The anatomy of friendship: neuroanatomic homophily of the social brain among classroom friends

Abstract

Homophily refers to the tendency to like similar others. Here, we ask if homophily extends to brain structure. Specifically: do children who like one another have more similar brain structures? We hypothesized that neuroanatomic similarity tied to friendship is most likely to pertain to brain regions that support social cognition. To test this hypothesis, we analyzed friendship network data from 1186 children in 49 classrooms. Within each classroom, we identified “friendship distance”—mutual friends, friends-of-friends, and more distantly connected or unconnected children. In total, 125 children (mean age = 7.57 years, 65 females) also had good quality neuroanatomic magnetic resonance imaging scans from which we extracted properties of the “social brain.” We found that similarity of the social brain varied by friendship distance: mutual friends showed greater similarity in social brain networks compared with friends-of-friends (β = 0.65, t = 2.03, P = 0.045) and even more remotely connected peers (β = 0.77, t = 2.83, P = 0.006); friends-of-friends did not differ from more distantly connected peers (β = −0.13, t = −0.53, P = 0.6). We report that mutual friends have similar “social brain” networks, adding a neuroanatomic dimension to the adage that “birds of a feather flock together.”

mutual friendship, neuroanatomic homophily, social networks, social neuroscience

Introduction

Peer relationships serve a number of functions for children, including reducing feelings of loneliness and providing the settings for the socialization of appropriate emotional, cognitive, and social development (Nangle et al. 2003; Erdley and Day 2017). However, despite repeated demonstrations of the importance of friendships for adaptive development and functioning in children and adolescents, little work to date has examined the biological underpinnings of reciprocal friendships in developmental populations (Vitaro et al. 2009; Bagwell and Bukowski 2018).

Friendships are built upon positive social interactions, which in turn depend upon shared intentionality; that is, they require the persons involved in a social exchange to represent, understand, and share each other’s cognitive, affective, and motivational states (Dunn and Cutting 1999; Rossignac-Milon et al. 2021). These processes themselves have been proposed to be highly dependent upon both the structure and function of the regions that we refer to in this study as the “social brain”—namely the temporo-parietal junction, dorsomedial prefrontal cortex, anterior prefrontal cortex/frontal pole, ventromedial prefrontal cortex, middle temporal gyrus, and amygdala (Molenberghs et al. 2016; Tso et al. 2018). This set of brain regions shows robust co-activation during false belief tasks, while viewing or reading about social interactions, during tasks requiring trait judgments on others, during games requiring the anticipation of a partner’s future decisions, and during the assessment of facial expressions for emotions and trustworthiness (Van Overwalle 2009; Mar 2011; Denny et al. 2012; Schurz et al. 2014; van Veluw and Chance 2014; Molenberghs et al. 2016; Adolfi et al. 2017; Tso et al. 2018). Studies using brain stimulation methods provide further support for a role of the “social brain” in social cognitive processes (Costa et al. 2008; Penton et al. 2020), with a particularly strong body of work demonstrating roles for frontal and temporo-parietal regions in social cognition over and above their roles in more domain general salience detection and spatial attention functions (Sowden et al. 2015; Sowden and Catmur 2015; Penton et al. 2020). Furthermore, these regions have been reported to learn and represent the structure of social networks, including the position of oneself and others in such networks (Zerubavel et al. 2015; Parkinson et al. 2017; Morelli et al. 2018; Parkinson and Du 2020; Tompson et al. 2020). The “social brain” therefore appears to play a key role in interpreting an individual’s interpersonal interactions and surrounding social environment (Tso et al. 2018; Nguyen et al. 2019). Moreover, people who have similar patterns of “social brain” activation while observing social scenarios also report similar interpretations of those scenarios (Nguyen et al. 2019) and demonstrate more successful social exchanges (Redcay and Schilbach 2019; Hoehl et al. 2020), further suggesting that similarities within underlying “social brain” regions allow people to be “on the same wavelength” in social situations.

Humans tend to befriend similar others, such as based on manifest characteristics including demographic, personality, and physical features, a preference known as homophily (McPherson et al. 2001). In recent work, Parkinson and colleagues showed that similarity in brain activation elicited by naturalistic movie stimuli designed to evoke social cognition and emotional processing was associated with social distance in a friendship network, such that mutual friends had the most similar patterns of brain activation (Parkinson et al. 2018; Hyon, Kleinbaum, et al. 2020). In a follow-up study, they showed that intersubject similarity in resting-state functional connectivity was positively associated with social network proximity, particularly for functional connections involving nodes of the default mode network, which shows significant anatomical overlap with what we refer to here as the “social brain” (Hyon, Youm, et al. 2020), although see McNabb et al. (2020) for an important negative finding using a similar study design. In other words, friends showed more similar brain phenotypes or neural homophily.

A separate body of work using structural neuroimaging has linked individual differences in gray matter volume and thickness within “social brain” regions to differences in social cognitive abilities in youths and adults (Lewis et al. 2011; Banissy et al. 2012; Rice et al. 2014; Eres et al. 2015; Rice and Redcay 2015; Sato et al. 2016; Batista et al. 2017; Yin et al. 2018; Baribeau et al. 2019; Wiesmann et al. 2020), indicating that structural differences within the “social brain” can also influence its functioning, as well as to social network properties such as social network size (Lewis et al. 2011; Bickart et al. 2012; Kanai et al. 2012; Powell et al. 2012; Von Der Heide et al. 2014; Kwak et al. 2018). This raises the possibility that homophily observed at the level of brain activation and connectivity may also apply to patterns of gray matter structure within brain regions subserving social cognition.

Therefore, in the present study, we aimed to build upon previous work by examining whether neural homophily extends from brain “function” to brain “structure” in a sample of elementary school children. To address this question, we first defined friendship networks formed within classrooms. We then used in vivo neuroanatomic imaging to define the “social brain” in a subset of the same children. Specifically, we calculated gray matter volume for each of the social brain regions for each child. We then placed each child’s “social brain” in 12D space, in which gray matter volume for each of the “social brain” regions defined one of 12 coordinates. Finally, we examined the relationship between friendship status for each dyad in the friendship network and intersubject similarity in “social brain” anatomy, defined as the pairwise Mahalanobis distance between the “social brains” of two children in 12D space. We hypothesized that children who like one another would show similar neuroanatomy within the “social brain.” Conversely, we predicted that friendship-associated anatomic similarity would not extend to nonsocial brain networks, such as brain regions supporting executive functioning and salience processing.

Materials and Methods

Social Network Characterization

Elementary schools in Rotterdam received a letter and booklet about the study, were invited to visit a website describing the study, and were provided a demonstration version of the peer nomination measure (details in Verlinden et al. 2014). Informed passive consent for participating in the peer nomination task was obtained from parents and children from the participating schools. Once parents had been informed about the study, they had an opportunity to withdraw their child from participation by informing a teacher or researcher before the assessment. Children were informed at school about the research and gave oral assent before the assessment. The decision to use passive consent was based on five considerations (Verlinden et al. 2014). First, passive consent procedures had been used in previous peer nomination studies of young children (Vermande et al. 2000; Monks et al. 2003; Perren and Alsaker 2006). Second, passive consent was expected to reduce the risk of failing to obtain representative reports from peers owing to selection bias and low participation rates. Third, positive feedback was received from schools, which took part in a pilot study (Verlinden et al. 2014). We also drew on the experience of Rotterdam’s City Public Health Services, which use passive consent in the administration of yearly questionnaires at local elementary schools. Fourth, we considered the nonexperimental nature of the study as well as its negligible health risks. Fifth, we ensured the opportunity for subjects to withdraw from the study. This included children opting not to participate on the day of testing, as well as the opportunity for parents to withdraw their child’s participation (Verlinden et al. 2014). Aside from participation in the peer nomination substudy, full written informed consent was acquired for participation in the main Generation R cohort study and Generation R imaging substudy (see Jaddoe et al. 2010; White et al. 2013 for details). Study procedures were approved by the Medical Ethics Committee at Erasmus University Medical Centre in Rotterdam, The Netherlands.

We formed classroom networks based on data collected from 1186 children from 49 classrooms, in which at least two children had usable neuroimaging data following quality assessments (see below). Each child nominated up to six classmates as friends using a web-based computer program within the same classroom. The program required children to imagine that they were going on a school trip. When prompted, they clicked on the photos of classmates that they wanted to invite on the trip, indicating friendship or peer acceptance (for details, see Verlinden et al. 2014, 2015). Networks were based on reciprocal friendship ties; that is, the children nominated one another for the trip. The social distance between two peers was taken as the geodesic distance, that is, the smallest number of intermediary social ties taken to connect two peers (determined using classroom networks). The distribution of geodesic distances is shown in Supplementary Figure 1. Three groups were then defined: mutual friends (peers who nominated one another—geodesic distance = 1); friends-of-friends (geodesic distance = 2); and minimally connected or unconnected peers (geodesic distance = 3 through no connection). Classrooms included an average (SD) of 24.02 (±3.62) students and a range from 11 to 31 students.

Magnetic Resonance Imaging Study Subjects

Under a separate medical ethics approval and consent, a subsample of children were invited for neuroimaging as part of the first wave of the larger Generation R imaging substudy (White et al. 2013). The mean amount of time between completing the peer nomination task and the magnetic resonance imaging (MRI) scan was 0.84 years (SD = 0.57), with 48 subjects first completing neuroimaging procedures and 77 subjects first completing the peer nomination procedures. Generation R is a longitudinal, population-based cohort of children born in Rotterdam between 2002 and 2006 with structured follow-up visits from fetal life to the present (Jaddoe et al. 2006, 2010). The first neuroimaging wave in Generation R included both children with normal range Child Behavior Checklist (CBCL) scores selected at random from the larger Generation R cohort, as well as a subsample of children selected owing to their CBCL scores falling into borderline or clinical ranges (see White et al. 2013 for details on the selection of some children based on CBCL scores). For the internalizing and externalizing subscales, scoring in the borderline range was defined as scoring between the 83rd and 91st percentile of respondents while scoring in the clinical range was defined as scoring at or higher than the 91st percentile of respondents, according to Dutch population norms (Achenbach and Ruffle 2000; Tick et al. 2007; White et al. 2013; Muetzel et al. 2018). In other words, in a sample representative of Dutch population norms, we would expect 8% of children to score in the borderline range and 9% to score in the clinical range for each subscale. However, for the present subset of subjects with usable neuroimaging and peer nomination data selected from the larger imaging cohort, the number of children scoring in the borderline range for the internalizing subscale was 12 out of 123 (8.94%), while 16 out of 123 children (13%) scored in the clinical range, meaning that borderline and clinical internalizing problems were slightly overrepresented in the present sample. In contrast, 6 out of 124 children (4.84%) scored in the borderline range for the externalizing subscale and 4 out of 130 children (3.23%) scored in the clinical range for this subscale. Therefore, high externalizing problem scores were slightly underrepresented in the present sample compared with Dutch population norms. Moreover, the majority of children with available CBCL data (94 out of 123, 76.42%) scored in the normal range on both of these CBCL subscales. Exclusion criteria for taking part in the Generation R imaging substudy included any MRI contraindications (e.g., pacemaker, ferrous metal implants), claustrophobia, motor or sensory disorders, moderate-to-severe head trauma with loss of consciousness, and neurological disorders (e.g., seizure disorder, neuromotor disorder, or a history of brain tumors).

Written informed consent for scanning was obtained from both parents and assent was given by the children (White et al. 2013). In total, N = 125 children with social network data and who shared a classroom with at least one other child with neuroimaging data were included in the present analysis. We used image quality ratings provided as part of the Generation R dataset, and which were assessed according to a standardized protocol by trained members of the Generation R research team (White et al. 2013, 2018). The quality control procedure first involved inspecting the unprocessed T₁ images. This visual inspection included assessments of the sharpness of the gray and white matter interface on the cortex, the presence of ringing in the image, and overall brain coverage. Based on this inspection, image quality was rated on a six-point Likert scale (0 = unusable, 5 = excellent quality). We opted to retain only images rated |$\ge$|3 (“good”) for analysis, leading to the exclusion of 31 scans. After the images were processed through the FreeSurfer pipeline, a subsequent visual inspection of the segmentation quality of the FreeSurfer constructed images was performed. FreeSurfer data that were rated as unusable, poor, or images that failed FreeSurfer construction were not used, leading to the removal of five subjects from the analysis (Mous et al. 2014). Full demographic details for the final 125 children are provided in Table 1.

Table 1

Demographic and behavioral characteristics for the 125 participants with phenotyping and neuroimaging scans

Measure	N	N (%) or Mean (sd)
Age (peer nomination task)	125	7.57 (0.8)
Age (MRI scan)	125	7.88 (1)
Gender	125	Female: 60 (48%) Male: 65 (52%)
Race/ethnicity	125	Dutch: 98 (78.4%) Other western: 6 (4.8%) Non-western: 22 (17.6%)
Maternal education	116	Up to secondary: 35 (30.17%) Higher education: 81 (69.82%)
Household Income	117	Below social security: 11 (9.4%) Medium income: 36 (30.77%) Above modal income: 72 (61.54%)
CBCL Internalizing problems	123	6.52 (5.06)
CBCL Externalizing problems	124	8.62 (6.71)
Mutual friendsa	125	0.11 (0.07)
In-degree centralitya	125	0.21 (0.12)

^aValues normalized for classroom network size (Szekely et al. 2016).

Imaging Procedures

The neuroanatomic data were collected on a 3 Tesla scanner (General Electric Discovery MR750) using an eight-channel head coil for signal reception. Following three-plane localizing and coil intensity calibration scans, a high-resolution T₁-weighted inversion recovery fast spoiled gradient recalled sequence was obtained with the following parameters: TR = 10.3 ms, TE = 4.2 ms, TI = 350 ms, NEX = 1, flip angle = 16°, readout bandwidth = 20.8 kHz, matrix 256 × 256, imaging acceleration factor of 2, and an isotropic resolution of 0.9 × 0.9 × 0.9 mm³.

Defining Brain Networks

We defined “social brain” anatomic regions based on prior imaging studies of core social cognitive skills and the literature on social networks. Specifically, findings of multiple meta-analyses of functional imaging studies of theory of mind have most consistently centered on the temporo-parietal junction, dorsomedial prefrontal cortex, ventromedial prefrontal cortex, and anterior medial prefrontal cortex/frontal pole, with recent meta-analyses further implicating lateral temporal regions (extending from the temporal pole along the middle gyrus to the posterior superior temporal gyrus) (Van Overwalle 2009; Mar 2011; Denny et al. 2012; Schurz et al. 2014; van Veluw and Chance 2014; Molenberghs et al. 2016; Adolfi et al. 2017; Tso et al. 2018). The amygdala has been implicated closely in assessing emotional facial expressions, as well as in assessments of trustworthiness, and is more activated when viewing social versus nonsocial scenes (Winston et al. 2002; Shaw et al. 2005; Mende-Siedlecki et al. 2013; Bickart et al. 2014; Norman et al. 2015; Tso et al. 2018). It also shares close structural connectivity with other regions of the social brain (Bickart et al. 2014). Gray matter volume within the temporo-parietal junction (Lewis et al. 2011; Sato et al. 2016; Yin et al. 2018; Wiesmann et al. 2020), dorsomedial prefrontal cortex (Banissy et al. 2012; Eres et al. 2015; Sato et al. 2016), ventromedial prefrontal cortex (Lewis et al. 2011), anterior medial prefrontal cortex/frontal pole (Rice and Redcay 2015), and amygdala (Rice et al. 2014; Batista et al. 2017; Baribeau et al. 2019) has also been noted to correlate with social cognitive performance. Finally, social network properties, such as social network size, have been linked to individual differences in gray matter structure within regions of the social brain including the dorsomedial prefrontal cortex (Kanai et al. 2012; Kwak et al. 2018), ventromedial prefrontal cortex (Lewis et al. 2011; Kanai et al. 2012; Powell et al. 2012; Von Der Heide et al. 2014; Kwak et al. 2018), middle temporal lobe (Kanai et al. 2012; Kwak et al. 2018), and amygdala (Lewis et al. 2011; Bickart et al. 2012; Von Der Heide et al. 2014). Altogether, these six brain regions were included bilaterally in the social brain and mapped onto the corresponding regions returned by the FreeSurfer segmentation algorithm (see Figure 1 and Supplementary Table 1).

Figure 1

Figures of the brain networks used. Pictured are the (a) social, (b) salience, and (c) executive brain networks. The salience and executive networks were defined by 10 regions and the social network included 12 regions. Brain figures were visualized with the BrainNet Viewer (Xia et al. 2013).

The executive and salience networks were based on definitions proposed by Power et al. (2011). The executive network supports sustained attention, planning, and working memory, and the salience network is implicated in the detection of salient stimuli pertinent to task performance in a flexible manner (Seeley et al. 2007; Uddin 2016; Norman et al. 2019). Power’s executive network included six bilateral cortical regions, of which we matched three to the FreeSurfer atlas: the pars triangularis, inferior parietal cortex, and lateral orbitofrontal cortex (Power et al. 2011). We then included the bilateral caudate and putamen, both because of their associations with executive functioning (Levy et al. 1997; Lewis et al. 2004; Norman et al. 2019) and because our social brain definition included the amygdala, it stood to reason that at least one of our control networks should include subcortical regions. Our definition of the salience network included the anterior cingulate cortex, insula, and middle frontal gyrus (Seeley et al. 2007; Power et al. 2011; Uddin 2016).

Statistical Analysis

Defining the Similarity of Brain Networks

Volumes of the 12 regions of the social brain (six in each hemisphere), adjusted for intracranial volume, were used to place each brain in a 12D space. That is, each of the volumes defined one of 12 coordinates. The pairwise distance (or similarity) between brains of children in the same classroom was taken as the Mahalanobis distance between brains in the 12D space. We used Mahalanobis rather than Euclidean distances as the volumes of brain regions are correlated. The Mahalanobis distance adjusts for the covariance between brain region volumes. For each pair, the Mahalanobis distance is calculated by

Distance = |$\sqrt{{\Big({x}_{1-12}-{y}_{1-12}\Big)}^T{S}^{-1}\Big({x}_{1-12}-{y}_{1-12}\Big),}$|

where S is the covariance matrix and |${x}_{1-12}\ \mathrm{and}\ {y}_{1-12}$| are the 12D of each child’s (child x and child y) social brain. The same approach was used to calculate the distances between the classmates’ executive and salience networks. Ten dimensions/volumes were included in the other networks and no network was allowed to have overlapping brain regions with another. The number of children with anatomic data within a classroom ranged from a minimum of two through a maximum of six. For each dyad, brain similarity was calculated for each of the three networks separately.

Regression

Random-intercept linear mixed effects models were used to test, across dyads, for associations between geodesic distance (mutual friends, friends-of-friends, and more distantly connected or unconnected peers) and Mahalanobis distance (Chen et al. 2017). These were conducted using the lme4 package (Bates et al. 2015) for R (http://www.r-project.org). Following recommendations by Chen et al. (2017), and in order to account for the nonindependence of observations, we included crossed random-effects terms for the children in each dyad, which were nested within a random-term for school class. Covariates included differences within pairs for age at scan, internalizing problems, externalizing problems (as measured by CBCL), and image quality, as well as concordance for race/ethnicity, gender, income, and maternal education.

Multiple imputation for missing covariate data was performed using the mice and mitml R packages (van Buuren and Groothuis-Oudshoorn 2011; Grund et al. 2019). Statistical models for missing data imputation used all variables included in the analyses. The amount of missing data was minimal, with less than 7.2% of included subjects missing data for any given covariate. Five datasets were imputed (von Hippel 2009; Van Buuren 2012). We then fit reduced models which included the dependent variable (e.g., Mahalanobis distance of the social brain) and all covariates but which did not include geodesic distance group, before fitting a full model which also included the geodesic distance group term (Grund 2018). These models were fit for each of the imputed datasets. The full and reduced models were then compared using a pooled version of the F-test (i.e., the Wald test), where a significant F-value would indicate a significant improvement of the full-model over the reduced model, and thus suggest a significant relationship between geodesic distance and Mahalanobis distance (Rubin 2004; Grund 2018). These analyses were repeated using the alternative methods available in the mitml package for comparing nested statistical models fitted to imputed datasets using pooled Wald-like or likelihood-ratio tests (i.e., D2, D3, and D4), with the pattern of findings remaining unchanged (Li et al. 1991; Meng and Rubin 1992; Chan and Meng 2017; Grund et al. 2019). In the event of a significant relationship between geodesic distance and Mahalanobis distance, follow-up pairwise tests were performed (i.e., comparing mutual friends vs. friends-of-friends, mutual friends vs. more distantly connected peers, and friends-of-friends vs. more distantly connected peers), by pooling the estimates and standard errors from each full model for each imputed dataset according to Rubin’s rules using the “pool” function in mice (Rubin 2004; van Buuren and Groothuis-Oudshoorn 2011).

We used this approach to examine the associations between geodesic distance and the Mahalanobis distances within the social, executive, and salience networks. Corrections for multiple comparisons were performed using the Benjamini–Hochberg method (Benjamini and Hochberg 1995). An overview of our methodological approach is given in Figure 2.

Figure 2

Methods overview. (a) Classroom networks defined by reciprocal friendship. Nominations were used to determine the geodesic distance between children who had neuroanatomic data. Child 1 and Child 2 are mutual friends (geodesic distance = 1). (b) Each child’s social brain is defined by the volumes of 12 regions supporting social cognition and friendship network formation. These 12 volumes provide 12 coordinates for each child’s social brain. Similarity in the social brains between two peers (such as Child 1 and Child 2) was taken as the Mahalanobis distance between the brains. (c) We predicted that the similarity (Mahalanobis distance) between social brains would be associated with geodesic distance. Mutual friends were predicted to have the most similar social brains, followed by friends-of-friends (geodesic distance = 2) and then more distant friends and isolates (geodesic distance = 3+ and infinite).

Based on previous reports in adult studies of significant associations between social network size and social brain structure, we also examined whether similar associations existed in our developmental sample as a supplementary analysis (Lewis et al. 2011; Bickart et al. 2012; Kanai et al. 2012; Powell et al. 2012; Von Der Heide et al. 2014; Kwak et al. 2018). Using linear mixed effects models, we examined whether gray matter volume in each of the “social brain” regions was associated with the number of mutual friendships for each child while controlling for age at scan, household income, maternal education, ethnicity, internalizing problems, externalizing problems, image quality, and intracranial volume. A random-term was included for classroom. For completion, we also examined potential associations between gray matter volume in the social brain and in-degree centrality, or the number of nominations that a given child received. Number of mutual friends and in-degree centrality values were normalized by dividing by the classroom network size for each classroom (Szekely et al. 2016). We further examined if friendship distance was associated with pairwise similarity on these classroom network metrics while controlling for similarity or concordance in age at scan, household income, maternal education, ethnicity, internalizing problems, and externalizing problems. Finally, we included the number of mutual friendships and in-degree centrality as additional covariates in a model examining the primary hypothesis of the paper that Mahalanobis distance of the social brain would show a significant association with geodesic friendship distance.

Results

There were 125 children with good-quality neuroanatomic data forming 111 dyadic (pairwise) geodesic distances within classroom friendship networks. Considering the friendship network structures of these children, 15 of the geodesic ties indicated mutual friendship (geodesic distance = 1), 20 ties indicated friends-of-friends (geodesic distance = 2), and 76 were more distantly connected friends or isolates (geodesic distance |$\ge 3$|⁠). Of the 76 more distant ties, 44 dyads were network isolates (geodesic distance = infinite).

We tested the hypothesis that mutual friends would have similar “social brains,” compared with their more remotely connected peers (friends-of-friends and even more distantly connected peers). We found a significant difference between groups defined on friendship degree, controlling for concordance on race/ethnicity, gender, age at scan, maternal education, household income, image quality, and internalizing and externalizing problems (F(2,79 452) = 4.03, adjusted P = 0.03). Pairwise contrasts indicated that mutual friends showed less distance (i.e., more similarity) between social brains than friends-of-friends (β = 0.65, t = 2.03, P = 0.045) and more distantly connected or unconnected peers (β = 0.77, t = 2.83, P = 0.006). The friends-of-friends and more remotely connected peers did not differ significantly from one another (β = −0.13, t = −0.53, P = 0.6). We found that the effects of mutual friendship did not extend to nonsocial brain networks. The association between geodesic distance and Mahalanobis distance was nonsignificant for the salience network (F(2, 11 997) = 0.48, adjusted P = 0.62). Regarding the executive network, a significant effect of geodesic distance on Mahalanobis distance was found (F(2, 442 565) = 4.02, adjusted P = 0.03). However, follow-up analyses indicated that this was driven by smaller Mahalanobis distances between distantly connected or unconnected peers compared with friends-of-friends (β = 0.78, t = 2.8, P = 0.006). In other words, this unexpected significant association was not driven by a pattern of pairwise group differences that is consistent with neuroanatomic homophily. Full details of these models are given in Supplementary Tables 2–4. See Figure 3. Analyses that removed isolated dyads did not change the pattern of findings (Supplementary Table 5). Further details on missing data and the results of the analyses repeated without imputation are provided in Supplementary Tables 6 and 7.

Figure 3

Associations between friendship distance and similarity in the social, executive, and salience brain networks.

Gray matter volume of the “social brain” was not associated with number of mutual friendships or with in-degree centrality (see Supplementary Tables 8 and 9). Furthermore, geodesic distance was unrelated to pairwise similarity in number of mutual friends (F(3, 99 211) = 0.14, P = 0.94). The three group ANOVA examining the relationship between geodesic distance and pairwise similarity in in-degree centrality did not reach significance (F(3, 5 472 000) = 1.75, P = 0.15), although exploratory follow-up pairwise comparisons did indicate that mutual friends had more similar in-degree centrality values than did more distantly connected or unconnected peers (β = 0.06, t = 2.12, P = 0.04). Importantly, the relationship between geodesic distance and Mahalanobis distance of the “social brain” remained significant when controlling for similarity in the number of mutual friends (F(2, 36 965) = 4.25, P = 0.01) or in-degree centrality (F(2, 48 253) = 4.46, P = 0.01).

Discussion

In the present study, we build upon previous work by showing that neural homophily extends from brain “functioning” to brain “structure” for regions believed to be pivotal to social cognitive processes. Specifically, we found that mutual friends exhibited greater anatomic similarity in brain regions known to support social cognition than dyads including friends-of-friends or more distantly related peers (Lewis et al. 2011; Bickart et al. 2012; Kanai et al. 2012; Powell et al. 2012; Von Der Heide et al. 2014; Sato et al. 2016; Kwak et al. 2018). This association was not driven by concordance between friends in race/ethnicity, gender, age, maternal education, household income, image quality, or internalizing and externalizing behaviors, which were included as covariates in the model, and did not extend to brain regions involved in nonsocial cognition, such as those supporting executive functions.

Since children are unable to observe one another’s brain structure, which manifest features might mediate this relationship? A plausible hypothesis is that children with more similar “social brains” may be better able to understand each other during social exchanges. Indeed, previous work has reported that subjects with similar interpretations of social scenarios also show similar patterns of activation within the “social brain” (Nguyen et al. 2019), while interpersonal functional synchrony in “social brain” regions during dyadic interactions has also been associated with successful communication, empathy, and social cooperation (Dumas et al. 2010; Stephens et al. 2010; Cui et al. 2012; Babiloni and Astolfi 2014; Jiang et al. 2015; Liu et al. 2016; Tang et al. 2016; Goldstein et al. 2018; Reindl et al. 2018). Therefore, similarities within the “social brain” may allow for easier mutual understanding, predictability, and communication, which are all features of social interactions proposed to be key to the formation of friendships in children and adults (Dunn and Cutting 1999; Erdley and Day 2017; Bagwell and Bukowski 2018; Parkinson et al. 2018). Although the present study examined brain structure rather than brain function, a number of structural neuroimaging studies have linked differences in gray matter volume and thickness within “social brain” regions to individual differences in social cognitive abilities (Lewis et al. 2011; Banissy et al. 2012; Rice et al. 2014; Eres et al. 2015; Rice and Redcay 2015; Sato et al. 2016; Batista et al. 2017; Yin et al. 2018; Baribeau et al. 2019; Wiesmann et al. 2020). Therefore, similarities in social brain structure may underlie similarities in brain functioning and social cognition and behavior. This potential role for similarities in social cognition and behavior as a mechanism underlying neuroanatomic homophily is further supported by the finding that mutual friends showed more similar gray matter structure within the social brain specifically, and these findings did not generalize to the other brain networks studied. However, the relationships between brain structure and function are still poorly understood. Consequently, collection of multimodal neuroimaging and behavioral data in the same subjects is required in order to assess how similarities in “social brain” structure in mutual friends relate to similarities in brain functioning, as well as how these similarities in brain structure and functioning are expressed as similarities in the social, emotional, and behavioral traits that play a role in emerging friendships between similar individuals.

We found that mutual friends were more similar with respect to in-degree centrality compared with more distantly connected or unconnected peers. That is, mutual friends tended to receive similar numbers of peer nominations to each other, indicating similarities in classroom peer acceptance and social network size. However, unlike previous work, we did not find that in-degree centrality was associated positively with gray matter volume within the social brain (Kwak et al. 2018). Reasons for these discrepant findings may include differences in the age of the samples studied, with our study focusing on elementary school-aged children, and previous work relating in-degree and other measures of social network size to “social brain” volume focusing on adults (Lewis et al. 2011; Bickart et al. 2012; Kanai et al. 2012; Powell et al. 2012; Von Der Heide et al. 2014; Kwak et al. 2018). Social cognitive abilities and their underlying brain circuitry continue to develop over the course of adolescence (Meinhardt-Injac et al. 2020). Moreover, the nature of mutual friendship and social networks changes with age, with childhood friendships largely forming within bounded neighborhood and classroom settings and centering around frequent opportunities for children to play together (Bukowski et al. 2009; Bagwell and Bukowski 2018). During the development into adulthood, friendships gain new functions as they move from being primarily activity-based to more complex relationships involving self-disclosure, emotional intimacy, identity formation, and mutual practical assistance (Bukowski et al. 2009; Bagwell and Bukowski 2018). Moreover, adult social networks are no longer constrained by daily classroom interactions, typically involve less frequent, face-to-face contact, and tend to be more widely distributed and include friendships made across multiple social contexts and stages of life (Johnson and Troll 1994; Wrzus et al. 2015; Dunbar 2018). It is therefore plausible that different social skills are needed to maintain social networks at different stages of life, meaning that associations between social network size and the structure of the “social brain” may change over the lifespan. Importantly, the primary finding in the present study of a significant relationship between Mahalanobis distance of the “social brain” and geodesic friendship distance remained after controlling for dyadic similarities in social network size.

This study has its limitations. First, only a subset of the children with peer nomination data also had neuroanatomic data. Second, we can only speculate on the behavioral characteristics that promote friendship, though it is plausible that similarity in social brain structure might be linked to similarity in neural functioning, and/or in the degree of social motivation and interest, sensitivity to social cues, and social astuteness (Stiller and Dunbar 2007; Parkinson et al. 2018). Finally, given our cross-sectional design, we cannot make causal inferences. Longitudinal data would be better poised to answer whether similar social brains predispose children to select one another as friends or if the experience of being a close friend engenders neuroanatomic similarity.

Conclusion

We found neuroanatomic similarity between classroom mutual friends compared with friends-of-friends and more distant individuals in a classroom friendship network. This similarity was found for a brain network supporting social cognition and the maintenance of social network size and did not extend to networks implicated in executive functions or salience processing. Our findings support the possibility of neural homophily in childhood social ties, which may be tied to facets of social cognition.

Notes

Conflict of Interest: Authors declare no competing or potential conflicts of interest.

Funding

National Human Genome Research Institute (ZIAHG200378); National Institute of Mental Health. Data collection for this study, a part of Phase I of the Generation R Study, was made possible by financial support from: Erasmus Medical Center, Rotterdam; Erasmus University Rotterdam; and the Netherlands Organization for Health Research and Development (ZonMw) TOP (91211021).

Author notes

Patrick D’Onofrio and Luke J. Norman have contributed equally to this work