Abstract
Previous research suggests that the proximity of individuals in a social network predicts how similarly their brains respond to naturalistic stimuli. However, the relationship between social connectedness and brain connectivity in the absence of external stimuli has not been examined. To investigate whether neural homophily between friends exists at rest we collected resting-state functional magnetic resonance imaging (fMRI) data from 68 school-aged girls, along with social network information from all pupils in their year groups (total 5066 social dyads). Participants were asked to rate the amount of time they voluntarily spent with each person in their year group, and directed social network matrices and community structure were then determined from these data. No statistically significant relationships between social distance, community homogeneity and similarity of global-level resting-state connectivity were observed. Nor were we able to predict social distance using a machine learning technique (i.e. elastic net regression based on the local-level similarities in resting-state whole-brain connectivity between participants). Although neural homophily between friends exists when viewing naturalistic stimuli, this finding did not extend to functional connectivity at rest in our population. Instead, resting-state connectivity may be less susceptible to the influences of a person’s social environment.
Significance statement The tendency of individuals to befriend others like themselves is a well-known feature of human social networks. This phenomenon, known as homophily, has even been shown to extend to neural activation patterns during viewing of naturalistic stimuli. If neural homophily is also present during rest, this would indicate that humans select friends who not only share similar neural patterns during socially relevant tasks but also during unprompted thought. In the current study, we found insufficient evidence for homophily of neural connectivity during rest in schoolchildren, suggesting that internally-generated brain function contributes little to friend selection.
Introduction
Homophily is the tendency of individuals to attract and interact with those who share similar traits. Homophilic selection is observed for gender and ethnicity (Shrum et al., 1988), motivation (Shin and Ryan, 2014b), personality (Ilmarinen et al., 2017), and academic achievement (Smirnov and Thurner, 2017). High school and university students have been found to rearrange their local social networks to form ties and clusters with students who have similar performance levels (Smirnov and Thurner, 2017) and this type of homophily has been observed even in polygenic scores for academic achievement (Domingue et al., 2018).
Given the predominance of social network homophily for behavioural, personality and cognitive traits, we can reasonably expect that this extends to similarities in brain function. In fact, neural responses observed during viewing of movie clips were significantly more similar among friends than those farther removed in a real-world social network (Parkinson et al., 2018). Social closeness also provides opportunities for behavioural contagion researchers have shown that social contagion modulates neural representations of risk in reward-related areas and that functional connectivity between the caudate and prefrontal cortex accounts for individual differences in susceptibility to risk-taking contagion (Suzuki et al., 2016).
Previously, neural similarity was assessed using intersubject correlation of blood oxygenation level-dependent (BOLD) timeseries across functionally derived regions of the brain. This method of intersubject correlation evaluates the externally generated stimulus-locked BOLD activation associated with the task but ignores the internally generated component of BOLD activity, which is cancelled out when correlating across participants (Nummenmaa et al., 2018). Patterns of brain connectivity elicited from internally generated resting-state BOLD activation are mirrored by activation networks found under explicit task-based activation (e.g. from attention, reasoning, and social cognition tasks) (Smith et al., 2009; Laird et al., 2011) and are associated with non-cognitive measures of motivation. Grit and growth mind-set were associated with functional connectivity between ventral striatal and bilateral prefrontal networks important for cognitive-behavioural control (Myers et al., 2016). Others have found that global connectivity of the left prefrontal cortex predicts individual differences in fluid intelligence and cognitive control (Cole et al., 2012). Connectivity at rest has also been used to successfully predict trait-level measures of personality, including openness to experience (Dubois et al., 2018), neuroticism and extraversion (Hsu et al., 2018). These findings highlight the utility of resting-state connectivity for identifying individual differences in cognition, behaviour and personality, all of which have exhibited homophily within social networks.
Researchers have also linked internally generated brain connectivity with a number of social behaviours. For example, resting-state subnetworks for motor, visual, speech and other language functions have been associated with quality and quantity of social networks in older adults (Pillemer et al., 2017). Others have demonstrated positive associations between functional connectivity and social network size and embeddedness (Joo et al., 2017). Social network size may also dictate the degree of connectivity within the default mode network (DMN). The DMN overlaps considerably with regions important for theory of mind and social cognition (Laird et al., 2011; Mars et al., 2012) and has been found to exhibit greater coupling with anterior cingulate and dorsolateral prefrontal cortex in those with larger social network size (Noonan et al., 2018). Despite investigations into the relationship between internally generated connectivity patterns and social behaviour, however, no study has investigated whether close social relationships are associated with similarities in resting-state connectivity.
In the current study, we collected resting-state fMRI data of pupils attending a single school, along with social network (friendship) information between them. This unique data allow us to test the hypothesis that friends exhibit greater similarity in internally generated connectivity compared with those farther removed in a school-based social network. We tested this hypothesis with a variety of methods based on resting-state network connectivity — whole-brain correlation, graph-theoretic metrics, and machine learning. We evaluated similarities in whole-brain connectivity as well as resting-state subnetworks associated with motivational and social behaviours, which are relevant to schoolchildren.
Methods
Participants
For the social network component of the study, individuals 12-14 years of age in years 8 (cohort 1: n=59; cohort 2: n=51) and 9 (cohort 3: n=65) were recruited from a private girls’ day and boarding school in the United Kingdom (as part of a larger study). Participants were recruited during 2017 and 2018 from total year group pools of 62 (cohort 1), 53 (cohort 2) and 75 (cohort 3) students, corresponding to inclusion rates of 95%, 96% and 87%, respectively. The study was approved by the University Research Ethics Committee, University of Reading. All students gave informed written assent to take part in the study and consent was obtained from parents.
Individuals from cohorts 1, 2 and 3 were also invited to take part in the functional magnetic resonance imaging (fMRI) component of the study. Twenty-eight students from cohort 1, 17 students from cohort 2 and 34 students from cohort 3 were recruited into the fMRI study (cohorts 1-fMRI, 2-fMRI and 3-fMRI, respectively); of these, 5 participants from cohort 1-fMRI and 6 from cohort 3-fMRI had unusable data due to artefacts caused by dental braces. Children with braces were excluded from cohort 2-fMRI at the time of screening. This resulted in the final inclusion of 68 students in the fMRI component of the study, consisting of 23 (cohort 1-fMRI, 12-13 years of age), 17 (2-fMRI, 12-13 years of age) and 28 (3-fMRI, 13-14 years of age) students in respective cohorts. Exclusion criteria pertaining to all groups consisted of standard safety-related contraindications for MRI. All students gave informed written assent and one parent per child gave informed written consent for their child to participate.
Experimental Design and Statistical Analysis
The social network data were acquired in class using an online survey administered with SurveyMonkey (SurveyMonkey Inc., San Mateo, California, USA). Social network data were acquired in October and November of 2017 (cohorts 1 and 3) and 2018 (cohort 2), one-to-two months after the start of the academic year. Students had been registered at the school for a maximum of one (cohorts 1 and 2) or two (cohort 3) years when the survey took place. A roster-and-rating method was used to assess the social networks of students. Specifically, participants were provided with a list of all students in their year group and asked to consider the question: “How much time do you spend interacting with this student?”. Students answered on a five-point Likert scale, which included options: “None”, “A rare amount”, “Some”, “More than some” and “Most”. Participants were told to consider time spent voluntarily interacting with other students but not time spent in planned seating situations, allocated group work or in classes without opportunities to talk amongst themselves. Investigators were blinded to the identity of students.
In addition, participants were asked to nominate up to five individuals from their cohort with whom they considered themselves to be “close with” (Coie et al., 1982). Peer nominations and roster-and-rating methods measure different aspects of peer interactions. Whereas rating assesses general acceptance of peers, nomination is thought to encourage the naming of “best friends” (Schofield and Whitley Jr, 1982). In the main text, we report results from roster-and-rating data only. However, nomination data provided the same conclusion as roster-and-rating data (see supplementary figures S1 and S2). Participants were asked to report on social interactions within their own year group only.
Structural and resting-state functional MRI data were acquired using a Siemens Magnetom Prisma_fit 3T scanner. All participants were imaged using a 32-channel head coil. A structural T1-weighted image was acquired using an MPRAGE sequence (Brant-Zawadzki et al., 1992) with repetition time (TR) 2300 ms; echo time (TE) 2.29 ms; inversion time 900 ms; flip angle 8°; in-plane acceleration (GRAPPA) factor of 2; field of view (FOV) 240 mm; voxel size 0.9 x 0.9 x 0.9 mm.
Resting-state functional images were acquired over 10 minutes using the Siemens two-dimensional multiband gradient-echo echo planar image (EPI) sequence with TR 1500 ms, TE 30 ms; multiband slice acceleration factor 4; GRAPPA 2; flip angle 66°; echo spacing 0.93 ms; EPI factor 96; phase-encode direction posterior >> anterior; slices 68; volumes 400; FOV 192 mm; voxel size 2 x 2 x 2 mm.
During resting-state data acquisition, participants were asked to lie still with eyes open and look at a blank screen in front of the scanner. Instructions were given to relax and think of nothing in particular.
Social network characterisation
Social network analysis was performed using the igraph package in R (Csardi and Nepusz, 2006; R Core Development Team, 2012). Data were included only for individuals who took part in the social network survey; any outgoing social ties with those not participating in the survey were removed. Mutually-reported (reciprocal) social ties are deemed to be more robust indicators of friendship than unreciprocated ties. Therefore, an unweighted, undirected graph containing only mutual social ties was used to determine social distance and community affiliation.
Social network data were processed as follows: Roster-and-rating data (5-point scale) were binarised using a threshold of 4 (i.e. only instances in which students spent “more than some” or “most” of the time with another student were included). This threshold was selected to mitigate central tendency bias often reported with Likert-type questionnaires (Stevens, 1971). Any non-mutual connections were then removed (i.e. if subji gave subjj a rating of 4 or greater but subjj gave subji a rating below 4, the connection would be lost). This yielded an unweighted (binary), undirected (reciprocal) adjacency matrix for each cohort, from which social networks graphs were derived. Adjusting the distance threshold to 5 (I spend “most of my time” with this person) or using non-reciprocated (directed) social ties (rather than mutual ties) did not change the overall results of the analyses (see supplementary figures S3-S6).
The distance between two individuals in a social network is an important predictor of behavioural tendencies (Apicella et al., 2012) and may relate to shared patterns in brain function (Parkinson et al., 2018). Social distances in the current study were relatively short compared with previous literature (e.g. Parkinson et al. (2018) reported a network diameter of 6 using only mutually reported social ties compared to a maximum diameter of 4 in the current study). The highly interconnected nature of our networks may affect how social distance and brain function similarities are associated with one another. Therefore, we also evaluated dyadic similarities as a function of community affiliation. By splitting social networks into smaller friendship communities and evaluating the differences in brain similarity between those inside and outside each community, we measured relationships between neural homophily and social behaviour at a friendship group level.
Measures of social proximity (social distance and community affiliation) were determined separately for each cohort. Social distance was calculated as the shortest path length between each pair of nodes (students) in the network. A dyad (student pair) with a social distance (shortest path length) of 1 represented a relationship in which both students had said they spent “more than some” or “most” of their time with the other student (i.e. they were friends). Social distances of 2 and 3 represented dyad pairs in which students did not possess a reciprocal friendship (i.e. did not have a mutual rating of “more than some” or “most” of the time) but reported a mutual friend or friend of a friend, respectively.
Community structure was ascertained using the Louvain method (Blondel et al., 2008). This method implements multi-level modularity optimisation to subdivide the network into non-overlapping groups of nodes that maximise the number of within-group (within-module) friendships (edges) and minimised the number of between-group friendships for each cohort.
Each cohort was described in terms of its network characteristics, in particular, its network diameter, modularity, mean path length, reciprocity and density. Network diameter is the length of the longest geodesic distance between two nodes in the network, i.e. the number of edges between subji and subjj when these individuals are the farthest from each other in the network.
Modularity (Q) is a measure of how easily a network segregates into smaller subnetworks and is defined by the equation: where m is the number of edges in the network, Aij is the element of the A adjacency matrix in row i and column j, ki is the degree (number of edges associated with a node) of i, ci is the community to which i belongs and δ(ci, cj) is 1 ci = cj if and 0 otherwise. Nonzero values of Q represent deviations from randomness; a value above 0.3 is an indicator of significant community structure in a network (Clauset et al., 2004).
Mean path length is the mean shortest path length (number of edges separating a pair of nodes) between all nodes in the network. Reciprocity defines the proportion of connections in a directed graph that are mutual connections. It is otherwise defined as the probability that the counterpart (j to i) of a directed edge (i to j) is included in the graph. Graph density is the ratio of actual connections (edges) to possible connections in the graph; larger values denote more densely connected networks.
Functional MRI data analysis
FMRI data processing was carried out using FEAT (FMRI Expert Analysis Tool, version 6.0 (Woolrich et al., 2001)) part of the FMRIB Software Library (FSL; Oxford, United Kingdom) (Smith et al., 2004; Jenkinson et al., 2012). Registration of functional images to high resolution structural and Montreal Neurological Institute (MNI-152) standard space images was carried out using FLIRT (Jenkinson and Smith, 2001; Jenkinson et al., 2002). Registration from high resolution structural to standard space was then further refined using FNIRT nonlinear registration (Andersson et al., 2007b, a).
The following pre-statistics processing was applied; motion correction using MCFLIRT (Jenkinson et al., 2002); non-brain removal using BET (Smith, 2002); multiplicative mean intensity normalization of the volume at each time point and high pass temporal filtering (Gaussian-weighted least-squares straight line fitting, with sigma=50.0s). Independent components analysis (ICA)-based exploratory data analysis was carried out using MELODIC (Beckmann and Smith, 2004), in order to investigate the possible presence of unexpected artefacts. FIX - FMRIB’s ICA-based Xnoiseifier (Griffanti et al., 2014; Salimi-Khorshidi et al., 2014) was used to auto-classify ICA components into “good” and “bad” components following hand-classification and training using a sample of 10 subjects’ data (4 from cohort 1-fMRI and 3 each from cohorts 2-fMRI and 3-fMRI, all randomly selected). “Bad” components were removed from the data and clean data were registered to standard space using warping parameters determined by FEAT.
Analysis of pre-processed data is illustrated in Figure 1. First, motion-corrected fMRI data were divided into 272 parcels using a whole-brain parcellation scheme (McNabb et al., 2018). This parcellation scheme combines parcels from the Human Brainnetome Atlas (Fan et al., 2016), a modification of the Desikan-Killiany (DK) atlas (Desikan et al., 2006) and the probabilistic MR atlas of the human cerebellum (Diedrichsen et al., 2009). Mean blood oxygen level dependent (BOLD) timeseries were extracted from each of the 272 parcellated regions (nodes). Next, the Pearson’s correlation coefficient (connection strength) between BOLD timeseries was determined for each pair of nodes in Matlab 2016a (MathWorks, USA). This produced a weighted, undirected whole-brain matrix of functional connectivity for each participant. All self-self nodal connections were removed prior to further analysis.
Z transformations of weighted, undirected whole-brain matrices were then used for inter-subject correlations. Data from each participant were compared with every other participant from the fMRI cohort using a Pearson’s correlation. For example, the Z-transformed whole-brain matrix of subji was vectorised and correlated with the (vectorised) Z-transformed whole-brain matrix of subjj, then subjk, then subjl, and so on. This method provides a measure of the similarity of connectivity strength in whole-brain networks within participant pairs. Prior to further analysis, correlation strengths were standardised within each cohort to have a mean of 0 and a standard deviation of 1.
Evaluating similarities in resting-state connectivity at a whole-brain level is informative but the inclusion of brain regions unlikely to impact or be impacted by social connectedness could lead to losses in signal to noise. Focusing on smaller groups of nodes, related to cognitive performance and social interaction would therefore be expected to increase functional similarities associated with friendship. Thus, in addition to whole-brain resting-state connectivity, we evaluated functional resting-state connectivity for subnetworks relevant for social and motivational processing. These included the DMN (Li et al., 2014), salience network (Silverman et al., 2015), and left and right frontoparietal networks (lFPN and rFPN, respectively) (Meyer et al., 2015). DMN and FPNs were extracted from (Smith et al., 2009). Z statistic images of ICA maps were then binarised using a threshold of z=4 and warped to 1 mm MNI-152 standard space. Our threshold of z=4 is larger than that reported by (Smith et al., 2009), who adopted a threshold of z=3 for presentation of resting-state subnetworks. Our larger z threshold was chosen to increase the probability that brain nodes overlapping DMN and FPN spatial maps were truly part of these networks. The salience network was derived using the online meta-analysis tool Neurosynth.org (Yarkoni et al., 2011) (accessed January 2019; seed positioned in the anterior insular [x=36, y=18, z=4]). The salience network was binarised using a threshold of 0.3, based on results from previous research (Menon, 2015). The overlap between each resting-state subnetwork and the whole-brain parcellation was determined by applying a binarised mask of the resting-state subnetwork over whole-brain parcellation data. Nodes from the whole-brain parcellation that overlapped with at least 50 voxels from the resting-state subnetwork were used to partition the whole-brain node-to-node matrix. A graphical depiction of the analysis is provided in the supplementary materials (figure S7).
Brain network characterisation
Graph metrics of functional connectivity were derived from whole-brain weighted, undirected matrices using the Brain Connectivity Toolbox (Rubinov and Sporns, 2010) in Matlab (2016a). Analysis steps are provided in Figure 2. Modularity and community structure were calculated using the Louvain method (Blondel et al., 2008), assigning higher values to positively, compared with negatively weighted connections. Modularity was defined as: where Q+ is modularity from positively weighted connections, Q− is modularity from negatively weighted connections, v+ is the sum of all positive connection weights of i, v−, is the sum of all negative connection weights of i, is the present within-module positive connection weights, is the present within-module negative connection weights, is the chance expected within-module positive connection weights, is the chance expected within-module negative connection weights and δ(Mi, Mj) is 1 if i and j are in the same module and 0 otherwise.
The absolute difference in brain modularity within a dyad pair (i.e. between subji and subjj) was calculated for every pair of dyads in a social network. Differences were then standardised within each cohort to have a mean of 0 and standard deviation of 1.
Nodal strength (a node-level measure of centrality, the importance of a node in its network) and diversity (a node-level measure of integration that takes into account the strength of a node within its own module) were also calculated using asymmetric values for positively and negatively weighted connections, as previously described by Rubinov & Sporns (2011). Strength was defined as: where is the normalised sum of positive connection weights associated with node i, and and are the raw sums of positive and negative connections weights, respectively, associated with node i.
Diversity was defined as: and where is the strength of node i within module u (the total weight of connections of i to all nodes in u) and m is the number of modules in modularity partition M.
Nodal strength and diversity data from each participant were compared with every other participant from the fMRI cohort using a Pearson’s correlation. Prior to further analysis, correlation strengths were standardised within each cohort to have a mean of 0 and a standard deviation of 1.
Dyadic similarities in functional connectivity as a function of social proximity
To test our hypothesis, we examined whether similarities in resting-state connectivity and brain network characterization were explained by social proximity. In this set of analyses, each student dyad (pair of students) in the fMRI cohort was described by two independent variables: social distance (i.e. shortest path length between the pair within the full year group cohort) and community affiliation (i.e. whether or not two students belonged to the same community, determined using the Louvain community detection method). Dependent variables were similarities in whole-brain connectivity, resting-state subnetwork (DMN, salience network, lFPN and rFPN) connectivity, nodal strength, nodal diversity (each determined by the Pearson’s correlation coefficient), and modularity (determined by the absolute difference between members in the dyad).
Using pairs of students (dyads) as the unit of analysis creates dependence in the data, caused by the involvement of every student in multiple dyads (cross-nesting) (Kenny et al., 2006) (Snijders and Kenny, 1999). Using ordinary least-square methods in such data potentially increases Type-1 error rates. We accounted for this dependence structure by including each dyad member (student) as a random variable in a linear mixed effects (LME) model with crossed random effects (see (Chen et al., 2017), for a similar model specification). Any subject-specific effects on the dyadic outcomes are therefore accounted for in the model (Baayen et al., 2008).
More specifically, the LME model is specified as:
Similarityij is the similarity in brain function between students i and j, where β00 is the intercept, and β01 represents the relationship between social distance and brain similarity; subji and subjjare crossed random effects (i.e. student-specific effects) of students i and j, respectively, and eji are residuals. We posited a Gaussian distribution for the random effects and residuals. No distribution assumption was made for social distance. Where social network data were described using community affiliation, social distanceij in the above equation was replaced with community Similarityij (i.e. whether or not two students belong to the same friendship community, as determined by the Louvain method).
Data were analysed using the lme4 package (Bates et al., 2015) in R. Dependent variables included similarities in whole-brain, resting-state subnetwork and graph theory measures of connectivity between students (dyads) in each cohort. For each dependent variable, we tested a model with social distance as the independent variable, and tested another model with community affiliation as the independent variable.
We also tested a model with random slopes (i.e. random slopes of the subjects were added to the model above). This model failed to converge in most of the analyses. The omission of random slopes could make the statistical test less conservative (Barr et al., 2013; Murayama et al., 2014). However, as our results all showed non-significant effects in the model without random slopes, we decided not to pursue the model with random slopes any further.
Effects of demographic data as well as their interactions with social distance or community affiliation were also included in a second series of models to determine whether boarding status (i.e. whether students matched in their boarding status [boarding or day student]) and ethnicity (i.e. whether or not students reported belonging to the same ethnic group) affected the relationship between social proximity and similarity in resting state connectivity between students. As only one student from each fMRI cohort was left-handed, we elected not to include handedness as a covariate.
Cohorts were analysed separately to preserve the independence of the three social networks. Then, we integrated the regression coefficients of the LME models from three cohorts using a random-effect meta-analysis with the Metafor package (Viechtbauer, 2010) in R.
Data-driven predictive model of social proximity from neural similarity
In the previous analysis, we assessed the overall similarity of the whole brain network or subnetworks between pairs of students and examined whether social distance is related to the overall similarity. Although this analysis gives us the most straightforward test of our hypothesis, the analysis does not address the possibility that social distance is represented by the collection of local-level similarities (i.e. similarity between a specific pair of nodes). To examine whether any local-level similarity in the brain functional connectivity encodes social distance, we employed regularised regression techniques to predict social distance between two students based on similarities in their functional brain connectivity of all pairs of nodes. Specifically, we computed the absolute difference in connection strength for each edge in the 272 node brain network (i.e. each node-to-node connection) for every student dyad in the fMRI cohort. This yielded Nnodes * (Nnodes −1)/2 = 36,856 similarity (absolute difference) measures for each student dyad, where Nnodes is the number of nodes in the whole-brain parcellation (i.e. 272). We then employed elastic-net regularised linear regression to predict social distance of student dyads from the 36,856 similarity measures of resting-state functional connectivity. This regression technique has been used to successfully predict a variety of outcomes from MRI data, including openness to experience (a Big Five personality trait) (Dubois et al., 2018), psychosis (Salvador et al., 2017), and progression to Alzheimer’s disease in people with mild cognitive impairment (Teipel et al., 2015), demonstrating its ability to cope with the high dimensionality of MRI data.
Elastic-net regression combines penalty features of lasso and ridge regression techniques to shrink the coefficients of some regressors toward zero to deal with high dimensional data. Whereas lasso regression can shrink unnecessary regressors to zero and thereby reduce the number of predictors, ridge regression retains all regressors for inclusion in the model. Both lasso and ridge regression techniques have been shown to perform well when dealing with high-dimensional data under various conditions (Sirimongkolkasem and Drikvandi, 2019). Combining the two approaches, elastic-net regression allows for adjustment of the lasso-to-ridge ratio (α), providing greater opportunity for better model fits (James et al., 2013).
Elastic-net regressions of connectivity and social distance data were conducted in R using the glmnet package (Friedman et al., 2010). Three regression models were trained using data from two fMRI cohorts each (see Table 1). One fMRI cohort’s data were withheld during training so that the performance of the regression model could be evaluated using a previously unseen set of data.
The best-performing regression model for each training set was determined by optimising the tuning parameters λ and α (see Table 1). λ is a tuning parameter for the shrinkage penalty used to adjust the regression coefficients in the elastic-net regression. When λ = 0, the penalty term has no effect but as λ tends toward infinity, the shrinkage penalty grows, and the regression coefficient estimates approach zero. The optimal value of λ for each regression model was determined using a 10-fold nested cross-validation within the training data. The largest value of λ such that the cross validation error was within one standard error of the minimum was selected and the model was re-fit using all available observations.
The tuning parameter α dictates the ratio of ridge to lasso in the elastic-net regression. α values between 0 and 1 (with iterations of 0.1) were evaluated for each regression model. The best performing α value was chosen based on the smallest root mean squared error (RMSE) between predicted and observed values of social distance in the training data. Best-performing α values were α = 0 (pure ridge regression) for models 1 & 2 and α = 1 (pure lasso regression) for model 3. The performance of the optimised regression model was then evaluated by predicting the social distance of dyads in the previously unseen testing dataset.
After obtaining the predicted social distance scores from the elastic-net regression models, we evaluated the accuracy of predictive models by examining the relationship between the predicted social distance and the observed (actual) social distance. As before, there is a dependency of the data structure, owing to the involvement of each student in multiple dyads, potentially inflating the test statistics assessing statistical significance of the accuracy scores. To account for the dyadic nature of the data, we again used an LME model to obtain p value using the observed social distance as the dependent variable and the predicted social distance as the independent variable, with subji and subjjincluded as crossed random effects.
Finally, to evaluate the overall predictability of social distance using similarity in resting-state connectivity, we conducted a meta-analysis of LME models from the three elastic-net regressions (models 1, 2 and 3). A positive beta weight with 95% confidence interval excluding zero would indicate that prediction of social distance from neural similarity is feasible.
Results
Demographic data for each cohort are presented in Table 2. fMRI cohorts were relatively well matched to full year group cohorts for most demographic characteristics. Network characteristics for each cohort are presented in Table 3. Relatively higher modularity and mean path length in cohort 3 suggest that this network was more segregated and less well integrated compared with the younger two cohorts (cohorts 1 and 2).
Social network characterisation
The social networks for each cohort are depicted in Figure 3; fMRI cohorts were well distributed within whole year group samples and include both highly influential and less influential students (determined by Eigenvector centrality). Community detection analysis identified four communities (modules) in cohorts 1 and 3 and three communities in cohort 2 (Figure 4).
Dyadic similarities in functional connectivity as a function of social distance
Similarities in functional connectivity were assessed using LMEs on a whole-brain network level as well as within individual resting-state subnetworks for each fMRI cohort. No statistically significant relationship between social distance and similarity in functional brain connectivity was observed for any fMRI cohort at any resting-state network level (see Figure 5 for example plot). Beta weights (slopes) and 95% confidence intervals for LME models are presented in Figure 6a. Corresponding t and p values for individual tests are provided in the supplementary materials (table S1). Meta-analyses of LME models did not reveal any significant effects of social distance on the degree of functional brain similarity between students. Likewise, there was no relationship between the similarity in strength, diversity or modularity of students’ brains and their distance from one another in the social network (Figure 7a). These results indicate that the minimum path length between two individuals in a social network is not associated with similarities in brain function at rest, either at a whole-brain network or subnetwork level.
Dyadic similarities in functional connectivity as a function of community affiliation
Students’ relationships were then defined by their affiliation to different communities (friendship groups), derived using the Louvain method for community detection. Dyads including students from the same community were scored 1 and dyads including students from different communities were scored 0. LME models predicting brain similarity as a function of community affiliation did not support an effect of friendship grouping as a predictor of connectivity similarity. This was true for whole-brain and resting-state subnetwork connectivity (Figure 6b), as well as for graph measures of connectivity (Figure 7b). These results support those of the social distance analysis. They indicate that the similarity in resting brain function is no greater between individuals in the same social community (which we used here to estimate friendship groups) than those spanning different communities.
Inclusion of demographic variables (i.e. similarities in ethnicity and boarding status) in LME models did not affect the statistical significance of the overall findings (see table S2 for t and p values of individual tests).
Prediction of social distance based on node-to-node neural similarities
Using elastic-net regression, we sought to determine whether similarities in node-to-node connectivity within the brain could predict social distance between students. Data were split into training and testing sets such that each cohort was used to train two models and test the predictive validity of a third model. Assignments of training and testing sets for the three models are provided in Table 1. Optimal parameters (i.e. α and λ values) for each model were determined using training data; these parameters were then used to predict social distance in the testing set (to which the model was naïve).
RMSE quantifies how much a set of predicted values differ from their observed counterparts by measuring the standard deviations of the prediction errors. Lower values indicate smaller errors. RMSEs were 0.60, 0.65 and 0.82 for models 1, 2 and 3, respectively. Correlations for the observed vs predicted distance with the p values of the beta weights obtained from the LMEs between each pair of participants are presented in Figure 8a. Meta-analysis of LME models showed poor predictive power of models to classify social distance of dyads based on whole-brain functional connectivity (Figure 8b). Our meta-analysis results suggest that our predictive models will not extrapolate well to predict social distance in previously unseen social networks.
Discussion
Our results provide little evidence for homophily of internally generated (resting-state) functional brain connectivity in school-based social networks. Neither whole-brain nor subnetwork-based analysis (i.e. subnetworks relevant to social and motivational processing) of resting-state connectivity resulted in significant differences in similarity between friends and those farther removed in their social network. Likewise, graph theoretical measures of brain connectivity, including modularity, diversity and strength, were no more similar among friends than other more distantly connected pairs of students. Results from elastic-net regression, using the whole collection of local-level connectivity to predict social distance, also provide minimal evidence for similarity in resting-state functional connectivity among friends. Our findings were robust across individual cohorts of students and demonstrated a consistent non-significant result for homophily of resting-state connectivity. The inclusion of a data-driven approach to analysis (elastic-net regression) suggests our lack of evidence for the hypothesis is not due to noise from irrelevant variables or poor a priori selection of resting-state subnetworks for dimensionality reduction. Likewise, results are not due to poor sampling from the overall population. Students participating in the fMRI component of the study exhibited demographic characteristics representative of their original class cohorts, from which 87% - 96% of enrolled students provided social network data.
Current literature supports a role for social closeness in synchronisation of neural activation. For example, students who report higher social closeness to one another who also engage in silent eye contact prior to class exhibit stronger pairwise brain-to-brain synchrony during class activities compared with those less close who engage in eye contact (Dikker et al., 2017). This increased synchronisation between friends is evident even when friends are in the absence of one another; using functional MRI of individual students in a real-world social network, Parkinson and colleagues (2018) investigated synchronisation of neural activation during video clip viewing and found evidence for homophily at the neural response level. Brain regions where response similarity was associated with social network proximity included areas implicated in motivation, learning, affective processing, memory, attention, theory of mind and language processing (Parkinson et al., 2018), some behavioural traits of which have exhibited homophily in previous non-imaging studies (Tesser et al., 1984; Shin and Ryan, 2014b; Shin and Ryan, 2014a; Ilmarinen et al., 2017). These results suggest that, at least in terms of cognitive processing, similarities in behaviour relate to similarities in brain function.
In contrast, the current study examined neural homophily during a resting state scan. Our findings suggest that neural homophily observed in previous work may be specific to stimulus-evoked activation, and may not extend to stimulus-free intrinsically-generated brain activities. Importantly, stimuli used by Parkinson and colleagues (2018) included video clips of comedy, debates and documentaries, intended to evoke social and emotional responses from participants. The homophily observed in their study may therefore be dependent on cognitive processes important for social interaction. This would also explain how our resting-state experiment, which was relatively devoid of social context, failed to elicit homophilic outcomes.
It should be noted that, in a stimulus-free environment, i.e. during rest, subjects are free to mind wander, providing no time-locked cue with which to directly compare activation between two subjects. Instead, simultaneous activation of spatially disparate brain regions is evaluated within each subject to identify networks of brain regions that exhibit highly correlated patterns of activity. Therefore, in the current study, we evaluated the correlational strength between every node in the brain for each individual participant and correlated the whole-brain or subnetwork connectivity pattern between every pair of participants in the social network. This was a powerful approach to evaluating dyadic similarities in resting-state brain function in a social network, but at the same time, this novel approach makes it difficult to directly compare the current findings with the previous one (Parkinson et al., 2018), which focused on the similarities of the activation pattern rather than the pattern of brain functional connectivity. Given the evidence that the architecture of task-based networks closely resembles networks seen at rest (Cole et al., 2014), it may be an interesting future inquiry to examine whether neural homophily is observed in the brain functional network connectivity triggered by external cues and stimuli.
A notable difference between our sample population and that of Parkinson et al. is the age at which social network and imaging data were collected (schoolchildren vs undergraduate students). Differences in DMN connectivity have been reported between adults and children (Fair et al., 2008) as well as within individuals throughout early adolescence (Sherman et al., 2014). In addition, researchers have found associations between pubertal development and strength of intrinsic functional connectivity (Ernst et al., 2019). Younger samples may therefore exhibit less intrinsic network homophily than more mature samples due to greater brain variability between subjects.
The current study benefits from several strengths. Most notably, we evaluated homophily of resting-state connectivity in three different social networks comprising students of the same gender and similar age and education level, eliminating by design these demographic variables as possible sources of confound for neural homophily. Cohorts were evaluated as independent samples and then as individual studies in meta-analyses, ensuring sufficient statistical power of the overall analysis to find neural homophily. Homophily based on cognitive ability has been reported at a higher rate among girls compared with boys (Ilmarinen et al., 2017) and polygenic scores for educational achievement are more homogenous in women’s female social networks compared with men’s male social networks (Domingue et al., 2018). As intelligence and cognitive control are reflected in the resting-state connectivity (Cole et al., 2012), we therefore anticipated that an all-female sample would exhibit stronger homophily of resting-state connectivity than a mixed-gender or all-male sample. Based on these attributes, our study was well designed to identify homophily at the resting-state connectivity level.
These results contribute to the homophily literature by suggesting that homophily at the neural level may require some external stimulus that engages individuals in social or cognitive thoughts before synchronisation or similarities in connectivity are evident. To further our understanding of how our brain functioning is shaped by social factors, future research should examine the exact conditions under which neural homophily can be observed.
Acknowledgements
This research was supported by the JSPS KAKENHI (Grant Number 16H06406 to Kou Murayama), F. J. McGuigan Early Career Investigator Prize from American Psychological Foundation (to Kou Murayama); and the Leverhulme Trust (Grant Number RL-2016-030 to Kou Murayama). This research is also supported by BrainCanDo charitable company. The authors would like to thank the students who gave their time to take part in the social network and imaging components of this study, as well as the parents, teachers and staff of Queen Anne’s School who made data collection possible. We also wish to thank the Centre for Integrative Neuroimaging and Neurodynamics at the University of Reading for assistance with scanning and timetabling. Many thanks also to the members of the Motivation Science Lab, University of Reading, who gave their time to assist with social network and imaging data collection.
Footnotes
Conflicts of interest: The authors declare they have no conflicts of interest.