Figure 2.

Results Stay-Switch Analysis

(A) Subjects' task behavior showed characteristics of both model-free and model-based influences, demonstrating that subjects combined both strategies in the task. The reward × transition interaction (a measure of the extent to which subjects consider the task structure) was significantly larger in L-DOPA compared to placebo, indicating stronger model-based behavior. Error bars represent SEM. (B) Difference in stay probability between L-DOPA and placebo condition, corrected for the main effect of drug. The observed interaction indicates a shift toward model-based choice (see F) but shows no resemblance to any of the three effects implicating the model-free system (see C–E). (C–F) Illustration of expected differences in stay probability for hypothetical drug effects. See Figure S1 and Table S1 for validation of these hypotheses. (C) Trials after uncommon transitions (second and fourth bar) are discriminatory between model-free and model-based choice, whereas both models make equal predictions for trials after common transitions (cf. Figure 1C). A shift toward model-free control would be indicated by an increased propensity to stay with the chosen pattern after uncommon rewarded trials and an increase in switching after uncommon unrewarded trials. (D) Stronger or faster model-free learning would increase the reward-dependent effect and be expressed as a general increase to stay after rewarded trials and general decrease to stay after unrewarded trials. (E) A selective enhancement of positive updating paired with impairment in negative updating might not change mean-corrected stay probabilities. This is because enhanced positive updating leads to a stronger propensity to stay after rewarded trials, while impaired updating of unrewarded trials decreases the propensity to switch after such trials. (F) Opposite to (C), a shift toward model-based control is expressed by enhanced sensitivity to the task structure.