Modeling decision dynamics disentangles working memory, cognitive control and reinforcement learning and reveals clinical differences

Abstract

Various neurocognitive processes contribute to reinforcement learning (RL) and decision making, requiring careful task designs and computational modeling to disentangle them. People rely on capacity-limited working memory (WM) to rapidly and flexibly adapt behavior during learning, together with slower incremental RL processes that support robust long-term retention. The RLWM paradigm was developed to isolate these contributions by manipulating WM demands during learning. Empirical and computational advances support the separability of such processes, but only imperfectly, because a given choice can often equally be attributed to use of WM vs. RL. Here, we show that jointly modeling decision dynamics – accounting not only for choices but also response time distributions, together with hierarchical Bayesian parameter estimation – improves the ability to disentangle separable learning and cognitive control processes. Despite added complexity, models with decision dynamics improved parameter recovery and yielded accurate out-of-sample prediction for choices in a held-out test phase after learning. In contrast, models fit only to choices produced an inflation of RL learning rates and failed the out-of-sample test. Moreover, joint modeling also revealed a novel neurocognitive process by which participants proactively widen decision boundaries to increase response caution under increased WM load. Applying our model to patients with schizophrenia revealed a deficit in proactive control mechanisms as well as slowed incremental RL, which were masked by previous choice-only models, while replicating previously established WM deficits. In summary, we show how joint modeling enables accurate, model-based and mechanism-oriented computational phenotyping of psychopathologies.