Behavioral and Neural Trajectories of Risk Taking for Peer and Parent in Adolescence

Adolescence is marked by rises in risk-taking behaviors; yet, most of adolescents' risks take place within a social context (e.g., in the presence of their peers; Chein, Albert, O'Brien, Uckert, & Steinberg, 2011). Although extant research on social contextual effects on risk taking has largely focused on risks that only affect the adolescents themselves, many instances of risk taking also impact individuals around them (Do, Guassi Moreira, & Telzer, 2017). For example, driving the family car and getting into an accident may result in parents' financial loss, and copying a friend's homework may result in getting the friend in trouble. In the current study, we sought to evaluate longitudinal changes in adolescents' risk taking that impacts their peers and parents, and the neural processes that support this development.

Adolescents show a remarkable ability to tune their risky behaviors to the context. For instance, adolescents take more risks for themselves than for a stranger (Crone, Bullens, van Der Plas, Kijkuit, & Zelazo, 2008), reflecting their ability to flexibly adjust their risk taking depending on who the behavior targets. Adolescents also take more risks for strangers who are perceived as “high risk-takers” than “low risk-takers” (Crone et al., 2008), demonstrating their ability to integrate others' perspectives into their decision-making process. In addition, adolescents shift who they are willing to take risks for over time. Early adolescents take a similar amount of risks when the recipient is them or their parent (Guassi Moreira & Telzer, 2018), but young adults prioritize their parents over their peers by taking more risks to earn rewards for their parents (Guassi Moreira, Tashjian, Galván, & Silvers, 2020). Thus, vicarious risk taking may change between early adolescence and early adulthood, and depending on the social agent for whom the risk targets.

Rises in risk taking occur in concert with shifting patterns in adolescents' social relationships. According to social identity theory, an individual's sense of self depends on the social group to which they belong and, thus, their behaviors (e.g., risk behaviors) are guided by the norms (e.g., risk prototypes) espoused by their social groups (Tajfel & Turner, 1979). Family and friends are the two most salient social groups with whom adolescents identify (see Telzer, van Hoorn, Rogers, & Do, 2018). Across the teenage years, these relationships evolve such that adolescents become hypersensitive to peers and form interdependent friendships (Furman & Rose, 2015; Brown & Larson, 2009). Adolescents also become more independent from their parents yet also maintain intimate connections and continue to rely on their parents for guidance (Nickerson & Nagle, 2005). As a result, adolescents may take distinct risks for their peers and parents, with this difference changing over time.

Although adolescents might take risks differently for their peers and parents, the relative influence of these partners is contingent on the type of decision-making (Wilks, 1986). One common type of risk taking is adaptive risk taking, whereby an individual effectively switches between safe and risky decisions by taking advantageous risks and avoiding disadvantageous ones. Whether a risk is economically advantageous or disadvantageous relies on the expected value (EV) of the reward that is associated with risk taking. A second common type of risk taking is one's risk preferences, which may be measured by risk taking in a neutral-reward context. That is, taking risks when the EVs of risky and safe options are equivalent (and thus there is no mathematically correct decision to make) may be reflective of one's proclivity for risks. For both forms of economic risk taking, adolescents are more prone to these risks than adults, highlighting adolescence as a developmental window during which risk and reward sensitivity normatively changes (Barkley-Levenson & Galván, 2014; Paulsen, Platt, Huettel, & Brannon, 2011). Given that sensitivity to social influence depends on the decision-making context, adolescents may take general (i.e., risk preferences) and adaptive (i.e., sensitivity to expected reward value) risks differently for their peers and parents over time.

Longitudinal neuroimaging methods can help us identify how neurodevelopment is associated with changes in vicarious risk taking during adolescence. In particular, adolescence is a time of a drastic reorganization of the neural systems (Nelson, Jarcho, & Guyer, 2016; Nelson, Leibenluft, McClure, & Pine, 2005), and thus, it is important to consider how the changing neural functions coincide with the changing adolescent behaviors. By contrast, some behaviors remain consistent during adolescence but these may nonetheless be paralleled by changing neural functions, suggesting that different neural strategies are needed to reach the same behavior (e.g., Eshel, Nelson, Blair, Pine, & Ernst, 2007). Taking a longitudinal approach to understand the development of vicarious risky behaviors will therefore elucidate how the neurobiological strategies or demands of adolescent social behaviors, like vicarious risky behaviors, change with maturation.

Functional changes in brain regions associated with valuation (ventral stratum [VS] and ventromedial prefrontal cortex [vMPFC]) are promising neural candidates. Both animal and human neuroscience research have repeatedly identified the role of VS in reward processing (e.g., Delgado, 2007; Schultz, Apicella, Scarnati, & Ljungberg, 1992). There are also developmental changes in VS functions, especially in the context of vicarious reward processing, such that there are age-related increases in the VS when processing money won for peers versus parents (Braams & Crone, 2017a). As such, the VS may subserve vicarious reward processing during risk taking for these two social agents as well. In addition, the vMPFC has also been linked to various reward processing such as computing subjective reward values and responding to stimuli of high personal value such as close others (e.g., Bartra, McGuire, & Kable, 2013; D'Argembeau, 2013; Grabenhorst & Rolls, 2011), indicating that the vMPFC may be involved in gauging values associated with peers and parents during vicarious risk taking. These two regions are also sensitive to EV during adolescents' risk taking (van Duijvenvoorde et al., 2015; Barkley-Levenson & Galván, 2014) and are thus involved in adaptive risk taking. Although the valuation system's sensitivity to EV is widely understood, it is unknown how this system tracks EV that may also hold social values. Taken together, neural patterns within the valuation system during general and adaptive risk taking may be differentially recruited for peers and parents over time.

The present preregistered study (https://osf.io/j7gsa/) utilized a longitudinal neuroimaging design to examine behavioral and neural trajectories of risk-taking behaviors that directly impact adolescents' best friend and parent. We hypothesized that at the behavioral level, adolescents will take general and adaptive risks differently for their best friend and parent across time. Similarly, we hypothesized that at the neural level, the VS and vMPFC will be differentially activated during general and adaptive risk taking for their best friend and parent across time. Given the divergent developmental trends in peer and parent influences, we did not have specific hypotheses regarding the direction of trajectories. That is, risk taking for best friend may increase faster than those for parents (Krosnick & Judd, 1982) or we may observe the opposite pattern (Cook, Buehler, & Henson, 2009). It is also possible that risk taking does not differentially change for best friend and parent across time (Chassin, Presson, Sherman, Montello, & McGrew, 1986). Multivariate growth models were used to simultaneously model best friend and parent trajectories; following, post hoc contrasts tested for differences in these trajectories. We conducted a priori ROI analyses, supplemented with longitudinal whole-brain analyses.

Adolescent participants were recruited as part of a larger study of 873 sixth and seventh grade students from three public middle schools to participate in a longitudinal fMRI study, based on interest and eligibility. For this current study, participants had to be at least 12 years old and in sixth or seventh grade, or within 2 months of turning 12 years old, at Wave 1 of data collection. Participants were excluded if they had any metal in their body including braces or permanent retainer. Other exclusion criteria included claustrophobia, history of seizure or head trauma, learning disability, and nonfluency in English. If participants regularly took medications, they were asked to do a 24-hr medication wash before the scan. One hundred seventy-three participants completed between one and three sessions annually across three waves. See Table 1 for demographic information about adolescent and parent participants. All participants were compensated for completing the session. In addition, all participants provided informed consent/assent and the university's institutional review board approved all aspects of the study.

To reach our target sample size of 150 participants after accounting for attrition and for excluded participants between waves of data collection, we recruited two cohorts of participants across 2 years of the study (e.g., Herd, King-Casas, & Kim-Soon, 2020, for similar study design). Power analysis was performed during grant submission, which identified a target size of 150 participants (which accounts for expected attrition over the course of the study). We recruited 148 participants at Wave 1 of the study (Cohort 1) and 30 additional participants at Wave 2 (Cohort 2). Across the three waves, 25 participants had 1 time point of behavioral data (34 for fMRI data), 36 had 2 (39 for fMRI data), and 112 had 3 (97 for fMRI data), leading to 433 behavioral and 403 fMRI data points.

At Wave 1, five participants were excluded because of exclusionary criteria that were revealed after recruitment. These participants were not invited back for subsequent study participation. Out of the remaining 143 participants (M_age = 12.8 years, SD_age = 0.52 years; range = 11.9–14.5 years; 73 girls), one participant was excluded for acute anxiety related to the fMRI scanner and three participants were excluded because of incomplete data (e.g., less than 60% of response on task; ending the scan early). Furthermore, three participants were excluded only from the neural analyses for completing the task behaviorally (i.e., outside of the scanner), 11 participants for not having enough behavioral data or variability across trial types to be modeled at the neural level, and nine participants for excessive motion (>0.9-mm framewise displacement on >10% of total volumes). The final Wave 1 sample size with behavioral and fMRI data included 139 and 116 adolescents, respectively.

At Wave 2, 27 participants from Cohort 1 indicated they did not want to participate or lost contact. With 30 new participants from Cohort 2, 146 adolescents participated (M_age = 13.7 years; SD_age = 0.58 years; range = 12.4–15.4 years; 78 girls). After using the same exclusionary criteria described above, the final Wave 2 sample size with behavioral and fMRI data included 143 and 115 adolescents, respectively. At Wave 3, seven participants indicated they did not want to participate or lost contact and six participants who participated in Wave 1 but skipped Wave 2 returned for Wave 3 of the study. As a result, 145 adolescents participated (M_age = 14.7 years; SD_age = 0.58 years; range = 13.4–16.3 years; 74 girls). After exclusion, the final Wave 3 sample size with behavioral and fMRI data included 144 and 100 adolescents, respectively. See Table 2 for more information about participants at each wave.

At each wave of data collection, participants completed behavioral and fMRI tasks as well as self-report questionnaires, totaling a 4-hr session with a 1.5-hr fMRI session. Before completing the fMRI scan, participants received training for the tasks, were acclimated to a mock scanner, and completed self-report measures. In the event the participant could not participate in the fMRI session after the first wave (e.g., braces), they completed the tasks behaviorally, outside of the scanner. At the end of the session, participants received monetary compensation ($90), prizes worth up to $20 for doing well in the scan (e.g., gift cards, headphones), and a meal after the scan. The participating parent/guardian received monetary compensation ($50), parking and gas reimbursement ($27), and a meal. At each subsequent wave, returning families received an additional $25 returning bonus (e.g., additional $25 for completing two waves). Adolescent participants also received additional money for themselves, their parent, and their best friend through the risky decision-making task.

Adolescents completed a modified cups task (Levin & Hart, 2003), which has previously been utilized to examine risky decision-making for others in developmental samples (e.g., Guassi Moreira & Telzer, 2018). Participants completed three runs of the cups task: one in which they made decisions for the self, one for parent, and one for best friend. The order in which participants completed each run was counterbalanced. Note that the self-run is part of another article (Kwon, Turpyn, Prinstein, Lindquist, & Telzer, 2022), and our focus here is on the relative roles of peers and parents in vicarious risky behaviors.

Each of the three runs consisted of 45 trials. On each trial, participants were presented with two scenarios of cups: The left side always had one cup with a guaranteed 15 cents hidden under the cup (Figure 1). On the right side, the number of cups (either two, three, or five cups) as well as the amount of money hidden (either 30, 45, or 75 cents) varied; however, the money was hidden under only one of the overturned cups. Participants were told that if they chose the right side (i.e., risky decision), then the computer would randomly select one of the cups and they may earn the higher amount or 0 cent, whereas if they chose the left side (i.e., safe decision), then they were guaranteed to earn 15 cents. After each decision, participants were shown the outcome.

On each trial, the cups were shown for 3000 msec, within which participants made their decision. Next, a fixation cross was jittered around an average of 2300 msec (range = 526.68–4017.12 msec), which was followed by the outcome for 1000 msec. Finally, there was an intertrial fixation cross that was jittered around an average of 2521.39 msec (range = 521.14–3913.31 msec). If participants did not make a decision within the given time, participants were told that they were “too late” and there was no change in the total points. Outcomes of each decision were added to the running total for that run, which was shown to the participant at the end of each run. At the end of each session, adolescents received the money they had earned for themselves, their parent was given the money their child had earned for them, and their best friend was provided with their earnings in cash. The participating best friend and parent did not know the adolescent was winning money for them until they received the award. For families who participated in more than one wave of data collection, 7.4% of participating parent changed at least once across their years of participation. For best friend nominations, 69.9% of adolescent participants changed their best friend at least once across their years of participation.

To operationalize adaptive vicarious risk taking, we assessed adolescents' likelihood of making a risky decision as a function of the EV of reward of the risky choice. To operationalize general vicarious risk taking, we assessed adolescents' likelihood of making a risky decision when the EVs of safe and risky choices are equal. Consistent with prior work, EV was composed of two factors: magnitude and probability of reward, both of which contribute to taking risks when rewards are at stake (Guassi Moreira & Telzer, 2018; van Duijvenvoorde et al., 2015). EV was calculated by dividing the amount of money under the cup (i.e., magnitude of reward) by the number of cups (i.e., probability of reward) for that trial. Given the parameters of the magnitudes and probabilities of reward, the EVs for risky decisions were: 6, 9, 10, 15, 22.5, 25, 37.5. The EV of safe decision was always 15. In this task, it is advantageous to take risks when the EV is greater than 15 (i.e., EV of safe decision), whereas it is disadvantageous to take risks when the EV is less than 15. It is therefore adaptive to take risks with increasing EV.

Imaging data were collected using a 3 Tesla Siemens Prisma MRI scanner. The cups task was presented on a computer screen and projected through a mirror. A high-resolution T2*-weighted EPI volume (repetition time [TR] = 2000 msec; echo time [TE] = 25 msec; flip angle = 90°; matrix = 92 × 92; field of view [FOV] = 230 mm; 37 slices; slice thickness = 3 mm; voxel size = 2.5 × 2.5 × 3 mm³) was acquired coplanar with a high-resolution T2*-weighted, matched-bandwidth, structural scan (TR = 5700 msec; TE = 65 msec; flip angle = 120°; matrix = 192 × 192; FOV = 230 mm; 38 slices; slice thickness = 3 mm). In addition, a T1* magnetization prepared rapid gradient echo (TR = 2400 msec; TE = 2.22 msec; flip angle = 8°; matrix = 256 × 256; FOV = 256 mm; 208 slices; slice thickness = 0.8 mm; sagittal plane) was acquired. The orientation for the EPI and matched-bandwidth scans was oblique axial to maximize brain coverage and to reduce noise.

Preprocessing was conducted using FMRIB's Software Library (Version 6.0; www.fmrib.ox.ac.uk/fsl) and included the following steps: skull stripping using the Brain Extraction Tool; motion correction with Motion Correction FMRIB's Linear Image Registration Tool; spatial smoothing with a 6-mm Gaussian kernel, FWHM; high-pass temporal filtering with a 128-sec filter width (Gaussian-weighted least-squares straight line fitting, with sigma = 64.0 sec); grand-mean intensity normalization of the entire 4-D data set by a single multiplicative factor; and individual level independent component analysis denoising for artifact signal using MELODIC (Version 3.15), combined with an automated signal classifier (Tohka et al., 2008; Neyman-Pearson threshold = .3). For spatial normalization, the EPI data were registered to the T1 image with a linear transformation, followed by a white-matter boundary-based transformation using FMRIB's Linear Image Registration Tool, linear and nonlinear transformations to standard Montreal Neurological Institute 2-mm brain using Advanced Neuroimaging Tools, and then spatial normalization of the EPI image to the Montreal Neurological Institute. Quality check during preprocessing and analyses ensured adequate signal coverage.

The task was modeled using an event-related design within the Statistical Parametric Mapping software package (SPM12; Wellcome Department of Cognitive Neurology, Institute of Neurology). Individual-level fixed-effects models were created for each participant using the general linear model with regressors for the following five conditions: trials for each decision (risky or safe) and trials for each outcome (15 cents, zero cent, or >15 cents). A parametric modulator (PM) was included for each decision, whereby the PM represented the EV (centered at 15) of the risky decision for that trial and served to examine neural activity that tracks EV when making decisions. Each condition was modeled using the onset of the cups (or outcome) and a duration equal to zero. It was also modeled separately for each run, totaling 15 conditions. The contrasts of interest for adaptive vicarious risk taking were risky decision-making for best friend and for parent, both with EV as the PM. The contrasts of interest for general vicarious risk taking were risky decision-making for best friend and for parent, both when EV was zero (i.e., when EV of safe and risky choices are equal).

Furthermore, trials in which participants did not respond, the final outcome trial, and volumes containing motion in excess of 0.9-mm framewise displacement were included as separate regressors of no interest. Six motion regressors were modeled as covariates of non-interest to control for head movement in six dimensions. Jittered intertrial periods (i.e., fixation cross) were not explicitly modeled and therefore served as the implicit baseline for task conditions.

We used bilateral ROIs of VS and vMPFC that were defined using the Harvard-Oxford Atlas (Harvard Center for Morphometric Analysis; Figure 2). We extracted parameter estimates from the VS and vMPFC ROIs for each contrast of interest, separately for each participant at each time point. Whole-brain group-level contrasts are available on Neurovault (https://neurovault.org/collections/13322/).

To compliment the ROI analyses, we conducted longitudinal whole-brain analyses using AFNI 3dLMEr models (Chen, Saad, Britton, Pine, & Cox, 2013). This program allows for voxel-level whole-brain analyses of linear mixed effects (maximum-likelihood, multilevel model). To assess the relative trajectories of vicarious risk taking at the whole-brain level, we modeled the interaction between risk taking for best friend versus parent and grade (for linear model; grade × grade for quadratic model) for each type of risk taking, totaling four models. To correct for multiple comparisons, we conducted a Monte Carlo simulation using the 3dFWHMx and 3dClustSim programs from the AFNI software package (Ward et al., 2000) and the group-level brain mask. Smoothness was estimated with the -acf option (−acf a, b, c parameters: 0.546, 4.574, 12.475), which used an average of individual-level autocorrelation function parameters that is obtained using each participant's residuals from the first-level model. This simulation indicated that a p < .05 FWE corrected would be achieved with a voxel-wise threshold of p < .001 and a minimum cluster size of 80 voxels. The codes and outputs for longitudinal whole-brain analyses using AFNI 3dlmer are available on GitHub (https://github.com/sehjookwon/BestFriendvsParent_3dlmer).

For each social target, we used a three-level univariate growth model with trials (i = 45 trials maximum) nested within time points (j = 3 time points maximum), nested within individuals (k = 173 participants). EV that is centered at 15 was added as a Level 1 predictor and grade that is centered at sixth grade as a Level 2 predictor, and binary decision (0 = safe decision; 1 = risky decision) of each trial was the dependent variable (DV). We first determined the functional form that best fits the development of behavior (see below) and then added maximal random effects to the model of chosen form (Barr, Levy, Scheepers, & Tily, 2013). For example, we estimated the following model to test for linear changes in behavior (lme4::glmer package in R; Bates, Mächler, Bolker, & Walker, 2015):

Here, the first set of parentheses denotes the fixed effects of general vicarious risk taking (i.e., the average log likelihood of risky decision-making when EV is zero, which is when the EVs of safe and risky choices are equal, at sixth grade and the average trajectory of this likelihood); the second set of parentheses denotes the fixed effects of adaptive vicarious risk taking (i.e., the average log likelihood of risky decision-making as a function of EV at sixth grade and the average trajectory of this likelihood); the last set of parentheses denotes random effects of intercept, EV, and grade. From the final chosen model, we examined the effect of “intercept” and “EV” for general and adaptive vicarious risk taking, respectively, at sixth grade, and the effect of “Grade” and “Grade × EV” for grade-related changes in general and adaptive vicarious risk taking, respectively. Note that random effects were removed serially if they led to issues with random effects estimations.

Next, we conducted a multivariate growth model to evaluate how changes in risk taking for best friend relate to those for parent. For example, we estimated the following model for linear changes in behavior:

For each social target, we used a two-level univariate growth model with time points (i) nested within individuals (j). We estimated separate models for each type of vicarious risk taking (general, adaptive) and ROI (VS, vMPFC). Parameter estimates from each ROI for each contrast of interest were the DV. Parameter estimates that were above or below 3 SDs from the mean (of each social target at each grade) were winsorized to 3 SDs to reduce the effect of extreme values. Like with the behavioral trajectory model, we first determined the best-fitting functional form and then, for example, estimated the following model to test for linear changes in neural activation (nlme::lme package; Pinheiro, Bates, & R Core Team, 2022):

For each univariate growth model above, we added a quadratic growth term: grade × grade × EV for behavioral trajectory model, grade × grade for neural trajectory models. The linear and quadratic models were formally compared using a log likelihood ratio test with difference in degrees of freedom (i.e., difference between two models in their degrees of freedom) and level of significance of p < .05. We also utilized Akaike's information criterion (AIC) values and Schwarz's Bayesian information criterion (BIC) for descriptive purposes. AIC and BIC values are standardized model-fit metrics, and preferred models have lower AIC and BIC values. AIC and BIC values and log likelihood ratio test results for all models are available on the Open Science Framework (OSF). All models were fit with full information maximum likelihood estimates. For neural trajectory models, each ROI was tested 6 times (2 social targets × 2 types of risks × 2 functional forms). We did not formally correct the significance level because significance was only tested on the final chosen model and prior longitudinal fMRI studies with similar analysis procedure did not do such correction (e.g., Braams & Crone, 2017b), but this analytic decision should be considered when interpreting the results.

Our current study has several deviations from our preregistration. First, we used grade instead of age as a predictor. Risk norms are often socialized, and parent and peer relationships shift in relative importance across school grades, and so vicarious risk taking may be better captured by one's social experiences (i.e., grade) than by the passage of time (i.e., age; O'Donnell, O'Donnell, & Stueve, 2001; McBride et al., 1995). Second, we winsorized extreme values instead of entirely excluding them. Winsorizing is a method often utilized by longitudinal neuroimaging studies to reduce the influence of extreme values while retaining as much data as possible (e.g., Baranger et al., 2021; Thijssen, Collins, Weiss, & Luciana, 2021). Third, we did not run our proposed VS-vMPFC functional connectivity analyses. Longitudinal changes in connectivity may vary depending on the analytic method, suggesting that changes in connectivity should be examined using a multiverse approach (Bloom et al., 2022).

A linear functional form best fit developmental changes in risk taking for both best friend and parent. There were no changes in general risk taking (i.e., main effect of grade on the log likelihood of making a risky decision when the EVs of risky and safe choices are equal) for best friend (γ_Grade = 0.065, p = .21) or for parent (γ_Grade = 0.042, p = .47), but there were increases in adaptive risk taking (i.e., interaction between grade and EV on the log likelihood of making a risky decision) for best friend (γ_EVxGrade = 0.019, p < .001) and parent (γ_EVxGrade = 0.009, p = .03). For best friend versus parent effects, adolescents' likelihood of taking general and adaptive risks were not significantly different between their best friend and parent at sixth grade (ps > .25) and across grade (ps > .26). Table 3 shows the predicted trajectories of each DV, R² values as a measure of effect size (MuMIn::r.squaredGLMM package for behavioral trajectory models and r2_nakagawa::lme4 package for neural trajectory models; Nakagawa, Johnson, & Schielzeth, 2017; Barton, 2009), and multivariate analyses results.

For ROI analyses, a linear functional form best fit developmental changes in both VS and vMPFC activation during both general and adaptive risk taking for both best friend and parent. For both best friend and parent, the VS and vMPFC did not significantly change across grade during equal-EV risk taking (i.e., general risk taking; γ_Grades > −0.046, ps > .09) and EV tracking (i.e., adaptive risk taking; γ_Grades > 0.0002, ps > .15). Furthermore, there were no significant differences between best friend and parent in VS and vMPFC activation during general risk taking at sixth grade (ps > .2) and across grade (ps > .13), and during adaptive risk taking at sixth grade (ps > .19) and across grade (ps > .44; see Table 3). Table 4 shows the mean parameter estimates of ROIs at each grade.

For longitudinal whole-brain analyses, no clusters survived the corrected cluster threshold of 80 voxels in both the linear and quadratic models of general vicarious risk taking. With a more liberal threshold (p < .001, k > 20 voxels), there were linear differences in the ventrolateral prefrontal cortex (vlPFC; x, y, z = 56, 30, 10; k = 59 voxels) and the SMA (x, y, z = 20, 4, 70; k = 53) during general risk taking for best friend versus parent across grade.

Similarly, in both the linear and quadratic models of adaptive vicarious risk taking, no clusters survived the corrected cluster threshold. With a more liberal threshold, there were linear differences in the intraparietal lobule (IPL; x, y, z = 38, −50, 46; k = 74) and quadratic differences in the cerebellum (x, y, z = −16, 66, −14; k = 48), occipital area (x, y, z = −26, −88, 24; k = 37), and primary motor area (x, y, z = 0, −16, 70; k = 37) during adaptive risk taking for best friend versus parent across grade. All results should be interpreted with caution given that these clusters do not survive the corrected cluster threshold.

A linear functional form best fits developmental changes in the total points won for both best friend and parent. There were increases in the total points won for best friend (γ_Grade = 10.52, p = .048), but no changes for parent (γ_Grade = 2.31, p = .64). Adolescents did not differently win total points for their best friend and parent at sixth grade (γ_{BF − Parent} = −4.29, p = .75) and across grade (γ_{Grade_BF − Grade_Parent} = 3.38, p = .64).

Collapsing across grade, whole-brain analyses on general risk taking for best friend and parent both yielded activations in the VS (x, y, z = 14, 8, −4; best friend: k = 54824, t(330) = 13.5; parent: k = 53277, t(330) = 14.2) and the vMPFC (x, y, z = 6, 42, −14; best friend: k = 7477, t(330) = −9.2; parent: k = 20422, t(330) = −11.1). These whole-brain results suggest that general vicarious risk taking indeed recruits these hypothesized regions. Furthermore, whole-brain analyses on adaptive risk taking yielded activations in the VS (best friend: x, y, z = −10, 8, −6, k = 68, t(330) = 4.1; parent: x, y, z = −8, 12, −2, k = 49, t(330) = 3.5), but the vMPFC was not activated for either best friend or parent.

A linear functional form best fits developmental changes in behavior, neural activations, and total points won for oneself. There were no changes in general risk taking (γ_Grade = 0.03, p = .51) and total points won (γ_Grade = 8.49, p = .09), but there were increases in adaptive risk taking (γ_EVxGrade = 0.019, p < .001). The vMPFC that tracked EV during risk taking marginally significantly decreased across grade (γ_Grade = −0.0059, p = .054). All other univariate growth models did not show significant grade-related effects (γ_Grades > −0.0014, ps > .26).

To be consistent with our preregistered analyses, we reran all analyses using age instead of grade. We first winsorized extreme values to 3 SDs above or below the mean of each social target at each age. Adaptive risk taking for best friend (γ_EVxAge = 0.015, p = .001) increased across age, but adaptive risk taking for parent (γ_EVxAge = 0.006, p = .16) and total points won for best friend (γ_Age = 6.67, p = .18) did not change. The vMPFC that tracked EV during risk taking for best friend significantly increased across age (γ_Age = 0.0072, p = .03). All other results of univariate growth models using age were consistent with those using grade.

The goal of this preregistered study was to investigate how behavioral and neural processing of vicarious risk taking for adolescents' close others develop. Adolescents did not take general risks (i.e., risk preferences) differently for their best friend and parent from early to mid-adolescence. Although adolescents increasingly took more adaptive risks (i.e., sensitivity to expected reward value) for both their best friend and parent, the two trajectories also did not differ. Despite the lack of behavioral differences, the adolescent brain differentially processed risks for best friend and parent during both forms of risk taking over time.

Although theory suggests there is a social reorientation away from parents and toward peers (Nelson et al., 2005, 2016), and adolescents take more risks in their peers' presence but less so in their parents' (Telzer, Ichien, & Qu, 2015; Chein et al., 2011), our behavioral findings demonstrate that peers and parents are not differentiated when risks are targeted at these two social agents. That is, peer and parent influences may differently shape youth's susceptibility to engage in risky behaviors, but not so in their ability to understand how their own risky actions have consequences on these close others. From a social identity perspective, perhaps adolescents similarly identify with their family and friend groups in vicarious risk situations and, therefore, do not differentially engage in vicarious risks. In summary, these findings together identify vicarious risk taking as a unique form of socially occurring risk in adolescence.

Consistent with previous studies that demonstrate that EV sensitivity during risk taking increases across development (van Duijvenvoorde et al., 2015) but risk taking in equal-EV situations are stable across age (Levin, Hart, Weller, & Harshman, 2007), our study extends these findings to a social context. Our study shows that adolescents are more likely to take adaptive risks, but not general risks, for close others from early to mid-adolescence. This result may be indicative of adolescents increasingly integrating objective contextual information, such as expected monetary reward, into their decision-making process. Thus, even in the presence of social information or under the pressure of their own risky actions having consequences on others, adolescents improve their ability to fine-tune their risky behaviors to reward information. However, although previous studies have shown age-related increases in nonsocial EV sensitivity (van Duijvenvoorde et al., 2015), our findings appear to be specific to grade-related increases in EV sensitivity that occur in a social context. Developmental changes in making strategic choices for close others are perhaps a socialized process, such that there are differences in such behaviors based on school-based, social experiences.¹

Longitudinal changes in the VS and vMPFC during both risk taking, however, were not different for best friend and parent. Only one other study to date has examined vicarious reward processing for best friend versus parent across adolescence (Braams & Crone, 2017a). According to this prior study, the VS was differentially activated when gaining rewards for best friend versus parent across adolescence (Braams & Crone, 2017a). By contrast, our findings posit that the VS is not differentially activated when evaluating potential reward for best friend versus parent during a decision-making process. The developmental changes in the VS's role during vicarious reward processing may therefore rely on the psychological context of the reward, such as anticipating versus receiving reward.

Last, we found subthreshold linear changes in the vlPFC and SMA during general risk taking for best friend versus parent, and the IPL during adaptive risk taking. Both the vlPFC and SMA are broadly implicated in cognitive and behavioral control (Levy & Wagner, 2011; Chen, Scangos, & Stuphorn, 2010), as well as in regulation following social feedback among adolescents (Jones et al., 2014; Masten et al., 2009). As a result, risk preferences toward close others may be subserved by differences in self-control or the need for control, with these differences shifting with grade in school. Furthermore, the IPL is involved in other-oriented processes such as viewing positive social interactions and taking others' perspectives (van de Groep et al., 2022; Perino, Miernicki, & Telzer, 2016; David et al., 2006). Making strategic choices for close others may thus be supported by differences in attunement to close others' perspectives when decisions can benefit or harm them, with these differences again shifting with grade in school. There were also subthreshold quadratic changes in the cerebellum, occipital area, and primary motor area, which demonstrate grade-dependent differences in the integration of sensorimotor information during thoughtful, calculated risks for close others. These results, however, do not survive corrected cluster threshold and should be interpreted with caution. Taken together, brain regions implicated in regulation during general risk taking and in social cognition during adaptive risk taking differentially process various social targets of adolescents' risky actions over time.

Our study used different types of risky behaviors, a longitudinal whole-brain method to gain a deeper understanding of neurodevelopment, and a large, diverse sample of adolescents. However, this study is not without limitations. First, peer and parent influences continue to change beyond ninth grade (Brown & Larson, 2009). Thus, it is important to investigate these trajectories across and beyond high school. Second, adolescents differentially take vicarious risks depending on the gain or loss of reward (Guassi Moreira & Telzer, 2018). Given the asymmetric development of reward and loss processing (Insel & Somerville, 2018), continued work is needed to test for differences when risks result in reward decrements for peers and parents.

Our longitudinal fMRI study investigated the developmental changes in risk taking for best friend and parent. Adolescents' heightened risky behaviors are often thought to be a byproduct of their brain processing risks and rewards in a certain way. However, neural correlates of vicarious risky behaviors are grade-dependent such that similar risky behaviors toward peers and parents are subserved by diverging neural correlates at each grade. In conclusion, adolescents may similarly identify with their peers and parents from early to mid-adolescence, yet these are modulated by diverse neural changes within this developmental window.

We greatly appreciate the assistance of the Biomedical Research Imaging Center at the UNC Chapel Hill, as well as members of the Developmental Social Neuroscience Lab for assistance with study design, data collection, and analysis.

Reprint requests should be sent to Eva H. Telzer, 235 E. Cameron Avenue, Chapel Hill, NC, USA, or via e-mail: [email protected].

The preregistration for this study is available on OSF (https://osf.io/j7gsa/). Behavioral data for this study are not publicly accessible, but whole-brain group-level contrasts are available on Neurovault (https://neurovault.org/collections/13322/). The codes and outputs for longitudinal whole-brain analyses using AFNI 3dlmer are available on GitHub (https://github.com/sehjookwon/BestFriendvsParent_3dlmer).

Seh-Joo Kwon: Conceptualization; Data curation; Formal analysis; Writing—original draft. Jessica E. Flannery: Formal analysis; Writing—Review & editing. Caitlin C. Turpyn: Formal analysis; Writing—Review & editing. Mitchell J. Prinstein: Conceptualization; Funding acquisition; Writing—Review & editing. Kristen A. Lindquist: Conceptualization; Funding acquisition; Writing—Review & Editing. Eva H. Telzer: Conceptualization; Formal analysis; Funding acquisition; Writing—Original draft.

Caitlin C. Turpyn, National Institutes of Health (https://dx.doi.org/10.13039/100000002), grant number: F32DA04946. Eva H. Telzer, National Institutes of Health (https://dx.doi.org/10.13039/100000002), grant number: R01DA039923. Eva H. Telzer, National Science Foundation (https://dx.doi.org/10.13039/100000001), grant number: SES 1459719.

All participants were compensated for completing the session. In addition, all participants provided informed consent/assent and the university's institutional review board approved all aspects of the study.

Retrospective analysis of the citations in every article published in this journal from 2010 to 2021 reveals a persistent pattern of gender imbalance: Although the proportions of authorship teams (categorized by estimated gender identification of first author/last author) publishing in the Journal of Cognitive Neuroscience (JoCN) during this period were M(an)/M = .407, W(oman)/M = .32, M/W = .115, and W/W = .159, the comparable proportions for the articles that these authorship teams cited were M/M = .549, W/M = .257, M/W = .109, and W/W = .085 (Postle and Fulvio, JoCN, 34:1, pp. 1–3). Consequently, JoCN encourages all authors to consider gender balance explicitly when selecting which articles to cite and gives them the opportunity to report their article's gender citation balance.