Identifying patterns and predictors of social health in adolescence using data from the Adolescent Brain Cognitive Development Study

Abstract

Social health, having an adequate quantity and quality of social relationships, is essential for well-being but understudied during adolescence compared to adulthood. We sought to identify patterns and predictors of social health by characterizing peer relationships among 10,050 adolescents (10- to 13-years old, 4,815 girls, 53.68% non-Hispanic White) in year 2 of the ABCD Study^Ⓡ. To characterize social health profiles, we applied latent profile analysis on peer variables collected in year 2: number of friends (close, general), aggression, victimization, relationships with prosocial and rule-breaking peers, and support. We then assessed whether loneliness (baseline, year 2), family conflict (baseline, year 2), and participant sex predicted profile membership. Fit indices supported a three-class solution: a “selective” class (~60% of sample) characterized by values below sample means but within population norms across variables (e.g., number of friends); a “robust” class (~30%) characterized by high numbers of friends; and a “concerning” class (~10%) characterized by high levels of peer aggression and victimization. Lonely adolescents were more likely to be in the concerning group and less likely to be in the robust group. Youth with more family conflict and boys were more likely to be in the concerning group; girls were more likely to be in the selective group. These findings reveal profiles of peer relationships in a large representative sample, providing a template for characterizing social health as adolescents begin to build intimate peer relationships. The results also highlight individual differences in social health profiles, which can inform targets to improve adolescent social health.

Introduction

Adolescence is a formative period in human social development, characterized by growing independence and an emerging focus on building close relationships with same-aged peers (Roisman et al., 2004). These new relationships not only build on social experiences in childhood (Benson et al., 2006) but also set the stage for adolescents’ future as functioning and independent adults. Positive relationships in adolescence predict positive mental health in adulthood, buffer against the effects of negative experiences, and help regulate the stress response system (Allen et al., 2022; Hostinar & Gunnar, 2015; Lamblin et al., 2017; Telzer et al., 2018). At the same time, adolescents are particularly vulnerable to experiencing social disconnection (Hang et al., 2024), with rising rates in the past ten years (Twenge et al., 2021). Adolescents face multiple barriers to social connection, including poor mental health, changes in their social context (e.g., place of education, more time spent either with friends or alone), and the challenge of maintaining high quality relationships with family while forming intimate peer relationships (Arnett, 2000; Nelson et al., 2016; Roisman et al., 2004; Sabato et al., 2021).

Given its developmental importance and the challenges adolescents face, it is crucial that researchers understand the nuances of social connection during adolescence, including how adolescents experience social health. Social health is defined as the “adequate quantity and quality of relationships in a particular context to meet an individual’s need for meaningful human connection” (Doyle & Link, 2024, p. 1). At its core, the notion of social health reflects the nature of an individual’s social relationships and the interactions that take place in these relationships. Social health has primarily been conceptualized in adults given assumptions of independence in determining social relationships (Doyle & Link, 2024; Larson, 1993), and related constructs like connectedness have primarily revolved around peer (as opposed to family) relationships (Lee & Robbins, 1995). However, because new peer relationships are a key characteristic of adolescence, assessing this developmental period is crucial for understanding social health at its foundation. Social health is person-specific: it is not only quantity and quality that matter, but also whether these two factors are adequate (i.e., satisfy social needs) as defined by the individual. However, existing work, including in adolescence, does not take a person-centered approach because it largely evaluates several different factors separately (e.g., number of friends, experiences with aggression) rather than looking across these measures holistically. This leaves open major questions for the study of social connection during adolescence: can we identify the patterns of social health during adolescence, and what factors predict social health during adolescence?

The goal of the current work was to 1) characterize social health during adolescence, focusing on peer relationships, and 2) examine key developmentally-relevant factors that may predict patterns of social health. To accomplish this, we conducted latent profile analysis (LPA) on social health data from over 10,000 early adolescents (aged 10–13 years old) from the Adolescent Brain Cognitive Development^SM (ABCD) Study. We focused on early adolescence, when the process of social reorientation is still in the early stages (Nelson et al., 2016). While the ABCD Study^Ⓡ was not designed with this goal in mind and therefore carries with it certain limitations (e.g., its measures are not directly mapped to the construct of social health), its breadth and scale offer a unique opportunity to characterize social health in a large, representative sample and can provide valuable insight into how adolescents begin to connect with peers during this crucial developmental stage.

Social health in early adolescence

Adolescents’ social health can be viewed from the lens of social re-orientation. Social re-orientation refers to the normative transition by which adolescents increasingly develop peer relationships alongside their existing family relationships and is accompanied by behavioral and neurological changes that aid in this process (Blakemore, 2008; Nelson et al., 2005, 2016; Picci & Scherf, 2016; van den Bos, 2013). Accordingly, building relationships with peers is considered a major developmental task of adolescence (Roisman et al., 2004), which aligns with models of social health that primarily focus on peer relationships in adulthood (Doyle & Link, 2024). Addressing this task (and the manner in which youth do so) matters not only for concurrent adolescent outcomes like adjustment (Knack et al., 2013; Waldrip et al., 2008) but also informs later developmental outcomes like rejection sensitivity, suicide risk, mental health, and life satisfaction (Landstedt et al., 2015; Marion et al., 2013; Masten et al., 2012; Victor et al., 2019). Therefore, understanding the nature of adolescents’ peer relationships during early adolescence may reveal key promotive and inhibiting factors that shape adolescents’ long-term development.

There are several underlying components of peer social health during adolescence, but it is unclear how these components intersect. For example, some work has focused on the number of friends, such as peer network size, given its protective associations against loneliness (Lodder et al., 2017; Ueno, 2005; Waldrip et al., 2008). However, there may be an upper limit on the number of friendships adolescents can manage before they start to become strained (Falci & McNeely, 2009). Such approaches characterize the construct of social health quantity, but do not consider the quality of those friendships. Other work has focused on measures on social support, conflict, and experiences of aggression and victimization, which align well with social health quality. For example, associations between both aggression towards peers and victimization by peers with internalizing symptoms during adolescence (Christina et al., 2021; Fite et al., 2014; Hoglund & Chisholm, 2014; Liao et al., 2022; Marshall et al., 2015; Reijntjes et al., 2010) suggest that difficulties connecting with peers amicably is a key indicator of maladjustment or poor social health. Lastly, with whom adolescents are connecting also matters. For example, adolescents who are aggressive tend to befriend adolescents who are also aggressive (Shin, 2017; Sijtsema et al., 2010), which bears important consequences for the frequency of negative interactions within relationships, whereas adolescents in prosocial classrooms tend to become more prosocial over time (Busching & Krahé, 2020), which may pave the way for positive social relationships.

However, these components of quantity, quality, and characteristics of peers with whom the adolescent is associating, are often studied separately, limiting understanding of how social health may coalesce across several different dimensions. We argue, therefore, that the study of adolescent social health must incorporate a person-centered approach. Such an approach may reveal unique characteristics or patterns of behavior that differentiate subgroups of adolescents. For example, while aggressive interactions during adolescence may signal difficulties building supportive relationships, there is also evidence that aggressors can also be prosocial (Olthof et al., 2011), yielding high rewards within friendships such as closer intimacy (Hawley et al., 2007). Similarly, while relationships with deviant or rule-breaking peers may increase the risk of internalizing symptoms and maladjustment in adulthood (Allen et al., 2019), it does not necessarily lead to social isolation (Brendgen et al., 2000) and some levels of delinquency may improve adolescents’ social status (Allen et al., 2005). These findings indicate that it is important to consider multiple variables together during adolescence to accurately characterize patterns of social health as different profiles or combinations of variables can lead to different outcomes.

Predictors of peer social health

In examining possible predictors of peer social health, we focused on three core variables that are closely tied to the adolescent social context. First, we focused on loneliness, which reflects adolescents’ evaluation of their social relationships (Cacioppo et al., 2015). Second, we focused on another vital source of social health during adolescence by examining the quality of family relationships. Third, we focused on participant sex, which, while not explicitly a “social” variable, is closely tied to how individuals learn to engage with others due to experiences like socialization (Rudolph & Dodson, 2022; Youniss & Haynie, 1992) and susceptibility to peer influence that is dependent on one’s gender or the gender of their friends (Conway et al., 2011; Deutsch et al., 2014). Thus, each of these variables offer valuable insight to begin to explore how social health emerges during adolescence.

Loneliness

Our focus on loneliness (defined as “the feelings of distress and dysphoria resulting from a discrepancy between a person’s desired and achieved levels of social relations”: Cacioppo et al., 2015, p. 202) as a predictor of social health stems from recent reports of a modern “epidemic” of loneliness (Breaux et al., 2023; Ernst et al., 2022; Jeste et al., 2020; Killeen, 1998; Sabato et al., 2021; Surgeon General, 2023) in adolescence (Hang et al., 2024; Twenge et al., 2021). Specifically, adolescence is a period of social-developmental (Laursen & Hartl, 2013) and neurobiological (Wong et al., 2018) susceptibility to perceived loneliness. While loneliness is frequently conceptualized as an outcome of social disconnection (e.g., Gallardo et al., 2018; Santini et al., 2021), it is plausible that extended periods of loneliness may lead to either extended problems connecting with others or renewed motivation to foster social connection (Hang et al., 2024). For example, loneliness has been associated with high levels of aggression and poor interpersonal skills, highlighting difficulties fostering new social connections that could alleviate this pain (Lee et al., 2001; Stenseng et al., 2014; Woodhouse et al., 2012), as well as continued withdrawal from social interaction (Goossens, 2018). On the other hand, existing work also suggests that individuals experiencing loneliness are more motivated to connect with others (Baumeister & Leary, 1995; Cacioppo et al., 2014; Gardner et al., 2005), thereby leading to the alleviation of loneliness through renewed social health. Given the crisis of loneliness facing today’s adolescents (Hang et al., 2024), it is crucial that researchers identify how and why some adolescents tend to become stuck in a vicious cycle of loneliness while others tend to experience only fleeting moments of loneliness, thereby considering loneliness not only as an outcome but also as a predictor of social health. Because loneliness data is available at baseline and year 2 in the ABCD dataset, we had the opportunity to directly test this possibility by exploring social health outcomes associated with loneliness during this developmental period.

Family relationships

Despite the emerging importance of peer relationships, family relationships continue to be important during adolescence. Relationships with family members often inform aspects of adolescents’ peer relationships (Benson et al., 2006; Delgado et al., 2022; Schulz et al., 2023), particularly by constituting the formative environment for attachment and contributing to social competence. For example, like adolescent peer relationships, adolescent family relationships also shape future romantic relationships in emerging adulthood (Crockett & Randall, 2006) and can offset feelings of social isolation (Heshmati et al., 2021). Thus, relationships within the family context are important to consider in the development of social health during adolescence.

It is also important to consider not only how family relationships predict peer relationships, but also how these two sources may coexist during adolescence as youth expand their social networks (Brown & Bakken, 2011), especially in early adolescence (Cooper & Ayers-Lopez, 1985). However, most of the existing research examining peers and family members during adolescence primarily focuses on how each group might independently influence adolescent behavior and outcomes (e.g., Branstetter et al., 2011; Farley & Kim-Spoon, 2014; Tatnell et al., 2014; van Harmelen et al., 2016; Williams & Anthony, 2015). Further, while some work focuses on changes in either family or peer relationships during adolescence (Branje, 2018; Larson et al., 1999; Laursen et al., 2017; Poulin & Chan, 2010; Veenstra & Dijkstra, 2012), researchers less frequently focus on both relationship sources together during the same period (e.g., Buhrmester, 2013; Tamm et al., 2014). The studies that assess both family and peers during this period tend to focus on attachment (e.g., Allen et al., 2007; Schneider et al., 2001) rather than the totality of social health. The inclusion of a measure of family conflict¹ at baseline and year 2 of the ABCD study allowed us to address this gap through an assessment of how family conflict may both inform and coexist with adolescent peer social health.

Participant sex

Lastly, we were also interested in how participant sex may be associated with differences in peer social health. For example, girls may experience stronger feelings of attachment and belonging to peers compared to boys (Ma & Huebner, 2008; Newman et al., 2007). Likewise, boys have been found to provide less social support to their friends (van Rijsewijk et al., 2016), although there are also sex differences in how boys and girls provide support (McNelles & Connolly, 1999). However, the support that boys do garner from peers may contribute more to their positive adjustment as compared to girls (Rueger et al., 2010). Boys also tend to perpetuate more aggressive behaviors towards peers (Card et al., 2008) and use fewer mitigating strategies to resolve conflict (Noakes & Rinaldi, 2006) than girls. Sex differences have also been found in the ways boys and girls demonstrate different socio-emotional pathways to peer acceptance (Oberle et al., 2010) and in susceptibility to peer influence, with boys being more susceptible to influence of rule-breaking peers (McCoy et al., 2019). On the other hand, evidence regarding sex differences in the number of friendships during adolescence is mixed (Almquist et al., 2014; Gest et al., 2007; Haines & Hurlbert, 1992; Rose & Rudolph, 2006; Vandervoort, 2000). These reported sex differences (or lack thereof) in social behaviors among adolescents guided our interest in understanding whether there are also discrepancies in social health by participant sex.

Current study

The goal of the present study was to identify patterns and predictors of peer social health during adolescence. In doing so, we evaluated how these patterns converge with predictions drawn from prior work about the potential overlap between multiple measures of peer relationships in adolescence (e.g., Allen et al., 2005; Olthof et al., 2011). We addressed this with data from the ABCD Study, a national ongoing effort to characterize adolescent psychosocial and neural development. This allowed us to draw inferences about adolescent social health from a large representative sample (Saragosa-Harris et al., 2022), which could have important implications for other investigations into adolescence. In addition, characterizing social health when social health data were first collected (year 2: 10- to 13-years old) may offer valuable insight into both predictors and downstream consequences associated with social health using ABCD data in future research.

The first aim of the study was to determine existing patterns of social health using latent profile analysis (LPA), which allowed us to classify participants into groups based on within-person similarities between adolescents across each of the peer variables included in the analysis. Two goals shaped how we selected measures for the LPA. First, given the developmental tasks of adolescence, including the process of neurobiological social reorientation towards peers, we aimed to focus on adolescents’ peers as core contributors to social health during this period. Second, given the emphasis of social health on indices of social relationships, including (at least) quantity, quality, and adequacy, we aimed to focus on measures that reflect adolescents’ peer relationships and how they engage with peers. Therefore, we selected seven measures that align with these goals: the number of friends, number of close friends, peer aggression and victimization, the proportion of rule-breaking friends and of prosocial friends, and peer network health. These dimensional measures each reflect aspects of adolescents’ peer relationships and the interactions that take place in these relationships, including measures of quantity (number of friends, network composition) and quality (experiences with aggression/victimization, peer network health).

In focusing on aspects of adolescents’ peer relationships as a reflection of their social health, our work adds to prior ABCD studies. Past studies have examined some of the same variables included in the present study, by largely focusing on them as individual measures such as (a) an outcome of other behaviors like screen time (e.g., Paulich et al., 2021; Wojciechowski, 2025), (b) a predictor of variables like functional brain development (e.g., Geckeler et al., 2022; Liu et al., 2025; Schiff & Lee, 2023; Shen et al., 2023), and (c) a mediator/moderator of outcomes (e.g., Conley et al., 2020; Ku et al., 2024; Liu et al., 2025; Schiff & Lee, 2023). However, many of these studies did not exclusively focus on peers, and collectively, these studies did not report how these multiple dimensions of social connection might coalesce to reveal insights about adolescents’ social health, which are core goals of the current study. Other studies that do examine multiple measures load the measures onto a single latent variable (Bates et al., 2025), rely heavily on parent-report measures (Pintos Lobo et al., 2025), and/or include individual characteristics (e.g., general prosocial behavior) alongside characteristics of adolescents’ relationships (Halbreich et al., 2023; Pintos Lobo et al., 2025). As a result, the current study addressed areas not covered in this prior work by focusing on peer relationships as a reflection of social health during early adolescence.

The use of this person-centered approach allowed us to comprehensively (i.e., across multiple dimensions) determine patterns of social health during adolescence drawn from the data. Broadly, we expected to identify social health relevant classes that varied in levels of “positive” indicators of social health (i.e., more close friends, more protective support) and of “negative” indicators (i.e., relationships with rule-breaking peers, high levels of aggression and/or victimization).

The second aim of the study was to examine predictors of adolescent social health. We focused on three predictors that are tied to key social-developmental processes of adolescence. First, we evaluated the potential effect of loneliness, allowing us to capture a vulnerability that is central to adolescence (Hang et al., 2024) and assess how this vulnerability may shape social health outcomes. Using both baseline and year 2 loneliness data, we predicted that high levels of loneliness would lead to membership in a poor social health class (Lee et al., 2001; Lin et al., 2025; Maner et al., 2007; Stenseng et al., 2014). Second, we examined individual differences in family relationship quality based on levels of family conflict, also assessed at two time points, allowing us to examine the process of social reorientation (cross-sectionally) and whether participants’ relationships with their family members shaped their relationships with peers (longitudinally). Based on prior work examining how early maternal attachment shapes later peer relationships (Seibert & Kerns, 2015) and work demonstrating within-person links between parent and peer conflict (Chung et al., 2011), we predicted that high levels of family conflict would predict membership in poor social health profiles. Leveraging the ABCD dataset, assessing loneliness and family conflict both at baseline and year 2 allowed us to capture both developmental and current processes that may influence social health during adolescence. We examined whether the confidence intervals for each effect overlapped across years to assess changes in associations over time. Thirdly, we examined the potential role of participant sex in shaping peer social health, allowing us to begin to assess how individual characteristics, and socialization associated with these characteristics, may shape social health. Given the age of the sample (early adolescence), we predicted that girls would be more likely to belong in strong social health profiles. In addition to these predictors, we also explored the potential role of the COVID-19 pandemic, which began during year 2 data collection, in shaping adolescent peer social health. Descriptions of this analysis and the findings are reported at the end of the results section and in Supplemental Materials.

Methods

Procedure

The ABCD study is a multi-site longitudinal investigation of adolescent brain development and health outcomes, beginning with baseline data collection spanning 2016 to 2018. Recruitment took place at 21 study sites, which were selected to match the national demographics of the United States. Year 2 data collection spanned 2018 and 2021 (Barch et al., 2021). In addition, recruitment procedures were designed to slightly oversample ethnic-racial minoritized youth. The variables of interest in this study were first collected at the two-year follow-up (year 2: ages 10–13). We used data from release 5.1, which included full data for baseline and year 2 assessments for the variables included in this analysis. For more information about the study design, protocol, and recruitment, please refer to https://abcdstudy.org/scientists/ and prior work (Garavan et al., 2018). Study approval was obtained prior to data collection from the institutional review boards at each of the 21 sites; centralized approval across sites was obtained from the University of California, San Diego. Parents provided written informed consent and adolescents written informed assent prior to data collection. Participants received monetary compensation for their participation.

Participants

Our initial sample included 11,868 adolescents. Because the dataset is well-powered to test sex differences for cisgender boys and girls, but not intersex children, we excluded intersex participants (N = 3) from the analysis. Further, because latent profile analysis does not have functionality to handle missing data, we excluded participants (N = 1815) with missing data among study variables at the appropriate time points (predictors: baseline and year 2; demographics: baseline; social health: year 2). Participants were 10,050 adolescents (4,815 girls, 5,235 boys; 53.68% non-Hispanic White, 17.69% multiracial, 13.63% Black or African American, 10.88% White Hispanic/Latino, 2.15% Asian or Asian American, <2.00% other). We did not exclude participants based on socioeconomic status, ethnic-racial group, substance use, mental health, or neurodevelopmental history.

Power analysis when using LPA is a complicated issue, especially when conducted in the absence of a-priori hypotheses about the number of true classes in the population and parameters are unknown (Nylund-Gibson & Choi, 2018; Tein et al., 2013). Common rule-of-thumb suggests that a sample of 500 participants is sufficient to identify the correct number of classes in a sample (Finch & Bronk, 2011; Spurk et al., 2020), suggesting that our analysis with more than 10,000 adolescents is well-powered in assessment of social health.

Measures

Peer variables

The number of friends (close and overall).

Quantitative measures of social health included a series of questions asking participants to identify the number of close and non-close (overall) friends in their life (Barch et al., 2018). Close friends were defined as “those you like spending time with, have fun with, and trust.” In each case, participants were asked to distinguish between friends who were boys (resiliency5a_y; resiliency5b_y), girls (resiliency6a_y; resiliency6b_y), and other (resiliency7a_y; resiliency7b_y). We summed across each sex group to identify the total number of close and overall friends (see Table 1 for descriptives). While data were also available at baseline, we included data only from the year 2 follow-up to maintain consistency across all social health measures.

Table 1.

Sample means with standard deviation for variables of interest

Measure	All Participants	Boys	Girls
N	10050	5235	4815
Social health (at year 2)
Close friends (#)	6.20 (5.60)	6.38 (6.05)	6.01 (5.07)
Friends (#)	20.77 (19.65)	21.77 (21.25)	19.69 (17.7)
Aggression [0–45]	10.27 (1.88)	10.42 (2.00)	10.10 (1.74)
Victimization [0–45]	12.27 (3.81)	12.27 (3.88)	12.26 (3.74)
Prosocial peers [2–15]	9.20 (2.71)	9.19 (2.68)	9.21 (2.74)
Rule-breaking peers [2–15]	3.57 (1.12)	3.68 (1.19)	3.46 (1.03)
Peer support [0–27]	11.91 (8.13)	11.21 (8.12)	12.66 (8.08)
Predictors of social health
Baseline family conflict [0–9]	2.00 (1.92)	2.12 (1.94)	1.88 (1.90)
Year 2 family conflict [0–9]	1.89 (1.79)	1.95 (1.78)	1.83 (1.81)
Baseline loneliness (% N)	11.86%	11.54%	12.21%
Year 2 loneliness (% N)	12.44%	11.35%	13.62%

Note: Cells show mean with standard deviation and proportion of sample as indicated.

Experiences with victimization and aggression.

Qualitative experiences with peers were indexed by the Revised Peer Experiences Questionnaire at year 2 (De Los Reyes & Prinstein, 2004; Prinstein et al., 2001). We specifically focused on reports of victimization from peers or aggression towards peers. The measure includes scales for overt (3 items, e.g., “a teen chased me like he or she was really trying to hurt me”), relational (3 items, e.g., “some teens left me out of an activity or conversation that I really wanted to be included in”), and reputational victimization (3 items, e.g., “A teen tried to damage my social reputation by spreading rumors about me”) as well as overt (3 items, e.g., “I threatened to hurt or beat up another kid”), relational (3 items, e.g., “I did not invite a kid to a party or other social event even though I knew the kid wanted to go”), and reputational aggression (3 items, e.g., “I gossiped about another kid so others would not like him/her”). Participants responded on a scale of how frequently they have had that experience (1 = Never; 2 = Once or twice; 3 = A few times; 4 = About once a week; 5 = A few times a week). Because our focus was on the overall experience of victimization and aggression rather than specific kinds of each, we then summed scores across each of the aggression and victimization categories (peq_ss_relational_victim + peq_ss_reputation_victim + peq_ss_overt_victim; peq_ss_relational_aggs + peq_ss_reputation_aggs + peq_ss_overt_aggression) to get total aggression and victimization scores (see Table 1 for descriptives). Cronbach’s alpha indicated high reliability for aggression (ɑ = 0.74) and victimization (ɑ = 0.84) subscales.

Relationships with prosocial and rule-breaking peers.

The characteristics of peers with whom adolescents form relationships were indexed by the Peer Behavior Profile survey at year 2 (Bingham et al., 1995). This measure allows adolescents to identify their levels of involvement with prosocial peers (3 items: how many peers 1) are athletes, 2) go to church, and 3) are good students) and with rule-breaking peers (3 items: how many peers 1) skip school, 2) have been suspended from school, and 3) have shoplifted). Participants responded on a scale about the proportion of their friends that fit into that category (i.e., 1 = None or almost none; 2 = A few; 3 = Half; 4 = Most; 5 = All or almost all). For the ABCD study, each set of prosocial (pbp_ss_prosocial_peers) and rule-breaking (pbp_ss_rule_break) items were summed to create summary scores (see Table 1 for descriptives). We treated involvement with each set of peers as separate variables for the analysis. Cronbach’s alpha indicated that the prosocial (ɑ = 0.45) and rule-breaking (ɑ = 0.55) subscales had only moderate reliability. However, the values were comparable across sex for both the prosocial (ɑ_boys = 0.42, ɑ_girls = 0.48) and rule-breaking (ɑ_boys = 0.53, ɑ_girls = 0.57) subscales and are comparable to prior reports of reliability in ABCD (Gonzalez et al., 2021). Our retention of these variables as-is, despite low reliability, followed prior work using ABCD data (Conley et al., 2023)².

Peer network health.

We used the five-item Peer Network Health questionnaire derived from the Adolescent Social Network Assessment (ASNA: Mason et al., 2015). The measure asks adolescents to complete three items to 1) identify whether their peers have suggested that they stay away from drugs (no = 0, yes = 3), 2) have given any type of support (i.e., money, transportation, emotional support) (no = 0, yes = 2), or 3) have given any type of encouragement to get or stay involved in prosocial activities (i.e., sports, school clubs, religious activities) (no = 0, yes = 2). Participants who answered yes on the support or encouragement items were then asked how much of each they received on a 10-point Likert scale (1 = a little, 10 = a lot) via follow-up items. We used the composite protective scale as computed for the ABCD study (pnh_ss_protective_scale), which sums across all items, to measure peer network protective health (see Table 1 for descriptives). To compute reliability, we applied a linear transformation so that for the first item (on suggestions to avoid drug use), an answer of no corresponded to 0 and an answer of yes corresponded to 10. For the remaining items, we collapsed across each dichotomous original item and its corresponding Likert item such that an answer of no corresponded to 0 and an answer of yes corresponded to the participant’s answer on the Likert scale. We again found moderate reliability (ɑ = 0.53) for the scale, which was consistent across gender (ɑ_boys = 0.53, ɑ_girls = 0.54). However, this estimate for reliability converges with prior work on ABCD data (Gonzalez et al., 2021), which may be related to the reduced number of items in this measure relative to the full ASNA. Given evidence for construct validity due to differences in this measure between high- and low-risk youth in line with hypotheses (Gonzalez et al., 2021), we retained this summary score as computed by the ABCD consortium.

Loneliness

Loneliness was measured at baseline and year 2 using an item (cbcl_q12_p) from the Child Behavior Checklist (CBCL), which is a parent-report measure assessing mental health symptoms in adolescents (Achenbach, 2018). Parents were asked whether their child “complains of loneliness” and responded on a 3-point scale (0 = not true, 1 = somewhat/sometimes true, 2 = very true or often true). We observed a high number of zeroes and fewer responses of 1 or 2. Therefore, we computed a dichotomous variable where 0 = no symptoms of loneliness, ≥1 = some symptoms of loneliness (see Table 1 for descriptives), which is a common approach in studies that use the CBCL (e.g., Go et al., 2022; Pandolfi et al., 2012; Sood et al., 2005; Supke et al., 2025). For implications and limitations associated with the use of a parent-report measure of loneliness in this age range, see the discussion section.

Family conflict

Family conflict was measured at baseline and year 2 with the Family Environment Scale (Moos & Moos, 1994), which is an adolescent-report 9-item scale of openly expressed conflict among family members. Participants were asked to determine whether a set of statements were true (0) or false (1); examples included, “Family members hardly ever lose their tempers” and “If there’s a disagreement in our family, we try hard to smooth things over and keep the peace”. Scores were computed in the ABCD study by summing across the items as a measure of total conflict (fes_y_ss_fc; see Table 1 for descriptives). Cronbach’s alpha indicated that the scale had moderate reliability at baseline (ɑ = 0.67) and at year 2 (ɑ = 0.64).

Data processing and analysis

All peer social health and family conflict scores were winsorized at 3 SD from the mean to account for extreme values. To account for remaining skew, we used a log transformation for the close and general friend measures, and an inverse transformation for the rule-breaking peers variable.

Our primary interests were to characterize the relationships among the peer social health variables via LPA and examine whether participant sex, family conflict, and loneliness predicted peer social health. First, we conducted LPA to identify latent patterns of overlap between variables across the participants. We used the tidyLPA package in R, which allowed us to compare models that varied in structure (parameters for variance around the mean of each variable and covariance between variables). The tidyLPA package compares solutions from four different model structures: model 1 (equal variances, covariances set to zero), model 2 (varying variances, covariances set to zero), model 3 (equal variances, covariances set to be equal), and model 6 (varying variances, covariances set to vary). These structures increase in flexibility of fit and complexity (see Rosenberg et al., 2018); model 1 assumes that classes differ only in mean levels of each indicator, model 2 allows classes to differ in both mean and variance for each indicator, model 3 restricts variances for each indicator but freely estimates covariances across indicators that are constrained to be equal for each class, and model 6 allows means, variances, and covariances to vary across classes. We considered solutions that evaluated between two and seven profiles, capping the maximum number of profiles at the number of peer social health variables included in the analysis to ensure that we did not overfit the data.

To determine the best fit for each analysis, we examined the model fit indices. Specifically, we focused on Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), which decrease as model fit improves; entropy, which represents the accuracy with which the model can assign each participant to a unique class on a scale of 0 to 1; class posterior probability, referring to the probability that a participant belongs to a given class based on their scores across the social health measures; class size, prioritizing models that do not result in disproportionately large or small classes; and results from the Bootstrap Likelihood Ratio Test (BLRT), with significant results indicating that the tested number of classes is a better fit for the data than a solution with one fewer class. Based on these criteria, we aimed to identify a model with comparatively low AIC and BIC, acceptable (> 0.70) or good (> 0.80) values for entropy, acceptable (> 0.80) or good (> 0.90) posterior probabilities across participants and classes, acceptable (> 5%) or good (> 10%) class sizes for the smallest class in the solution, and a significant BLRT for the selected class in contrast to the solution with one fewer class. These cutoff values stem from prior LPA work (Clark & Muthén, 2009; Jung & Wickrama, 2008; Nylund et al., 2007; Pastor et al., 2007). We also assessed the interpretability of each model by plotting and examining the mean scores for each variable for each profile. Once deciding on the most suitable model, we used the predicted class probabilities to assign each participant to their most likely profile of peer social health.

Next, we tested whether the three participant characteristics predicted social health. To do so, we entered the predicted class membership for each participant as a dependent variable in multinomial regressions, with separate models to evaluate the roles of loneliness, family conflict, and sex as predictors. We also conducted separate models for baseline levels of loneliness and family conflict. For each model, we set the largest profile (by class size) as the reference level. This allowed us to identify a main effect for each non-reference level corresponding to the odds of being in that non-reference level relative to the odds of being in the reference level. We report and interpret odds ratios as effect sizes for these analyses. We then evaluated the predicted probabilities extracted from the regression to determine the likelihood of being sorted into each class based on levels in the independent variable.

Transparency and openness

We report the determination of sample size, data exclusions, and measures in the study. All processing and analyses were done in R (R Core Team, 2025: version 4.5.1) using RStudio (Posit Team, 2025: version 2025.09.1–401). We used the following packages in R: corrplot (Wei & Simko, 2024: version 0.95), cowplot (Wilke, 2024: version 1.1.3), effects (Fox & Hong, 2010; Fox & Weisberg, 2019: version 4.2–2), nnet (Venables & Ripley, 2002: version 7.3–19), psych (Revelle, 2024: version 2.4.12), rstatix (Kassambara, 2023: version 0.7.2), tidyLPA (Rosenberg et al., 2018: version 1.1.0), tidymodels (Kuhn & Wickham, 2020: version 1.2.0), and tidyverse (Wickham et al., 2019: version 2.0.0). This study was not preregistered. The data are available to the public through the ABCD Data Repository (https://nda.nih.gov/abcd). Code used in this analysis is available upon request of the corresponding author.

Results

Descriptives

Mean scores for each social health variable are presented in Table 1. Before conducting the LPA, we assessed the inter-relatedness of the variables using Pearson’s correlation (see Table 2. The strongest associations were between the overall number of friends and the number of close friends, r(10050) = 0.61, p < 0.001, 95% CI [0.60, 0.63]. Further, we observed positive correlations among the friends, close friends, prosocial peers, and protective support variables (see Table 2 for estimates), perhaps indicating that these measures reflect positive indices of peer social health. Regarding negative indices of peer social health, we found that experiences with victimization tended to relate to experiences with aggression, r(10050) = 0.53, p < 0.001, 95% CI [0.52, 0.54]. In addition, victimization (r(10050) = 0.17, p < 0.001, 95% CI [0.16, 0.19]) and aggression (r(10050) = 0.23, p < 0.001, 95% CI [0.21, 0.25]) were positively associated with the number of rule-breaking friends; participants with more rule-breaking friends reported more victimization and aggression.

Table 2.

Correlation matrix of variables used for peer social health profiles at year 2

Measure	1	2	3	4	5	6	7
1. Close friends	-	0.61***	0.05***	0.00	0.21***	−0.12***	0.19***
2. Friends	-	-	0.10***	0.02	0.23***	−0.17***	0.18***
3. Aggression	-	-	-	0.53***	−0.05***	−0.23***	0.05***
4. Victimization	-	-	-	-	−0.06***	−0.17***	0.10***
5. Prosocial peers	-	-	-	-	-	−0.00	0.23***
6. Rule-breaking peers	-	-	-	-	-	-	−0.11***
7. Peer support	-	-	-	-	-	-	-

Note: N = 10050. Cells show Pearson’s product moment correlation.

^{^}p < .10

^*p < .05

^**p < .01

^***p < .001.

There was one set of perhaps counterintuitive associations between positive and negative indices of peer social health: aggression (r(10050) = 0.05, p < 0.001, 95% CI [0.03, 0.07]), victimization (r(10050) = 0.10, p < 0.001, 95% CI [0.09, 0.12]), and the proportion of rule-breaking peers (r(10050) = 0.11, p < 0.001, 95% CI [0.09, 0.13]), were positively associated with peer network health. This may reflect the specificity of the latter measure in targeting protective behaviors against substance abuse rather than other types of support (e.g., emotional support). This is particularly relevant in the case of the rule-breaking peers variables, which does not include items about drug use and instead focuses on school attendance, disciplinary record, and shoplifting. Further, because two of the items in the protective scale measure intensity (i.e., how much help or encouragement was received from peers), it is possible that this variable also reflects some degree of general peer involvement, which could explain small but positive associations with both aggression and victimization.

Latent profile analysis

Model selection

The full list of fit indices is included in Table 3. The model 2 and model 6 structures did not converge in our sample, suggesting that simpler variance structures (i.e., model 1, model 3) are better fits for the data. Further, while model 3 (equal variances, covariances set to be equal) solutions offered slightly better solutions by AIC and BIC, the values for entropy, posterior probabilities, and class sizes were much smaller and generally not acceptable. Thus, estimating covariances among indicators but constraining them to be equal across classes (i.e., model 3) limits model stability across solutions in this analysis.

Table 3.

Latent profile fit indices

Model	Classes	AIC	BIC	Entropy	Probability [min, max]	Proportion [min, max]	BLRT	BLRT p
1	1	199665.65	199766.67	1.00	[1.00, 1.00]	[1.00, 1.00]	NA	NA
1	2	192490.56	192649.30	0.94	[0.92, 0.99]	[0.11, 0.89]	4233.79	0.01
1	3	189164.94	189381.40	0.74	[0.81, 0.93]	[0.10, 0.57]	6299.78	0.01
1	4	187897.86	188172.05	0.74	[0.70, 0.93]	[0.10, 0.60]	1274.58	0.01
1	5	185868.90	186200.81	0.75	[0.73, 0.92]	[0.07, 0.54]	190.90	0.01
1	6	185489.82	185879.44	0.77	[0.69, 0.95]	[0.03, 0.54]	2090.43	0.01
1	7	185356.42	185803.77	0.70	[0.61, 0.93]	[0.06, 0.32]	172.06	0.01
3	1	188964.95	189217.48	1.00	[1.00, 1.00]	[1.00, 1.00]	NA	NA
3	2	187850.67	188160.93	0.69	[0.89, 0.93]	[0.46, 0.54]	3777.05	0.01
3	3	184738.25	185106.23	0.73	[0.84, 0.91]	[0.11, 0.48]	46.47	0.01
3	4	184759.69	185185.40	0.61	[0.00, 0.92]	[0.00, 0.47]	3465.05	0.01
3	5	184732.95	185216.38	0.50	[0.40, 0.93]	[0.12, 0.37]	35.74	0.01
3	6	181429.65	181970.80	0.60	[0.30, 0.92]	[0.03, 0.37]	−2.91	0.66
3	7	169309.11	169907.98	0.69	[0.40, 1.00]	[0.03, 0.25]	12431.59	0.01

Note: Cells include fit indices for each possible model identified in the latent profile analysis. Bolded/shaded cells correspond to the selected model fit for this analysis. AIC = Akaike Information Criterion. BIC = Bayesian Information Criterion. BLRT = Bootstrap Likelihood Ratio Test.

Across the model 1 (equal variances, covariances set to 0) solutions, we identified the best model solution as three profiles. While the two-profile solution had strong entropy (0.94), it had higher AIC and BIC values than subsequent solutions and offered limited interpretability. The three-profile solution was a significant improvement over the two-profile solution and exhibited acceptable entropy (0.74) with lower AIC and BIC relative than the two-profile solution. After the three-profile solution, entropy remained in the moderate range (Δ_entropy ~ 0.02); while entropy did improve for the five-profile (0.75) and six-profile (0.77) solutions, both solutions also revealed smaller classes (7% and 3% of sample, respectively). Indeed, class posterior probability was also highest for the three-profile solution compared to subsequent solutions, which were all below the threshold of acceptability, and smaller classes tended to be revealed after the three-profile solution, raising concerns about the stability of these solutions. Furthermore, decreases in AIC and BIC were much smaller after the three-profile solution (3 to 4: Δ_BIC = 1209.35; 4 to 5: Δ_BIC = 1971.24; 5 to 6: Δ_BIC = 321.37; 6 to 7: Δ_BIC = 75.67) compared to the contrast between one and two profiles (Δ_BIC = 7117.37) and two and three profiles (Δ_BIC = 3267.90). Lastly, while BLRT tended to support improved model fit as the number of profiles increased, we also considered prior work indicating that with large samples, BLRT more easily detects small differences between solutions, even for minor improvements (Nylund et al., 2007). Therefore, because there was not uniformly strong statistical evidence in support of extracting additional profiles across these criteria, we selected the three-profile solution as the best fit.

Mean scores for each variable for each profile are visualized in Figure 1. The first class (57.66% of sample) was characterized by lower scores for each measure compared to the means of the other classes. The second class (32.01%) exhibited the highest values among the classes for overall friends, close friends, prosocial peers, and peer support. The final class (10.33%) exhibited the highest values among classes for aggression, victimization, and numbers of rule-breaking peers.

Figure 1. Standardized means (a) and raw means (b) with confidence intervals for the three-class solution of the latent social health profile analysis.

Note: Standardized means with confidence intervals (a) and raw means with confidence intervals (b) for each variable across the three classes identified in the latent profile analysis solution using Year 2 assessments of the Adolescent Brain Cognitive Development Study.

To examine whether these differences between classes were meaningful, we conducted a series of ANOVAs with each measure (using raw scores as opposed to standardized) as a dependent variable and class as the independent variable, allowing us to determine significant differences between classes for each measure. Results revealed a significant main effect of class membership for each variable, and pairwise comparisons indicated that each comparison between classes was significant (see Table 4). Thus, these results indicate that there were meaningful differences between classes in each of the selected social health variables.

Table 4.

Class differences in study measures

Measure	ANOVA			Profile 1 vs Profile 2	Profile 1 vs Profile 3	Profile 2 vs Profile 3
	F(2, 10047)	p	η2g	p	p	p
Close friends	3123.85	<0.001	0.38	<0.001	<0.001	<0.001
Overall friends	4014.46	<0.001	0.44	<0.001	<0.001	<0.001
Aggression	8206.21	<0.001	0.62	<0.001	<0.001	<0.001
Victimization	1716.36	<0.001	0.26	<0.001	<0.001	<0.001
Prosocial peers	567.83	<0.001	0.10	<0.001	0.025	<0.001
Rule-breaking peers	459.41	<0.001	0.08	<0.001	<0.001	<0.001
Peer support	453.69	<0.001	0.08	<0.001	<0.001	<0.001

Note: Cells include test statistics for ANOVAs and pairwise comparisons designed to test differences between classes for each of the social health measures. ANOVA and pairwise comparison p-values have been separately corrected for false discovery rate.

In labelling and characterizing the profiles, we drew from extant literature. For example, while the number of friends (close and overall) is lower in the first profile compared to the other profiles, the estimated profile mean for close friends (M = 3.66) is within estimates of a prediction interval around the number of best friends from a recent meta-analysis ([2.06, 7.54]: Neal, 2025). Similarly, scores for aggression and victimization in this group are also close to means reported in prior work (Schacter et al., 2023; Tarlow & La Greca, 2021). Therefore, given that these values, while below the sample mean, are not outside population norms, we labelled this the “socially selective” group. By contrast, the second profile had numbers of friends and close friends that were outside their corresponding prediction intervals (Neal, 2025). Therefore, we labelled this group the “socially robust” group given that this profile was also high in proportion of prosocial peers and in network health. Lastly, given high scores in aggression, victimization, and the proportion of rule-breaking peers, we labelled the third profile as the “socially concerning” group.

Because there was also some evidence in support of a four-profile solution (identical entropy, lower AIC and BIC, similar class sizes, and significant BLRT), we also conducted the following analyses using the four-profile classification instead of three profiles. Descriptions of these profiles and the corresponding analyses are reported in Supplemental Materials. To summarize, this solution identified a fourth profile with values hovering around the mean for all variables, and we found similar associations with loneliness, family conflict, and gender.

Predictors of social health

Differences by loneliness

First, we ran multinomial regressions to predict peer social health class membership by parent-reported loneliness, separately for baseline and year 2. The first model revealed that baseline loneliness predicted membership in the robust profile relative to the selective profile, such that the odds of being in the robust profile were 20% lower for lonely adolescents compared to not lonely adolescents (OR = 0.80, p < 0.001, 95% CI [0.69, 0.92]). We also found an effect corresponding to differences in the likelihood of belonging to the concerning relative to the selective profile, such that the odds of being in the concerning profile were 56% higher for lonely adolescents compared to not lonely adolescents (OR = 1.56, p < 0.001, 95% CI [1.31, 1.87]). The predicted probabilities indicated that moving from the not lonely to the lonely group corresponded to a 6 percentage point reduction in the probability of belonging to the robust profile and a 6 percentage point increase in the probability of belonging to the concerning profile (Figure 2). These estimates were larger two years later (robust vs selective: OR = 0.76, p < 0.001, 95% CI [0.66, 0.87], 8 percentage point reduction; concerning vs selective: OR = 1.73, p < 0.001, 95% CI [1.46, 2.06], 7 percentage point increase), although the confidence intervals overlapped.

Figure 2. Results of loneliness predicting social health from the multinomial regression.

Note: Predicted probability with confidence intervals of belonging to one of the three classes identified in the latent profile analysis solution as predicted by levels of loneliness using baseline and Year 2 assessments of the Adolescent Brain Cognitive Development Study. See the online article for the color version of this figure.

Differences by family conflict

Next, we conducted multinomial regressions to predict peer social health class membership by family conflict. We ran separate models for baseline and year 2 levels of conflict. We found that baseline family conflict predicted year 2 social health class membership for the robust profile relative to the selective profile, such that an increase of 1 in log-transformed family conflict scores corresponded to a 11% reduction in the odds of belonging to the robust relative to the selective profile (OR = 0.89, p = 0.001, 95% CI [0.84, 0.95]). Further, we also found an effect for the concerning profile relative to the selective profile, such that an increase of 1 in log-transformed family conflict scores corresponded to a 60% increase in the odds of belonging to the concerning relative to the selective profile (OR = 1.60, p < 0.001, 95% CI [1.44, 1.77]). To interpret the predicted probabilities, we un-transformed the log-transformed family conflict scores and found that an increase of 1 on the conflict scale corresponded to a 2 percentage point decrease in the probability of belonging to the robust profile and a 2 percentage point increase in the probability of belonging to the concerning profile (Figure 3). In year 2, we no longer saw a difference in the odds of belonging to the robust relative to the selective profiles across the levels of year 2 family conflict: OR = 1.02, p = 0.518, 95% CI [0.96, 1.10]. However, we found a stronger effect of year 2 family conflict (compared to baseline family conflict) predicting membership in the concerning profile relative to the selective (OR = 3.18, p < 0.001, 95% CI [2.83, 3.58], 4 percentage point increase), again indicating that stronger family conflict corresponded to increased likelihood of belonging to the concerning profile (Figure 3).

Figure 3. Results of family conflict predicting social health from the multinomial regression.

Note: Predicted probability with confidence intervals of belonging to one of the three classes identified in the latent profile analysis solution as predicted by levels of family conflict using baseline and Year 2 assessments of the Adolescent Brain Cognitive Development Study. See the online article for the color version of this figure.

Differences by sex

Lastly, we ran multinomial regressions predicting class membership by sex. We found that sex did not predict membership in the robust profile relative to the selective profile (i.e., profile distribution across each sex group was consistent: OR = 0.95, p = 0.238, 95% CI [0.87, 1.03]). However, sex did predict membership in the concerning profile relative to the selective profile, such that girls had a 37% reduction in the odds of belonging to the concerning relative to the selective profile compared to boys: OR = 0.63, p < 0.001, 95% CI [0.55, 0.72]. The predicted probabilities indicated that there was a 4 percentage point difference for belonging to the concerning profile for girls relative to boys (Figure 4). We also observed that girls were 4 percentage points more likely to belong to the selective profile.

Figure 4. Results of sex predicting social health from the multinomial regression.

Note: Predicted probability with confidence intervals of belonging to one of the three classes identified in the latent profile analysis solution as predicted by participant sex using baseline and Year 2 assessments of the Adolescent Brain Cognitive Development Study. See the online article for the color version of this figure.

Exploring COVID-19 Pandemic Effects

Given the fact that data collection for year 2 included the onset of the COVID-19 pandemic, we were also interested in exploring the possible effects of the pandemic on adolescent’s peer relationships. Prior research indicates that pandemic-related lockdowns and social isolation made it even harder for early adolescents to reach out to others also experiencing loneliness (Sabato et al., 2021). Approximately 31% of our sample completed the year 2 assessment during the COVID-19 global pandemic (on or after March 1, 2020). To assess the potential impact of COVID-19 on these findings, we conducted a post-hoc analysis using a multinomial regression with date (before vs on or after March 1) as a predictor to assess profile membership for participants prior to or after the onset of the pandemic. We found that survey date significantly predicted the odds of being in the robust vs the selective profile, such that adolescents who completed social health measures during the pandemic showed a 42% reduction in the odds of being in the robust group (OR = 0.58, p < 0.001, 95% CI [0.52, 0.64]). Similarly, we also found that survey date predicted the odds of being in the concerning vs the selective profile, such adolescents who completed the survey during the pandemic showed a 32% reduction in the odds of being in the concerning profile (OR = 0.68, p < 0.001, 95% CI [0.58, 0.78]). The predicted probabilities indicated that participating after the pandemic’s onset coincided with a 10 percentage point decrease in the probability of being in the robust profile and a 2 percentage point decrease in the probability of being in the concerning profile (see Supplemental Figure S1).

Discussion

Adolescence is characterized by significant social transitions that facilitate peer relationships as adolescents begin to branch out from their family. The form and function of peer relationships that emerge early in adolescence likely influence their social health, a concept understudied in adolescent development. For example, aggressive adolescents may exhibit difficulties forming large friendship networks and victimized adolescents may have few prosocial friends. Using a large representative dataset from the ABCD Study, the present study aimed to characterize patterns of early adolescents’ peer relationships to advance understanding of social health as applied to adolescence as well as identify key individual differences factors that may predict different peer social health profiles concurrently and over time.

Social health during adolescence

Applying the social health framework (Doyle & Link, 2024), we utilized a latent profile analysis to examine how the number of overall and close friends, proportion of prosocial and rule-breaking peers, peer support, aggression with peers, and victimization by peers would coalesce within participants at this age to determine groups who were similar across multiple metrics. While previous studies have used LPA to characterize adolescent social behavior, including some aspects of peer relationships, few studies have characterized the multitude of ways that adolescents interact with their peers. For example, prior work has focused on types of peer popularity and status (Favre et al., 2022; Marinucci et al., 2023), types of peer aggression and victimization (Garthe et al., 2021; Gini et al., 2019; Shao et al., 2014), types of peer support (Narainsamy et al., 2024), types of peer conformity (Kosten et al., 2013), and types of peer prosocial behavior (Arbel et al., 2023). Because this prior work did not look across these dimensions, our analysis here offers new insight into how adolescents can be clustered together based on within-person similarities across multiple dimensions of peer relationships. Other studies (e.g., Y. Chen et al., 2021), including in ABCD (e.g., Geckeler et al., 2022; Halbreich et al., 2023; Liu et al., 2025; Pintos Lobo et al., 2025), begin to address this gap but do not aim to comprehensively characterize how multiple indices of adolescents’ relationships intersect, include measures of individual or community characteristics alongside characteristics of adolescents’ relationships to characterize broader level social experiences rather than social health, and do not focus exclusively on peers, limiting understanding of how adolescents’ reorientation to peers in this period is reflected in their social health. As a result, our work in this study makes an important contribution to characterizing how adolescents experience social health as they begin to expand their social networks and focus on close peer relationships.

The largest profile in this solution was the socially selective profile, so named because despite having values that were below the sample means across all variables included in the LPA, there is evidence suggesting that these participants are within population norms for this age group. For example, a recent meta-analysis indicates that this group of participants had means for the number of close friends and the number of overall friends that were within prediction intervals for the number of best friends and number of overall friends for adolescents (Neal, 2025). Thus, these participants were selective in that they had relatively low numbers of friends, but their friendship network sizes were still within population norms. The same can be said for levels of aggression and victimization compared to non-ABCD samples (e.g., Schachter et al., 2023). However, non-ABCD data for other measures (e.g., prosocial peers) are limited in availability because many of the measures used here were adapted specifically for use in ABCD (see Gonzalez et al., 2021). Even so, these comparisons to extant samples suggest that this group, while having “low” values in the context of this sample, demonstrated normative scores across multiple measures of social health. Interestingly, this attribute of the selective class may be an artifact of this specific ABCD sub-sample and solution. Specifically, in this sub-sample, this profile was the largest profile, likely including many participants who were close to the mean across variables, and therefore, within population norms. By contrast, exploratory analyses with the four-profile solution (see Supplementary Materials) reveal that many of the participants identified as “selective” in the three-profile solution are grouped into a profile that has values closer to the mean in the four-profile solution. Under this solution, the size of the selective group is much smaller, and the means for variables like the number of friends are now outside population norms. Therefore, we recommend that future work with ABCD data, especially work that uses a different subsample than what was included here, consider exploring this solution in more detail (provided it is the best fit for the data) to examine how this profile may be alternatively conceptualized in contrast to the normative profile.

The finding that participants in the robust profile had high numbers of both friends and close friends suggest that these adolescents are not only exposed to many peers, but also that they can extend these relationships to also successfully form many intimate relationships. This finding, alongside the fact that this profile also reported the highest levels of peer network health, can be contextualized alongside prior work investigating whether larger social networks may be associated with increased relationship intimacy (e.g., whether having a large broad network is associated with having more close or intimate relationships). For example, the social brain hypothesis (Dunbar, 2008) suggests that social network size is at least partially constrained by cognitive load, such that there is an upper limit on the number of close relationships individuals can maintain. Similar trends have been reported in adolescence as well, as both too few and too many friends may be associated with depression (Falci & McNeely, 2009). For this reason, individuals with large networks often still focus on a smaller group of individuals as their close relationship partners (Hill & Dunbar, 2003). The identification of a profile with numbers of friends and close friends well above the mean may begin to reveal how this constraining process originates in early adolescence, especially given recent work suggesting that early adolescents are more likely to have more “best friends” than older adolescents (Neal, 2025). Given the research indicating that there is an upper limit on the number of friends, and the finding in this study that the number of friends reported for the robust profile exceeds prior population estimates, it is also worth examining whether adolescents consistently remain in this profile over time, or if there begin to be adverse consequences (e.g., for mental health) associated with having a large network.

Lastly, the concerning profile highlights participants who experience particularly high levels of both aggression and victimization. Prior work focusing on these constructs, including using LPA in adolescents, has also identified profiles of adolescents that are high in both aggression and victimization, as well as profiles that are uniquely victimized (with lower levels of aggression) and uniquely aggressive (with lower levels of victimization) (Bettencourt et al., 2013; Graham et al., 2006; Williford et al., 2011). Thus, the fact that our LPA did not reveal profiles of uniquely aggressive or uniquely aggressive adolescents stands out among prior work. It is possible that other latent profile solutions would have also resulted in profiles that were uniquely aggressive or victimized alongside our identification of a co-occurring profile (although this is not the case for the four-profile solution: see Supplemental Materials). In our sample, these participants did not exhibit similarly high or low scores in most of the other measures, including in the number of friends. This potentially contrasts with prior work that has reported difficulties for both aggressive and victimized adolescents in connecting with others (Berger & Rodkin, 2009; Ellis & Zarbatany, 2007). The lone exception was that these participants also reported relatively high proportions of rule-breaking peers. This aligns with prior work indicating that aggression and/or victimization are positively associated with engaging in deviant behavior and having friendships with deviant peers across early and middle adolescence (Benson & Buehler, 2012; Jiang et al., 2016; Sullivan et al., 2006; Thompson et al., 2020; Zhu et al., 2016). As a result, we conclude that this group of adolescents present key vulnerabilities in their social health as indicated by high levels of negative experiences and interactions with their peers and relatively high involvement with rule-breaking peers, which may have important long term implications for outcomes like levels of depression (Blain-Arcaro & Vaillancourt, 2017), substance abuse (Herrenkohl et al., 2009; Price et al., 2019), physical health (Ames et al., 2019), and suicidal ideation (Z. Chen et al., 2025).

Together, our identification of these three profiles begins to provide insight into how adolescents experience social health in their peer relationships. For example, recall that social health entails the adequate quantity and quality of relationships in a particular context to meet an individual’s need for meaningful social connection (Doyle & Link, 2024); because adequacy is defined on an individual basis, one should expect that individuals vary in their constellation of how measures of social health quantity and quality align. Accordingly, in this study we found a profile that was primarily characterized by extreme quantity (i.e., robust profile) as well as a profile that was primarily characterized by quality (i.e., concerning profile). As such, the characteristics of the profiles from these solutions align with the social health model. At the same time, it should be noted that because of the measures used in the study, our approach may have been limited in identifying the unique ways adolescents may differentially assign weight to quantity and quality. For example, while many of the selected measures reflected quality (e.g., proportion of peers, aggression/victimization), only two reflected quantity. Even further, the degree to which these measures accurately reflect social health quantity and quality may vary. Thus, while these data begin to answer questions about social health in early adolescence, more work is needed to better uncover this phenomenon.

Associations with social health

After we had identified the three social health profiles, we examined three separate predictors of profile membership: loneliness, family conflict, and participant sex. Each of these three factors are closely tied to youth’s social experiences during this developmental period. Given the design of the ABCD study, we leveraged the opportunity to take a two-pronged approach in the case of loneliness and family conflict: we focused on between-persons associations between these variables and social health to determine how social health may correlate with other social indices, and we examined within-persons associations between these variables and social health to determine predictors of peer social health profiles.

Regarding loneliness, we hypothesized that lonely adolescents would report worse social health. We found that while the distribution of profiles (i.e., selective > robust > concerning) was consistent across the lonely and not-lonely groups, lonely participants were less likely to belong to the robust profile and more likely to belong to the concerning profile compared to their peers (Figure 2). This pattern was consistent when evaluating both baseline and year 2 loneliness data, especially given the overlapping confidence intervals for the odds ratios from both associations. Thus, the fact that there is an effect for both baseline and concurrent loneliness reveals insight about processes associated with loneliness in adolescence, indicating that loneliness has important implications for both concurrent social health and for social health prospectively. The fact that loneliness is tied to both aggression and victimization in this sample fits well with prior work investigating the vicious cycle of loneliness (Hang et al., 2024). On the one hand, this pattern aligns with the aggression hypothesis of loneliness, which argues that individuals who experience loneliness are socially frustrated and may therefore become aggressive (Stenseng et al., 2014). On the other hand, it also aligns with work suggesting that victimization and loneliness are positively associated (Almeida et al., 2021; Catterson & Hunter, 2010; Machimbarrena et al., 2019; Matthews et al., 2022; Storch et al., 2003), such that the perceived absence of close social relationships may result in lonely adolescents becoming targets for more victimization (Cappadocia et al., 2013; Espinoza et al., 2020; González-Abaurrea et al., 2025). Previous work has hypothesized about a variety of neurobiological (e.g., autonomic nervous system activity), physiological (e.g., stress reactivity), personality (e.g., intraversion), and contextual (e.g., parental relationship quality) factors that may lead to lonely adolescents responding with aggression and/or social withdrawal (Hang et al., 2024). While the contributing factors to this pattern were not considered here, we advise future researchers to help clarify the mechanisms by which loneliness appears prospectively associated with aggression/withdrawal in adolescence.

Second, we evaluated whether experiences with family conflict would predict profile membership using both baseline and year 2 data. This approach allowed us to assess how family and peer relationships might intersect throughout adolescence. While family conflict was associated with membership in the concerning profile both concurrently and longitudinally, the association was much stronger concurrently (i.e., non-overlapping confidence intervals), indicating that while prior family conflict matters, current family conflict matters much more. Conversely, the association between membership in the robust profile with baseline but not concurrent family conflict (with non-overlapping confidence intervals) indicates that this association may be cumulative and emerges developmentally over time, such that an association between conflict and membership in the robust profile is not tied to adolescents’ current context. The fact that we found different associations over time depending on the profile of interest may also highlight the complexity of assessing social health during this developmental period, given the dynamic changes in both peer (Nicolaisen & Thorsen, 2017; Wrzus et al., 2013) and family (Tsai et al., 2013; Whiteman et al., 2011) relationships during this period, and the importance of taking a person-centered rather than unidimensional approach to understand social health. Given the differences between the robust and concerning profiles, the differing associations with family conflict over time may have important implications for explaining how certain patterns of peer behavior are related to the adolescents’ family context.

Lastly, we examined whether participant sex would predict social health profile membership. This follows prior work investigating how boys and girls may differentially form and experience their peer relationships during adolescence (e.g., Oberle et al., 2010). Results in this study indicate that a larger proportion of participants in the selective profile were girls, whereas a larger proportion of participants in the concerning profile were boys. There was no effect of sex for the robust profile. In other words, these results indicate that boys may be more likely to experience aggression and victimization and to associate with rule-breaking peers. There is some precedent for this idea, particularly emerging from work investigating the aggression hypothesis. While evidence regarding sex differences in rates of aggression and victimization are mixed, there have been reports that boys and girls perpetrate and/or experience aggression differently (e.g., boys tend more towards physical aggression: Björkqvist et al., 1992; Zimmer-Gembeck et al., 2013) and that boys are more likely to rate aggressive behaviors as normative relative to girls (Crick et al., 1996). There has also been work suggesting that boys are more likely to respond to rejection with stronger agentic social goals, corresponding to traits related to dominance (e.g., aggressive behavior, narcissism, high self-esteem) while girls respond with weaker agentic goals (Findley-Van Nostrand & Ojanen, 2022). Therefore, these findings add to this prior work by suggesting that boys tend towards aggression/victimization and girls tend towards withdrawn behavior. The fact that there were no sex differences for membership in the robust profile may also align with prior work. For example, the aforementioned meta-analysis on the number of friends in adolescents’ social networks indicated that sex does not moderate the number of self-reported best friends (Neal, 2025). Interestingly, there has also been some work on how boys and girls may differentially value certain aspects of friendships (e.g., social support vs companionship: Rudolph & Dodson, 2022) and define what it means to be a friend (Kitts & Leal, 2021). Because these ideas could not be assessed in the current analysis, it is possible that future LPA work that looks at not only the existence of friends but also in what these friends may mean to boys and girls may reveal more insights about sex differences beyond the mere number of friends.

Given prior reports of negative psychosocial outcomes in adolescence during the pandemic (X. Chen et al., 2021; Magis-Weinberg et al., 2025; Sabato et al., 2021), we also assessed whether the onset of the pandemic predicted profile membership via a post-hoc analysis. Results confirmed that adolescents who participated during the pandemic were disproportionately likely to belong to the selective profile. Thus, this preliminary finding may align with the broader literature. However, it is important to note that the relative distribution of profile membership (e.g., selective > robust > concerning) was consistent prior to and during the pandemic (see Supplemental Figure S1). This suggests that while the pandemic may have influenced peer social health, the findings we described here are likely applicable to adolescents before and after the pandemic. This could be confirmed via a follow-up analysis focusing on the approximately 3000 participants in this sample who participated after the pandemic. We recommend future research extend this preliminary analysis to assess the effect of the pandemic on adolescent social outcomes in the ABCD Study.

Limitations

Our study is not without limitations. First, it is worth noting that the solution obtained in this study represents a proxy of social health because the measures included in the ABCD Study were not designed a priori to reflect our definition of social health (Doyle & Link, 2024) definition. We selected variables in the dataset that correspond to aspects of adolescents’ relationships with peers, which do map well onto the definition. For example, the number of friends is a direct reflection of social health quantity, and experiences of aggression and victimization are likely also strong reflections of social health quality. However, because we used secondary data analysis in this study, this analysis is limited in that the measures we used were not designed to correspond to social health. For example, the variables included in this study do not have a strong analog to social health adequacy (i.e., a measure of whether quantity and quality together satisfy adolescents’ need to connect with others). Even further, there are limitations regarding the measures themselves, some of which had low reliability in this sample, potentially indicating concerns about whether the items fit together (e.g., prosocial and rule-breaking peers measures). Similarly, because a small portion of youth only completed some of the social health measures, their data were excluded due to the analytic approach used, raising the possibility for data not missing at random and potential limits on the generalizability of results. At the same time, we argue that despite these limitations, this work offers valuable insight into the nature of social health in adolescence because we assessed key aspects of adolescents’ peer relationships during a time in which these peer connections are becoming more central. Even further, doing so in the ABCD dataset offers valuable opportunities for future work to examine how social health might be implicated in the core focuses of the project at large, including substance abuse and brain development. Our goal in this work, in sum, is to extend ongoing conversations in the literature about the nature of adolescent peer relationships to the ABCD dataset and evaluate how researchers can leverage the dataset to begin to explore social health. Therefore, future research in ABCD and other adolescent samples would benefit from considering other variables that might more precisely track the components of the social health concept, especially adequacy.

Second, it should also be noted that our measurement of loneliness was limited in that 1) we relied on parent-report because ABCD does not include self-report measures of loneliness, 2) the scale used a small range (three possible responses), and 3) the distribution was skewed such that most adolescents were categorized as “not lonely” when we dichotomized the variable. Only one prior ABCD study (Harman et al., 2021) has reported using this item to assess loneliness, and did not report information about the distribution or about any transformations done to handle skew. These limitations are important to consider because we found that approximately 12% of participants could be categorized as lonely (see Table 1); while this does align with recent work using this same parent-report single-item measure (9–13%: Supke et al., 2025), there has also been prior work suggesting that rates of adolescents who report feelings of loneliness can exceed 50% (Farrell et al., 2023; Hang et al., 2024). In addition, there has been prior work showing discrepancies between how parents and youth may rate adolescent loneliness (Farrell et al., 2023; Go et al., 2022), with parents tending to underestimate (Supke et al., 2025). For these reasons, it is possible that by using parent-report via the CBCL, we identified only a subset of participants who would have been categorized as lonely using a self-report measure³ (i.e., there were many false negatives). Even so, our approach to use this variable was consistent with the goals of the study and the data that is available through ABCD. Further, the effect sizes (i.e., odds ratios) were robust in revealing effects of loneliness on social health, even with this possibly reduced sample size. Therefore, we suggest that our analyses with these data in the current study serve as an initial assessment of the associations between loneliness and social health in adolescence, and we advise future researchers to either 1) consider combining the single loneliness item with other CBCL items to measure constructs like social withdrawal in the ABCD dataset, or 2) integrate self-report measures in other studies.

Third, it is worth noting that latent profile analysis necessarily involves a partially-subjective approach, particularly when fit indices across multiple different solutions are similar. In the case of this study, we selected the solution that appeared to be most appropriate based on the indices evaluated and on interpretability. However, there were other solutions that also appeared suitable (e.g., comparable entropy, strong BLRT values, etc.) that may have led to slightly different conclusions. While this may be considered a limitation, we also view it as an opportunity for other researchers to conduct similar analyses with data from the ABCD Study or other studies and evaluate alternative ways to characterize peer social health during adolescence. These alternate approaches may inform future work using similar or more nuanced measures. For transparency and to guide future work, we have included the results of our analyses that revealed a four-profile solution in the Supplemental Materials.

Future implications

As previously noted, the size and scope of the ABCD study presents a unique opportunity for researchers to thoroughly investigate the period of adolescence along a wide array of dimensions. In the context of this study, we aimed to understand how peer relationships are informed by loneliness, family conflict, and participant sex. Given the breadth of the ABCD study, we encourage future researchers to consider other potential factors shaping social health. For example, prior work indicates not only that family social health may change throughout adolescence, but also within peer relationships, adolescents may eventually tend to draw more and more support from romantic partners (Furman & Buhrmester, 1992), suggesting differences in social health by age. Systemic barriers associated with position factors like socioeconomic status (Grütter et al., 2021) and ethnicity-race (Kao & Joyner, 2004; Rivas-Drake et al., 2017) likely also inform social health because they may restrict opportunities to engage with peers. Other variables that may influence adolescent peer social health include mental health symptomology, which is a particularly pressing concern in adolescence and may negatively impact social functioning (Long et al., 2020), physical health, which may influence ability to participate in activities to connect with peers (De La Haye et al., 2011), neurodevelopmental history, as youth with disorders like autism may experience difficulty connecting with their peers (Locke et al., 2010) and rates of brain development can correspond to changes in social cognition (Becht et al., 2021), personality traits, which may influence how adolescents engage with their peers (Yu et al., 2014), and other contextual factors like school environment (Way & Greene, 2006). Given the breadth of coverage provided by the ABCD study, we suggest that future research utilize this dataset to better understand how social health may be informed by these factors and others that are central during adolescent development.

Beyond the ABCD study, the results of the present study also set a foundation for further investigating the combined roles of family and peer relationships in adolescence. The current study utilized one measure of adolescent family relationships, focusing on family conflict. In future work, researchers should expand this assessment of family relationships to include other measures like family cohesion (which was collected starting in the year 4 follow-up of the ABCD study), relationships with extended family members (grandparents, aunts and uncles, etc.), and differentiate between parental and sibling relationship quality. Each of these measures may be differentially associated with peer social health in adolescence. Furthermore, additional measures of family relationships would provide an opportunity to conduct a more comprehensive latent profile analysis of social health that incorporates both peer and family relationships. Therefore, we recommend that future researchers incorporate a range of both peer and family relationship measures during adolescence to better determine within-person similarities as well as characterize social health.

Further, it should be noted that our focus on social health as a reflection of how adolescents connect with their peers does not account for other ways to empirically assess these connections. Specifically, a recent review, aimed at identifying ways to measure social connection, included measures of social health (e.g., subjective reports of relationship quality, measures of network size) alongside measures like interpersonal biobehavioral synchrony (i.e., heart rate synchrony: Baek et al., 2025). As such, we recognize that there are other important ways to understand how adolescents can connect with their peers and argue that future researchers should carefully consider how these levels of analysis may or may not intersect.

Finally, our primary focus in the present study was on identifying patterns and predictors of social health in adolescence. As a result, we did not explore social health as a predictor. For example, loneliness may result from poor social health (e.g., Gallardo et al., 2018; Santini et al., 2021), and several studies using ABCD data have examined how adolescents’ relationships may inform their neurobiological development (e.g., Geckeler et al., 2022; Liu et al., 2025; Schiff & Lee, 2023; Shen et al., 2023). Even further, some of the measures we included in the LPA, like aggression, may pique researchers’ interest as outcomes of poor social health rather than as direct indices of social health. We encourage future researchers to explore these directions to create a more comprehensive picture of how social health may operate during this developmental period, especially as ABCD data continues to be released.

Conclusions

In this study, we characterized social health in early adolescents by identifying how they experience peer relationships and interactions across multiple different dimensions, and evaluating how loneliness, family relationship quality, and participant sex were associated with adolescents’ peer relationships. We implemented the social health framework (Doyle & Link, 2024), from which we derived a multifaceted definition of social health that emphasizes the variety of ways in which individuals may navigate their social contexts. To this end, we used a person-centered approach by applying latent profile analysis on several measures related to adolescent peer relationships. We found that a three-profile solution was the best fit for the data, including selective, robust, and concerning patterns of peer social health. Even further, we found that loneliness, family conflict, and sex predicted profile membership concurrently and over time, with the strongest association emerging between the cross-sectional assessment of family conflict and social health. These findings, stemming from a large and representative dataset, serve as an important step in identifying the various facets of how adolescents are beginning to build close relationships with peers. Future research should advance these findings by expanding our operationalization of social health, including by assessing whether adolescents perceive their relationships to be adequate, and by assessing how social health might develop alongside mental health symptoms and the adolescent brain. Such knowledge can inform more targeted approaches to supporting adolescent social development and well-being and preventing social isolation during this critical period of development.

Public significance statement:

By analyzing multiple aspects of adolescents’ peer connections as a marker of their social health, we revealed that adolescents cluster into distinct relationship profiles that differ meaningfully as a function of feelings of loneliness, conflict with family members, and sex assigned at birth. These findings, therefore, hold promise for understanding the ways teens interact with other peers and the factors that contribute to these interactions. In turn, this can help parents, educators, and mental health professionals better identify adolescents who are struggling socially and tailor interventions accordingly.