Volatility Estimates Increase Choice Switching and Relate to Prefrontal Activity in Schizophrenia

Background: Reward-based decision making is impaired in patients with schizophrenia (PSZ), as reflected by increased choice switching. The underlying cognitive and motivational processes as well as associated neural signatures remain unknown. Reinforcement learning and hierarchical Bayesian learning account for choice switching in different ways. We hypothesized that enhanced choice switching, as seen in PSZ during reward-based decision making, relates to higher-order beliefs about environmental volatility, and we examined the associated neural activity. Methods: In total, 46 medicated PSZ and 43 healthy control subjects performed a reward-based decision-making task requiring flexible responses to changing action–outcome contingencies during functional magnetic resonance imaging. Detailed computational modeling of choice data was performed, including reinforcement learning and the hierarchical Gaussian filter. Trajectories of learning from computational modeling informed the analysis of functional magnetic resonance imaging data. Results: A 3-level hierarchical Gaussian filter accounted best for the observed choice data. This model revealed a heightened initial belief about environmental volatility and a stronger influence of volatility on lower-level learning of action–outcome contingencies in PSZ as compared with healthy control subjects. This was replicated in an independent sample of nonmedicated PSZ. Beliefs about environmental volatility were reflected by higher activity in dorsolateral prefrontal cortex of PSZ as compared with healthy control subjects. Conclusions: Our study suggests that PSZ inferred the environment as overly volatile, which may explain increased choice switching. In PSZ, activity in dorsolateral prefrontal cortex was more strongly related to beliefs about environmental volatility. Our computational phenotyping approach may provide useful information to dissect clinical heterogeneity and could improve prediction of outcome.