Skip to main content
NCBI home page
Human Brain Mapping logoLink to Human Brain Mapping
. 2026 May 19;47(8):e70545. doi: 10.1002/hbm.70545

Striatal Prediction Error Tracking Moderates the Influence of Social Exposure on Adolescent Substance Use Curiosity

Jessica S Flannery 1,✉, Elizabeth S Escalante 2, Ruofan Ma 2, Kristen A Lindquist 3, Eva H Telzer 2
PMCID: PMC13184721  PMID: 42153343

ABSTRACT

Adolescence is a period of social development marked by increased sensitivity to social feedback and substance use experimentation. Although reinforcement learning (RL) models provide a powerful framework for examining how individuals learn from experience, they have rarely been applied to understand how adolescents learn from social experiences or how these processes relate to real‐world behavioral outcomes. In a sample of 261 youth (11.0 ± 1.6 years old), we applied computational modeling to a novel social RL fMRI paradigm. Q‐learning models estimated individual differences in learning and trial‐by‐trial prediction errors, which were used as a parametric modulator to assess brain activity that tracked the degree to which social feedback was better or worse than expected. Among older participants (n = 73; 12.9 ± 0.9 years old), we assessed associations between RL metrics and measures of substance use propensity. Greater substance use curiosity and household exposure to substance use were both linked to weaker striatal prediction error tracking of better‐than‐expected outcomes. However, among youth with substance‐using peers, curiosity was associated with elevated striatal prediction error signals and better positive RL performance. Findings suggest that both hypo‐ and hyper‐sensitivity to positive reinforcement learning signals may confer an increased propensity for substance use, possibly through distinct pathways.

Keywords: adolescence, insula, prediction error, reinforcement learning, social, striatum, substance use

Weaker striatal prediction error tracking among adolescents was associated with greater curiosity about substance use. In contrast, striatal tracking also moderated peer influence such that adolescents with stronger social prediction error tracking showed greater curiosity when peers used substances.

graphic file with name HBM-47-e70545-g003.jpg

1. Introduction

Adolescence is marked by heightened sensitivity to social feedback (Blakemore and Mills 2014; Somerville 2013; Smith et al. 2015) and increased experimentation with substances such as alcohol, nicotine, and cannabis (Miech et al. 1975–2023; Sussman et al. 2008; Wagner and Anthony 2002; DeWit et al. 2000; Kandel and Logan 1984). Heightened social sensitivity can facilitate adaptive social learning and relationship building during a time of increasing autonomy (Blakemore and Mills 2014; Telzer et al. 2023; Crone and Dahl 2012). However, it may also increase susceptibility to peer influences toward health risk behaviors, such as substance use, especially when social norms or peer contexts promote such behavior (Steinberg and Monahan 2007; Albert et al. 2013). Although experimentation with substances during adolescence is considered normative (Gray and Squeglia 2018; Siebenbruner et al. 2006), understanding the mechanisms involved in how these behaviors begin and progress could help identify factors that elevate risks for prolonged substance use and its associated harms. Social influence research suggests that peer and familial environments play a critical role in early substance use (Duncan et al. 1995; Pomery et al. 2005; Ashby Wills and Yaeger 2003; Boyle et al. 2001; Dishion and Tipsord 2011; Chassin et al. 1990). Accordingly, understanding how youth learn from social information may help clarify how social exposure translates into substance use curiosity and ultimately early initiation. Prior research has documented the dual effects of social selection (i.e., gravitating toward peers with shared interests or predispositions) (Simons‐Morton 2007), and socialization (i.e., being influenced by exposure, access, peer pressure, and explicit feedback) (Henneberger et al. 2021). Social feedback can positively and negatively reinforce adolescents' behaviors and attitudes. As such, examining sensitivity to social forms of reinforcement may further clarify phenotypes linked to early substance use. Social rewards, such as peer approval, smiles, or inclusion, and social punishments, such as disapproval or exclusion, are particularly salient during adolescence (Kwon et al. 2019; Braams et al. 2015; Chein et al. 2011; Foulkes and Blakemore 2016). Further, individual differences in adolescents' sensitivity to, and learning from, positive and negative social feedback may underlie differential susceptibility to peer and familial influences toward early substance use (Telzer et al. 2021; Turpyn et al. 2021).

Computational reinforcement learning models offer a powerful framework for characterizing how individuals learn from positive and negative feedback over time, providing both behavioral and neural indices of learning mechanisms (Gläscher and O'Doherty 2010). However, relatively few studies have applied reinforcement learning models to understand the social feedback learning mechanisms involved in early substance use. These models can estimate how expectations are updated based on prediction errors, the degree to which outcomes are better or worse than expected (Lerner et al. 2021; Diederen and Fletcher 2021), and can disentangle learning from rewards versus learning from punishments (Frank et al. 2004, 2007; Dombrovski et al. 2019). Dopaminergic signaling within frontostriatal circuits is thought to encode prediction errors (Lerner et al. 2021; Diederen and Fletcher 2021) with increased neuronal firing rates when outcomes are better than expected and decreased firing when outcomes are worse than expected (Schultz 1998). By signaling outcomes based on their predictability, prediction errors support reinforcement learning and thus play a key role in guiding future behavior (Cohen and Frank 2009). Empirical support for this framework comes from both pharmacological (Frank et al. 2004) and genetic studies (Frank et al. 2007) suggesting that sensitivity to reward and punishment, which impacts how people learn from their environments, may be partly rooted in neurobiological variability in frontostriatal functioning.

This framework is particularly relevant for understanding risk for substance use (Keiflin and Janak 2015) as variation in reward‐related brain function has been linked to use propensity (Schneider et al. 2012; Squeglia et al. 2015; Urošević et al. 2015; Flannery et al. 2023; Bart et al. 2021). Supporting a neural risk factor hypothesis, evidence has shown adolescents (Büchel et al. 2017) and young adults (Bart et al. 2021) with relative striatum hypoactivity during non‐drug reward anticipation are more likely to engage in later substance use, potentially as a means of compensating for decreased activation of reward circuits. However, other studies have linked heightened striatal reward responsivity with propensity for early substance use (Radoman et al. 2021; Tervo‐Clemmens et al. 2020; Bjork 2020), suggesting the possibility of multiple risk phenotypes. For instance, blunted reward responses may contribute to elevated substance use curiosity as a compensatory mechanism for reduced reward sensitivity, whereas heightened reward responses may be associated with greater susceptibility to social influences by increasing sensitivity to the perceived social rewards of engaging in risky behaviors with peers (Griffin et al. 2001; Wang et al. 2025). Reinforcement learning models may help reconcile these findings by modeling individual differences in learning signals, rather than relying on broader measures of reward responsiveness (Keiflin and Janak 2015; Tolomeo et al. 2021).

In the present study, participants (N = 261; 11.0 ± 1.6; 54.7% female) completed a probabilistic social reinforcement learning task during a functional magnetic resonance (fMRI) scan, in which feedback took the form of adolescent facial expressions. In the learning phase of the task, participants learned through trial and error which adolescent faces were the “friendliest” (i.e., more likely to deliver positive [smiles] vs. negative [scowls] feedback). Participant responses in the learning phase of the task were fit to reinforcement learning computational models using variational Bayesian methods to obtain individual‐level posterior parameter estimates and trial‐by‐trial prediction errors. Based on prior meta‐analytic work, we focused on prediction error‐related activity in the striatum and anterior insula, regions implicated in prediction error processing in social contexts (Corlett et al. 2022). Among a subsample of older participants (n = 73; 12.9 ± 0.9 years old), we examined whether behavioral and neural indices of feedback learning were associated with electronic‐cigarette (e‐cigarette), cannabis, or alcohol use curiosity and whether this learning moderated the relationship between social exposure to substance use (e.g., household and friend group use) and substance use curiosity.

Consistent with a hypo‐reward responsivity risk hypothesis (Trifilieff and Martinez 2014; Casey et al. 2014; Volkow et al. 2007), we predicted substance use curiosity would be associated with diminished sensitivity to positive social reinforcement and blunted striatal prediction error responses. We further hypothesized that household and friend group substance use would be associated with greater substance use curiosity and this association would be strongest among those with heightened sensitivity to social reinforcement and greater neural responsivity to positive prediction errors.

2. Methods

2.1. Participants and Procedures

Participants (N = 261; 54.7% female at birth; 7.4–14.6 years old) were recruited in 2nd through 8th grade in the southeast United States. Exclusion criteria for fMRI scanning included braces, permanent retainers, claustrophobia, metal in the body, learning disabilities, seizures, and head trauma. The sample was racially and ethnically diverse and varied in socioeconomic status indicators (Table S1). All study procedures were approved by the University of North Carolina at Chapel Hill's Institutional Review Board. Participants provided consent/assent, self‐reported demographic information (verified against parental reports for accuracy), and underwent a 1.5‐h fMRI scan. During the scan, participants completed a series of tasks, including a social reinforcement learning task called the Finding Friendliness Task. A subsample of older participants additionally self‐reported their use of, curiosity about, and household exposure to addictive substances. Older participants also reported frequency of substance use among their five closest friends.

2.2. Social Reinforcement Learning Task (Finding Friendliness Task)

Participants completed two runs of an adapted probabilistic selection task (Frank et al. 2004, 2007), the “Finding Friendliness Task”, used to assess the behavioral and neural processes involved in learning from social feedback. This task includes a ~6‐min learning phase (Figure 1A) and a ~6‐min testing phase (Figure 1B). In the learning phase, participants are presented with pairs of neutral faced adolescents from the National Institute of Mental Health (NIMH) Child Emotional Faces Picture Set (Egger et al. 2011). On each trial, they chose one of the adolescents and then received social feedback from the adolescent they chose: a picture of that adolescent either smiling or scowling. Feedback was presented based on predetermined probabilities (see Figure 1A). Participants learn, through trial and error, which adolescents are the “friendliest”, meaning which adolescents are more likely to smile and less likely to scowl. In the learning phase, participants are instructed to always try to pick the friendliest adolescent. In the testing phase, participants again choose the adolescents whom they learned to be the friendliest, but in this phase, they do not receive feedback.

FIGURE 1.

FIGURE 1

Open in a new tab

Social reinforcement learning task (Finding Friendliness Task). (A) During the learning phase, three different adolescent pairs (AB, CD, EF) are presented 20 times each for a total of 60 learning phase trials. Through ‘trial and error’ participants learn to choose which adolescent is the “friendliest” based on probabilistic positive and negative social feedback. Positive feedback is a smiling picture of the adolescent they chose while negative feedback is a scowling picture of the adolescent they choose. In AB trials, adolescent A smiles 80% of the time whereas adolescent B smiles 20% of the time. CD and EF pairs are less reliable, such that adolescent C smiles on 70% of CD trials and adolescent D smiles on 30% of CD trials and adolescent E smiles on 60% of EF trials and adolescent F smiles on 40% of EF trials. Participants learn to choose adolescent A, adolescent C, and adolescent E more often than B, D, or F. Adolescent pairs are neutral faced pictures from The NIMH Child Emotional Faces Picture Set (Egger et al. 2011). Each time the task is started, six adolescent stimuli (3 girls and 3 boys) are randomly assigned as A–F, with their respective probabilistic outcomes, to minimize stimulus‐specific effects. Each trial in the learning phase consists of a choice screen in which an adolescent pair is presented, and participants use their left and right index fingers to pick one of the neutral adolescent faces (FRIENDLY?: 2500 msec). Each adolescent is presented on the left and right side an equal number of times to account for any screen side biases. Following a fixation screen (jittered duration: 1000–3200 msec), participants receive social feedback from the adolescent they chose based on the trial's probabilistic contingencies (FEEDBACK: 1000 msec). If participants did not choose quickly enough, they receive a “too slow” message. (B) During the testing block, novel combinations of adolescent pairs that include either adolescent A (Approach A: AC, AD, AE, AF) or adolescent B (Avoid B: BC, BD, BE, BF) are presented and no feedback is provided. In total, there are 24 Approach A trials, 24 Avoid B trials and 24 Control trials in which C, D, E and F adolescents are paired with each other and A and B adolescents are not choice options (Control: CE, CF, DE, DF). Each trial in the testing phase consists of presenting a pair of neutral faces (FRIENDLY?: 2500 msec) and the participants pick one of the faces. However, in the testing phase, no feedback is provided such that the next trial begins following a fixation screen (jittered duration: 1000–3200 msec).

Performance in the testing phase was used to separately assess the ability to learn from positive social feedback (smiles) and negative social feedback (scowls). Positive reinforcement learning performance was operationalized as the ability to approach adolescent A, the adolescent with the highest probability of providing a smiling outcome (80% of the trials) during the learning phase. Negative reinforcement learning performance was operationalized as the ability to avoid choosing adolescent B, the adolescent with the highest probability of providing a scowling outcome (80% of trials) during learning.

To assess behavior in the learning phase, we calculated the proportion of trials in which the participant made the “optimal” choice (chose the adolescent with the higher probability of a positive outcome) for each trial type (i.e., AB, CD, EF), and across trial types. To estimate each participant's positive and negative social feedback learning rates, as well as choice stochasticity, behavior during the learning phase was fit to an asymmetrical Q‐learning model (Frank et al. 2004, 2007; Frydecka et al. 2016). This computational modeling approach estimated each participant's trial‐specific expected value (i.e., expected probability of receiving a smile), as well as the corresponding prediction error (i.e., difference between expected and actual feedback), which was used as a parametric modulator in fMRI analyses.

Nine participants were excluded from behavioral analyses due to learning phase responses not meeting quality control thresholds (> 60% no responses or > 80% screen side bias). In addition to these nine, 14 more participants were excluded from fMRI analyses due to fMRI data loss or high motion. Further, 16 participants' testing phase behavioral data were excluded from behavioral analyses having not met the same quality control cutoffs during this run.

2.3. Q‐Learning Models

We characterized underlying components of participants' decision‐making in the task using a model selected from a family of established Q‐learning algorithms fit to participants' trial‐by‐trial responses (variational Bayesian analysis (VBA) toolbox, MATLAB (Daunizeau et al. 2014)). This approach allowed us to (1) estimate model parameters that represent distinct aspects of reinforcement learning, (2) examine individual differences in such parameters, and (3) investigate how trial‐by‐trial variation in prediction errors relates to brain responsivity during the task. Ultimately, an asymmetric two‐learning rate Q‐learning model, used in prior implementations of probabilistic selection tasks (Frank et al. 2004, 2007), was selected and estimated model parameters were extracted for further analyses. The Q value or expected friendliness of each adolescent and the corresponding prediction error (difference between the expected value of the chosen stimulus and feedback) was tracked across trials. Further details on model features, fitting, and selection can be found in the supplement (Figure S1). Forward simulations for each participant, based on their best fit model parameters, validated model selection, showing that parameters from the selected model generate simulated data that reproduces key aspects of actual behavioral performance (see Figure S2 for details).

2.4. MRI Data Processing and Analysis

Acquisition and preprocessing of all imaging data is reported in the supplement. Following pre‐processing, participant‐level general linear models (GLM; AFNI's 3dDeconvolve and 3dREMLfit) were conducted for the learning run of the Finding Friendliness task and included the following 16 nuisance regressors: cerebral spinal fluid (CSF) and white matter (WM) signal and their temporal derivatives, as well as six head‐motion estimates and their temporal derivatives. Timepoints with a framewise displacement exceeding 0.9mm3 were censored. Participants with greater than 25% of timepoints censored were excluded from fMRI analyses. On average, 2.8% ± 4.5% of TRs were censored among participants that were included.

Three task‐related regressors were included: choice, feedback, and no response. We used a parametric modulator to examine brain activity that tracked increasing prediction error for each feedback presentation. Trial‐by‐trial prediction error was calculated for each participant by subtracting the current expected value of their choice (a number between −1 and 1) from the actual feedback they received (1 for positive feedback and −1 for negative feedback) such that prediction error was always greater than 0 for positive feedback and less than 0 for negative feedback. Increasingly positive prediction error values represent the degree to which the outcome was better than expected, whereas decreasingly negative prediction error values represent the degree to which the outcome was worse than expected.

Feedback and no response regressors were modeled as impulse response functions while choice was modeled as a duration modulated stimulus based on participant response times. The main effect of prediction error was assessed (AFNI's 3dMEMA). We also considered whole brain activity that significantly tracked prediction error for positive and negative feedback separately. Whole‐brain statistical maps are available on NeuroVault. Average prediction error‐related brain activity during feedback presentation was extracted from a priori ROIs: bilateral striatum and anterior insula, defined via the Brainetome Atlas (Fan et al. 2016).

2.5. Substance Use Measures

A subset of older participants (n = 73; ages:10.9–14.6, 52.1% female at birth) reported the degree to which they have ever been curious about using an e‐cigarette, using cannabis, or drinking alcohol on a 6‐point Likert scale ranging from “never” to “very curious”. We considered curiosity ratings for each substance as well as the participants' maximum curiosity rating across all three substances. Participants additionally reported whether they had ever tried the substances and how many of the adults (e.g., parents, grandparents, stepparents, aunts, uncles, other caregivers) and children/teenagers (e.g., siblings, stepsiblings, cousins, foster siblings) currently living in their home use the substances on four‐point scales from “none” to “all”. We averaged across household use ratings to get an approximate measure of how many people living in the participant's household use substances. Participants also reported how frequently their five closest friends use each substance on a four‐point scale ranging from “never” to “always”.

2.6. Analyses

We began by examining social reinforcement learning behavior in the Finding Friendliness task. We characterized task performance based on learning phase behavior: proportion of trials in which the participant made the optimal choice (i.e., chose the adolescent with the higher probability of a positive outcome), and testing phase behavior: proportion of correct approach and avoid trials. Next, to identify learning parameters (positive learning rate [α₊], negative learning rate [α₋], and inverse temperature [β]) that were linked to better overall task performance, we built a natural‐neighbor surface from observed data and defined the optimum as the surface maximum of task performance: a composite of learning phase and testing phase accuracy.

Then we characterized whole‐brain activity that tracked trial‐by‐trial prediction error values during feedback delivery and examined associations between behavioral task performance and average prediction error tracking in a priori ROIs. We then conducted Pearson correlations to assess relationships between substance use curiosity and social exposure to substance use (i.e., household and friend group use). Spearman's rho was applied for skewed variables. Subsequently, we examined links between substance use propensity measures and social reinforcement learning behavior, as well as prediction error‐related brain activity (Pearson correlations). One striatal prediction error activity outlier was observed among the older sub‐sample. However, this value was not an outlier among the full sample, suggesting that it did not necessarily reflect measurement error. As such, we ran all sub‐sample analyses that included this variable with both raw values and with the outlier winsorized.

Finally, to examine whether individual differences in reinforcement learning processes moderated the relationship between substance use exposure and substance use curiosity, we conducted two multiple moderation analyses using PROCESS (Version 5.0; Model 2; (Hayes 2017)). In analyses examining learning from positive social feedback, the predictor variable was exposure to substance use (peer or household exposure), and the outcome variable was substance use curiosity. To examine whether neural and behavioral reinforcement‐learning processes jointly influenced the association between exposure and curiosity, two moderators were simultaneously included in the model: striatal prediction error activity during the learning phase and behavioral approach performance during the test phase. Including both moderators allowed the unique moderating effect of each variable to be evaluated while accounting for the other.

Another multiple moderation model, examining learning from negative social feedback, tested whether insular tracking of worse‐than‐expected outcomes and avoidance learning moderated the association between substance use exposure and curiosity. For both moderation analyses, friend group substance use was treated as a binary variable (friend use reported vs. not reported) to address its skewed distribution. Additional follow‐up moderation analyses were conducted in which each moderator was examined independently to confirm that the observed moderation effects were not dependent on the simultaneous inclusion of both moderators in the model.

3. Results

3.1. Prediction Error‐Related Brain Activity and Relationships With Social Learning Behavior

A parametric modulator was used to examine sample average brain activity that tracked prediction error during feedback (degree to which feedback was better than expected vs. worse than expected). Trial‐by‐trial prediction error was calculated for each participant by subtracting the current expected value of their choice (a number between −1 and 1) from the actual feedback they received (1 for positive feedback and −1 for negative feedback) such that prediction error was always greater than 0 for positive feedback and less than 0 for negative feedback. Brain activity that significantly increased as prediction error increased was observed in the bilateral putamen, nucleus accumbens, ventral medial prefrontal cortex, and paracentral gyrus, as well as bilateral clusters in the fusiform face area and occipital lobe (Figure 2A). Brain activity in the bilateral supplementary motor area, bilateral anterior insula, bilateral temporal pole, and right superior temporal gyrus, as well as an epithalamic cluster encompassing the habenula, was negatively correlated with prediction error values, such that activity increased as outcomes were increasingly worse‐than‐expected. Observed prediction error tracking during feedback was largely consistent with prior work (Corlett et al. 2022; Colas et al. 2017; Cohen 2007). To further characterize the valence‐related specificity of these effects, we examined whole‐brain prediction error tracking separately for positive and negative feedback (Figure S3). During positive feedback, activity in the bilateral anterior insula, amygdala, thalamus, midbrain, and occipital lobe increased as feedback was increasingly better‐than‐expected. During negative feedback, activity in the bilateral putamen, supplementary motor area, and occipital lobe decreased as feedback was increasingly worse‐than‐expected (cluster coordinates reported in Table S2). These analyses may suggest that the anterior insula encodes deviations from expectations regardless of valence, but with a bias toward worse‐than‐expected outcomes.

FIGURE 2.

FIGURE 2

Open in a new tab

Prediction error‐related brain activity and associations with task performance. (A) Sample average whole brain maps of activity tracking prediction error (PE) value (degree to which feedback was better than expected vs. worse than expected). As PE increased (warm colors), brain activity increased in the bilateral striatum, including the putamen and nucleus accumbens, the ventral medial prefrontal cortex, the paracentral gyrus and in bilateral clusters in the fusiform face area and occipital lobe. As PE values decreased (cool colors), brain activity increased in the bilateral anterior insula, supplementary motor area, temporal pole, and right superior temporal gyrus, as well as an epithalamic cluster encompassing the habenula. Unthresholded statistical brain maps available at https://neurovault.org/collections/21902/. (B1) Stronger tracking of worse‐than‐expected outcomes in an atlas‐defined bilateral anterior insula mask was significantly associated with a higher percentage of AB trials in which the participant selected the “friendliest” adolescent (% Optimal Choice). However, approach A (App A) and avoid B performance in the testing phase were not significantly associated with PE‐related insula activity. (B2) PE‐related activity in the bilateral striatum was not significantly associated with learning phase or testing phase performance. Cluster coordinates are reported in Table S3.

Next, we examined relationships between task behavior and prediction error‐related brain activity from the primary analysis within atlas‐defined a priori regions of interest (ROIs): the bilateral striatum and anterior insula. Sample average social learning behavioral performance is reported in the supplement (Figure S4). Stronger anterior insula tracking of worse‐than‐expected outcomes was associated with selecting the friendliest adolescent (over the less friendly adolescent) on a greater proportion of AB trials (r[236] = −0.2, p = 0.010). However, insula prediction error tracking was not significantly associated with learning phase choices across all trial types (r[236] = −0.1, p = 0.068) or testing phase performance (approach: r[234] = −0.1, p = 0.123; avoid: r[234] = −0.04, p = 0.574; Figure 2 B1). Contrary to expectations, prediction error‐related striatal activity was not significantly associated with any task performance (p's > 0.5; Figure 2 B2). These results did not change when controlling for age.

3.2. Substance Use Curiosity, Household Use, and Friend Group Use

Overall, 27.4% of the sample reported being at least a little curious about using e‐cigarettes, cannabis, or alcohol, and 16.4% reported that they had tried one of the substances. Substance use curiosity was significantly higher among participants that had tried at least one of the substances (t [71] = −3.0, p = 0.012). 71.2% of the sample reported that at least some of the individuals living in their household use e‐cigarettes, cannabis, or alcohol, and 12.3% reported substance use among their five closest friends. We examined relationships between social exposure to each substance and curiosity about using that substance. Household e‐cigarette use (r [71] = 0.6, p < 0.001) was positively associated with participants' e‐cigarette curiosity, but friend group e‐cigarette use was not (ρ [70] = 0.2, p = 0.1; Table S4). Household cannabis use (r [71] = 0.3, p = 0.025) and friend group cannabis use (ρ [70] = 0.3, p = 0.017) were both positively associated with participants' cannabis curiosity. However, both household alcohol use and friend group alcohol use were not significantly associated with alcohol curiosity. We did not observe any significant age or sex effects on substance use curiosity, household use, or friend group use in this sample.

3.3. Associations Between Prediction Error‐Related Brain Activity and Substance Use Propensity Measures

More substance using people living in the participant's household (r [70] = −0.4, p = 0.001) and higher curiosity for using substances (r [70] = −0.3, p = 0.010) were significantly associated with weaker tracking of better‐than‐expected outcomes in the atlas‐defined bilateral striatum ROI. However, friend substance use was not (ρ [70] = 0.5, p = 0.687). To confirm associations were not driven by an outlier, analyses were re‐run with the outlier winsorized and associations with household substance use (r [70] = −0.4, p = 0.001) and curiosity (r [70] = −0.3, p = 0.016) remained significant (Figure 3A). Stronger prediction error tracking of worse‐than‐expected outcomes in the anterior insula was also significantly associated with a higher amount of household substance users (r [70] = −0.3, p = 0.033), but not with friend group substance use, or substance use curiosity (Figure 3B). Further, given that voxel‐wise analyses also identified prediction error tracking in the ventromedial prefrontal cortex (vmPFC), a region implicated in corticostriatal circuits (Haber 2016) and thought to play a key role in value representation (Roy et al. 2012), we additionally conducted post hoc exploratory analyses of prediction error tracking within vmPFC subregions. These analyses revealed a negative correlation between anterior vmPFC tracking of better‐than‐expected outcomes and substance use curiosity (r [70] = −0.4, p = 0.001; details in the supplement Figure S5).

FIGURE 3.

FIGURE 3

Open in a new tab

Associations between prediction error‐related brain activity and substance use propensity measures. (A) Lower prediction error (PE) related activity extracted from an atlas‐defined bilateral striatum mask was significantly associated with more substance‐using people living in the participant's household and higher curiosity for using substances. (B) Stronger anterior (ant.) insula tracking of worse‐than‐expected outcomes was also significantly associated with a higher amount of household substance users but not substance use curiosity. Data points jittered for visualization.

Task performance and learning rates were not significantly related to general household or friend group substance use or substance use curiosity. These results did not significantly change when controlling for age. Although in exploratory follow‐up analyses considering each substance separately, higher household alcohol use exposure was significantly associated with higher avoidance performance during testing (r [62] = 0.3, p = 0.010).

3.4. Social Reinforcement Learning and Brain Activity Moderate the Effect of Friend Group Substance Use on Curiosity

The multiple moderation model examining whether learning from positive social feedback moderated the association between friend group substance use exposure and substance use curiosity was significant and explained approximately 30% of the variance in substance use curiosity (R 2 = 0.30, F [5, 63] = 5.4, p = 0.0003; Figure 4A). When holding approach learning performance at the mean, striatal prediction error activity significantly moderated the association between friend group substance use and curiosity (t = 3.0, p = 0.004). Among individuals with stronger striatal tracking of better‐than‐expected outcomes, those with substance‐using friend groups reported significantly greater curiosity (+1 SD; t = 3.2, p = 0.002), but this same effect of friend use was not seen among those with weaker prediction error striatum tracking (Figure 4B). A similar moderation effect was observed for behavioral approach learning performance (t = 2.2, p = 0.031) when holding striatal prediction error activity at the mean, friend use predicted greater curiosity only among adolescents with higher approach learning performance (+1 SD; t = 2.6, p = 0.012) but not among those with average or low approach learning (Figure 4C). The effect of friend group substance use on curiosity was further amplified among adolescents showing both elevated striatal prediction error tracking and higher approach learning performance. Specifically, among adolescents high in both moderators (+1 SD), friend use was even more strongly associated with increased substance use curiosity (t = 4.1, p < 0.001). Follow‐up analyses, testing each moderator independently, produced similar results. Further, in exploratory analyses prediction‐error tracking of better‐than‐expected outcomes in the vmPFC also moderated the effect of friend group use on curiosity (details in the supplement).

FIGURE 4.

FIGURE 4

Open in a new tab

Social reinforcement learning and brain activity moderate the effect of friend group substance use on substance use curiosity. (A) Social reinforcement learning mechanisms significantly moderate the effect of friend group substance use on substance use curiosity (PROCESS v5.0; Model 2). For each analysis, friend group substance use was entered as a binary independent variable and substance use curiosity was the dependent variable. Two moderators were included simultaneously: Striatal prediction error (PE) activity during the learning phase and behavioral approach learning performance during the test phase. Conditional effects of friend group substance use were probed at three levels of the moderator (−1 standard deviation [SD], mean, and +1 SD). Clustered bar plots display predicted values of substance use curiosity by friend group use at each level of the moderator, with error bars representing the standard error of the model‐estimated conditional effects. (B) Striatal prediction error activity significantly moderated the association between friend group substance use exposure and curiosity. When approach learning performance was held at the mean, friend group substance use predicted greater curiosity only among adolescents showing strong striatal tracking of better‐than‐expected outcomes (+1 SD), but not among those with average or low striatal PE tracking. (C) Behavioral approach learning performance during the test phase also significantly moderated this association. When striatal PE activity was held at the mean, friend group substance use predicted greater curiosity only among adolescents with higher approach learning performance (+1 SD), but not among those with average or low performance. Icon credit: Flaticon/Eucalyp, Flaticon/Skyclick.

The multiple moderation analysis examining whether learning from negative social feedback moderated the association between friend group substance use and curiosity was not significant (R 2 = 0.07, F [3, 67] = 1.7, p = 0.19). Neither anterior insula prediction error tracking nor avoidance learning moderated the effect of friend group substance use on curiosity. Further, none of the learning metrics significantly moderated the effect of household substance use on curiosity (p's > 0.05). Moderation results were robust to heteroskedasticity, with all effects remaining significant when re‐estimated with heteroskedasticity‐consistent standard errors. Further, results did not significantly change when controlling for age.

4. Discussion

We utilized a novel social reinforcement learning task to investigate how adolescents learn from positive and negative social feedback. We explored how these learning processes relate to real‐world outcomes, including substance use curiosity and exposure to household and peer group use. By integrating task behavior, prediction error‐related brain activity, and self‐reported substance use propensity measures, this study shows individual differences in reinforcement learning mechanisms that may underlie susceptibility to social influences and substance use curiosity.

4.1. Prediction Error‐Related Brain Activity and Relationships With Social Learning Behavior

Theory suggests that learning occurs when an outcome differs from expectations. The magnitude and valence of this difference can be quantified as a prediction error (Fouragnan et al. 2018). Consistent with prior work (Schultz 1998; O'Doherty et al. 2003; Daw et al. 2011; Preuschoff et al. 2008; Palminteri et al. 2012), we observed that ventral striatal activity was positively correlated with prediction error values, such that activity increased as outcomes were increasingly better than expected, whereas anterior insula activity was negatively correlated with prediction error values, such that activity increased as outcomes were increasingly worse than expected.

Like prior work (Frömer et al. 2021; Lohse et al. 2020; Hester et al. 2009), individual differences in prediction error–related neural responsivity during learning were not associated with test‐phase performance. However, attenuated insula tracking of worse‐than‐expected outcomes was associated with poorer learning‐phase performance. This dissociation suggests that variability in prediction error tracking may influence ongoing adjustment of choice behavior during acquisition without necessarily producing enduring differences in value representations that persist into the test phase.

Notably, this relationship with learning phase behavior was specific to the anterior insula. Striatal prediction error tracking was not associated with learning‐phase or test‐phase behavior. One possible interpretation of these findings may be that the anterior insula and striatum contribute differently to reinforcement learning. Prior work suggests that anterior insula activity is particularly sensitive to the salience of negative outcomes (e.g., (Jensen et al. 2007; Ullsperger et al. 2010)), whereas striatal prediction error signals are thought to be more involved in value‐updating processes (Jensen et al. 2007; Lejarraga and Hertwig 2017). Signals reflecting the salience of negative outcomes may therefore exert a more immediate influence on trial‐by‐trial behavioral adjustments during learning, whereas variability in striatal value‐updating processes may not always translate into observable differences in task performance.

Developmental findings suggest that this distinction may be especially relevant during adolescence. For example, adolescents have displayed stronger anterior insula tracking of social prediction‐errors (Jones et al. 2014) but reduced coupling between striatal prediction‐error signals and behavioral performance (Christakou et al. 2013) when compared to adults. Thus, the present findings in a relatively young sample, together with prior work, raise the possibility that insula‐mediated prediction‐error signals may play a particularly important role in guiding behavioral adjustment during social learning in adolescence, whereas the coupling between striatal value signals and task behavior may still be strengthening during this developmental period. However, as the present study was not designed to directly test age‐related differences, additional work will be needed to characterize prediction‐error tracking changes across development.

4.2. Substance Use Curiosity, Household Use, and Peer Exposure

Our findings support longstanding social influence models of adolescent substance use (Henneberger et al. 2021; Telzer et al. 2017; Graham et al. 1991), demonstrating that youth curiosity about substance use is positively associated with both household (Pomery et al. 2005; Ashby Wills and Yaeger 2003; Merikangas et al. 1998) and friend group use (Reed and Rountree 1997). While this cross‐sectional study design cannot establish causality, these observed associations align with prior work and suggest that, when peers engage in substance use, mechanisms of social influence, including social selection and socialization (Simons‐Morton 2007; Henneberger et al. 2021), may lead to higher substance use curiosity. Further, as environmental exposure and heritable vulnerability have been shown to independently contribute to substance use (Henneberger et al. 2021), the observed link between household use and substance use curiosity may also be due to multiple interacting mechanisms.

4.3. Associations Between Prediction Error‐Related Brain Activity and Substance Use Propensity Measures

Our results support prior work linking reduced striatal responsivity during reinforcement processing to substance use risk (Flannery et al. 2023, 2019; Bart et al. 2021; Büchel et al. 2017; Peters et al. 2011). We found that adolescents with weaker positive prediction error‐related striatal and vmPFC tracking reported higher curiosity about using substances. One interpretation of this finding is that diminished reward responsivity increases vulnerability to substance use by making especially potent reinforcers, like addictive substances, more appealing. While blunted reward responses have been associated with early substance use in adolescence (Flannery et al. 2023; Büchel et al. 2017; Peters et al. 2011), prediction error tracking has not previously been examined as an early marker of use propensity. Extending prior evidence from substance‐dependent adults showing weaker striatal prediction error tracking (Tolomeo et al. 2021; Tanabe et al. 2013), our findings link this neural signature to substance use propensity among youth. Addictive substances are thought to appropriate the brain's natural learning mechanisms (Redish 2004) and repeated drug exposure has been linked to disrupted learning‐related neural signals that are hypothesized to perpetuate drug‐taking (Monterosso et al. 2012; Takahashi et al. 2007). Our findings suggest that similar disruptions, previously observed among substance dependent individuals (Tanabe et al. 2013), may already be present among youth with elevated use propensity, even before heavy or chronic use begins.

Further, as more substance‐using people living in the participant's household were also significantly associated with weaker striatal tracking of better‐than‐expected outcomes, but peer group substance use was not, blunted striatal responsivity to unexpectedly positive outcomes may represent a preexisting phenotype, potentially shaped by inherited factors (Stokes et al. 2013) and/or early rearing environment (Kamkar et al. 2017; Hanson et al. 2021; Robinson et al. 2013). However, more research is needed to parse genetic and environmental contributions to striatal tracking of prediction errors and implications for substance use propensity. Overall, measures of substance use propensity were related to individual differences in neural signals of social learning in the absence of differential learning behavior.

4.4. Social Reinforcement Learning Mechanisms Moderate the Effect of Friend Substance Use on Use Curiosity

Weaker striatal and vmPFC tracking of better‐than‐expected social reinforcement was generally associated with higher curiosity for using substances. However, among adolescents who reported substance use among their five closest friends, the opposite pattern was observed. These findings suggest that variability in striatal and vmPFC reinforcement learning signals may be associated with substance use risk through multiple, mechanisms. Specifically, blunted neural tracking of better‐than‐expected outcomes may contribute to elevated substance use curiosity because reduced reward sensitivity may increase the tendency to seek out more potent alternative rewards. Whereas heightened tracking of better‐than‐expected social outcomes may indicate an increased susceptibility to social influences due to greater sensitivity to social rewards. Supporting this interpretation, participants who demonstrated task behavior indicative of better learning from positive social feedback (i.e., greater approach learning performance) also showed a greater positive effect of friend group use on substance use curiosity. Our findings may offer preliminary evidence for a subgroup of adolescents that is particularly sensitive to positive social reinforcement and, consequently, is more vulnerable to peer influences toward substance use. Alternatively, this heightened susceptibility could also reflect a general sensitivity to reinforcement which, given the heightened motivational salience of social contexts during this developmental period, could also amplify the influence of peers during adolescence. Interestingly, individual differences in social reinforcement learning did not moderate the influence of household substance use, suggesting that mechanisms probed by our social reinforcement learning task may be more relevant to peer‐based, rather than familial, social learning. Overall, these findings suggest the complex ways social reinforcement processing may influence early substance use risk.

4.5. Limitations

Several limitations of the current study should be noted. First, our measures of household substance use cannot disentangle effects of social learning from inherited predispositions, nor can our measure of peer substance use fully distinguish effects of socialization from peer selection. Second, while our focus on prediction error‐related activity in the striatum and anterior insula was theory‐driven, activity in other brain regions likely interacts with these regions and contributes to reinforcement learning. Further, while the social reinforcement delivered in the Finding Friendliness Task is not as quantifiable as monetary reinforcement, such ambiguity may better replicate how social feedback is processed in real‐world environments. Next, although participants were instructed to always try to select the higher‐probability option during the learning phase, the present task design cannot distinguish whether suboptimal responding during this phase reflects differences in learning, purposeful exploratory sampling, or choice stochasticity. Future studies using paradigms designed to dissociate exploration from exploitation might help clarify the mechanisms underlying individual differences in learning‐phase choice behavior. Furthermore, substance use data were available only for a smaller subset of the sample, and lastly, the distribution of ages represented in the sample was not sufficient for age‐based comparisons.

4.6. Conclusions

This study suggests that adolescents with weaker striatal reinforcement learning signals may be interested in seeking out more potent rewards like substances, whereas adolescents with relatively higher neural responsivity to positive social feedback may be more susceptible to peer influences. This dual‐risk framework may help shed light on prior research that has linked both hyper‐ and hypo‐responsivity in reward‐related brain systems to substance use vulnerability (Büchel et al. 2017; Tervo‐Clemmens et al. 2020; Hammond et al. 2015). Further, these results may suggest the value of tailoring prevention strategies to specific neurocognitive profiles. For instance, interventions that emphasize peer resistance skills may be especially useful for youth who are highly sensitive to social rewards. Whereas those with reduced reward responsivity may instead need more help finding other sources of reinforcement.

Funding

This work was supported by the National Institute on Drug Abuse, DA051127.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Demographic information.

Table S2: Whole brain social reinforcement prediction error tracking during positive and negative feedback cluster coordinates. Sample average whole‐brain activity during positive and negative feedback presentation that was significantly associated with trial‐specific prediction error values. All subjects prediction error related images (n = 238) were entered into whole‐brain one‐sample t‐tests for positive and negative feedback (3dMEMA, p voxel‐level = 0.001, p cluster‐level< 0.05, cluster extent threshold = 56 voxels for positive feedback and 51 voxels for negative feedback).

Table S3: Whole brain social reinforcement prediction error tracking cluster coordinates. Sample average whole‐brain activity during feedback presentation that was significantly positively, or negatively, associated with trial‐specific prediction error values. All subjects’ prediction error related images (n = 238) were entered into a whole‐brain one‐sample t‐test 3dMEMA, p voxel‐level< 0.1e −7, p cluster‐level< 0.05, cluster extent threshold = 10 voxels.

Table S4: Substance use curiosity and exposure correlation table.

Figure S1: Q‐Learning model comparison and selection. Random‐effects Bayesian model comparison across 12 candidate models. Bars show posterior model probability r(sorted). Labels indicate exceedance probability (EP). The asymmetric Q‐learning (Base) model (AsymBase) has the highest rand exceedance probability (EP) compared to all other models (Bayesian Omnibus Risk [BOR] = 0.001) and was selected as the best fit model. The AsymBase model had an EP of 1.0 indicating strong evidence that the selected model is the best across participants.

Figure S2: Simulation‐based validation of selected asymmetric Q‐learning model. Forward simulation was conducted using each participant's posterior mean parameters (learning rates and softmax inverse‐temperature). (A) Pearson correlations between actual and simulated performance per pair. (B) Visualized correspondence between simulated and actual learning curves for each stimuli pair across trials.

Figure S3: Whole brain prediction error activity for positive and negative feedback. We considered whole brain activity that significantly tracked prediction error for positive and negative feedback separately. (A) During positive feedback, activity in the bilateral anterior insula, amygdala, thalamus, midbrain, and occipital lobe increased as feedback was increasing better than expected. (B) During negative feedback, activity in the bilateral putamen, supplementary motor area, and occipital lobe decreased as feedback was increasingly worse than expected. Warm colors denote activity with a significant positive association with prediction error values and cool colors denote activity with a significant negative association with prediction error values. Cluster coordinates reported in Table S2.

Figure S4: Social reinforcement learning and approach/avoidance performance. (A) On average participants’ performance decreased as difficulty level of stimuli pairs increased such that participants made the optimal choice on 68.0% ± 18.2% of AB trials, 63.8% ± 19.4% of CD trials and 57.1% ± 19.7% of EF trials. (B) Across trials in the learning phase, participants’ expected friendliness of adolescent A (QA) increased whereas the expected friendliness of adolescent B (QB) decreased. (C) In the testing phase, participants displayed better negative social reinforcement learning ability (percent correct on avoid B trials; 70.1% ± 19.8%) than positive social reinforcement learning ability (percent correct on approach A trials; 60.5% ± 23.2%). Approach and avoidance learning performance were not significantly associated. (D) On average, learning rates for positive feedback (0.34 ± 0.2) were significantly lower than learning rates for negative feedback (0.56 ± 0.2). Positive and negative learning rates were negatively correlated.

Figure S5: vmPFC prediction error (PE) tracking and associations with substance use measures. (A) Ventromedial prefrontal cortex (vmPFC) regions of interest defined from an independent meta‐analysis of subjective value representation, including central (cvmPFC; red), ventral (vvmPFC; green), and anterior (avmPFC; blue) subregions (9‐mm radius spheres). (B) Prediction error (PE)‐tracking of better‐than‐expected outcomes in the anterior vmPFC is significantly negatively associated with substance use curiosity. (C) PE‐tracking in the anterior vmPFC was not significantly associated with household substance use. PE‐tracking in ventral and central vmPFC subregions was not associated curiosity and exposure (not shown).

HBM-47-e70545-s001.docx (13.7MB, docx)

Data Availability Statement

The authors have released code and tabular data associated with this manuscript on GitHub (https://github.com/Flanneryg3/FF_ProjectCode), and unthresholded brain maps are available on NeuroVault (https://neurovault.org/collections/21902/).

References

  1. Albert, D. , Chein J., and Steinberg L.. 2013. “The Teenage Brain: Peer Influences on Adolescent Decision Making.” Current Directions in Psychological Science 22, no. 2: 114–120. 10.1177/0963721412471347. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Ashby Wills, T. , and Yaeger A. M.. 2003. “Family Factors and Adolescent Substance Use: Models and Mechanisms.” Current Directions in Psychological Science 12, no. 6: 222–226. 10.1046/j.0963-7214.2003.01266.x. [DOI] [Google Scholar]
  3. Bart, C. P. , Nusslock R., Ng T. H., et al. 2021. “Decreased Reward‐Related Brain Function Prospectively Predicts Increased Substance Use.” Journal of Abnormal Psychology 130, no. 8: 886–898. 10.1037/abn0000711. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Bjork, J. M. 2020. “The Ups and Downs of Relating Nondrug Reward Activation to Substance Use Risk in Adolescents.” Current Addiction Reports 7, no. 3: 421–429. 10.1007/s40429-020-00327-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Blakemore, S. J. , and Mills K. L.. 2014. “Is Adolescence a Sensitive Period for Sociocultural Processing?” Annual Review of Psychology 65, no. Volume 65, 2014: 187–207. 10.1146/annurev-psych-010213-115202. [DOI] [PubMed] [Google Scholar]
  6. Boyle, M. H. , Sanford M., Szatmari P., Merikangas K., and Offord D. R.. 2001. “Familial Influences on Substance Use by Adolescents and Young Adults.” Canadian Journal of Public Health 92, no. 3: 206–209. 10.1007/BF03404307. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Braams, B. R. , Van Duijvenvoorde A. C. K., Peper J. S., and Crone E. A.. 2015. “Longitudinal Changes in Adolescent Risk‐Taking: A Comprehensive Study of Neural Responses to Rewards, Pubertal Development, and Risk‐Taking Behavior.” Journal of Neuroscience 35, no. 18: 7226–7238. 10.1523/JNEUROSCI.4764-14.2015. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Büchel, C. , Peters J., Banaschewski T., et al. 2017. “Blunted Ventral Striatal Responses to Anticipated Rewards Foreshadow Problematic Drug Use in Novelty‐Seeking Adolescents.” Nature Communications 8, no. 1: 1. 10.1038/ncomms14140. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Casey, K. F. , Benkelfat C., Cherkasova M. V., Baker G. B., Dagher A., and Leyton M.. 2014. “Reduced Dopamine Response to Amphetamine in Subjects at Ultra‐High Risk for Addiction.” Biological Psychiatry 76, no. 1: 23–30. 10.1016/j.biopsych.2013.08.033. [DOI] [PubMed] [Google Scholar]
  10. Chassin, L. , Presson C. C., Sherman S. J., and Edwards D. A.. 1990. “The Natural History of Cigarette Smoking: Predicting Young‐Adult Smoking Outcomes From Adolescent Smoking Patterns.” Health Psychology 9, no. 6: 701–716. 10.1037/0278-6133.9.6.701. [DOI] [PubMed] [Google Scholar]
  11. Chein, J. , Albert D., O'Brien L., Uckert K., and Steinberg L.. 2011. “Peers Increase Adolescent Risk Taking by Enhancing Activity in the Brain's Reward Circuitry.” Developmental Science 14, no. 2: F1–F10. 10.1111/j.1467-7687.2010.01035.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Christakou, A. , Gershman S. J., Niv Y., Simmons A., Brammer M., and Rubia K.. 2013. “Neural and Psychological Maturation of Decision‐Making in Adolescence and Young Adulthood.” Journal of Cognitive Neuroscience 25, no. 11: 1807–1823. 10.1162/jocn_a_00447. [DOI] [PubMed] [Google Scholar]
  13. Cohen, M. X. 2007. “Individual Differences and the Neural Representations of Reward Expectation and Reward Prediction Error.” Social Cognitive and Affective Neuroscience 2, no. 1: 20–30. 10.1093/scan/nsl021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Cohen, M. X. , and Frank M. J.. 2009. “Neurocomputational Models of Basal Ganglia Function in Learning, Memory and Choice.” Behavioural Brain Research 199, no. 1: 141–156. 10.1016/j.bbr.2008.09.029. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Colas, J. T. , Pauli W. M., Larsen T., Tyszka J. M., and O'Doherty J. P.. 2017. “Distinct Prediction Errors in Mesostriatal Circuits of the Human Brain Mediate Learning About the Values of Both States and Actions: Evidence From High‐Resolution fMRI.” PLoS Computational Biology 13, no. 10: e1005810. 10.1371/journal.pcbi.1005810. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Corlett, P. R. , Mollick J. A., and Kober H.. 2022. “Meta‐Analysis of Human Prediction Error for Incentives, Perception, Cognition, and Action.” Neuropsychopharmacol 47, no. 7: 1339–1349. 10.1038/s41386-021-01264-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Crone, E. A. , and Dahl R. E.. 2012. “Understanding Adolescence as a Period of Social–Affective Engagement and Goal Flexibility.” Nature Reviews. Neuroscience 13, no. 9: 636–650. 10.1038/nrn3313. [DOI] [PubMed] [Google Scholar]
  18. Daunizeau, J. , Adam V., and Rigoux L.. 2014. “VBA: A Probabilistic Treatment of Nonlinear Models for Neurobiological and Behavioural Data.” PLoS Computational Biology 10, no. 1: e1003441. 10.1371/journal.pcbi.1003441. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Daw, N. D. , Gershman S. J., Seymour B., Dayan P., and Dolan R. J.. 2011. “Model‐Based Influences on Humans' Choices and Striatal Prediction Errors.” Neuron 69, no. 6: 1204–1215. 10.1016/j.neuron.2011.02.027. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. DeWit, D. J. , Adlaf E. M., Offord D. R., and Ogborne A. C.. 2000. “Age at First Alcohol Use: A Risk Factor for the Development of Alcohol Disorders.” American Journal of Psychiatry 157, no. 5: 745–750. 10.1176/appi.ajp.157.5.745. [DOI] [PubMed] [Google Scholar]
  21. Diederen, K. M. J. , and Fletcher P. C.. 2021. “Dopamine, Prediction Error and Beyond.” Neuroscientist 27, no. 1: 30–46. 10.1177/1073858420907591. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Dishion, T. J. , and Tipsord J. M.. 2011. “Peer Contagion in Child and Adolescent Social and Emotional Development.” Annual Review of Psychology 62: 189–214. 10.1146/annurev.psych.093008.100412. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Dombrovski, A. Y. , Hallquist M. N., Brown V. M., Wilson J., and Szanto K.. 2019. “Value‐Based Choice, Contingency Learning, and Suicidal Behavior in Mid‐ and Late‐Life Depression.” Biological Psychiatry 85, no. 6: 506–516. 10.1016/j.biopsych.2018.10.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Duncan, T. E. , Tildesley E., Duncan S. C., and Hops H.. 1995. “The Consistency of Family and Peer Influences on the Development of Substance Use in Adolescence.” Addiction 90, no. 12: 1647–1660. 10.1046/j.1360-0443.1995.901216477.x. [DOI] [PubMed] [Google Scholar]
  25. Egger, H. L. , Pine D. S., Nelson E., et al. 2011. “The NIMH Child Emotional Faces Picture Set (NIMH‐ChEFS): A New Set of Children's Facial Emotion Stimuli: Child Emotional Faces Stimuli.” International Journal of Methods in Psychiatric Research 20, no. 3: 145–156. 10.1002/mpr.343. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Fan, L. , Li H., Zhuo J., et al. 2016. “The Human Brainnetome Atlas: A New Brain Atlas Based on Connectional Architecture.” Cerebral Cortex 26, no. 8: 3508–3526. 10.1093/cercor/bhw157. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Flannery, J. S. , Jorgensen N. A., Kwon S. J., Prinstein M. J., Telzer E. H., and Lindquist K. A.. 2023. “Developmental Changes in Habenular and Striatal Social Reinforcement Responsivity Across Adolescence Linked With Substance Use.” Biological Psychiatry 94, no. 11: 888–897. 10.1016/j.biopsych.2023.04.018. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Flannery, J. S. , Riedel M. C., Poudel R., et al. 2019. “Habenular and Striatal Activity During Performance Feedback Are Differentially Linked With State‐Like and Trait‐Like Aspects of Tobacco Use Disorder.” Science Advances 5, no. 10: eaax2084. 10.1126/sciadv.aax2084. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Foulkes, L. , and Blakemore S. J.. 2016. “Is There Heightened Sensitivity to Social Reward in Adolescence?” Current Opinion in Neurobiology 40: 81–85. 10.1016/j.conb.2016.06.016. [DOI] [PubMed] [Google Scholar]
  30. Fouragnan, E. , Retzler C., and Philiastides M. G.. 2018. “Separate Neural Representations of Prediction Error Valence and Surprise: Evidence From an fMRI Meta‐Analysis.” Human Brain Mapping 39, no. 7: 2887–2906. 10.1002/hbm.24047. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Frank, M. J. , Moustafa A. A., Haughey H. M., Curran T., and Hutchison K. E.. 2007. “Genetic Triple Dissociation Reveals Multiple Roles for Dopamine in Reinforcement Learning.” National Academy of Sciences of the United States of America 104, no. 41: 16311–16316. 10.1073/pnas.0706111104. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Frank, M. J. , Seeberger L. C., and O'Reilly R. C.. 2004. “By Carrot or by Stick: Cognitive Reinforcement Learning in Parkinsonism.” Science 306, no. 5703: 1940–1943. 10.1126/science.1102941. [DOI] [PubMed] [Google Scholar]
  33. Frömer, R. , Nassar M. R., Bruckner R., Stürmer B., Sommer W., and Yeung N.. 2021. “Response‐Based Outcome Predictions and Confidence Regulate Feedback Processing and Learning.” eLife 10: e62825. 10.7554/eLife.62825. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Frydecka, D. , Drapala J., Kłosińska E., Krefft M., and Misiak B.. 2016. “Computational Modeling of Reinforcement Learning Using Probabilistic Selection Task and Instructional Probabilistic Selection Task.” European Psychiatry 33, no. S1: S138. 10.1016/j.eurpsy.2016.01.225. [DOI] [Google Scholar]
  35. Gläscher, J. P. , and O'Doherty J. P.. 2010. “Model‐Based Approaches to Neuroimaging: Combining Reinforcement Learning Theory With fMRI Data.” WIREs Cognitive Science 1, no. 4: 501–510. 10.1002/wcs.57. [DOI] [PubMed] [Google Scholar]
  36. Graham, J. W. , Marks G., and Hansen W. B.. 1991. “Social Influence Processes Affecting Adolescent Substance Use.” Journal of Applied Psychology 76, no. 2: 291–298. 10.1037/0021-9010.76.2.291. [DOI] [PubMed] [Google Scholar]
  37. Gray, K. M. , and Squeglia L. M.. 2018. “Research Review: What Have we Learned About Adolescent Substance Use?” Journal of Child Psychology and Psychiatry 59, no. 6: 618–627. 10.1111/jcpp.12783. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Griffin, K. W. , Epstein J. A., Botvin G. J., and Spoth R. L.. 2001. “Social Competence and Substance Use Among Rural Youth: Mediating Role of Social Benefit Expectancies of Use.” Journal of Youth and Adolescence 30, no. 4: 485–498. 10.1023/A:1010449300990. [DOI] [Google Scholar]
  39. Haber, S. N. 2016. “Corticostriatal Circuitry.” Dialogues in Clinical Neuroscience 18, no. 1: 7–21. 10.31887/DCNS.2016.18.1/shaber. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Hammond, C. J. , Mayes L. C., and Potenza M. N.. 2015. “Neurobiology of Adolescent Substance Use and Addictive Behaviors: Prevention and Treatment Implications.” [PMC free article] [PubMed]
  41. Hanson, J. L. , Williams A. V., Bangasser D. A., and Peña C. J.. 2021. “Impact of Early Life Stress on Reward Circuit Function and Regulation.” Frontiers in Psychiatry 12: 744690. 10.3389/fpsyt.2021.744690. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Hayes, A. F. 2017. Introduction to Mediation, Moderation, and Conditional Process Analysis, Second Edition: A Regression‐Based Approach. Guilford Publications. [Google Scholar]
  43. Henneberger, A. K. , Mushonga D. R., and Preston A. M.. 2021. “Peer Influence and Adolescent Substance Use: A Systematic Review of Dynamic Social Network Research.” Adolescent Research Review 6, no. 1: 57–73. 10.1007/s40894-019-00130-0. [DOI] [Google Scholar]
  44. Hester, R. , Madeley J., Murphy K., and Mattingley J. B.. 2009. “Learning From Errors: Error‐Related Neural Activity Predicts Improvements in Future Inhibitory Control Performance.” Journal of Neuroscience 29, no. 22: 7158–7165. 10.1523/JNEUROSCI.4337-08.2009. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Jensen, J. , Smith A. J., Willeit M., et al. 2007. “Separate Brain Regions Code for Salience vs. Valence During Reward Prediction in Humans.” Human Brain Mapping 28, no. 4: 294–302. 10.1002/hbm.20274. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Jones, R. M. , Somerville L. H., Li J., et al. 2014. “Adolescent‐Specific Patterns of Behavior and Neural Activity During Social Reinforcement Learning.” Cognitive, Affective, & Behavioral Neuroscience 14, no. 2: 683–697. 10.3758/s13415-014-0257-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Kamkar, N. H. , Lewis D. J., van den Bos W., and Morton J. B.. 2017. “Ventral Striatal Activity Links Adversity and Reward Processing in Children.” Developmental Cognitive Neuroscience 26: 20–27. 10.1016/j.dcn.2017.04.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Kandel, D. B. , and Logan J. A.. 1984. “Patterns of Drug Use From Adolescence to Young Adulthood: 1.” Periods of Risk for Initiation, Continued Use, and Discontinuation 74, no. 7: 7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Keiflin, R. , and Janak P. H.. 2015. “Dopamine Prediction Errors in Reward Learning and Addiction: From Theory to Neural Circuitry.” Neuron 88, no. 2: 247–263. 10.1016/j.neuron.2015.08.037. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Kwon, S. J. , Ivory S. L., McCormick E. M., and Telzer E. H.. 2019. “Behavioral and Neural Dysregulation to Social Rewards and Links to Internalizing Symptoms in Adolescents.” Frontiers in Behavioral Neuroscience 13: 158. 10.3389/fnbeh.2019.00158. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Lejarraga, T. , and Hertwig R.. 2017. “How the Threat of Losses Makes People Explore More Than the Promise of Gains.” Psychonomic Bulletin & Review 24, no. 3: 708–720. 10.3758/s13423-016-1158-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Lerner, T. N. , Holloway A. L., and Seiler J. L.. 2021. “Dopamine, Updated: Reward Prediction Error and Beyond.” Current Opinion in Neurobiology 67: 123–130. 10.1016/j.conb.2020.10.012. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Lohse, K. R. , Miller M. W., Daou M., Valerius W., and Jones M.. 2020. “Dissociating the Contributions of Reward‐Prediction Errors to Trial‐Level Adaptation and Long‐Term Learning.” Biological Psychology 149: 107775. 10.1016/j.biopsycho.2019.107775. [DOI] [PubMed] [Google Scholar]
  54. Merikangas, K. R. , Stolar M., Stevens D. E., et al. 1998. “Familial Transmission of Substance Use Disorders.” Archives of General Psychiatry 55, no. 11: 973–979. 10.1001/archpsyc.55.11.973. [DOI] [PubMed] [Google Scholar]
  55. Miech, R. A. , Johnston L. D., and National Survey Results on Drug Use . 1975. –2023. Monitoring the Future National Survey Results on Drug Use, 1975–2023: Secondary School Students. Monitoring the Future Monograph Series. Institute for Social Research, University of Michigan. https://monitoringthefuture.org/results/annual‐reports/. [Google Scholar]
  56. Monterosso, J. , Piray P., and Luo S.. 2012. “Neuroeconomics and the Study of Addiction.” Biological Psychiatry 72, no. 2: 107–112. 10.1016/j.biopsych.2012.03.012. [DOI] [PubMed] [Google Scholar]
  57. O'Doherty, J. P. , Dayan P., Friston K., Critchley H., and Dolan R. J.. 2003. “Temporal Difference Models and Reward‐Related Learning in the Human Brain.” Neuron 38, no. 2: 329–337. 10.1016/S0896-6273(03)00169-7. [DOI] [PubMed] [Google Scholar]
  58. Palminteri, S. , Justo D., Jauffret C., et al. 2012. “Critical Roles for Anterior Insula and Dorsal Striatum in Punishment‐Based Avoidance Learning.” Neuron 76, no. 5: 998–1009. 10.1016/j.neuron.2012.10.017. [DOI] [PubMed] [Google Scholar]
  59. Peters, J. , Bromberg U., Schneider S., et al. 2011. “Lower Ventral Striatal Activation During Reward Anticipation in Adolescent Smokers.” American Journal of Psychiatry 168, no. 5: 540–549. [DOI] [PubMed] [Google Scholar]
  60. Pomery, E. A. , Gibbons F. X., Gerrard M., Cleveland M. J., Brody G. H., and Wills T. A.. 2005. “Families and Risk: Prospective Analyses of Familial and Social Influences on Adolescent Substance Use.” Journal of Family Psychology 19, no. 4: 560–570. 10.1037/0893-3200.19.4.560. [DOI] [PubMed] [Google Scholar]
  61. Preuschoff, K. , Quartz S. R., and Bossaerts P.. 2008. “Human Insula Activation Reflects Risk Prediction Errors as Well as Risk.” Journal of Neuroscience 28, no. 11: 2745–2752. 10.1523/JNEUROSCI.4286-07.2008. [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. Radoman, M. , Crane N. A., Gorka S. M., et al. 2021. “Striatal Activation to Monetary Reward Is Associated With Alcohol Reward Sensitivity.” Neuropsychopharmacol 46, no. 2: 343–350. 10.1038/s41386-020-0728-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  63. Redish, A. D. 2004. “Addiction as a Computational Process Gone Awry.” Science 306, no. 5703: 1944–1947. 10.1126/science.1102384. [DOI] [PubMed] [Google Scholar]
  64. Reed, M. D. , and Rountree P. W.. 1997. “Peer Pressure and Adolescent Substance Use.” Journal of Quantitative Criminology 13, no. 2: 143–180. 10.1007/BF02221306. [DOI] [Google Scholar]
  65. Robinson, O. J. , Overstreet C., Charney D. R., Vytal K., and Grillon C.. 2013. “Stress Increases Aversive Prediction Error Signal in the Ventral Striatum.” National Academy of Sciences of the United States of America 110, no. 10: 4129–4133. 10.1073/pnas.1213923110. [DOI] [PMC free article] [PubMed] [Google Scholar]
  66. Roy, M. , Shohamy D., and Wager T. D.. 2012. “Ventromedial Prefrontal‐Subcortical Systems and the Generation of Affective Meaning.” Trends in Cognitive Sciences 16, no. 3: 147–156. 10.1016/j.tics.2012.01.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  67. Schneider, S. , Peters J., Bromberg U., et al. 2012. “Risk Taking and the Adolescent Reward System: A Potential Common Link to Substance Abuse.” American Journal of Psychiatry 169, no. 1: 39–46. 10.1176/appi.ajp.2011.11030489. [DOI] [PubMed] [Google Scholar]
  68. Schultz, W. 1998. “Predictive Reward Signal of Dopamine Neurons.” Journal of Neurophysiology 80, no. 1: 1–27. 10.1152/jn.1998.80.1.1. [DOI] [PubMed] [Google Scholar]
  69. Siebenbruner, J. , Englund M. M., Egeland B., and Hudson K.. 2006. “Developmental Antecedents of Late Adolescence Substance Use Patterns.” Development and Psychopathology 18, no. 2: 551–571. 10.1017/S0954579406060287. [DOI] [PubMed] [Google Scholar]
  70. Simons‐Morton, B. 2007. “Social Influences on Adolescent Substance Use.” American Journal of Health Behavior 31, no. 6: 672–684. 10.5993/AJHB.31.6.13. [DOI] [PubMed] [Google Scholar]
  71. Smith, A. R. , Steinberg L., Strang N., and Chein J.. 2015. “Age Differences in the Impact of Peers on Adolescents' and Adults' Neural Response to Reward.” Developmental Cognitive Neuroscience 11: 75–82. 10.1016/j.dcn.2014.08.010. [DOI] [PMC free article] [PubMed] [Google Scholar]
  72. Somerville, L. H. 2013. “The Teenage Brain: Sensitivity to Social Evaluation.” Current Directions in Psychological Science 22, no. 2: 121–127. 10.1177/0963721413476512. [DOI] [PMC free article] [PubMed] [Google Scholar]
  73. Squeglia, L. M. , Sorg S. F., Jacobus J., Brumback T., Taylor C. T., and Tapert S. F.. 2015. “Structural Connectivity of Neural Reward Networks in Youth at Risk for Substance Use Disorders.” Psychopharmacology 232, no. 13: 2217–2226. 10.1007/s00213-014-3857-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  74. Steinberg, L. , and Monahan K. C.. 2007. “Age Differences in Resistance to Peer Influence.” Developmental Psychology 43, no. 6: 1531–1543. 10.1037/0012-1649.43.6.1531. [DOI] [PMC free article] [PubMed] [Google Scholar]
  75. Stokes, P. R. A. , Shotbolt P., Mehta M. A., et al. 2013. “Nature or Nurture? Determining the Heritability of Human Striatal Dopamine Function: An [18F]‐DOPA PET Study.” Neuropsychopharmacol 38, no. 3: 485–491. 10.1038/npp.2012.207. [DOI] [PMC free article] [PubMed] [Google Scholar]
  76. Sussman, S. , Silvana S., and Ames S. L.. 2008. “Substance Abuse Among Adolescents.” Substance Use and Misuse 43, no. 12–13: 1802–1828. 10.1080/10826080802297302. [DOI] [PubMed] [Google Scholar]
  77. Takahashi, Y. , Roesch M. R., Stalnaker T. A., and Schoenbaum G.. 2007. “Cocaine Exposure Shifts the Balance of Associative Encoding From Ventral to Dorsolateral Striatum.” Frontiers in Integrative Neuroscience 1: 1. 10.3389/neuro.07.011.2007. [DOI] [PMC free article] [PubMed] [Google Scholar]
  78. Tanabe, J. , Reynolds J., Krmpotich T., et al. 2013. “Reduced Neural Tracking of Prediction Error in Substance‐Dependent Individuals.” American Journal of Psychiatry 170, no. 11: 1356–1363. 10.1176/appi.ajp.2013.12091257. [DOI] [PMC free article] [PubMed] [Google Scholar]
  79. Telzer, E. H. , Jorgensen N. A., Prinstein M. J., and Lindquist K. A.. 2021. “Neurobiological Sensitivity to Social Rewards and Punishments Moderates Link Between Peer Norms and Adolescent Risk Taking.” Child Development 92, no. 2: 731–745. 10.1111/cdev.13466. [DOI] [PMC free article] [PubMed] [Google Scholar]
  80. Telzer, E. H. , Kwon S. J., and Jorgensen N. A.. 2023. “Neurobiological Development in Adolescence and Early Adulthood: Implications for Positive Youth Adjustment.” In APA Handbook of Adolescent and Young Adult Development. APA Handbooks in Psychology, 629–643. American Psychological Association. 10.1037/0000298-038. [DOI] [Google Scholar]
  81. Telzer, E. H. , Rogers C. R., and Van Hoorn J.. 2017. “Neural Correlates of Social Influence on Risk Taking and Substance Use in Adolescents.” Current Addiction Reports 4, no. 3: 333–341. 10.1007/s40429-017-0164-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  82. Tervo‐Clemmens, B. , Quach A., Calabro F. J., Foran W., and Luna B.. 2020. “Meta‐Analysis and Review of Functional Neuroimaging Differences Underlying Adolescent Vulnerability to Substance Use.” NeuroImage 209: 116476. 10.1016/j.neuroimage.2019.116476. [DOI] [PMC free article] [PubMed] [Google Scholar]
  83. Tolomeo, S. , Yaple Z. A., and Yu R.. 2021. “Neural Representation of Prediction Error Signals in Substance Users.” Addiction Biology 26, no. 3: e12976. 10.1111/adb.12976. [DOI] [PubMed] [Google Scholar]
  84. Trifilieff, P. , and Martinez D.. 2014. “Blunted Dopamine Release as a Biomarker for Vulnerability for Substance Use Disorders.” Biological Psychiatry 76, no. 1: 4–5. 10.1016/j.biopsych.2014.04.017. [DOI] [PMC free article] [PubMed] [Google Scholar]
  85. Turpyn, C. C. , Jorgensen N. A., Prinstein M. J., Lindquist K. A., and Telzer E. H.. 2021. “Social Neural Sensitivity as a Susceptibility Marker to Family Context in Predicting Adolescent Externalizing Behavior.” Developmental Cognitive Neuroscience 51: 100993. 10.1016/j.dcn.2021.100993. [DOI] [PMC free article] [PubMed] [Google Scholar]
  86. Ullsperger, M. , Harsay H. A., Wessel J. R., and Ridderinkhof K. R.. 2010. “Conscious Perception of Errors and Its Relation to the Anterior Insula.” Brain Structure and Function 214, no. 5: 629–643. 10.1007/s00429-010-0261-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  87. Urošević, S. , Collins P., Muetzel R., Schissel A., Lim K. O., and Luciana M.. 2015. “Effects of Reward Sensitivity and Regional Brain Volumes on Substance Use Initiation in Adolescence.” Social Cognitive and Affective Neuroscience 10, no. 1: 106–113. 10.1093/scan/nsu022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  88. Volkow, N. D. , Fowler J. S., Wang G. J., Swanson J. M., and Telang F.. 2007. “Dopamine in Drug Abuse and Addiction: Results of Imaging Studies and Treatment Implications.” Archives of Neurology 64: 1575–1579. 10.1001/archneur.64.11.1575. [DOI] [PubMed] [Google Scholar]
  89. Wagner, F. A. , and Anthony J. C.. 2002. “From First Drug Use to Drug Dependence: Developmental Periods of Risk for Dependence Upon Marijuana, Cocaine, and Alcohol.” Neuropsychopharmacology 26, no. 4: 479–488. 10.1016/S0893-133X(01)00367-0. [DOI] [PubMed] [Google Scholar]
  90. Wang, W. , Evans K., and Schweizer S.. 2025. “Social and Non‐Social Risk‐Taking in Adolescence.” Scientific Reports 15, no. 1: 6880. 10.1038/s41598-025-90050-y. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Table S1: Demographic information.

Table S2: Whole brain social reinforcement prediction error tracking during positive and negative feedback cluster coordinates. Sample average whole‐brain activity during positive and negative feedback presentation that was significantly associated with trial‐specific prediction error values. All subjects prediction error related images (n = 238) were entered into whole‐brain one‐sample t‐tests for positive and negative feedback (3dMEMA, p voxel‐level = 0.001, p cluster‐level< 0.05, cluster extent threshold = 56 voxels for positive feedback and 51 voxels for negative feedback).

Table S3: Whole brain social reinforcement prediction error tracking cluster coordinates. Sample average whole‐brain activity during feedback presentation that was significantly positively, or negatively, associated with trial‐specific prediction error values. All subjects’ prediction error related images (n = 238) were entered into a whole‐brain one‐sample t‐test 3dMEMA, p voxel‐level< 0.1e −7, p cluster‐level< 0.05, cluster extent threshold = 10 voxels.

Table S4: Substance use curiosity and exposure correlation table.

Figure S1: Q‐Learning model comparison and selection. Random‐effects Bayesian model comparison across 12 candidate models. Bars show posterior model probability r(sorted). Labels indicate exceedance probability (EP). The asymmetric Q‐learning (Base) model (AsymBase) has the highest rand exceedance probability (EP) compared to all other models (Bayesian Omnibus Risk [BOR] = 0.001) and was selected as the best fit model. The AsymBase model had an EP of 1.0 indicating strong evidence that the selected model is the best across participants.

Figure S2: Simulation‐based validation of selected asymmetric Q‐learning model. Forward simulation was conducted using each participant's posterior mean parameters (learning rates and softmax inverse‐temperature). (A) Pearson correlations between actual and simulated performance per pair. (B) Visualized correspondence between simulated and actual learning curves for each stimuli pair across trials.

Figure S3: Whole brain prediction error activity for positive and negative feedback. We considered whole brain activity that significantly tracked prediction error for positive and negative feedback separately. (A) During positive feedback, activity in the bilateral anterior insula, amygdala, thalamus, midbrain, and occipital lobe increased as feedback was increasing better than expected. (B) During negative feedback, activity in the bilateral putamen, supplementary motor area, and occipital lobe decreased as feedback was increasingly worse than expected. Warm colors denote activity with a significant positive association with prediction error values and cool colors denote activity with a significant negative association with prediction error values. Cluster coordinates reported in Table S2.

Figure S4: Social reinforcement learning and approach/avoidance performance. (A) On average participants’ performance decreased as difficulty level of stimuli pairs increased such that participants made the optimal choice on 68.0% ± 18.2% of AB trials, 63.8% ± 19.4% of CD trials and 57.1% ± 19.7% of EF trials. (B) Across trials in the learning phase, participants’ expected friendliness of adolescent A (QA) increased whereas the expected friendliness of adolescent B (QB) decreased. (C) In the testing phase, participants displayed better negative social reinforcement learning ability (percent correct on avoid B trials; 70.1% ± 19.8%) than positive social reinforcement learning ability (percent correct on approach A trials; 60.5% ± 23.2%). Approach and avoidance learning performance were not significantly associated. (D) On average, learning rates for positive feedback (0.34 ± 0.2) were significantly lower than learning rates for negative feedback (0.56 ± 0.2). Positive and negative learning rates were negatively correlated.

Figure S5: vmPFC prediction error (PE) tracking and associations with substance use measures. (A) Ventromedial prefrontal cortex (vmPFC) regions of interest defined from an independent meta‐analysis of subjective value representation, including central (cvmPFC; red), ventral (vvmPFC; green), and anterior (avmPFC; blue) subregions (9‐mm radius spheres). (B) Prediction error (PE)‐tracking of better‐than‐expected outcomes in the anterior vmPFC is significantly negatively associated with substance use curiosity. (C) PE‐tracking in the anterior vmPFC was not significantly associated with household substance use. PE‐tracking in ventral and central vmPFC subregions was not associated curiosity and exposure (not shown).

HBM-47-e70545-s001.docx (13.7MB, docx)

Data Availability Statement

The authors have released code and tabular data associated with this manuscript on GitHub (https://github.com/Flanneryg3/FF_ProjectCode), and unthresholded brain maps are available on NeuroVault (https://neurovault.org/collections/21902/).

Articles from Human Brain Mapping are provided here courtesy of Wiley

RESOURCES