’ friendship networks: Students who join a school become part of the within-school friendships network, but students who depart leave it (i.e., any continued friendships become “out-of-school”). Siennick and colleagues (2017) found that students who entered or left the PROSPER schools systematically differed from students who were continuously enrolled in them. Thus, a sample limited to the latter would provide a misleading picture. Retention was high, however, with students who participated in the initial wave contributing data on a mean of 5.9 of the eight waves.

Cohort size declined considerably at many schools during the 12^th grade, likely reflecting rates of dropout, early graduation, and decreased willingness to complete a voluntary survey. As a result, any change in similarity for friends versus nonfriends could be due to the unique changes in the pool of respondents and potential school friends for this final year of high school, and therefore we omit the 12^th grade data from our analyses. Thanks to the Pennsylvania State University’s Institutional Reviewer Board’s approval of passive consent procedures (PROSPER Student Surveys, IRB #14226), participation rates of eligible students ranged from 80% to 92% across included waves, with a mean of 87%. Data were collected from over 9,000 students per wave and more than 15,500 students overall.

To collect social network data at each wave, students were asked, “Who are your best and closest friends in your grade?” and allowed to provide two names for best friends and five names for “other close friends.”¹ Limiting the question to friends in the same grade and school, while truncating the friendship pool, enables full network analysis of a consistent, clearly defined, and interacting population across this age span. The first and last names of each friend were matched to student rosters. The proportion of eligible students who provided social network data ranged from 74% to 87% across waves (mean 81%), and the overall name matching rates by wave ranged from 71% to 85% (mean 79%). These network data enable us to assess similarity through friends’ independent reports, rather than respondents’ reports about friends’ attributes, which are biased toward overestimating similarity (Bauman & Ennett, 1994; Wilcox & Udry, 1986). The mean number of matched friendship nominations was 3.98 per respondent, per wave.

Measures of Individual Attributes

The present study examines how similarity between friends versus nonfriends develops across 12 attributes potentially relevant to adolescents’ friendships. The four demographic attributes are coded as binary variables. Gender and race are operationalized as male versus female and White versus non-White, respectively. We could not differentiate race more finely because too few schools had sufficient representation of multiple nonwhite groups. Free lunch captures social class in terms of low family income, assessed by whether the respondent reported receiving a free or reduced-cost school lunch. Our indicator of family structure is whether the respondent resided with two biological parents.

Our measures of orientations toward major institutions concern family, school, and religious participation. The family relations measure (Redmond et al., 2009) derives from 26 items that addressed child monitoring, inductive reasoning, joint activities, and affection between the respondent and his or her parents. This index is the mean of four standardized subscales (α = 0.78 for combining subscales); the full list of items used in the construction of our scales is found in Supplement S1. The measure of school bonds (α = 0.80) is the mean of eight items reflecting the respondents’ attitudes toward school, adapted from Simons, Whitbeck, Conger, and Conger (1991). The respondents were asked, for example, whether they like school, how much effort they put into school, and their sense of belonging at school. School grades is a self-reported measure of academic achievement and ranges from 1 (mostly D’s) to 5 (mostly A’s). Religious participation was captured with a single item measuring the frequency of attending church or religious services on a scale from 1 (hardly ever or never) to 4 (more than once a week).

The final four attributes are problem behaviors. The first of these, delinquency, was measured by a series of 12 questions about respondents’ participation in delinquent activity, with the answers for each item ranging from 1 (never) to 5 (five or more times). The measure was based on Elliott and colleagues (1985) widely used self-reported delinquency scale, for which Huizinga and Elliott (1986) reported evidence of reliability and validity. Questions included, for example, how many times during the past year respondents had taken something from a store that they did not pay for, purposely damaged or destroyed property, and beat up someone or physically fought with someone. The graded response IRT model was used to generate composite scores for delinquency that are less skewed and more suitable for computing correlation coefficients (Osgood et al., 2002; summative scoring α = 0.89, IRT marginal reliability = 0.60). Items for alcohol use, cigarettes, and marijuana represent past 30-day use of alcohol, tobacco, and marijuana, respectively. These measures come from Botvin and colleagues (1997), and comparable self-report items are widely used for research on adolescent substance use (e.g., Monitoring the Future, Bachman et al., 2014). We dichotomized each substance use measure to indicate “no use” vs. “any use.” Table 1 presents descriptive statistics, including rates of missing data, both for the individual attributes, measured for individual respondents, and for the Moran’s I correlations between friendship pairs, measured for school-grade cohorts, described next.

Table 1.

Descriptive Statistics

	Survey Responses across All Waves						Correlations across Friendship Pairs
Attribute	N	% Missing1	Mean	Std. Dev.	Min.	Max.	N	% Missing	Mean	Std. Dev.	Min.	Max.
Demographic Gender	65,788	0.2%	0.485	—	0	1	454	0.0%	0.816	0.090	0.511	1.031
Race	64,313	2.4%	0.848	—	0	1	454	0.0%	0.129	0.157	−0.128	0.652
Family Structure	65,114	1.2%	0.604	—	0	1	454	0.0%	0.098	0.082	−0.156	0.419
Free Lunch	63,946	3.0%	0.276	—	0	1	454	0.0%	0.148	0.107	−0.104	0.633
Orientations to Institutions
Family Relations	65,089	1.2%	−0.061	0.480	−2.354	1.004	454	0.0%	0.077	0.071	−0.148	0.394
School Grades	63,563	3.5%	4.024	0.910	1	5	454	0.0%	0.256	0.108	−0.142	0.534
School Bonds	65,749	0.2%	3.731	0.756	1	5	454	0.0%	0.178	0.087	−0.074	0.523
Religious Participation	65,066	1.3%	2.592	1.312	1	4	454	0.0%	0.118	0.088	−0.194	0.439
Problem Behaviors
Delinquency	65,527	.06%	0.297	0.820	−0.341	3.943	454	0.0%	0.154	0.089	−0.139	0.467
Alcohol Use	65,646	0.4%	0.261	—	0	1	449	1.1%	0.122	0.108	−0.129	0.544
Cigarettes2	48,119	0.3%	0.164	—	0	1	292	0.0%	0.189	0.105	−0.017	0.521
Marijuana2	48,007	0.6%	0.104	—	0	1	285	2.4%	0.135	0.106	−0.050	0.575

¹Percent missing for individual is among respondents who completed the friendship network section of the questionnaire.

²Moran’s I adjusted for bias dependent on sample size, −1 / (N – 1).

³Waves 1 & 2 omitted because there was no variance in a substantial portion of networks

Plan of Analysis

Index of Similarity

Our study concerns the difference in similarity between friends and nonfriends, but not the interpersonal processes that create it. Thus, the powerful network analytic approaches available for estimating dynamic relational processes (e.g., Steglich et al., 2010) are not needed or applicable. Instead, we require a simple index of association that reflects the extent to which friends are more similar than nonfriends. Further, it must be computed separately for each network, both to limit attention to friendship within school-grade cohorts (inbreeding homophily) and to be able to relate this index of association to the setting-level explanatory variables.

Our index of differential similarity of friends versus nonfriends is Moran’s I (Moran, 1950), a standard indicator of network or spatial autocorrelation. For network data such as ours, Moran’s I is nearly identical to the Pearson product-moment correlation between the scores of pairs of friends, computed across all friendships in a network. The minor difference between them is that the formula for Moran’s I standardizes the data by the attribute’s mean and standard deviation across individuals, while the formula for the Pearson correlation standardizes the data based on the mean and standard deviation across friendships. In computing the correlations, cases are friendship ties between one respondent and a nominated friend, so reciprocated friendships count twice, once for each friend naming the other. We calculated the Moran’s I correlation between friends separately for all 12 attributes in each school/cohort/wave combination.

Correlations such as these, between pairs of friends or partners, play a special role in research on relationships and networks because they reflect the association between the presence of the relationship and the degree of similarity on a variable. They are analogous to the intraclass correlation coefficient for nested or grouped data, which expresses within-group similarity as the proportion of total variance that is between groups rather than within them. Indeed, in the special case of a dataset of mutually exclusive dyads (such as twins or couples), the two are identical (Kenny, Kashy, & Cook 2020). The intraclass correlation coefficient is not applicable to network data because network ties do not divide individuals into mutually exclusive groups. The correspondence of the two types of correlations does, however, suggest two useful interpretations for our use of Moran’s I as an indicator of the association between similarity and friendship: 1) how closely students’ scores on an attribute track those of their friends, as a standardized regression within the network, 2) how much smaller the variance for differences between friends is than the overall variance of the sample (as an approximate, proportionate reduction).

Prior to analysis, we correct the Moran’s I coefficients for an inherent negative bias that decreases with sample size. The expected value of I across randomly paired individuals is −1 / (N − 1) rather than zero (with N equal to the number of individuals in the network, Griffith, 2010). This bias arises from the constraint that people cannot choose themselves as friends, and it occurs for Pearson correlations computed from network data as well. To ensure that this bias does not distort our findings about the relationship of network size to similarity of friends versus nonfriends, we correct for it adding 1 / (N − 1) to the Moran’s I values.

Examination of the distributions of Moran’s I for each attribute, prior to beginning analyses, revealed that three of the 5112 observations were outliers that might unduly distort results (Judd et al., 2017). These datapoints were more than four standard deviations above the mean and more than 1.5 standard deviations above the next most extreme case. We recoded the outliers to the next highest observed value, thereby making their influence on results less exceptional, while retaining them in the analysis.

Statistical Model

We address our research questions thorough multi-level regression analyses in which the outcome measures are the correlation coefficients that serve as our index of the association between friendship and similarity on an attribute. This network-level index is defined for each school and grade cohort combination in seven waves of data. To address the dependence among network-level indices inherent in this research design, we apply a three-level random effects model with random intercepts for waves (level 1), schools (level 2), and school districts (level 3). A random coefficient for cohort allows for cohort differences specific to each school district.

This approach also is a meta-analysis because the outcome measure (i.e., the Moran’s I correlation between friends) is a statistic computed for each of a set of samples. Each data point includes error variance corresponding to the variance of the statistic’s sampling distribution, which is largely dependent on sample size. Griffith (2010) demonstrated that, across a wide variety of data distributions and network patterns, the value 2 / N_F (with N_F equal to the total number of friendships in the network) closely approximates the sampling distribution variance of Moran’s I in networks of 25 or more individuals. Therefore, our analyses include this quantity as a fixed meta-analytic error variance component (implemented in MLwiN, see Rasbash et al., 2019, Chapter 7).

Network Level Explanatory Variables.

The analyses capture age-related change in correlations between friends through a set of dummy variables representing the waves, thereby allowing for an unconstrained, curvilinear time trend. To obtain the adjusted mean correlations that appear in Table 2, we use a version of the model that includes dummy variables for all seven waves, omits the intercept, and mean centers all other variables (as described below). Significance tests for change across all waves come from deviance test comparisons² of this model to one that omits these dummy variables but has an intercept.

Table 2.

Changes Across Grades in Correlations Between Friends: Results from Three-Level Random Effects Meta-Analyses

	Adjusted Mean Correlation1 Between Friends							Developmental Trend
	6th grade	6th Grade	7th Grade	8th Grade	9th Grade	10th Grade	11th Grade	Deviance Test2
Attribute	Fall	Spring	Spring	Spring	Spring	Spring	Spring	χ2	p
Demographic
Gender	0.888	0.860	0.840	0.805	0.782	0.756	0.730	151.05	<0.001
	(0.010)	(0.010)	(0.009)	(0.009)	(0.010)	(0.010)	(0.011)
Race	0.089	0.095	0.118	0.141	0.178	0.145	0.139	45.15	<0.001
	(0.025)	(0.025)	(0.025)	(0.025)	(0.025)	(0.026)	(0.026)
Family Structure	0.090	0.088	0.103	0.100	0.122	0.131	0.102	15.58	0.016
	(0.010)	(0.010)	(0.010)	(0.010)	(0.010)	(0.011)	(0.012)
Free Lunch	0.133	0.152	0.160	0.162	0.138	0.145	0.139	9.23	0.161
	(0.015)	(0.015)	(0.014)	(0.014)	(0.015)	(0.016)	(0.016)
Orientations to Institutions
Family Relations	0.051	0.069	0.089	0.095	0.091	0.086	0.071	23.69	0.001
	(0.009)	(0.008)	(0.008)	(0.008)	(0.008)	(0.009)	(0.010)
School Grades	0.191	0.201	0.266	0.302	0.307	0.304	0.305	117.04	<0.001
	(0.011)	(0.010)	(0.010)	(0.010)	(0.011)	(0.012)	(0.013)
School Bonds	0.142	0.165	0.193	0.209	0.198	0.195	0.180	43.57	<0.001
	(0.010)	(0.010)	(0.009)	(0.009)	(0.010)	(0.011)	(0.012)
Religious Participation	0.099	0.095	0.111	0.129	0.145	0.151	0.153	23.70	0.001
	(0.011)	(0.010)	(0.010)	(0.010)	(0.011)	(0.011)	(0.012)
Problem Behaviors
Delinquency	0.126	0.137	0.146	0.184	0.180	0.183	0.151	37.57	<0.001
	(0.011)	(0.010)	(0.010)	(0.010)	(0.011)	(0.012)	(0.012)
Alcohol Use	0.034	0.069	0.101	0.141	0.167	0.195	0.209	135.18	<0.001
	(0.010)	(0.010)	(0.010)	(0.010)	(0.011)	(0.012)	(0.012)
Cigarettes	—	—	0.121	0.197	0.208	0.215	0.216	47.33	<0.001
	—	—	(0.012)	(0.012)	(0.013)	(0.013)	(0.014)
Marijuana	—	—	0.066	0.131	0.145	0.163	0.185	48.24	<0.001
	—	—	(0.011)	(0.011)	(0.012)	(0.013)	(0.014)

Mean correlations adjusted for state, cohort, and treatment status.

All adjusted mean correlations p < .005.

Standard errors in parentheses.

¹Moran’s I adjusted for bias dependent on network sample size, −1 / (N – 1).

²For deviance tests, 4 df for cigarettes and marijuana use and 6 df for all other attributes.

To examine whether the association of similarity with friendship is related to the variability of an attribute, the network-specific standard deviations of attributes serve as our explanatory variable. When analyzing network size, we expected that the relationship to similarity of friends versus nonfriends would be roughly proportional to each network’s size. For example, the increase from 50 to 100 students should be more consequential than from 350 to 400. Accordingly, we use the natural log of the number of students within the school grade cohort as our explanatory variable. We allow for additional curvilinearity in that relationship by also including that value squared (after subtracting the mean to reduce collinearity). Correlations between network size and attribute variabilities were largely positive, but not so strong as to create collinearity problems (from −.03 to .45, with a mean across attributes of .29). Our statistical models add attribute variability and network size together, thereby evaluating the independent contributions of each controlling for the other.

Other Control Variables.

Finally, the analyses also control for features of the research design not related to our research questions. A dummy variable designated the school districts randomly assigned to the PROSPER prevention model, versus comparison group.³ Two more dummy variables controlled for differences between the two grade cohorts and between schools in Iowa and Pennsylvania. The dummy variables for wave were uncentered, the cohort variable was centered within school districts, and all other explanatory variables were grand mean centered. This study was not preregisted, and readers should contact the corresponding author about data files and model specifications necessary for replication.

Missing Data and Sample Restrictions

To learn how data missing due to non-participation likely differed from the data included in the analyses, we compared means for respondents who participated every wave they appeared on the school rosters with means for respondents who did not participate for one or more waves. This comparison is informative because the large majority of survey-wave non-responses (84.4%) were from students who did participate in one or more of the other waves. Students who never participated constituted only 3.1% of the total individual sample and were responsible for only 15.6% of the missing data at the survey-wave level.

The mean differences between “sometimes” participants and “always” participants, presented in Supplemental Table S1, correspond to small but meaningful effect sizes. This analysis is based on the initial wave of data for each of the twelve attributes. Values of Cohen’s D (i.e., mean differences in standard deviation units) range from a low absolute value of .051 for family relations to a high of .206 for school grades; all are significant at p < .05 and most at p < .001. Respondents who sometimes did not participate were more likely to be male, to be nonwhite, not to live with two parents, to receive free school lunch, to be less oriented to institutions, and to engage in problem behaviors. We do not consider missingness to be a serious problem for our work given the magnitude of these differences, the high overall rates of participation, and that most non-participants in each wave did contribute data on at least some other occasions.

Table 1 shows that there was little variable-level missing data in completed questionnaires. The mean percent missing was 1.2%, with a high of 3.5% for school grades. Computations of Moran’s I are limited to friendship pairs for which both the respondent and nominated friend provided data on the attribute. Further, we unavoidably lack the friendship choices of non-respondents. To the best of our knowledge, more advanced procedures for addressing missing data, such as multiple imputation, are not available for an analysis such as ours, which uses network data to compute an outcome variable for a multi-level meta-analysis.

Analyses were limited to school-grade cohorts of at least 25 students to meet the assumptions underlying our estimates of meta-analytic error variance (Griffith, 2010). Eliminating smaller networks also limited the size of adjustments to Moran’s I needed to compensate for its negative bias to no more than .04. This sample size restriction eliminated 18 sixth grade networks, all from five small elementary schools in two school districts. Thirty-five students were enrolled in the smallest network included in the analysis, and they reported 91 friendships.

Up to 454 school-cohort-wave specific networks contributed to the analyses. Correlations could not be computed and therefore are missing when students within a school were uniform on an attribute. Rates of cigarette smoking and marijuana use were quite low in sixth grade, resulting in many missing correlations, so analyses of those attributes omitted the two sixth grade waves. Otherwise, missing data for Moran’s I were limited to alcohol and marijuana use for a few smaller schools in the early waves.

Results

Table 2 demonstrates that the connection between friendship and similarity, ubiquitous in previous research, holds for our data as well. Our large sample of adolescent friendship networks yielded precise estimates of these mean correlations between friends (Moran’s I, corrected for sample size bias), with almost all standard errors between .009 and .015 and none over .026. The association of similarity with friendship was significantly positive (p < .005, two-tailed) on every attribute at all seven time points. As expected, gender was the strongest basis of similarity between friends versus nonfriends throughout the study (mean I > .72 every wave). For all other attributes, the mean correlation between friends was meaningful but not especially powerful, never surpassing I = .31. The connection between similarity and friendship was consistently weak for family relations (both mean I < .10 for all waves). In contrast, the association of similarity with friendship was stronger for school grades than most other attributes (mean I = .191 − .307).

Developmental Trends

The association between similarity and friendship significantly changed across this age span for every attribute except receiving free or reduced-price lunch, though the amount of change varied considerably (see Table 2). Deviance test χ² values (6 df) for the developmental trend across grades ranged from 15.58, p = .016, for family structure to 151.05, p < .001, for gender. Notably, the correlation between friends for gender declined substantially from 6^th to 11^th grade. Thus, cross-gender friendships were increasingly common at later grades, though within-gender friendships continued to predominate. For most other attributes, the correlation between friends increased with age, as we anticipated due to adolescents’ greater independence and access to a broader range of settings. Correlations for these 10 attributes were at least 40 percent higher at later waves than at the initial assessment. We observed marked increases of 75 percent or more in the correlations between friends for race, family relations, and the three forms of substance use. Increases in correlations were most notable for sixth through eighth or ninth grade, after which they were more limited and less consistent. Though correlations for some attributes appear to decline in later waves, decreases from eighth through eleventh grades were no greater than chance: Of 60 potential pairwise decreases (6 possible differences on 10 attributes), the minimum p value was .037.

Figures 1 through 4 present these changing mean correlations between friends graphically. Mean correlations over time for the demographic attributes appear in Figure 1 and underscore the prominence of gender in friendships during adolescence. Figure 2 presents the demographic attributes other than gender to give a clearer picture of the smaller but meaningful change for those attributes. Figure 3 presents mean correlations for the variables reflecting orientations to the institutions of family, school, and religion, and Figure 4 presents mean correlations for problem behavior.

Figure 1.

Correlations between friends for demographic attributes

Figure 4.

Correlations between friends for problem behaviors

Figure 2.

Correlations between friends for demographic attributes, excluding gender

Figure 3.

Correlations between friends for orientations to institutions

Network Level Explanatory Variables

The next analyses examine the associations of the network explanatory variables with correlations between friends. Table 3 reports the independent relationships of attribute variability and network size to correlations between friends, using a model that includes both, along with grade level differences and the control variables. The full results for these analyses are in supplemental Tables S2 through S4. For ten of the twelve attributes, the correlation between friends was stronger when variability among students in a network was higher. This pattern is reflected in statistically significant positive regression coefficients for the network standard deviations as predictors. The relationship is quite strong for many of the attributes: The coefficients of Table 3 imply correlation differences between the least and most variable networks of .23 to .32 for race, free lunch, delinquency, smoking cigarettes, and marijuana use, and of .16 to .21 for school grades, school bonds, and drinking alcohol. The relationship between network variability and correlations between friends failed to reach significance only for gender and religious participation. Notably, the variability of these two attributes differed little across networks, giving them limited potential to affect correlations between friends.⁴

Table 3.

Relationship of Similarity Among Friends with Network Size and Attribute Variability, from Three-Level Random Effects Meta-Analyses of Correlations¹ Between Friends

	Explanatory Variable
	Attribute Variability				Size of Network
Attribute	b Net Std Dev	z	p	b logn	b logn2	Wald χ2 (2 df)	p
Demographic
Gender	0.839	0.69	0.492	0.004	−0.012	1.338	0.512
	(1.222)			(0.009)	(0.011)
Race	0.795	7.29	<0.001	0.040	0.041	7.727	0.021
	(0.109)			(0.017)	(0.020)
Family Structure	0.789	2.74	0.006	0.045	−0.004	26.392	<0.001
	(0.288)			(0.009)	(0.011)
Free Lunch	0.840	6.68	<0.001	0.036	0.004	8.151	0.017
	(0.126)			(0.013)	(0.015)
	(0.126)			(0.013)	(0.015)
Orientations to Institutions
Family Relations	0.248	2.70	0.007	0.003	0.005	0.599	0.741
	0.092)			(0.008)	(0.009)
School Grades	0.255	6.28	<0.001	0.053	−0.035	36.778	<0.001
	(0.041)			(0.009)	(0.011)
School Bonds	0.336	6.07	<0.001	0.014	−0.012	3.158	0.206
	(0.055)			(0.009)	(0.011)
Religious Participation	0.098	1.45	0.146	0.042	−0.009	18.088	<0.001
	(0.067)			(0.010)	(0.012)
Problem Behaviors
Delinquency	0.255	5.68	<0.001	0.017	−0.005	3.117	0.210
	(0.045)			(0.010)	(0.012)
Alcohol Use	0.565	5.56	<0.001	0.006	−0.008	0.854	0.653
	(0.102)			(0.009)	(0.011)
Cigarettes2	0.731	6.35	<0.001	0.013	−0.022	0.735	0.693
	(0.115)			(0.020)	(0.026)
Marijuana2	0.738	6.63	<0.001	0.047	−0.047	5.090	0.078
	(0.111)			(0.021)	(0.026)

¹Moran’s I adjusted for bias dependent on network sample size, −1 / (N – 1).

²Waves 3–7

b Net Std Dev = unstandardized regression coefficient for the network specific standard deviation of the attribute

b log_n = unstandardized regression coefficient natural log of the number of students in this school grade cohort

b log_n² = unstandardized regression coefficient square of the log_n, after subtracting mean log_n

Standard errors in parentheses.

Wald χ² = Joint significance test for the overall relationship of network size to similarity between friends.

Table 3 indicates that the correlation between friends is significantly associated with network size for five of the twelve attributes: race, family structure, free lunch, school grades, and religious participation. This relationship is curvilinear due to both the logarithmic coding and quadratic term, so we include graphs of these relationships in Supplemental Figures S1 through S2. These regression coefficients correspond to mean correlation differences of .08 to .15 across the range of network sizes (34 to 544). The especially strong trend for school grades might well reflect a potential for more extensive tracking based on academic performance as grade cohort size increases.

Mediation

Finally, we examine the extent to which the developmental changes in correlations between friends (shown in Table 2 and Figures 1–4) may be mediated by the two ecological factors, grade cohort size and variability of the attribute. Complexities of our analyses, such as multiple random effects and curvilinear relationships, rule out the most accurate significance tests for indirect effects, and instead we rely on the safely conservative joint significance approach (Hayes & Scharkow, 2013). Fortunately, this test yielded plenty of evidence for mediation, despite its lower statistical power. The joint significance approach specifies that indirect effects are significant if both the path from the explanatory variable (grade level) to the mediator and the path from the mediator to the outcome (correlation between friends) are statistically significant. The relationships of grade level to both network size and variability (SD) were statistically significant at p < .001 for all attributes other than gender. (Supplemental Table S5 provides the grade-specific means for the explanatory variables.) Thus, all statistically significant associations in Table 3 correspond to statistically significant indirect effects between developmental trends and differential similarity of friends versus non-friends.⁵ Accordingly, attribute variability mediates indirect effects of grade level on correlations between friends for all attributes other than gender and religious participation, and network size does so for race, family structure, school grades, and religious participation.

How much of the developmental trends across grades do network size and attribute variability explain? Table 4 addresses this question for the 11 attributes with significant developmental trends by comparing the strength of these trends before versus after taking the mediators into account. For this comparison, we use Osgood and colleague’s (1996: 647) coefficient, which expresses the magnitude of a curvilinear relationship using a metric equivalent to unstandardized regression coefficients.⁶ This table shows that the significant indirect effects indicate substantial mediation of the developmental trends for correlations between friends. The values for total reduction in the relationship indicate that, together, the two mediators accounted for at least 35% of 8 of the 11 significant developmental trends. Both the amount and proportion explained were especially large for the three forms of substance use (alcohol use, 73.1%; cigarettes, 50.6%; marijuana, 58.8%), followed by school grades (, 35.8%), race, 35.1%), and school bonds (, 39.6%).

Table 4.

Mediating Effects of Attribute Variability and Network Size: Results from Three-Level Random Effects Meta-Analyses of Correlations¹ Between Friends

	Original Relationship Between Grade & Similarity	Reduction in Relationship Due to Controlling for Potential Mediators
Attribute		Total		Attribute Variability Only		Size of Network Only		Shared
Demographic	$\tilde{\mathit{b}}$	Amount Decr. $\tilde{\mathit{b}}$	Percent Decr. $\tilde{\mathit{b}}$	Amount Decr. $\tilde{\mathit{b}}$	Percent Decr. $\tilde{\mathit{b}}$	Amount Decr. $\tilde{\mathit{b}}$	Percent Decr. $\tilde{\mathit{b}}$	Amount Decr. $\tilde{\mathit{b}}$	Percent Decr. $\tilde{\mathit{b}}$
Gender	0.028	0.000	−1.1%	0.000	−0.1%	0.000	−0.9%	0.000	−0.2%
Race	0.015	0.005	35.1%	0.002	14.6%	0.002	13.1%	0.001	7.5%
Family Structure	0.008	0.003	43.7%	0.000	5.3%	0.002	28.4%	0.001	10.0%
Orientations to Institutions
Family Relations	0.008	0.003	36.4%	0.003	33.0%	0.000	0.3%	0.000	3.0%
School Grades	0.025	0.009	35.8%	0.005	18.5%	0.004	15.5%	0.000	1.9%
School Bonds	0.011	0.005	39.6%	0.003	29.0%	0.001	9.5%	0.000	1.1%
Religious Participation	0.012	0.002	13.1%	−0.001	−11.0%	0.003	25.9%	0.000	−1.8%
Problem Behaviors
Delinquency	0.012	0.003	22.6%	0.001	11.7%	−0.001	−4.8%	0.002	15.7%
Drinking	0.032	0.023	73.1%	0.023	71.3%	0.000	0.8%	0.000	1.1%
Cigarettes2	0.025	0.013	50.6%	0.007	28.4%	0.000	0.1%	0.006	22.0%
Marijuana2	0.028	0.017	58.8%	0.010	35.2%	0.000	0.3%	0.007	23.3%

¹Moran’s I adjusted for bias dependent on sample size, −1 / (N – 1).

²Waves 3–7

$\tilde{\mathit{b}}$ = Osgood et al. (1996) index of strength of relationship for curvilinear relationships, in a metric comparable to an unstandardized regression coefficient.

We speculated that increases in cohort size across grade levels would at least partially account for the increasing correlations between friends across grades. Table 4 shows that network size accounted for a meaningful share of developmental trends in the correlations for race, family structure, school grades, and religious participation (13.1% to 25.9% above and beyond what is explained by size of network alone). This degree of mediation is notable, considering that the average increase in grade cohort size was only moderate. Due to the predominance of middle schools in sixth grade and relatively few school mergers in the transition to high school (Felmlee et al., 2018), average grade cohort size rose only from a low of 138.6 in sixth grade to a high of 223.2 in tenth grade.

Table 4 further reveals that attribute variability uniquely accounted for a moderate to large share (14.6% to 71.3%) of change across grade levels in correlations between friends on nine of the eleven attributes with statistically significant developmental trends. Attribute variability played a large role for family relations (33.0% reduction), school bonds (29.0%), and substance use (drinking 71.3%, cigarettes 28.4%, marijuana 35.2%). The smaller amounts explained for race (14.6%) and school grades (18.5%), are more impressive in light of the limited differences in their mean standard deviations between grades (for race, from a low of .317 in fall of sixth grade to a high of .331 spring of tenth grade, and for grades, from .813 to .941 over the same period; see Supplement Table S5). As noted above, the standard deviations of gender and religious participation were nearly constant across schools and waves, effectively precluding mediation.

Because this mediation concerns a curvilinear developmental trend, we need to consider an additional feature not captured in Table 4‘s focus on change in the strength of the relationship. Controlling for a mediator might alter the developmental trend in ways that would not be reflected in its strength (i.e., ), such as reversing its direction or changing its shape from convex to concave. To see if controlling for mediators produced such changes in the form of relationships, we examined fitted trends across grade levels for correlations between friends, comparing results from models that did and did not include the mediators. For mediation of the significant developmental trends in correlations between friends on the demographic variables and orientations toward institutions, a straightforward pattern was evident: the trend across grades had the same shape and direction before and after introducing the controls, changing only in magnitude. For these attributes, therefore, Table 4 adequately represents the mediation findings.

For the problem behaviors, however, the pattern of mediation involved features not reflected in Table 4. The potential for a mediating role of attribute variability was substantial because it increased dramatically for the three measures of substance use, which were rare at the start of the study and then steadily increased; the increase in variability was quite large for delinquency as well. For these attributes, the fitted trends across grade levels revealed that controlling for attribute variability was even more consequential than suggested by in Table 4. Not only did attribute variability account for the bulk of the strong trend in correlations between friends for alcohol use, it entirely eliminated the overall increases for smoking delinquency, cigarette smoking, and marijuana use (see Supplemental Figures S3 – S6). For those three, the indirect effects of attribute variability were strong enough that the original upward trends (see Figure 4) reversed and became weak downward trends.

Discussion

The present study examined the evolution across adolescence of a key feature of the peer context for development, namely, the greater similarity of friends than nonfriends. Most research on similarity and friendship has concerned the mechanisms that underly this differential. In contrast, we studied it to map the evolving social organization of adolescents’ peer experiences during this period of emerging independence from adults and increasing prominence of peers. Although the link between friendship and similarity, or homophily, is ubiquitous in human relations (McPherson et al., 2001), our results demonstrate that it is much greater for some attributes than others, in a pattern largely consistent with Kandel’s (1978) early findings. We also found that the association evolved considerably across grade levels. Further, this evolution was strongly associated with age-related change in a social institution (the size of school grade cohorts) and in distributions of individual attributes (their variability).

We expected that the profile of similarity between friends versus nonfriends would change across adolescence as growing autonomy from adults gave adolescents more freedom of choice among peers and opportunity for peer influence, and at the same time specialized classes and activities heighten exposure to peers who have similar interests, backgrounds, and skills. As we anticipated, the correlation between friends grew for most attributes, replacing the singular dominance of gender similarity during middle childhood. In the United States, adolescents’ transitions through elementary, middle, and high school correspond to increasing individuation in many life domains. The progression through the American educational system is analogous to societal-level patterns first articulated by Durkheim (1893/2013) and Simmel (1922/1955). They argued that the division of labor in the industrial age provides individuals greater freedom to rationally choose their group affiliations. This process fosters greater individuation as people find themselves at the intersections of membership in multiple groups that only partially overlap. Our findings suggest a similar developmental pattern is reproduced within individuals as they progress through contemporary age-graded schooling in the United States.

Evidence of a Changing Profile of Similarity with Friends Versus Nonfriends

The findings support our expectations about change across adolescence in the way individual attributes align with friendships. The initial correlation of .854 for friends’ gender is the highest among the attributes we examined, far above those for any of the other attributes, and of a magnitude rarely seen in social and behavioral science. Gender is also the only attribute for which there is a sustained and meaningful decrease in the correlation between friends through mid-adolescence. Even following this anticipated decrease (Poulin & Pedersen, 2007; Shrum et al., 1988), however, the correlation for friends’ gender is still over twice that for any other attribute. The correlation between friends grows across grade levels for all other attributes, except receiving free or reduced-price lunch. Most of these increases are modest, but the changes related to drug and alcohol use are striking: Despite having some of the lowest correlations in the Fall of 6^th grade, they are among the strongest correlations during high school. The correlation of friends’ school grades also increases considerably, becoming greater than any other attribute except gender.

Thus, although the importance of gender to friendship formation declines, it remains central. Meanwhile, other attributes come to play more meaningful roles in organizing the social lives of teenagers, particularly attributes related to school and to drug and alcohol use. In other words, the near total gender homogeneity of friendships in middle childhood has diminished in favor of moderate associations between similarity and friendship on a broader profile of attributes. Consistent with previous findings for alcohol use (Burk et al., 2012) and delinquency (Richmond et al., 2019), most of the increase in the association came in earlier adolescence. Our results suggest there may be some decline in correlations between friends beyond ninth grade, as Richmond et al. (2019) predicted would arise from increased involvement in romantic relationships (DeLay et al., 2016). Yet those differences do not appear statistically significant, even in our very large sample.

Ecological Mediators

We examined two ecological factors as potential mediators of age-related increases in the extent to which friends are more similar than nonfriends. Building on a study by Bahns and colleagues (2012), we reasoned that both would promote a stronger association by providing opportunities for more highly differentiated friendships. The first, attribute variability, reflects how wide a range of matches is available for an attribute (holding school size constant), while the second, size of the school grade cohort (Barker & Gump, 1964), reflects the number of alternatives for potential friends (holding attribute variance constant). We found strong evidence that both ecological factors play mediating roles. In combination, these two variables reduced change across grade levels in correlations between friends by 22.6% to 73.1% for ten of the twelve attributes, with the exceptions being gender, as expected, and free school lunch, which had no developmental trend. The strongest mediation we observed was for attribute variability as a mediator of increased correlation between friends on substance use.

These results confirm that the peer context of development is shaped by both externally imposed organization of social contexts (school cohort size) and the collective consequences of individual development (attribute variability). Friendships represent voluntary, personal connections between individuals, which combine to generate the social structure of adolescent friendships. These correlations between friends summarize systematic patterns of exposure to peers, as illustrated in Moody’s (2001) work on the racial segregation of friendships. Our results reveal that the evolution of these systematic, structural patterns in the composition of dyads is shaped by the number of potential friends provided by the organization of the local school and by the developmentally evolving variation of an attribute in this peer community.

Limitations and Future Directions

Our analyses draw on data collected as a part of the evaluation of the PROSPER program, and these data are well-suited to answering our research questions. The study is longitudinal and spans much of adolescence, following grade cohorts that constitute consistently relevant populations of potential friends. The dataset also contains direct (rather than perceptual) measures of peer behaviors and attributes. The two cohorts of students are nested within multiple, independent school districts in different communities, which allowed us to test how cohort size and attribute variability contribute to changes in correlations between friends. In addition, we structured our analyses to avoid confounding between-school differences with correlations between friends within schools.

All research designs carry limitations, and we acknowledge three characteristics of ours that limit the generalizability of our findings and suggest useful avenues for future research. First, the PROSPER sample is predominantly non-Hispanic White and drawn from non-urban communities in the United States in which a substantial portion of students were eligible for free or reduced-price school lunches. Future research should consider how the association between similarity and friendship develops within other, more diverse samples, as well as in other countries that organize schooling differently across this age range (Neal, 2020). We suspect, however, that many of our primary conclusions will not be particularly dependent on the nature of our sample. For instance, in a recent meta-analysis of peer influence on delinquency by Gallupe and colleagues (2019), estimates for our sample (Osgood et al., 2015) were exceptionally close to the composite results across diverse samples. We anticipate that different results would be most likely to emerge for aspects of our results tied to features of adolescent life that vary across settings and cultures. For instance, we found that larger grade cohorts at older ages account for some of the growth in the correlations between friends, which would not apply in communities where cohort size is constant or even declines. We would expect comparable differences in results for a sample with differing profiles of attribute variability across ages and community.

Second, our data from the PROSPER study are limited to friendships between individuals who are both within the same school and the same grade. Although this constraint usefully provides a consistent comparison across ages, it leaves out an interesting part of the picture: out-of-school and out-of-grade friendships (Jose et al., 2021; Kiesner et al., 2003; Neal, 2020). Further, such friendships are likely to increase in importance as age brings growing mobility and independence, along with the mixed-grade classes and activities typical of high schools in the United States. Our understanding of similarity and friendship would benefit from future research that also encompasses friendships that occur outside of the school grade cohort.

Third, we were able to examine the link between similarity and friendship for attributes pertaining to three important aspects of adolescent life, namely demographics, orientations to institutions, and problem behavior. Yet, many other attributes may be important as well, and friendship is not the only consequential relationship between adolescents. For instance, our research did not cover personality traits (Selfhout et al., 2007), values, political orientation, tastes, or school activities, and research is needed about the association of similarity with relationships other than friendship, such as romance, antagonism, helping, and leisure companionship (Neal, 2020). Another promising direction for future work would be to investigate the interplay of different attributes through a multidimensional approach to similarity, as demonstrated in several interesting recent studies of network ties (Block & Grund, 2014; Hooijsma et al., 2020; Meissner & Vertovec, 2015).

Focusing on the organization of peer relations provides a valuable counterpoint to the abundant research on the peer processes that create it, such as selection and influence. We have studied within-school differences in the similarity of friends versus nonfriends, but many other aspects of peer organization also merit attention. It would be valuable to consider our results about correlations between friends within schools (inbreeding homophily) in combination with information about the contribution of differences between schools (baseline homophily). Doing so would provide a means to understand patterns of friendships across the broader society (McPherson et al., 2001). The organization of peer networks also could be studied in terms of structural features such as integration (e.g., network centralization) or network level associations of various attributes with individuals’ status in the network or engagement in it (e.g., centrality based on nominations received or given).

Our investigation of similarity and friendship has shown that the social organization of the peer context of development changes across adolescence in systematic and predictable ways. Results from this project are consistent with longstanding theoretical conceptualizations of adolescent social development, they contribute to this literature using rich longitudinal peer network data. Although investigating this topic is challenging, we encourage future investigations of the structure of peer relations during this important developmental stage.