32131
Model-Based Approach Reveals Differences in Children’s Cooperation during Social and Non-Social Exchange
Cooperation is a complex and inherently social phenomenon, relying on our ability to learn about other humans and adapt to their strategies. In early childhood, humans become increasingly skilled collaborators as they develop more sophisticated models of other people’s intentions and behaviors. In contrast, children with autism spectrum disorder (ASD) have core deficits in inferring others’ mental states from subtle, implicit social cues that may significantly impair cooperation. Computational modelling, in particular Reinforcement learning (RL) algorithms, could reveal the mechanisms underlying cooperation in typically developing (TD) children and how these differ in children with ASD.
Objectives:
We use a computational modeling approach to characterize how TD children and children with ASD sustain cooperation with a peer (social) versus a computer partner (non-social). We hypothesize that cooperation strategies will differ between the social and non-social conditions in TD children; conversely, cooperative strategies will not differ between conditions in children with ASD.
Methods:
In two visits, children play child-friendly versions of the multi-round trust game (TG). In the TG, children have the option to either keep five coins or share any amount with their partner. Their partner receives four-times the amount of coins and can either keep everything or share half back. Children were unaware of the number of trials each game comprises.
In the initial visit, children (TD-group: N=37, age=13 years, SD=3.26) perform an IQ test and play a non-adaptive computer partner who shares 70% of the time (fixed non-social condition). In the second visit, children (TD-group: N=25, age=12 years, SD=2.6), are paired with a peer based on age, sex and IQ. Children do not know each other before the visit. They are told that they would sometimes play with the peer (social condition) and sometimes with a computer algorithm (adaptive non-social condition). In fact, they only play with their peer once.
Results:
During the fixed non-social condition, TD children share with the computer partner 61% of times on average. Children share similar amounts when knowing they play the adaptive non-social computer partner (59%). Compared to non-social conditions, children share more often when they think they play with the peer in the social condition (68%). Children are also more forgiving, re-initiating cooperation after the trustee defects, in the social condition compared to the computer conditions.
Bayesian model comparison shows that TD children’s decisions to share in the non-social and social conditions rely on different strategies. In both non-social conditions, RL models best describe their behavior. Notably, in the social condition (that only differs in the information given to the child) a simpler tit-for-tat strategy describes children’s behavior more accurately.
Conclusions:
Results indicate that a computational modeling approach can detect selective social strategies for establishing trust and cooperation in TD children. In a next step, we will test the how these strategies differ in children with ASD. Our approach could yield objective and quantifiable markers of cooperation in children with ASD that improve phenotypic classification of the disorder.