Early-Childhood Social Reticence Predicts Brain Function in Preadolescent Youths During Distinct Forms of Peer Evaluation

Abstract

Social reticence is expressed as shy, anxiously avoidant behavior in early childhood. With development, overt signs of social reticence may diminish but could still manifest themselves in neural responses to peers. We obtained measures of social reticence across 2 to 7 years of age. At age 11, preadolescents previously characterized as high (n = 30) or low (n = 23) in social reticence completed a novel functional-MRI-based peer-interaction task that quantifies neural responses to the anticipation and receipt of distinct forms of social evaluation. High (but not low) social reticence in early childhood predicted greater activity in dorsal anterior cingulate cortex and left and right insula, brain regions implicated in processing salience and distress, when participants anticipated unpredictable compared with predictable feedback. High social reticence was also associated with negative functional connectivity between insula and ventromedial prefrontal cortex, a region commonly implicated in affect regulation. Finally, among participants with high social reticence, negative evaluation was associated with increased amygdala activity, but only during feedback from unpredictable peers.

Neuroscience-based models suggest that early-life social behavior influences long-term social competence by altering neural circuits (Nelson, Jarcho, & Guyer, 2016). Thus, delineating relations between early-childhood behavior and subsequent brain function may broadly inform theories of social cognition and development. While early behavioral traits may affect neural-response tendencies, these tendencies may be more prominent when experimental contexts are constrained (Mischel & Shoda, 1995). Hence, in the present study, we aimed to link early social behavior to brain function measured when attention focuses on salient social information. Social reticence reflects competing drives to interact and withdraw from peers (Degnan et al., 2014; Fox, Henderson, Rubin, Calkins, & Schmidt, 2001). Socially reticent children often have poor social competence (Rubin, Coplan, & Bowker, 2009). The underlying neural correlates of such behavior may be best detected in late childhood through preadolescence, a unique developmental phase when complex response patterns to social contexts are established (Crone & Dahl, 2012). We used a novel functional MRI (fMRI) paradigm to determine whether longitudinally assessed early-childhood social reticence predicts neural response among preadolescents as they interact with distinct types of peers.

Social reticence emerges during preschool, when peer-based interactions increase. Social reticence often develops in young children with behaviorally inhibited temperaments (Coplan, Rubin, Fox, Calkins, & Stewart, 1994). Behaviorally inhibited infants with childhood social reticence are at particularly high risk for poor social competence (Degnan et al., 2014). Although many behaviorally inhibited children exhibit social reticence during preschool, temperament during infancy is likely only one pathway that leads to childhood social reticence (Rubin et al., 2009).

Overt expressions of infants’ behavioral inhibition and preschoolers’ social reticence diminish with maturation (Degnan & Fox, 2007). However, early-childhood social reticence remains a long-term latent risk factor for poor social competence, which may be reflected in brain function. Indeed, such persistent effects manifest in adolescent and adult brain function during the processing of faces (e.g., Perez-Edgar et al., 2007; Schwartz, Wright, Shin, Kagan, & Rauch, 2003), monetary rewards (e.g., Bar-Haim et al., 2009), and cognitive conflict (e.g., Jarcho, Fox, et al., 2013). However, stimuli eliciting these differences are only weakly related to the construct of social competence. In the only fMRI study to test for lasting effects on brain function during social processing (Guyer et al., 2014), we showed that young adults with a history of behavioral inhibition as infants and social reticence as preschoolers had enhanced striatal responding when anticipating feedback from high-value peers. Here, we addressed several limitations of this prior work by using a paradigm (the virtual-school paradigm) with three unique features (for a review, see Jarcho, Leibenluft, et al., 2013). First, this paradigm models unpredictable, anxiogenic social situations (Grupe & Nitschke, 2013). Second, it includes ongoing, real-time social interactions. Finally, it elicits behaviorally meaningful, flexible responses from participants.

We used the virtual-school paradigm to evaluate whether longitudinally assessed early-childhood social reticence predicts altered brain function during preadolescence. One potent elicitor of aberrant behavior in socially reticent youth is anticipating social evaluation (Coplan et al., 1994; Rubin et al., 2009), a circumstance that engages brain regions implicated in threat detection, social distress, and salience processing, including amygdala (Guyer et al., 2008), dorsal anterior cingulate cortex (dACC; Moor, van Leijenhorst, Rombouts, Crone, & Van der Molen, 2010), and mid-to-anterior insula (Liljeholm, Dunne, & O’Doherty, 2014), respectively. During social evaluation, this pattern of engagement is typically heightened among individuals with poor social competence, including those with social anxiety or chronic rejection experiences (see Jarcho, Leibenluft, et al., 2013, for a review). These alterations are often accompanied by negative functional connectivity with brain regions implicated in self-reflection, valuation, and emotion regulation, such as ventromedial prefrontal cortex (vmPFC; e.g., Bruhl, Delsignore, Komossa, & Weidt, 2014; Gee et al., 2013), perigenual anterior cingulate cortex (ACC) and orbital medial frontal gyrus (Brodmann’s area, or BA 10/32). Thus, we hypothesized that for preadolescents anticipating unpredictable relative to predictable social evaluation, a history of childhood social reticence will predict greater engagement of amygdala, dACC, and mid-to-anterior insula, and negative functional coupling between these regions and vmPFC.

Our secondary goal in this study was to evaluate whether early-childhood social reticence predicts altered brain function in preadolescents as they receive feedback from peers. Prediction errors associated with unexpectedly positive outcomes typically heighten activity in the striatum, a structure implicated in reward-based learning (Schultz, Dayan, & Montague, 1997), whereas unexpectedly negative feedback can engage amygdala, a structure implicated in threat-based learning (McHugh et al., 2014). Moreover, we recently linked poor social competence, particularly during early adolescence, to social prediction-error signaling (Jarcho et al., 2015). Thus, we hypothesized that early-childhood social reticence will differentially predict engagement of striatum and amygdala during unpredictable positive and negative social feedback, respectively.

Method

Participants

Participants were recruited from the community at 2 years of age and enrolled in an ongoing longitudinal study. Individuals in this unselected sample served as age- and gender-matched comparisons for another sample targeted in the longitudinal study. Those in this second sample were recruited as infants and enrolled on the basis of their expression of behaviorally inhibited temperament (see the Supplemental Material available online for details). This second sample did not complete the current study. No neuroimaging data from either sample have been reported previously, and participants did not complete additional fMRI-based tasks in conjunction with this study.

At 11 years of age, participants from the unselected sample were eligible for the current study if they were free of psychiatric impairment, did not use psychotropic medication, and showed no contraindications for neuroimaging. We chose to study participants at this age for three reasons. First, preadolescence represents a unique inflection point during development when complex peer-based relationships begin to gain salience (e.g., Nelson et al., 2016). Second, while preadolescents with a history of high social reticence may be at risk for developing social anxiety disorder, impairing social anxiety symptoms most commonly emerge around 13 years of age (Kessler et al., 2005). Studying participants at 11 years of age, before overt expression of such symptoms, provides the opportunity to quantify mechanisms related to risk without the confounding presence of current impairment. Third, this study was performed in the context of a larger longitudinal program of research in which participants undergo laboratory assessments at 1- to 3-year intervals that map onto distinct phases of development. The time between late childhood and preadolescence (10–11 years of age) is one such phase.

Participants were divided into high- (n = 30) and low- (n = 23) social-reticence groups on the basis of maternal-report questionnaires and behavioral observations collected at 2, 3, 4, 5, and 7 years of age (see Table 1 for sample demographics). As in previous neuroimaging work on the temperamental trait of behavioral inhibition (e.g., Bar-Haim et al., 2009; Jarcho, Fox, et al., 2013; Perez-Edgar et al., 2007), measures were standardized and averaged to create a social-reticence composite. The current sample was not selected for the presence or absence of the specific temperamental trait of behavioral inhibition, since making such a determination would have required studying the participants in infancy.

Table 1.

Demographic Statistics

	Low-social-reticence group (n = 23)		High-social-reticence group (n = 30)
Variable	M	SD	M	SD
Age (years)	11.11	0.49	11.06	0.39
Pubertal development stagea	1.49	0.28	1.35	0.38
Intelligence scoreb	114.17	11.20	118.73	10.00
Social-reticence composite scorec	−0.47	0.29	0.46	0.43
Current social anxiety symptomsd	16.57	13.73	15.24	12.78
Current shynesse	1.82	0.69	3.02	0.80

Note: There were 13 males and 10 females in the low-social-reticence group and 16 males and 14 females in the high-social-reticence group.

^aPubertal development stage was assessed by determining participants’ Tanner stage on the Pubertal Development Scale (Petersen, Crockett, Richards, & Boxer, 1988). Data for this measure were available only for a subset of the full sample (low-social-reticence group: n = 18; high-social-reticence group: n = 29). ^bParticipants’ intelligence scores were measured using the second version of the Weschler Abbreviated Scale of Intelligence (Wechsler, 2011). ^cThe social-reticence composite score was derived from parents’ report of shyness and observed social reticence when participants were between 2 and 7 years of age. ^dThe current number of social anxiety symptoms reflects the Child Total score on the Screen for Child Anxiety Related Disorders questionnaire (Muris et al., 1998). ^eCurrent shyness reflects the shyness score on the Early Adolescent Temperament Questionnaire (Ellis & Rothbart, 2001). Data for this measure were available only for a subset of the full sample (low-social-reticence group: n = 18; high-social-reticence group: n = 30). The two groups’ means for this measure are significantly different (p < .05).

A composite measure of social reticence was utilized because combining data from different contexts, informants, and ages better reflects behavioral tendencies than a single measure at one time point. A categorical approach to data analysis was used so results could be interpreted in light of prior neuroimaging studies of behavioral inhibition, which have largely used this approach (e.g., Bar-Haim et al., 2009; Jarcho, Fox, et al., 2013; Perez-Edgar et al., 2007; Schwartz et al., 2003). However, all analyses reported using categorical measures were repeated using fully continuous measures of social reticence, and the same pattern of results emerged.

During a visit to the National Institute of Mental Health (NIMH) at 11 years of age, participants’ psychopathology, pubertal status, symptoms of social anxiety, and shyness were assessed; intelligence was assessed at an earlier visit the same year at the University of Maryland (UMD), College Park (see the Supplemental Material for assessment details). A structured clinical interview (Schedule for Affective Disorders and Schizophrenia for School-Age Children: Present and Lifetime Version; Kaufman et al., 1997) determined that 2 preadolescents with low social reticence and 1 with high social reticence met diagnostic criteria for generalized or social anxiety disorder. Follow-up questioning confirmed that participants suffered only minor degrees of impairment from their anxiety, thus they were retained in the sample. On average, participants with high and low social reticence did not differ on intelligence (assessed using the Weschler Abbreviated Scale of Intelligence; Wechsler, 2011), questionnaire-based measures of social anxiety (the Screen for Child Anxiety Related Emotional Disorders; Muris et al., 1998), or pubertal status (assessed using the Pubertal Development Scale; Petersen, Crockett, Richards, & Boxer, 1988). The majority of participants were prepubertal (Tanner Stage 1 = 91.49%, Tanner Stage 2 = 8.51%). All results remained significant after we controlled for pubertal stage (see the Supplemental Material for further details). Of note, overall rates of psychopathology were low, likely not only because of the relatively young age of the sample, but also because we excluded participants who were taking medication or who had symptoms considered clinically impairing. Social reticence was defined using a longitudinal, multi-informant composite index, but this index was not acquired at 11 years of age. However, shyness, an important component of social reticence, was assessed at age 11 via questionnaire-based maternal report (Early Adolescent Temperament Questionnaire; Ellis & Rothbart, 2001). Shyness at age 11 and the early-childhood social-reticence composite were correlated (r = .68, p < .01), and preadolescents with high social reticence were shyer than those with low social reticence (p < .001).

Data from 11 additional participants were excluded from analysis because of excessive head motion during the fMRI scan (n = 5), failure to complete the fMRI scan (n = 3), technical failure (n = 2), and a structural brain abnormality (n = 1). One further participant was excluded from analysis for being an outlier on IQ (score = 84). (Total sample size was constrained by the number of participants enrolled in the original longitudinal study; see the Supplemental Material for further details.)

Procedure

Institutional review boards at NIMH and UMD approved all procedures. We employed the virtual-school paradigm, which models uncertainty during anticipated social evaluation in real-time interactions (see Jarcho, Leibenluft, et al., 2013, for details). Participants completed the virtual-school task during two visits to NIMH. At the first visit, participants were told they were going to be the “new kid” at our virtual school. They generated a cartoon avatar and personal profile they believed would be shown to “other students” who were purportedly age and gender-matched peers. These other student peers had, in fact, been preprogrammed to maintain experimental control.

During a second visit to NIMH, just prior to scanning, participants learned that two of these peers have a reputation for being nice, two for being unpredictable, and two for being mean (Fig. 1a). Providing this information minimized between-participants variability in learning during social interactions and modeled real-world contexts involving peers with distinct, known personalities. Participants confirmed that they understood their peers’ reputations via self-report. Specifically, participants were asked to rate the kind of person they thought each peer was on a 10-point Likert scale from 0, mean, to 10, nice, with 5, can’t tell, as the midpoint (Fig. 1b). To eliminate confounds, we randomly assigned peer reputations to the other-student avatars across participants.

Fig. 1.

Overview of the virtual-school paradigm. Participants first learned the reputation of six gender-matched “other students” (a) by viewing their avatars and chat names, as well as comments and ratings purportedly provided by “prior subjects.” A 10-point scale (b) was then used to assess participants’ comprehension of the other students’ reputations. Next, the virtual-school task (c) was completed in the functional MRI scanner. Each trial was preceded by an intertrial interval (0–8 s; M = 4 s). Participants began to anticipate social evaluation when the word “Typing . . .” appeared in a balloon above one of the other students (2–4 s; M = 3 s). This was followed by a social-evaluative comment addressed to participants (2–10 s; M = 6 s); these comments could be positive or negative. Finally, participants used a button box to select a response (5 s), which was purportedly delivered to the other student when it appeared on screen (2 s).

At this second visit to NIMH, participants then completed the virtual-school task (Fig. 1c) during an fMRI scan, which consisted of three 9-min runs (see the Supplemental Material for fMRI acquisition parameters). Each run contained two blocks. Each block began with the participant being randomly assigned to one of several classrooms (9 s) populated by other-student peers, who were randomly assigned to seats. Each run consisted of 24 trials (8 randomly presented trials per reputation type, evenly split across the two blocks). This resulted in 72 trials total (24 trials per reputation type). Each trial contained three events. First, participants anticipated social evaluation from predictably nice, predictably mean, and unpredictable peers. Second, participants received a social evaluation, which was always positive (100%) from predictably nice peers, positive half of the time (50%) and negative half of the time (50%) from unpredictable peers, and always negative (100%) from predictably mean peers. Finally, participants selected one of six prespecified responses to social evaluation.

Each trial was preceded by an intertrial interval (0–8 s; M = 4 s). At the start of each trial, participants were cued to anticipate social evaluation when the word “Typing . . .” appeared above one of the other students (2–4 s; M = 3 s). Because each other student had an established reputation, imaging procedures modeled brain function as participants anticipated predictably positive evaluations from nice peers, predictably negative evaluations from mean peers, and unpredictable evaluations from unpredictable peers. The anticipation period was followed by an evaluative comment directed at the participant (2–10 s; M = 6 s). All participants received the same set of pregenerated social-evaluative comments, which were either positive (e.g., “Cool avatar!”) or negative (e.g., “You’re lame.”). To strengthen the perception that participants were interacting with real peers, half of the comments were programmed to automatically reference information from the participant’s personal profile (e.g., “You like Justin Bieber? You’re lame.”).

To establish an interactive context that modeled how participants might react to real-world experiences, we asked them to respond to positive or negative social evaluation using a button box (5 s). When the row of six response options appeared at the bottom of the screen, participants used the left and right buttons to navigate to their preferred response and a third button to make their selection. Response options were positive (“You’re nice,” “That’s nice”), negative (“That’s mean,” “You’re mean”), sarcastic (“Thanks!!!”), or avoidant (⦸). The selected option then appeared on the screen for purported peers to see or was omitted in the case of avoidant responses (2 s).

The sarcastic response option was included on the basis of behavioral testing of the paradigm by Jarcho, Leibenluft, et al. (2013). In that study, a notable number of positive responses were generated following negative social evaluation from mean and unpredictable peers. During debriefing, participants reported using this response pattern to express sarcasm. To improve the interpretability of behavioral data, we instructed participants in the present study to use “Thanks!!!” to convey sarcasm instead of the positive response options.

After scanning, deception was assessed, and participants were debriefed. All participants were deceived by the task (see the Supplemental Material for further details), and no adverse events occurred.

Data analyses

Validation analyses

Measures of reputation comprehension were averaged across each pair of other students embodying each type of reputation (nice, mean, unpredictable). To confirm that participants understood peer reputations, we ran a 3 (peer reputation: nice, mean, unpredictable) × 2 (social reticence: high, low) between-participants repeated measures analysis of variance (ANOVA) to evaluate personality ratings of peers.

Participant response frequency

Behavioral response frequencies provided additional data that confirmed participant understanding of peer reputations. Response frequency was averaged across each pair of other students embodying each type of reputation. A subset of participant responses could not be clearly interpreted: negative and sarcastic responses to positive evaluation from nice and unpredictable peers, as well as positive responses to negative evaluation from mean and unpredictable peers. Overall, the use of such ironic responses was uncommon (M = 4.38 across all trials, SD = 6.43) and did not differ between high- and low-social-reticence groups (see Table S1 in the Supplemental Material). This response type was therefore excluded or treated as a nuisance variable in subsequent analyses.

To confirm that the valence of social-evaluative feedback (positive, negative) and the reputation of peers (nice, mean, unpredictable) engendered distinct social experiences, we ran a repeated measures ANOVA on the frequency with which participants selected each of the six response options. Follow-up paired-samples t tests assessed differences in response frequency to (a) positive feedback compared with negative comments from all peer types, (b) positive feedback from nice compared with unpredictable peers, and (c) negative feedback from mean compared with unpredictable peers. To determine whether social reticence biased frequency of participant response selection, the between-participants factor of social reticence (high, low) was added to the repeated measures ANOVA. Each pairwise comparison was then tested using separate repeated measures ANOVAs, with the between-participants factor of social reticence added to the analysis.

fMRI data preprocessing and individual-level fMRI analyses

Preprocessing and fMRI analyses were conducted using AFNI software (Cox, 1996). Standard preprocessing steps were implemented with afni_proc.py; these steps included slice timing, coregistration, smoothing to 6-mm full-width half maximum (FWHM), spatial normalizing to standard Talairach space, and resampling, which resulted in 2.5 mm³ voxels. Separate regressors were created for three groups of events: (a) anticipated social evaluation based on nice, mean, and unpredictable peer reputations; (b) receipt of evaluation based on peer reputation (positive evaluation from nice peers, positive evaluation from unpredictable peers, negative evaluation from mean peers, negative evaluation from unpredictable peers); and (c) participant response to evaluation. There was insufficient power to group response events by each of the six prespecified response types elicited by each of the four types of social evaluation, particularly since the response patterns tended to differ as a function of both the valence and predictability of the feedback (see Fig. 2). However, recent data suggest that patterns of neural engagement during behavioral responding can be meaningfully interpreted in terms of response valence, regardless of the event that elicits the behavior (e.g., Guitart-Masip, Duzel, Dolan, & Dayan, 2014). To explore these relations, we modeled events during participant response to social evaluation, without respect to peer feedback, as positive response valence (“You’re nice,” “That’s nice”), negative response valence (“You’re mean,” “That’s mean”), and responses of no interest (avoidant and sarcastic responses, which did not occur often enough to be modeled separately, and ironic responses). The classroom selection process at the beginning of each block was modeled but treated as a regressor of no interest to eliminate potential noise generated during that time.

Fig. 2.

Pie charts showing the percentage of times participants chose each of the six responses after being evaluated by a peer. Results are shown separately for positive evaluations by (a) nice peers and (b) unpredictable peers and for negative evaluations by (c) mean peers and (d) unpredictable peers. Dotted sections indicate ironic responses (sarcastic and negative responses to positive peer evaluation, and positive responses to negative peer evaluation).

Individual-level regression analyses were carried out with AFNI’s 3dDeconvolve function, which automatically flags regressors with medium (r > .40) to high (r > .76) levels of collinearity. Task-specific events (spanning the duration of each event) were convolved with a duration-modulated boxcar regressor. An additional six regressors modeled motion residuals (corresponding to translation and rotation in each of the Montreal Neurological Institute, or MNI, x, y, z directions), and four regressors modeled low-frequency baseline drift. Temporally adjacent repetition times (TRs) with a euclidean-norm motion derivative greater than 1.5 mm were omitted from the model via censoring. No additional nuisance covariates were included. This analysis produced a β coefficient and t statistic for each voxel and regressor. Whole-brain percentage signal-change maps were generated by dividing signal intensity at each voxel by the mean voxel intensity, and multiplying by 100. Participants (n = 5) with censoring rates of 30% or more TRs were omitted from analysis. Motion rates did not differ between participants with low and high social reticence, who had an average of 4.76% TRs censored.

Collinearity was uniformly low between events that occurred while anticipating social evaluation and all other event types. Medium and high levels of collinearity were detected among a subset of events that occurred during social evaluation and participant response. Given this, results for Social Reticence × Response Valence analyses are reported in the Supplemental Material (see the Exploratory Analyses: Social Reticence × Response Valence section) and should be considered exploratory.

Primary analysis for anticipated social evaluation: Social Reticence × Peer Reputation interactions

Primary analyses focused on group differences in brain function during the anticipation of social evaluation, since this is typically the most distressing phase of social interaction for children with social reticence (Coplan et al., 1994). AFNI’s 3dMVM software (Chen, Adleman, Saad, Leibenluft, & Cox, 2014) was used to conduct a repeated measures ANOVA with social reticence (high, low) as a between-participants factor and peer reputation (nice, mean, unpredictable) as a within-participants factor. All events were included in these models, thus inferences were based on results from the highest-order interactions, with significance set using an overall false detection probability based on 10,000 Monte Carlo simulations calculated by AFNI’s AlphaSim software (mean estimated spatial correlation of 9.21 mm × 9.09 mm × 8.20 mm FWHM). Simulations determined that for a whole-brain analysis, a cluster size (ke) of 73 contiguous voxels was needed to achieve a significance threshold of less than .005, with an overall family-wise error rate (α) of .05. This approach reflects the call for more rigor in statistical analyses of neuroscience data (e.g., Nieuwenhuis, Forstmann, & Wagenmakers, 2011). Applied to fMRI data, this can be achieved by basing inferences on analyses that include all experimental conditions and corresponding results that survive relatively conservative cluster-extent thresholds. For all analyses, significant clusters were extracted and plotted. To facilitate interpretation and explicate factors driving interactions, we conducted post hoc analyses in SPSS on extracted data.

Functional connectivity: interactions of social reticence with the anticipation of unpredictable versus predictably positive social evaluations

To investigate participants’ functional activity while they anticipated unpredictable versus predictably positive social evaluations, we used a generalized psychophysiological-interaction (PPI) approach (McLaren, Ries, Xu, & Johnson, 2012) that tested condition-dependent covariation in brain activity between one brain region (the “seed”) and the rest of the brain. For each participant, mean-adjusted eigenvariate time-series data were extracted from a 6-mm seed, centered at the peak voxel (MNI space coordinates) of significant clusters in the primary analysis that are most commonly linked with social threat and peer evaluation: dACC (x = −1, y = −1, z = 39), left insula (x = −44, y = −1, z = 4), and right insula (x = 49, y = −6, z = 6). Given the size of each cluster (kes > 125), a 6-mm sphere was used to ensure that each seed reflected functional connectivity associated with a single brain region. Time-series data were deconvolved with the hemodynamic response function before a PPI term was generated for anticipated social evaluation from nice, mean, and unpredictable peers. Although all three events were modeled, we focused on specific contrasts of interest, selected on the basis of results from the primary analysis, to avoid issues related to multiple comparisons. Therefore, the analyses focused on differences in functional connectivity during the anticipation of social evaluation from unpredictable compared with nice peers, the contrast that showed the largest between-groups differences. (See the Supplemental Material for results from the other two contrasts; i.e., unpredictable vs. mean and mean vs. nice peers.)

A group-level analysis for each seed was performed with 3dMVM (Chen et al., 2014). Specifically, for each seed, we conducted a repeated measures ANOVA with social reticence (high, low) as a between-participants factor and PPI term for evaluation type (nice, mean, unpredictable) as a within-participants factor. The same whole-brain-corrected statistical threshold was applied to the PPI analyses (p < .005, ke > 73; see the Supplemental Material for further information).

Secondary analysis for receipt of social evaluation: Social Reticence × Social Evaluation

Secondary analyses focused on group differences in brain function during the receipt of distinct types of social evaluation. We used 3dMVM (Chen et al., 2014) to conduct a repeated measures ANOVA with social reticence (high, low) as a between-participants factor and evaluation type (predictably positive, unpredictably positive, predictably negative, unpredictably negative) as a within-participants factor. To test hypotheses regarding positive and negative prediction error, we performed region-of-interest (ROI) analyses using anatomical masks of striatum (caudate, putamen, and nucleus accumbens; bilateral mask size = 1,375 voxels) and amygdala (bilateral mask size = 137 voxels). Significance was set using an overall false detection probability based on 10,000 Monte Carlo simulations for each set of bilateral masks. Simulations determined that the cluster size needed to achieve a significance threshold of less than .005, with an overall family-wise error rate (α) of .05 was 18 for striatum and 3 for amygdala.

Other data

Given that the present sample is enrolled in a longitudinal study, considerable data are available for these participants. However, analyses were restricted to isolating differences in brain engagement and behavior as a function of peer reputation, participant social reticence, and participant exposure to peer victimization (to be reported elsewhere). Group differences in demographic data, and measures of mental health and social competence, were also assessed to more fully characterize the sample but were not used in fMRI analyses.

Results

Validation analyses

As expected, before participants completed the virtual-school task in the fMRI scanner, there was a main effect of peer reputation on participants’ ratings of peers, F(1, 51) = 341.04, p < .001, η_p² = .87. Specifically, nice peers (M = 8.86, SD = 1.31) were rated as “nicer” than unpredictable peers (M = 4.65, SD = 1.30), who were in turn rated as “nicer” than mean peers (M = 2.47, SD = 1.28). Contrasts were significant (ps < .001) for each pairwise comparison. This confirmed that peer reputations were well learned.

There was also a Social Reticence × Peer Reputation interaction on personality ratings, F(2, 50) = 4.57, p = .013, η_p² = .08. However, only trends for group-based pairwise comparisons emerged (see Table 2). Response variance was more restricted for participants with high relative to low social reticence on personality ratings for unpredictable and nice peers (Levene’s test of equality for variance; ps < .01).

Table 2.

Mean Reputation Rating for Each Type of Peer

Peer type	Low-social-reticence group	High-social-reticence group
Mean	2.78 (1.35)	2.23 (1.20)
Unpredictablea	4.22 (1.57)	4.98 (0.94)
Nicea	8.48 (1.58)	9.15 (0.99)

Note: Participants rated peer reputations prior to undergoing functional MRI. Standard deviations are given in parentheses.

^aThe means for the two groups are marginally different (p < .10).

Participant response frequency

As depicted in Figure 2, there was a robust interaction between peer reputation and valence of social evaluation on participant response frequency, F(15, 38) = 83.35, p < .001, η_p² = .62. Positive comments yielded more positive than negative responses, and negative comments yielded more negative than positive responses (ts > 9.50, ps < .001). There were also effects of reputation. Compared with positive comments from unpredictable peers, positive comments from nice peers elicited more frequent “You’re nice” responses, t(52) = 11.77, p < .001, and less frequent “That’s nice” responses, t(52) = −6.76, p < .001, and avoidant responses, t(52) = −5.42, p < .005. Compared with negative comments from unpredictable peers, negative comments from mean peers elicited more frequent “You’re mean” responses, t(52) = 4.52, p < .001, and less frequent “That’s mean” responses, t(52) = −3.80, p < .001. Frequency of avoidant and sarcastic responses did not vary. No significant group differences emerged.

Primary analysis for anticipated social evaluation: Social Reticence × Peer Reputation

As expected, Social Reticence × Peer Reputation interactions manifested in dACC and left and right insula (Table 3; Figs. 3a–3c), which included extensive activation clusters extending from mid to anterior aspects of the structure (see Fig. S1 in The Supplemental Material). Additional activity was observed in lingual gyrus, premotor, and posterior cingulate cortices (see Fig. S2 in the Supplemental Material). In each region, participants with high social reticence had significantly greater activity while anticipating unpredictable feedback than while anticipating predictably positive feedback from nice peers or predictably negative feedback from mean peers. A similar but less consistent pattern emerged in left insula and premotor cortex when contrasting anticipation of predictable feedback from mean relative to nice peers.

Table 3.

Clusters Whose Activation During Anticipation of Social Evaluation Was Significantly Affected by the Social Reticence × Peer Reputation Interaction

Brain region	MNI coordinates			Cluster size (voxels)	F(2, 50)	ηp2
	x	y	z
Right insula	49	−4	4	138	11.21	.18
Left insula	−44	−1	4	170	14.78	.23
Dorsal anterior cingulate cortex (BA 24)	−1	−1	39	215	9.57	.16
Lingual gyrus	−14	−46	4	130	10.77	.17
Right premotor cortex (BA 6)	41	−8	47	166	10.01	.16
Posterior cingulate cortex (BA 31)	11	−27	43	179	10.58	.17

Note: The table presents results from a corrected whole-brain analysis with a significance threshold of less than .005 and a cluster size of at least 73 voxels. BA = Brodmann’s area. MNI = Montreal Neurological Institute.

Fig. 3.

Functional activity during anticipated social evaluation. The brain maps depict clusters with significant activation for the Social Reticence × Peer Reputation interaction. The graphs decompose the interaction for the following clusters: (a) dorsal anterior cingulate cortex (dACC), (b) left insula, and (c) right insula. In each graph, mean percentage signal change is shown for contrasts between anticipating feedback from unpredictable relative to nice peers, unpredictable relative to mean peers, and mean relative to nice peers, separately for participants with high and low social reticence (SR). Error bars show ±1 SE. Asterisks directly above the data bars indicate significant differences for contrasts within a group, and asterisks above the brackets indicate significant differences for contrasts between groups (^†p < .10, *p < .05, **p < .01, ***p < .005). Unpred = unpredictable.

Functional connectivity: interactions of social reticence with the anticipation of unpredictable versus predictably positive social evaluations

Separate PPI analyses used left and right insula and dACC as seeds. For the right-insula seed (Fig. 4), group differences emerged in an area of vmPFC (x = 5, y = 48, z = −4; ke = 439), F(1, 51) = 21.49, p < .005, η_p² = .296, that extended from rostral ACC to orbital aspects of medial frontal gyrus (BA 10/32) and in right premotor cortex (BA 6; x = 44, y = −7, z = 44; ke = 203), F(1, 51) = 23.84, p < .005, η_p² = .319. Functional connectivity during the anticipation of unpredictable peers’ feedback, relative to nice peers’ predictably positive feedback, was negative for participants with high social reticence and positive for participants with low social reticence, which indicates a relative decoupling in the high-social-reticence group. No significant effects emerged for left insula or dACC seeds.

Fig. 4.

Functional connectivity during anticipated social evaluation. The brain maps depict clusters with significant psychophysiological interactions (PPIs) with right insula during anticipated social evaluation from unpredictable relative to nice peers. The graphs decompose mean PPI connectivity (a) between ventromedial prefrontal cortex (vmPFC) and right insula and (b) between right premotor cortex and right insula for the contrast of unpredictable relative to nice peers, separately for participants with high and low social reticence (SR). Error bars show ±1 SE. Asterisks directly above the data bars indicate significant PPI differences within a group, and asterisks above the brackets indicate significant PPI differences between groups (^†p < .10, **p < .01, ***p < .005).

Secondary analysis of receipt of social evaluation: Social Reticence × Social Evaluation

ROI analyses of the Social Reticence × Social Evaluation interaction revealed group differences in left amygdala in response to negative social evaluation (Fig. 5; x = −26, y = −9, z = −14; ke = 28), F(3, 49) = 7.71, p < .005, η_p² = .131. Specifically, participants with high social reticence had less activity when they received negative social evaluation from unpredictable relative to mean peers, t(29) = −2.59, p < .05. The opposite pattern emerged for those with low social reticence, t(22) = 3.53, p < .005. There were no group differences in striatal ROI or whole-brain analyses.

Fig. 5.

Functional activity during receipt of social evaluation. The brain map depicts a cluster within the amygdala with significant activation for the Social Reticence × Peer Evaluation interaction. The graph decomposes the interaction for this cluster: Mean percentage signal change in left amygdala is shown for contrasts between positive evaluation from unpredictable relative to nice peers and between negative evaluation from unpredictable relative to mean peers, separately for participants with high and low social reticence (SR). Error bars show ±1 SE. Asterisks directly above the data bars indicate significant differences for contrasts within a group, and asterisks above the brackets indicate significant differences for contrasts between groups (*p < .05, ***p < .005). Unpred = unpredictable.

Discussion

There were three main findings of the present study. First, high relative to low social reticence in early childhood predicted heightened engagement in dACC as well as mid-to-anterior insula as preadolescents anticipated unpredictable social evaluation. Second, high social reticence was associated with negative functional connectivity between insula and both vmPFC and premotor cortex in similar circumstances. Finally, high social reticence was associated with greater amygdala response to predictable relative to unpredictable negative social evaluation, while the opposite pattern emerged for participants with low social reticence. Thus, using a novel paradigm with high ecological validity, we demonstrated that early-childhood social reticence uniquely affects neural circuits engaged by unpredictable social interactions. This suggests that early-life social behavior has a lasting effect on neural mechanisms that support social cognition during subsequent phases of development (e.g., Nelson et al., 2016).

Among preadolescents with high social reticence, dACC and insula were engaged by anticipating unpredictable social evaluation. Distress and motivation to avoid social contexts are primary manifestations of social reticence (Coplan et al., 1994). This is consistent with data on functions supported by dACC, mid-to-anterior insula (Craig, 2009; Shackman et al., 2011) as well as prior research on social anxiety (Bruhl et al., 2014; Jarcho, Leibenluft, et al., 2013). Thus, this pattern may reflect brain-based expression of childhood social reticence that confers long-term risk for social anxiety disorder. The link between neural response and psychopathology could not be evaluated in the current sample, since participants were younger than the typical age of onset for pathological social anxiety. Thus, we will test whether brain function engaged in the current study predicts social competence in longitudinally assessed adolescents at 14 years of age. We hypothesize that the heightened activity in dACC and mid-to-anterior insula documented here will predict poor social competence in midadolescence.

Interplay between insula and interconnected brain regions may guide responses to socially distressing or salient stimuli (Uddin, 2015). Consistent with this idea, our findings revealed that functional connectivity between insula and vmPFC, commonly implicated in self-reflection, valuation, and inhibitory control (Jenkins & Mitchell, 2011; Nelson & Guyer, 2011), and premotor cortex, which is closely linked with planning behavioral responses (Nachev, Kennard, & Husain, 2008), varied as a function of social reticence. Specifically, during the anticipation of unpredictable, relative to predictable, positive social evaluation, functional connectivity with insula was negative for preadolescents with high childhood social reticence, but positive for those with low social reticence. However, group differences in functional connectivity did not relate to group differences in behavior.

Distinct results emerged during the receipt of social evaluation. High social reticence was associated with greater amygdala activity during receipt of predictable relative to unpredictable negative evaluation; the opposite pattern occurred for participants with low social reticence. These data can be interpreted in light of prediction-error models of brain function in which predictability differentially interacts with outcome valence (e.g., McHugh et al., 2014). Such models suggest that people with low social reticence exhibit normative prediction-error signaling during receipt of negative social evaluation, whereas those with high social reticence may exhibit aberrant signaling. An emerging literature maps the neural correlates of social prediction error (e.g., Jones et al., 2014; Somerville, Heatherton, & Kelley, 2006). However, few studies relate individual differences in social prediction error to real-world behaviors, development, or symptoms. Preliminary findings link social anxiety to prediction error in the social domain (Jarcho et al., 2015), and adaptations of the current task could extend this literature. Unfortunately, the current paradigm is not optimized to do so, as prediction-error signaling was not linked to explicit expectations about social-evaluation outcomes; modifications to the paradigm’s design may address these limitations and more clearly link differences in prediction-error signaling to social deficits.

The primary limitation of this study is the relatively small sample size, which requires results to be replicated. A secondary limitation is that some participant responses were ironic. Despite instructions to use “Thanks!!! ” to convey sarcasm, participants occasionally used this response to thank their peers for positive social evaluations. Inclusion of a sarcastic response option that cannot be misinterpreted (e.g., “Thanks . . . NOT!!!”) is therefore needed. Another limitation is that there were no group differences in behavioral responding. This is not entirely unexpected given that the lasting effects of behavioral inhibition in infancy and social reticence in childhood are often more evident for brain responses constrained by behavior on an experimental task than in the rate or magnitude of behavioral response on the task (e.g., Bar-Haim et al., 2009; Jarcho, Fox, et al., 2013). However, failure to isolate group differences in behavior, despite differences in the engagement of premotor cortex, suggest that further gradients in valenced response options may be needed to increase task sensitivity. Future studies will determine whether the paradigm elicits behavioral biases among socially anxious youth, whose poor social competence is linked to restricted response flexibility (Crick & Dodge, 1994).

A final limitation is that stronger between-groups differences emerged for brain function than task-based behavior. This raises concerns about “reverse inference.” In fMRI studies, reverse inference occurs when the psychological meaning of between-groups differences in brain function (here, during the anticipation and receipt of social evaluation) are inferred in the absence of behavioral differences (Foland-Ross et al., 2013; Lythe et al., 2015; Singh et al., 2014; Stringaris et al., 2015). However, the presence of between-groups behavioral differences introduces “performance confounds,” which place other limitations on data interpretation (Carter, Heckers, Nichols, Pine, & Strother, 2008; Church, Petersen, & Schlaggar, 2010). Complementary problems of reverse inference and performance confounds underscore the need for two types of fMRI studies: those with group differences in behavior, which address concerns about reverse inference, and those without such differences, which address concerns about performance confounds. This second type of study can make the strongest contribution when experimentally manipulated conditions elicit distinct patterns of behavior. Distinct patterns of behavior can reflect distinct psychological states and may therefore help constrain interpretation of concurrent brain function; this partially addresses the reverse-inference problem. In this study, behavior differed in response to peers with distinct reputations; this constrained our interpretation of brain function because distinct psychological states were mapped to specific experimental conditions. Finally, methodological issues that resulted in collinearity during the participant response phase of the task must be addressed to appropriately model brain function during behavioral responding itself.

These limitations are offset by other strengths. First, large between-groups effects were elicited by brain function engaged by anticipating distinct forms of social evaluation, a nuance not found in most fMRI tasks. Second, although fMRI data were obtained at only one time point, participants were categorized as having high and low social reticence on the basis of longitudinal data obtained during multiple points in childhood. Moreover, brain function will be linked to forthcoming data collected during midadolescence. Finally, studying preadolescents with high early-childhood social reticence, yet free of psychiatric impairment, isolates neural correlates associated with risk from neural correlates associated with the expression of psychopathology.

In conclusion, preadolescents with early-childhood social reticence exhibit distinct patterns of brain function. This may inform attempts to identify and treat underlying risk for social anxiety disorder. The utility of this novel paradigm is not limited to pediatric samples and can be used to study various mechanisms of social cognition across the life span. The social context can be shifted to be relevant for adults (e.g., in the workplace), and peers can be easily modified to embody specific age, gender, weight, ethnic, or affiliative groups. The complex social world is difficult to represent in the confines of an fMRI scanner. The virtual-school paradigm begins to fulfill the critical need for neuroimaging paradigms with high levels of ecological validity.