Model-Based and Model-Free Decision-Making Vary According to Stage of Illness in Anorexia Nervosa

Background: Individuals with anorexia nervosa (AN) are hypothesized to experience impairments in flexible, computationally-intensive model-based decision-making and to over-rely on simpler, more automatic strategies, such as model-free decision-making. However, research investigating model-based versus model-free learning disturbances in AN has been mixed and limited by cross-sectional investigations at a single illness stage. Methods: We administered a Markov Two Stage Decision Task capturing model-based and model-free decision-making to AN (n=23), weight-restored AN (WR; n=34), and non-eating disorder control (NC; n=50) participants. Decision-making differences were examined using Reinforcement Learning and One-Trial-Back Stay/Switch models. Regressions examined cross-sectional and longitudinal relations between model-based and model-free estimates and clinical variables. Results: There were no between-group differences across model-based or model-free estimates (ps=.099-.901). However, the WR group showed some indications of model-based deficits (e.g., difficulty distinguishing rare versus common task transitions). Cross-sectionally, higher model-based estimates were associated with lower BMIs (ps=.014-.017), driven by NC participants. There was an interaction between group and One-Trial-Back Stay/Switch model-free estimates on test meal intake (p=.039); higher model-free estimates were associated with greater intake in NC, but lower intake in WR. Longitudinally, there was a significant group by model-based learning estimate interaction on follow-up BMI (ps=.027-.006); higher model-based estimates predicted higher BMIs for WR, but lower BMIs for AN at follow-up. Conclusions: Model-based and model-free decision-making and their relations with clinical variables differed according to stage of illness, as well as cross-sectional or longitudinal measurement, in AN. Greater attention to contextual variables is needed in future computational models examining decision-making in AN.