Introduction
Peers impact almost all aspects of adolescent lives, from the more trivial, such as taste in music and clothing, to the more serious, such as the use of illicit drugs or engaging in unprotected sex [Steinberg, 2008]. These latter, riskier, choices may carry life-long consequences for the adolescent and bring significant cost to society. It is empirically well established that the presence of peers influences risky behavior in adolescence [Gardner and Steinberg, 2005; Chein et al., 2011; Pfeifer et al., 2011; Smith et al., 2014], but the underlying developmental processes remain poorly understood. Understanding these processes, however, is important for at least two reasons. First, empowering adolescents to become more competent decision-makers will be more effective if we succeed at tailoring interventions to their developmental affordances. Second, we can only identify these affordances if we succeed at linking adolescent neuronal and cognitive development with adolescent behavior across different social contexts.
Here we argue that this link cannot be made without formal models of adolescent peer influence. In this article we thus aim to take the first steps toward a quantitative and testable framework of adolescent social influence. Adolescence is marked by several developmental changes which offer multiple biological explanations of social influence on adolescent decision making. We refer to the current theoretical perspectives of these changes as “verbal models.” Verbal models are distinct from formal models in that they do not make quantitative predictions. In order to establish formal models that do make quantitative predictions, we first review existing verbal models and the associated empirical findings about social influence in adolescents, focusing on risky decision-making. We identify three verbal models of social influence which can be subject to developmental change; these are then formalized by grounding them in expected utility theory. Next, we show that our formal models can reliably be recovered and therefore can be used to compare hypotheses via quantitative model comparison. Finally, we fit these models to existing data and reveal previously overlooked patterns of peer influence. We conclude with a discussion on how the specificity provided by this formal approach contributes to a deeper understanding of the developmental processes behind social influence.
Social Influence on Adolescent Decision-Making
We identify three main families of verbal models in the existing literature, hereafter named as follows: [i] social motivation model, [ii] reward sensitivity model, and [iii] distraction model. These three models focus on two distinct neurodevelopmental explanations of altered decision-making during adolescence. Social motivation verbal models stress the importance of the developing “social brain.” The other two verbal models [reward sensitivity and distraction] both emphasize the relatively slow maturation of cognitive control systems. Previous works that fall into the reward sensitivity family of verbal models often refer to it as “dual-systems” models, as they also stress the relatively fast maturation of reward-processing brain regions and explain adolescent behavior with the maturational imbalance between reward processing and cognitive control brain regions [Casey et al., 2008; Steinberg, 2008; Geier et al., 2010; Shulman et al., 2016]. By contrast, the distraction model has a single focus on the development of cognitive control. Our subsequent review of the existing experimental evidence shows that all three of these families of verbal models are currently equally well supported in the literature, even though each model provides a different explanation for similar observations.
Verbal Models: Social Motivation
The first verbal model we consider states that adolescents have increased social motivation. Demonstrating risky behavior, or conforming to behavior of the peer group, are considered ways to reach these social goals. In other words, social motivation models assume that during adolescence there are situations where a high social value is attributed to displaying risky behavior [Crone and Dahl, 2012; Ruff and Fehr, 2014] which is independent from the non-social value of the outcome [e.g., money].
Verbal Models: Reward Sensitivity
The verbal reward sensitivity model is based on research which suggests that adolescence is the time where fast maturation of reward processing brain systems coincides with relatively slow maturation of cognitive control systems. According to the reward sensitivity model, the biological imbalance between these two systems gives rise to risky adolescent decision-making [Casey et al., 2008; Ernst et al., 2015; Shulman et al., 2016]. Here we will not address the debate concerning the validity [Pfeifer and Allen, 2016] or the different variants of these models [Casey et al., 2008; Steinberg, 2008; Larsen and Luna, 2018]. Instead, we focus on the element that is suggested to be most relevant for understanding developmental changes in peer influence: reward sensitivity. Reward sensitivity states that social influence has such dramatic effects on adolescent risk-taking because a social context “may sensitize the incentive processing system to respond to cues signaling the potential rewards of risky behavior” [Chein et al., 2011, p. 2]. Indeed, Chein et al. [2011] showed that while being observed during a risk-taking task, brain regions related to reward processing were more active in adolescents than in adults. This was interpreted as evidence for a reward sensitivity model as it suggests that, in adolescents, the social context itself leads to changes in the processing of rewards in general.
Verbal Models: Distraction
The relatively slow maturation of cognitive control brain regions forms the basis of a third verbal model that we call “distraction model.” Here, maturational imbalance and arousal is not only specifically associated with altered representations of reward but more generally with poor self-control and diminishing cognitive skills in emotionally salient situations [Dumontheil, 2016]. This lack of self-control can lead adolescents to show more erratic or distracted behaviors in a social as compared to a solitary context. The distraction model does not assume any changes in value computation, but rather suggests that behavioral changes are due to stochasticity in the decision process.
Social motivation, reward sensitivity and distraction models do not assume mutually exclusive processes. Although it is plausible that the defining processes emphasized in each of these models simultaneously impact peer influence, it is important to examine which are most relevant in a particular context.
This is essential because different models provide different footholds for interventions. For example: if adolescent risk-taking is subject to social motivation it can be fruitful to provide other, less risky, means to acquire social status for instance by using meaningful roles interventions [Ellis et al., 2016, see also: Yeager et al., 2018]. Adolescent reward sensitivity suggests it is useful to prohibit teens from gathering in risky situations. For instance, many states in the United States and Canada prohibit teenage drivers from taking other teenage passengers along. Distraction suggests that training in mindfulness and meditation are a good prospect for increasing desirable behaviors in adolescence [Kuyken et al., 2013]. These implications for interventions underscore how crucial it is to comprehend the most relevant determinants of adolescent behavior in a given context. We therefore inspect experimental work which manipulated aspects of social contexts with respect to the three verbal models of adolescent social influence: [i] social motivation, [ii] reward sensitivity, and [iii] distraction.
Seeing and Being Seen – Empirical Studies of Social Influence
Despite the complexity of social exchange, studies investigating social influence can be roughly divided into two types of situations: those where the participant observes others and those where the participant is being observed. In the light of this distinction, we review experimental studies about peer influence in adolescent risky decision-making.
Observing Others
When uncertain of what to do, observing the behavior of others can help with making a decision. Monetary lotteries are often used as an experimental setting with uncertain prospects, wherein the effect of observing the behavior of others can be investigated. In such experiments, participants observe others’ previous decisions [Blankenstein et al., 2016; Reiter et al., 2019] or receive explicit advice [Haddad et al., 2014] while making private decisions. These studies suggest that the impact of social information is greatest in early to mid-adolescence and then declines with age. Notably, in a recent study, adolescents were influenced more by safe than by risky advice [Braams et al., 2019]. However, currently evidence seems most in line with models that emphasize social motivation, as an increase in safe decisions is not predicted by reward sensitivity models. A small increase of participant safe choices in studies such as Braams et al. [2019] however, could also be attributed to a greater distraction during adolescence. Notably, none of these studies provided adolescents with information about the outcomes of others’ decisions. In real life, such outcomes are observable; there is evidence that observing others’ risky real-world behaviors, such as smoking or drug use, increases the likelihood of adolescents to adopt these behaviors themselves [Clark and Lohéac, 2007; Liu et al., 2017]. This can reasonably be explained using social motivation models, if adolescents anticipate peer approval. It can also be explained with reward sensitivity models when assuming that the rewarding properties of risky behaviors [smoking] themselves become subjectively more rewarding in this social context.
In sum, experimental results from paradigms in which participants observe the choices of others are sometimes more consistent with the social motivation model, and sometimes more consistent with reward sensitivity. Paradigms designed for testing distraction models when observing others are underrepresented, so their pertinence here cannot yet be sufficiently evaluated. As such, which verbal model family best accounts for adolescent behavior when they observe others remains unclear.
Being Observed
When a decision maker is observed by another individual, risk-taking also sends a social signal to the observer [Baker and Maner, 2009]. For instance, adolescents can show how “cool” they are by taking extreme risks, or signal that they are or want to be part of a group by mimicking its members’ risk-taking behavior. Thus, if adolescent behavior in peer contexts is sending a social signal to their peers, their beliefs about the risk-norms of observing peers should impact their behavior. In line with this, one study found that exposing teenagers to risk-accepting peers increased their risky driving while exposure to a risk-averse peers did not [Shepherd et al., 2011]. Further, there is evidence that risk perception and understanding of social norms are important predictors of adolescent risky driving [Carter et al., 2014]. Social motivation models can therefore explain increased risk-taking in paradigms when participants are being observed.
However, even without assuming complex social motivation, behavior change in a social context was traditionally explained with social facilitation theory [Zajonc, 1965], which foreshadowed both reward sensitivity, and distraction models by one principled observation: Being observed induces arousal.
The reward sensitivity model suggests that arousal leads to altered reward processing, making risk-taking more appealing. Indeed, most developmental studies of how being observed impacts risk-taking report an increase in the number of risky choices made by adolescents in social contexts [Gardner and Steinberg, 2005; Chein et al., 2011; Smith et al., 2014; Somerville et al., 2018]. In the context of social facilitation theory this increase in risk-taking can be seen either as facilitation, for example by increasing explorative behaviors and socially acceptable risk-taking, or impediment, when the risks are illegal and dangerous [Duell and Steinberg, 2019]. In one remarkable neuroimaging study along these lines [Chein et al., 2011], found evidence for the reward sensitivity model. The presence of another person increased activity in the ventral striatum when adolescents received rewards, as compared to a solitary reward condition. This was true for adolescents but not for adults.
However, in another variant of social facilitation theory [Sanders et al., 1978], social arousal is thought to result in distraction from the task at hand, thus mostly resulting in detrimental or sub-optimal behavior. In fact, there is evidence that arousal leads to decreased cognitive control, which results in more distracted behavior in decision-making tasks [Starcke and Brand, 2012]. There is also evidence that distraction accounts for typical adolescent behavior in some experimental paradigms. For instance, Dumontheil et al. [2016] demonstrated reduced reasoning abilities in adolescents when monitored by peers. Similarly, another study found that adolescents who showed poor conflict monitoring in an emotionally arousing Stroop task also turned out to be risky drivers in a driving simulator [Botdorf et al., 2017].
Consequently, changes in risky choice while being observed could be the result of the motivation for social signaling, of arousal-based reward sensitive decisions, or distraction, and each of these three processes possibly has a different developmental trajectory. Merely observing an increase in risky decisions in adolescents seems insufficient to specify which underlying psychological process is most relevant.
In sum, different studies have emphasized different models and found results in favor of each. This holds for paradigms when adolescents are observing others and even more for paradigms in which they are observed. These mixed results may be due to the fact that each study has used different experimental paradigms with large variations of the key variables [e.g., known risk vs. uncertainty, best friend vs. unknown peer] and most studies do not directly compare different social contexts in order to identify if they are subject to different psychological processes [but see Somerville et al., 2018]. Another reason for the diversity of experimental findings, which can also be attributed to variations in key variables, is that studies likely differ in their affective content. For instance, the affective content of a study on social influence which only uses information about choices of strangers who are not currently present is fundamentally different from a study wherein social influence is examined by looking at changes in behavior in the presence of a close friend. The distinction between affectively “hot” and “cold” contexts is a useful heuristic to understand adolescent risk-taking. There is evidence that adolescents make more risky choices in “hot” contexts. Notably, reward sensitivity and distraction models explain behavior change via affect [arousal] as well [Blakemore and Robbins, 2012; Rosenbaum et al., 2018]. In order to comprehend adolescent socio-emotional development, we need to better understand how affect and social processing interact and impact each other. We argue that the specificity provided by formal modeling might help to disentangle these important components in developmental research, similar to the field of computational psychiatry [Montague et al., 2012; Huys et al., 2015; Jolly and Chang, 2018].
However, before further elaborating on the benefits of formal models in developmental research we first want to pay credit to the neuroscience of adolescent development. Neuroimaging studies may provide better clues to what extent different processes underlie behavior. In addition, it may be possible to generate more specific hypotheses about which psychological processes are involved based on the localization of neural activation.
Social Influence and Brain Development
Most verbal models of adolescent social influence are inspired by recent findings from developmental neuroimaging. Here we will review some of those findings and indicate to what extent they support existing models. Given that neural activation is a more direct reflection of the processes underlying behavior, neuroimaging may be instrumental to identify which process is most relevant in which context.
Adolescent social motivation models are supported by findings about the development of a network of brain regions associated with social cognition. This network, sometimes subsumed as the “social brain,” continues to develop during adolescence [Mills et al., 2014]. The most prominent regions of this network are the temporo parietal junction [TPJ], the posterior superior temporal sulcus [pSTS], the anterior temporal cortex [ATC], and the medial prefrontal cortex [mPFC]. When reasoning about others, the social network seems more active in adolescents than in adults or children [van den Bos et al., 2011]. Further, in a study by Somerville et al. [2013] observed by others resulted in increased mPFC activity in adolescents. However, activity in these regions is not unique to social processing. For instance the same study found an adolescent increase in connectivity of the mPFC with striatal brain regions, which are relevant for processing rewards. Further, the mPFC itself is also involved in basic reward processing [Harris et al., 2007; Silverman et al., 2015]. Taken together, the increased mPFC activity when being observed can also be interpreted as supporting the reward sensitivity model.
Neural correlates of the role of adolescent reward sensitivity in non-social contexts were recently examined in a meta-analysis [Silverman et al., 2015]. This study estimated an increased likelihood of activation in adolescents within a broad range of regions associated with reward processing. These comprise the ventral and dorsal striatum, subcallosal cortex, insula, and amygdala as well as the anterior cingulate cortex [ACC], the posterior cingulate cortex [PCC], and the paracingulate region and the medial prefrontal cortex [mPFC]. One study found increased activity in the ventral striatum when adolescents where taking risks in a social but not in a solitary context, whereas this activity difference was not found in adults [Chein et al., 2011]. These results are evidence in favor of the reward sensitivity model, but there are multiple possible interpretations. For instance, increased reward related neural activity could either be the result of altered reward perception or of an orthogonal, social value of conforming to a norm. Both social and non-social value is represented in the striatum [Ruff and Fehr, 2014]; both mechanisms can lead to more risky behavior in certain tasks.
Distraction models emphasize the development of the lateral prefrontal cortex [lPFC] and the inter parietal sulcus, which make up the main regions of the cognitive control network. Studies based on the distraction model consistently found increased IPS activation during cognitive control in adolescents, whereas lPFC findings were mixed [Dumontheil, 2016]. One study investigating the effects of social context on neural processing while performing a relational reasoning task found that adolescents recruited this cognitive control network more strongly than adults when an audience was present, while performance changed in a similar magnitude for both age groups [Dumontheil et al., 2016]. This result also allows for multiple interpretations. Adolescents may be more distracted, but on the other hand it may also be that they exert more control to counteract their distraction, and thus stay on par with adult’s behavior. The fact that they exert more control could potentially be the result of an increased motivation to perform well while observed by others.
In summation, we have seen that all verbal models are supported by neuroimaging research. Different models emphasize the development of different brain networks, but these networks often overlap with respect to functional and structural components. As long as a one-to-one mapping between cognitive and neural processes is not given, it is not justifiable to make the reverse inference about the presence or absence of a cognitive process purely on the basis of observed, or unobserved neural activity [Poldrack, 2006, 2011].
We do not wish to discredit the existing studies on neural correlates of adolescent peer influence; On the contrary, we believe that these are excellent and well-designed neuroimaging studies. In combination with appropriate experimental control conditions, reverse inference is valid and insightful [Hutzler, 2014]. However, experimentally isolating a cognitive process becomes exponentially difficult when the processes in question increase in complexity. Different attempts have been suggested to attenuate the issue, such as large scale brain decoding [Poldrack, 2011; Yarkoni et al., 2011], using functional localizers [Saxe et al., 2006], and formal modeling [Marr and Poggio, 1976; Montague et al., 2012; Stephan et al., 2015; van den Bos and Eppinger, 2016; Hauser et al., 2018]. None of these strategies will completely solve the problem of reverse inference, however, each may increase our confidence in reliably identifying the neural correlates of a particular cognitive processes. This article is motivated by the advantage of formal models; in what follows, we will illustrate how verbal models of social influence in adolescence might be translated into formal ones.
Formal Models of Social Influence
Here we demonstrate how the three verbal models about adolescent socioemotional development which we introduced earlier can be formalized as variations of expected utility models. We then show that model comparison can be used to infer underlying social mechanisms. The rationale behind formal modeling of cognition is that in order to identify if behavior is consistent with a proposed cognitive process, we need to formulate algorithms that represent the process mathematically. Comparing the behavior of the algorithms with actual behavior observed in participants can subsequently be used to quantify support for the hypothesis which is represented by the algorithm. In this section, we aim to translate verbal models of adolescent development into formal ones. However, current models often lack the details required in order to be directly translated into formal models. To formalize the models, we have therefore made several assumptions rooted in expected utility theory. The model space that we present here is not exhaustive. Nevertheless, the current framework illustrates how formal modeling can be used in developmental science, and provides a strong starting point for developing more elaborate models. More importantly, it enables precise discussions on which models are favored by existing experimental data. To formalize models of adolescent decision making First, we address how risk seeking behavior is understood within the expected utility framework in order to familiarize the reader with its’ assumptions. Then we extend these models with parameters that can be read as social sensitivity, reward sensitivity, and distraction. This finally enables us to test models of adolescent development against one another, even within the same experiment.
Expected Utility
The first assumption of expected utility theory is that people have a subjective experience of objective rewards. For instance, the first dollar someone ever earns is worth more to them than the hundredth. The change in wealth from nothing to $1 feels different from the change in wealth from $99 to $100. This transformation of objectively equal values [$1 in both cases] into a subjective utility is often modeled by a power function borrowed from psychophysics [Helmholtz, 1896], where it is used to describe the non-linear relationship between subjective psychological experience of a stimulus intensity and the objective physical intensity of the stimulus:
U=Vρ,[1]
Where V denotes the objective value of a reward and ρ determines the convexity of the utility function [Figure 1]. Often times this parameter is referred to capturing “outcome” or “reward sensitivity” of an individual [Kellen et al., 2016]. When considering risky choices rewards are not certain; they occur probabilistically. The subjective utility of a probabilistic reward is then simply described as:
Figure 1. Verbal models of social influence during adolescence, and how they map to our taxonomy of formal models.
Where V denotes the objective value of a reward and ρ determines the convexity of the utility function [Figure 1]. Often times this parameter is referred to capturing “outcome” or “reward sensitivity” of an individual [Kellen et al., 2016]. When considering risky choices rewards are not certain; they occur probabilistically. The subjective utility of a probabilistic reward is then simply described as:
EU=p*V ρ,[2]
where p denotes the probability of the reward. Note that in more elaborate models, such as cumulative prospect theory, the probability itself is also transformed to a subjective probability weight [Tversky and Kahneman, 1992]. Although this would allow for even more detailed insights in developmental differences in risky behavior [Engelmann et al., 2012], we do not further consider subjective probability here, as it would exponentially increase our model space and thus not serveour purpose.
When individuals are repeatedly presented with the same choice options, their decisions will most likely differ from oneanother. Consequently, we need to account for this probabilistic nature of choice in a model of behavior. To achieve this, a model for choosing between two rewards feeds the difference between reward utilities into a sigmoid function, through which we obtain an estimate of the probability that a decision maker chooses one option over another
pCho oseRisk=11+e-[EUrisk-EUsafe]τ *.[3]
Here, τ accounts for individual differences in choice sensitivity. The smaller τ the less sensitive the decision maker is to the expected utility differences [and the more random the choice pattern appears]. We now turn to examine how different models of social decision making can be represented within this framework.
Modeling Social Influence
In our earlier example, we used the subjective value of objective monetary amounts as the key variable for decision making, but there is ample evidence that people also attribute utility to social outcomes such as fairness [Fehr and Schmidt, 1999] and social status [van den Bos, 2009]. Furthermore, there is evidence that humans integrate value information from social and non-social sources into a common currency when making a choice [see Ruff and Fehr, 2014, for a review]. Consequently, the expected utility framework can be extended to include social rewards and represent social behavior.
Social Sensitivity
Social rewards, such as belonging or expected status gains, can add to the expected utility associated with a non-social decision, because the prospects of social and non-social rewards are combined by the brain when making a choice [Ruff and Fehr, 2014].Within expected utility theory, the changed valuation of an option due to the presence of social information can be expressed as a single parameter that shifts subjective utility. For example, if we consider a typical experiment where there are two options, a relatively safe option and a risky option [defined by outcome variance differences]. A social signal, for instance seeing that a peer chose the risky [safe] option, contributes to the utility of the risky [safe]option, while the expected value of the choice option and reward sensitivity remains the same [Chung et al., 2015]. This can be implemented with a single additional parameter:
EUSocial=p*Vρ+ψ, [4]
where ψ corresponds to the impact of social information on risky and safe choice options. We call this model “symmetric social influence model.” The larger ψ the more likely the participant is to move into the direction of the social information [see Figure 2A].
Figure 2. Two utility functions which are used to model reward sensitivity and risk-taking. The x axis depicts the expected value of potential choice options. The y axis shows the subjective utility of these expected values given different reward sensitivity parameters. [A] A convex utility function generated by ρ = 1.7. The difference between reward magnitudes is subjectively amplified, which makes it more attractive for the individual to take risks in order to obtain higher rewards. [B] A concave utility function generated by ρ = 0.3. Risk aversion occurs here because potential rewards are compressed, therefore more similar to each other and in turn it will be less attractive to take a risk in order to obtain the higher reward. The black lines illustrate that while the difference in expected values is equal in both graphs, the difference in subjective utility of these options is smaller in the right figure. Axis ticks and labels are not shown to, to emphasize the relative, not the absolute difference as exponential functions scale very differently.
It is likely that social information has asymmetric effects on behavior depending on whether social information favorsrisk aversion or risk seeking. For instance, Braams et al. [2019] showed that risky advice had less impact than safe advice. This can be captured by adding two independent parameters to the utility function that vary depending on whether social information favors safe or risky choices [see Figure 2B].
EUSocialRisk=p*Vρ+ψrisky∀SocialSignal=Risky,
EUSocialsafe=p*Vρ+ψsafe∀SocialSignal=Safe.[5 ]
We call this model “asymmetric social influence model.” Note that the precise interpretation of ψ depends on the specifics of the experiment. In an experiment where the participant is observed it could represent the expected value of gaining status by taking more risks. In an experiment where the participant observes, social information can reduce the participants uncertainty about what to choose, which will then be reflected in ψ and in yet another experiment, Ψ can represent the value attributed to conforming to the behavior of others [e.g., status vs. belonging motivation]. In addition, such a framework offers insight in how different aspects of the outcomes are weighted [e.g., money vs. social gains].
Reward Sensitivity
Developmental theories on social impact that focus on imbalance suggest that in a social context, rewards are valued more by adolescents because the socially induced arousal triggers reward-processing brain regions [Chein et al., 2011]. Reward sensitivity is a basic feature of expected utility models; it is governed by parameter ρ [see Equation 1]. This parameter has already been used to characterize individual and developmental differences in risk attitudes [e.g., Blankenstein et al., 2016; van den Bos and Hertwig, 2017]. To capture changes in reward sensitivity due to social facilitation one can add a parameter ω to the “reward sensitivity” part of the utility function:
EU social=p*V[ρ+ω]|ω∈ℝ:ω>0. [6]
The larger ω the more risk seeking an individual becomes [see Figures 1, 2C]. This equation will henceforth be called “reward sensitivity model.” In our reading of verbal reward sensitivity models, ω will never be smaller than 0 given that it is the expectation that is there is an increase, not a decrease, in risky behavior due to arousal.
Distraction
Other work emphasizes that arousal in social situations creates distracting goal conflicts, especially for adolescents [Dumontheil, 2016; Dumontheil et al., 2016; Botdorf et al., 2017; Breiner et al., 2018]. For choices that are value- or preference-based, it is hard to judge whether a decision results from distraction or inattentiveness; there is no objectively correct benchmark to evaluate correct and incorrect responses. However, formal modeling provides the means of unmasking choice stochasticity unique to social contexts that could otherwise be falsely interpreted as an increase or decrease in risk taking. Distraction or inattentiveness would lead to an increase in choices that are less determined by expected value. In decision models this kind of behavior is often captured by a “trembling hand” choice rule [Loomes et al., 2002]. This rule modifies the choice function by adding a fixed probability that the individual does not use expected utility to guide their choice, but rather chooses randomly. To capture this increase in distraction we can estimate how this probability of choosing randomly increases in the social context:
pChooseRiskSocial =[1-ζ]11+e-[EUrisk-EU safe]*τ+ζ2|ζ∈ℝ: 0