Computational mechanisms underlying the impact of Pavlovian bias on instrumental learning in problematic social media users

Abstract

Background and aims: Problematic social media use (PSMU), a potential behavioral addiction, has become a worldwide mental health concern. An imbalanced interaction between Pavlovian and instrumental learning systems has been proposed to be central to addiction. However, it remains unclear whether individuals with PSMU also over-rely on the Pavlovian system when flexible instrumental learning is required.

Methods: To address this question, we used an orthogonalized go/no-go task that distinguished two axes of behavioral control during associative learning: valence (reward or punishment) and action (approach or avoidance). We compared the learning performance of 33 individuals with PSMU and 32 regular social media users in this task. Moreover, latent cognitive factors involved in this task, such as learning rate and reward sensitivity, were estimated using a computational modeling approach.

Results: The PSMU group showed worse learning performance when Pavlovian and instrumental systems were incongruent in the reward, but not the punishment, domain. Computational modeling results showed a higher learning rate and lower reward sensitivity in the PSMU group than in the control group.

Conclusions: This study elucidated the computational mechanisms underlying suboptimal instrumental learning in individuals with PSMU. These findings not only highlight the potential of computational modeling to advance our understanding of PSMU, but also shed new light on the development of effective interventions for this disorder.

Keywords: Pavlovian bias; computational modeling; instrumental learning; problematic social media use; reward sensitivity.

Fig. 1.

Task design overview. (A) A schematic illustration of the orthogonalized go/no-go task. Four types of stimuli were presented. Two conditions were Pavlovian-congruent: go action to win reward and no-go action to avoid punishment. The other two conditions were Pavlovian-incongruent: go action to avoid punishment and no-go action to win reward. (B) Valence-action contingencies in the four experimental conditions. The optimal action (go or no-go), possible outcomes, and their probabilities after a correct response to the target (in each cell)

Fig. 2.

Model-free results of the orthogonalized go/no-go task. (A) The PSMU group showed decreased accuracy in the no-go to win condition, reflecting a higher Pavlovian bias in the reward domain compared with the HC group. (B) The PSMU group showed a higher behavioral Pavlovian bias index (congruent accuracy - incongruent accuracy) in the reward domain, but not in the punishment domain, compared with the HC group. (C) Learning curves across trials indicated that the two groups performed similarly in most conditions, except for the no-go to win condition. Error bars indicate standard error of mean

Fig. 3.

Computational modeling results of the orthogonalized go/no-go task. (A) Model comparison showed that Model 5 was the best for both the HC and PSMU groups. Lower LOOIC values indicated better model performance. ξ: irreducible noise; ε: learning rate; ρ: sensitivity to outcomes; ρ_rew: sensitivity to reward outcomes; ρ_pun: sensitivity to punishment outcomes; b: go bias; π: Pavlovian bias. π_rew: Pavlovian bias in the reward domain; π_pun: Pavlovian bias in the punishment domain. (B–C) The PSMU group showed a higher learning rate and lower reward sensitivity compared with the HC group. Posterior distributions of the group-level parameters were taken from Model 5