Similar neural responses predict friendship

Human social networks are overwhelmingly homophilous: individuals tend to befriend others who are similar to them in terms of a range of physical attributes (e.g., age, gender). Do similarities among friends reflect deeper similarities in how we perceive, interpret, and respond to the world? To test whether friendship, and more generally, social network proximity, is associated with increased similarity of real-time mental responding, we used functional magnetic resonance imaging to scan subjects’ brains during free viewing of naturalistic movies. Here we show evidence for neural homophily: neural responses when viewing audiovisual movies are exceptionally similar among friends, and that similarity decreases with increasing distance in a real-world social network. These results suggest that we are exceptionally similar to our friends in how we perceive and respond to the world around us, which has implications for interpersonal influence and attraction.


Supplementary Note 1 Defining social distance based on both reciprocated and unreciprocated social ties.
Our main analyses defined social ties based only on reciprocated ties. We reasoned that some unreciprocated ties may be the result of some participants tending to nominate large numbers of classmates as friends (out-degree ranged from 2 to 146), and that mutually reported ties were most likely to correspond to meaningful friendships. The same pattern of results as is reported in the main text was achieved when defining social distance based on both reciprocated and unreciprocated ties. An ordered logistic regression model revealed a significant effect of neural similarity (ordered logistic regression: ß = -0.26, SE = 0.12, p = .029; N = 861 dyads) on social distance that was comparable in magnitude to our main results: holding other covariates constant, compared to a dyad at the mean level of neural similarity and at any given level of social distance, a dyad one standard deviation more similar is 23% more likely to have social distance that is one unit shorter. Of the control variables also included in the model, dyadic dissimilarities in gender (ordered logistic regression: ß = 0.37, SE = 0.10, p = .0003; N = 861 dyads), nationality (ordered logistic regression: ß = 0.78, SE = 0.15, p = 1.2 x 10 -7 ; N = 861 dyads), and ethnicity ordered logistic regression: ß = 0.26, SE = 0.053, p = 7.6 x 10 -7 ; N = 861 dyads) were also related to social distance, whereas age (ordered logistic regression: ß = 0.0255, SE = 0.124, p = .84; N = 861 dyads) and handedness (ordered logistic regression: ß = 0.18, SE = 0.12, p = .14; N = 861 dyads) were not. A likelihood ratio test indicated that neural similarity added significant predictive power, above and beyond observable demographic similarity, χ 2 (1) = 9.61, p = .0019.

Supplementary Note 2
Testing whether neural similarity is associated with social network proximity without normalizing neural similarities within brain region. Prior to conducting the analyses reported in the main text, correlation coefficients were z-scored for each brain region across dyads in order to have a mean of 0 and a standard deviation of 1. This normalization step was performed to account for the fact that brain regions would likely vary in the extent to which they would become coupled across participants overall, as well as in the extent to which that coupling would vary across dyads, and we sought to characterize how similar neural responses were for a given pair of participants for a given brain region, relative to the similarity of all dyads' responses for that brain region. We also repeated our main analyses without z-scoring the Pearson correlation coefficients, and found the same pattern of results that is reported in the main text. Specifically, in an ordered logistic regression using social distance as the dependent variable and the dissimilarities in control variables (handedness, ethnicity, nationality, age, gender) and weighted (by ROI volume) average neural similarity (based on the Pearson correlation coefficients between preprocessed time series for each brain region for each unique pair of participants) as predictor variables, there was a significant effect of neural similarity on social distance (ordered logistic regression: ß = -0.232, SE = 0.108, p = .03; N = 861 dyads) similar to the results reported in the main text. There was also a significant relationship between A likelihood ratio test comparing the model described above to a model that did not include a term corresponding to neural similarity indicated that (un-normalized) average overall neural similarity added additional predictive power, above and beyond similarity in terms of the observed demographic variables, χ 2 (1) = 11.987, p = .0005.

Supplementary Note 3
Testing whether neural similarity is associated with social network proximity without weighting neural similarities by brain region volume. The analyses reported in the main text that probe the relationship between social network proximity and overall neural similarity, collapsed across brain regions, involve weighting each region of interest (ROI) by volume prior to averaging. As noted in the main text, a similar pattern of results was obtained when weighting every ROI equivalently, irrespective of its volume. In an ordered logistic regression using social distance as the dependent variable and the dissimilarities in control variables (handedness, ethnicity, nationality, age, gender), unweighted average neural similarity as predictor variables, and multi-way clustering to account for the non-independence of dyadic observations, there was a marginally significant effect of neural similarity on social distance predictive of social distance. A likelihood ratio test comparing the model described above to a model that did not include a term corresponding to neural similarity indicated that (unweighted) average overall neural similarity added additional predictive power, above and beyond similarity in terms of the observed demographic variables, χ 2 (1) = 8.477, p = .0036.

Supplementary Note 4
Accounting for participants' previous familiarity with videos. As reported in Supplementary Table 1 and the Methods section, post-scan interviews indicated that the majority of participants had no previous familiarity with the video stimuli used in the neuroimaging study.
Five of the 14 videos had been seen by both members of one or more dyads (please see the 'Prior familiarity with stimuli' sub-section of the Methods section for further details). After excluding any dyads whose members had both seen the same clips prior to participating in the study, the effect of neural similarity on social distance remained significant (ordered logistic regression: ß = -0.218, SE = 0.107, p = .042; N = 848 dyads) in our main ordered logistic regression analysis.

Supplementary Note 5
Permutation testing based on network randomization. We also performed permutation testing of the data to supplement the analyses described in the main text. We adopted the topological clustering methods employed by Christakis and Fowler 1 to test if there was a greater degree of clustering of particular neural response patterns than would be expected based on chance (i.e., if there was exceptionally high neural similarity among individuals close together in the social network). This method entailed iteratively computing the neural similarity between all individuals in the network in 1,000 randomly generated datasets in which the topology of the social network and the prevalence of particular neural response patterns were held constant while the assignment of neural data to individuals was randomly shuffled.
More specifically, a distribution of Pearson correlation coefficients corresponding to the null hypothesis that no relationship exists between social distance and neural similarity was obtained by randomly shuffling the neural time series data among participants 1,000 times, then computing the weighted (by ROI volume, as described in the main text) average neural similarity for dyads in each social distance category for each of the 1,000 randomly generated permutations of the dataset. Each participant's neural time series data consists of an 80 (brain regions) x 1,010 (time points) matrix -i.e., a set of 80 time series, each consisting of 1,010 time points. These neural time series datasets were randomly shuffled among the 42 fMRI study participants 1,000 times while keeping the social network data characterizing connections between participants constant. The magnitude of the weighted average neural similarity for each social distance category within each of the randomly permuted datasets was compared to that of the original, non-permuted data.
Results of these permutation tests revealed a similar pattern of results to those described in the main text and are illustrated in Fig. 6. Distance 1 dyads' (N = 63) neural response time series were, on average, exceptionally more similar to one another than would be expected based on chance, p = .03. There was a non-significant trend such that distance 2 dyads (N = 286) were marginally more similar to one another than would be expected based on chance, p = .06. Distance 3 dyads (N = 412) were exceptionally less similar to one another than would be expected based on chance, p = .003. Distance 4 dyads (N = 100) were neither more or less similar to one another than would be expected based on chance, p = .5. We note that that the fact that distance 3 dyads were significantly less similar to one another than would be expected based on chance alone does not imply that members of these dyads had anti-correlated neural response time series. Rather, members of distance 3 dyads were characterized by neural response similarities that were smaller in magnitude than would be expected if there were no relationship between neural response similarity and proximity in the social network.