Impaired arbitration between reward-related decision-making strategies in Alcohol Users compared to Alcohol Non-Users: a computational modeling study.

Reinforcement learning studies propose that decision-making is guided by a tradeoff between computationally cheaper model-free (habitual) control and costly model-based (goal-directed) control. Greater model-based control is typically used under highly rewarding conditions to minimize risk and maximize gain. Although prior studies have shown impairments in sensitivity to reward value in individuals with frequent alcohol use, it is unclear how these individuals arbitrate between model-free and model-based control based on the magnitude of reward incentives. In this study, 81 individuals (47 frequent Alcohol Users and 34 Alcohol Non-Users) performed a modified 2-step learning task where stakes were sometimes high, and other times they were low. Maximum fitting of a dual-system reinforcement-learning model was used to assess the degree of model-based control, and a utility model was used to assess risk sensitivity for the low- and high-stakes trials separately. As expected, Alcohol Non-Users showed significantly higher model-based control in higher compared to lower reward conditions, whereas no such difference between the two conditions was observed for the Alcohol Users. Additionally, both groups were significantly less risk-averse in higher compared to lower reward conditions. However, Alcohol Users were significantly less risk-averse compared to Alcohol Non-Users in the higher reward condition. Lastly, greater model-based control was associated with a less risk-sensitive approach in Alcohol Users. Taken together, these results suggest that frequent Alcohol Users may have impaired metacontrol, making them less flexible to varying monetary rewards and more prone to risky decision-making, especially when the stakes are high.