Fig. 4.

Compulsivity-associated state transition learning rates are suboptimal. (a) Compulsivity is associated with sub-optimally fast state transition learning in a stable environment (Study 2). We simulated agents that played the exact same task as described in Gillan et al. (2016) and plot the min-max normalized average reward. We instantiated agents with a model-based β weight that maximized reward earned but was still within the tail of the empirical range, using a selection procedure for extreme values in skewed distributions (Rousseeuw & Hubert, 2011). We set all other parameters to their group-fitted medians for distributions that were highly skewed (z-score >4, which was the lowest among statistically significant z-scores), and group-fitted means otherwise. Agents played the game 100 000 times with different state transition learning rates [sampling from (0,1) in increments of 0.1; main plot], and 1 000 000 times within a region of interest around the optimal and empirical learning rates (denoted by red box) for increased precision. The low- and high-compulsivity groups included participants scoring <–1 and >1 on the standardized scale of compulsivity factor derived in Gillan et al. (2016). Medians from each group (due to their skew) are plotted on the inset plots. (b) Compulsivity is associated with sub-optimally slow state transition learning in a changing environment (Study 3). Plots were generated using the procedure described in panel A, here applied to the Study 3 model.