Children are full of optimism, but those rose-tinted glasses are fading-Reduced learning from negative outcomes drives hyperoptimism in children.

Introduction

Learning and knowing what is good for us is crucial for survival and a key part of development in human and non-human animals. Organisms are known to use trial-and-error learning to acquire and continuously adjust their knowledge and behaviour ( Schultz et al., 1997 ). This reinforcement learning - the process of forming and adjusting expectations based on prediction errors ( Sutton & Barto, 1998 ) - is known to converge to the optimal behaviour ( Watkins & Dayan, 1992 ). This means that trial-and-error learning will allow the formation of correct and unbiased knowledge that maximises rewards ( Körding & Wolpert, 2004 ; Sutton & Barto, 1998 ). Nevertheless, humans are known to be subject to many cognitive biases that distort these optimal representations ( De Martino et al., 2006 ; Tversky & Kahneman, 1974 ).

A prominent distortion is an optimism bias, the tendency to see the world through rose-tinted glasses and to expect it to be better than reality actually is ( Sharot, 2011 ). Even though this optimism bias may corrupt an adequate representation of one’s environment, it has been suggested to be beneficial because it boosts motivation and thus increases the likelihood of overcoming obstacles to achieve ambitious goals. Indeed, beneficial effects of optimism bias have been found in multiple domains, such as physical health ( Rasmussen et al., 2009 ), mental health ( Nolen-Hoeksema et al., 1992 ; Strunk et al., 2006 ) and professional development ( Puri & Robinson, 2007 ).

The optimism bias might be of particular importance during development in childhood and adolescence. For example, an overly optimistic view of a dream job will allow a child to pursue their ambition and overcome the many obstacles along the way. In fact, the first studies investigating abstract concept knowledge revealed that children show features of optimism biases about the future in general (Bamford & Lagattuta, 2020; Fischer & Leitenberg, 1986 ), as well as their own future knowledge ( Lockhart et al., 2017 ), and overestimate positive traits in themselves and others ( Boseovski, 2010 ; Lockhart et al., 2002 ).

However, little is known about how this optimism bias arises mechanistically, especially in youths. A pioneering study in adults has investigated this process and has shown that an optimism bias indeed arises from a bias in learning, in which better-than-expected (positive) outcomes are weighted more strongly than worse-than-expected (negative) outcomes ( Lefebvre et al., 2017 ). Using computational modelling, the authors showed that this optimism bias could be explained by an ‘optimistic learning’ bias, by means of a decreased learning rate for negative compared to positive prediction errors ( Lefebvre et al., 2017 ).

Such an optimistic learning bias could be of particular importance when investigating optimism biases during development because a growing body of evidence indicates that prediction error related learning changes substantially during development ( Cohen et al., 2020 ; Hauser et al., 2015 ; Nussenbaum & Hartley, 2019 ). In particular, several studies have suggested that overall learning from prediction errors increases during youth, although not ubiquitously so (for review cf ( Nussenbaum & Hartley, 2019 )). Whether a bias towards learning from positive compared to negative events exists during development and whether this optimistic learning bias leads to an optimism bias is unclear.

In this study, we investigate the development of optimism bias in childhood and adolescence, and how it is related to biases in reinforcement learning. We have chosen a sample that covers childhood (8-9 year-olds), early adolescence (12-13 year-olds) and late adolescence (16-17 year-olds) to understand how and when the optimism bias mechanistically changes over these key developmental stages. Using an ecologically realistic learning task ( Hauser et al., 2017 ) that assessed youth’s ability to predict and learn from effortful attainment of reward, we show that children are hyperoptimistic with an increased optimism bias compared to early and late adolescents, and this bias is driven by a depleted learning rate for worse-than-expected outcomes.

Methods

We recruited 108 participants from multiple schools across Greater London, UK. We only included participants who were in a certain age range (8-9 years old (yo), or 12-13 yo or 16-17 yo), and fluent in English. We allowed all children who provided consent to take part in the study but excluded from analysis those who had a history of neurological or psychiatric disorders, and visual impairment that was not corrected by the use of glasses or contact lenses. In total, nine participants were excluded from the analysis: 1 due to a pre-existing neurological condition (children group), 6 due to not understanding the task when asked questions about the task or not paying attention (as observed by experimenter; 2 from children group; 3 from early adolescents group; 1 from late adolescents group), 1 due to a technical problem (late adolescents group) and 1 due to low effort success rate (37.5%; children group). The final sample included 27 children (17 females; mean age = 9.32 ± 0.27), 38 early adolescents (20 females; mean age = 13.13 ± 0.31) and 34 late adolescents (21 females, mean age = 17.17 ± 0.28). All the behavioural findings reported were present when including those participants. These age ranges were selected to span late childhood, early- and late-adolescence. The groups did not differ in their age-adjusted IQ estimates (c.f. Table 1 ). We deliberately recruited participants from schools in socially diverse areas with lower socioeconomic status (SES) to counteract the currently overrepresented recruitment bias towards youth with higher SES ( Fakkel et al., 2020 ). The sample size was selected assuming similar, medium to large effect sizes based on previous developmental studies and our own study using a similar task (Bamford & Lagattuta, 2020; Decker et al., 2015 , 2016 ; Hauser et al., 2017 ; Lockhart et al., 2002 , 2017 ). All participants provided written informed consent and participants under 16 provided written consent from a parent or legal guardian in addition to their own consent. Each participant was given a gift voucher of £7. The study was approved by the Research Ethics Committee of University College London (study number: 14261/001). Different tasks from the same participants are reported elsewhere ( Bowler et al., 2021 ; Dubois et al., 2020 ; Moses-Payne et al., 2020 ). The current study was not preregistered. The data, analysis code and modelling toolbox are publicly available at https://github.com/DevComPsy/EL-development .

The testing took place in a quiet room in the participants’ school in groups of three to four students. The adolescent groups completed the experiment in about 1.5 hours. For the youngest group the testing was spread over two sessions to reduce fatigue and took about 2 hours (including age-appropriate short breaks) in total to complete. Participants completed four tasks (other three reported elsewhere; e.g., ( Bowler et al., 2021 ; Dubois et al., 2020 ; Moses-Payne et al., 2020 )), a battery of questionnaires (used in conjunction with the other tasks) and short form of the WASI-II including the Vocabulary and Matrix Reasoning subtests ( Wechsler, 1999 ) to estimate age-adjusted IQ. The participants were given both verbal and written instructions of the task before completing it. Children received longer instructions with examples to ensure that they understood the task. The order in which the tasks, questionnaires and WASI-II were administered, was pseudo-randomised across participants.

Task

The goal of this study was to investigate how optimism bias and learning changed throughout development using an ecologically realistic effortful reward attainment task. We used a modified and child-friendly version of a previously established task ( Hauser et al., 2017 ), where participants were helping an astronaut to fly a rocket to planets across the universe. They had to learn about the reward (1 to 7 gold coins) and an effort threshold that needed to be surpassed (amount of fuel the battery needed, which ranged between 42% and 92% of maximal button presses) in order to reach the planet to collect coins. Both reward magnitude and effort threshold slowly changed over time in a Gaussian random-walk-like manner. These trajectories were constructed so that reward and effort were decorrelated (cf Fig. 1 ). Importantly, the exerted effort was calibrated individually for each participant’s maximum number of presses. The maximum button presses were obtained in the practice session where the participants had to perform as many button presses as they could in 5 seconds. The calibration also included a staircase procedure, where the maximum effort was updated when the participant had exceeded their previous maximum effort in a trial.

In the beginning of each trial, the participants rated their belief about the amount of fuel needed to reach the next planet and their belief about how many coins they will get, ranging from 1 to 7. There was no time limit for reporting one’s beliefs. This was followed by 5 seconds of rapid manual button presses to fill up the battery. If the exerted effort was above the effort threshold, participants received the coins that were on display during outcome. If the participants’ effort did not exceed the threshold, a cross appeared above the number on display, which indicated that the participant did not receive any coins for that trial. Participants did not receive any explicit information about the threshold but had to learn from trial and error. After training, participants completed 40 trials.

Multiple regression analyses

To assess the factors that influenced reward belief ( Fig. 2C ), we used multiple regression (fitglm function in Matlab) to predict the reward belief at each trial, respectively. As predictors, we entered the number of points (displayed during feedback) on the previous trial (previous reward; range:1-7); and whether the force threshold was successfully surpassed on the previous trial (previous failure; failure coded as 1, success as -1). The last predictor was the participant’s reported reward belief on previous trial (previous reward belief; range: 1-7). The regression weights of the predictors were obtained for each individual and then tested for consistency across subjects using t-tests in a summary statistics approach.

Statistical analyses

We compared behavioural measures using one-way ANOVAs with age group as between-subject factor (children, early adolescents, late adolescents). Significant effects were further explored using (independent samples) t-tests. We report effect sizes using partial eta squared (η 2 ) for ANOVAs and Cohen’s d (d) for t-tests.

Computational model

To investigate the computational mechanisms underlying age-related change in reward learning, we fitted six different variants of Rescorla-Wagner models ( Rescorla & Wagner, 1972 ). We provide a summary of the models with key equations, parameter and model recovery in the Supplementary Material. Here we provide a brief description of the winning model and the key model parameters.

In this model, the subject starts with a prior belief μ 0 about how big a reward will be, which is then adjusted based on the task feedback. This parameter can be seen as a static optimism bias, reflecting subjects’ prior expectation about how good a reward will be. Subsequently, the subject will learn from the task feedback by using two learning rates: if the outcome is better-than-expected (i.e. positive prediction error), the subject will update their belief using positive learning rate (α + ), and when the outcome is worse-than-expected (i.e. negative prediction error), then the subject will update the belief using a negative learning rate (α − ). If α + >α − , then the subject learns more from positive outcomes compared to negative outcomes, in line with the precious account of optimistic learning bias ( Lefebvre et al., 2017 ). The model also incorporates a noise parameter ξ that captures the noisiness of responding.

Mediation analysis

We used mediation analysis to evaluate whether the effect of age on optimism bias was mediated by negative learning rate. We used standard notation to report mediation paths, where X represents the independent variable (age), Y represents the dependent variable (optimism bias) and M represents the mediating variable (negative learning rate). Relationship between X and M is expressed by path a , and relationship between M and Y is expressed by path b . The overall/total effect of X on Y is defined by path c and the direct effect of X on Y controlling for M is defined by path c ’. The product ab defines the indirect effect of X on Y through M. If M mediates the relationship between X and Y, then the product ab should be significantly different from zero.

We used Mediation Toolbox in Matlab ( https://github.com/canlab/MediationToolbox ; Wager, Davidson, Hughes, Lindquist, & Ochsner, 2008 ; Wager et al., 2009 ) to perform the analysis. This toolbox is used to calculate mediation analysis based on a standard 3-variable path model ( Baron & Kenny, 1986 ) with a bootstrap test for the statistical significance of the product ab ( Efron & Tibshirani, 1993 ; Shrout & Bolger, 2002 ). This toolbox tests the significance of ab using the accelerated, bias-corrected bootstrap test ( Efron & Tibshirani, 1993 ; Shrout & Bolger, 2002 ) with 10,000 bootstrap samples to test each of the a , b and ab path coefficients.

Results

We examined the developmental differences in learning and optimism bias testing three groups of young people: 27 children (8-9 year-olds), 38 early adolescents (12-13 year-olds) and 34 late adolescents (16-17 year-olds). All subjects played a child-friendly, gamified version of a previously developed learning task ( Hauser et al., 2017 ), in which subjects need to predict and learn from effortful attainment of reward. In essence, subjects needed to exert physical effort (button presses) to obtain a time-varying reward. Because both effort and reward changed over time (independently from each other), they had to constantly track and learn the changing effort demands and reward magnitudes through a process of trial and error ( Fig. 1 ).

On every trial, participants were asked to report their beliefs about the reward they could obtain (number of gold coins, ranging from 1 to 7) and the effort needed to do so (battery fuel level) using a visual analogue scale ( Fig. 1A ). This allowed us to assess the subjects’ beliefs about both effort and reward. After reporting their belief, subjects needed to charge the battery of a space rocket to the level they believed was needed to reach the next planet quickly alternating between two button presses (physical effort exertion; fixed total charging time of 5 seconds). Subsequently, subjects were informed whether they exerted enough effort (i.e. reached the planet), and how many points they had won (or would have won). In the remainder of this paper, we will focus on reward learning and optimism bias, but detailed analyses on effort learning can be found in the supplementary material.

Before investigating optimism bias, we were interested in the overall performance of this task. We constructed the task so that actual performance was equal across all ages, thus preventing performance from confounding optimism bias. Concretely, we used a staircase procedure for each individual’s physical effort exertion to account for differences in physical or other abilities (cf Methods for details). This allowed the task to be challenging and achievable for everyone, whilst ensuring similar performance across groups. Indeed, we found that the number of successful trials ( F (2,96) = 1.14, p = 0.323, η 2 = 0.023, Table S1 ), and the total points won in the task ( F (2,96) = 1.33, p = 0.268, η 2 = 0.027) did not differ between the age groups. Furthermore, the average effort exerted (relative to the individual’s maximum effort) did not differ between groups ( F (2,96) = 2.16, p = 0.121, η 2 = 0.043), and neither did the variability of the effort exerted as measured by the effort standard deviation ( SD ; F (2,96) = 1.65, p =0.198, η 2 = 0.033). These findings thus allow us to compare optimism bias and reward learning between groups without having to account for biases in their performance.