Fig. 4.

Top: The fraction of choices correctly predicted by the best-fitting model (the Pruning ‘rho’ model). (a) All participants combined. (b) Healthy participants only. (c) Depressed participants only. Each bar depicts this as a function of the number of choices remaining on each trial. For example, the right most bar (i.e. bar ‘8’) depicts the fraction of choices at a depth of 1 on eight-choice trials that were correctly predicted by this model; the third rightmost bar (i.e. bar ‘6’) depicts both the fraction of choices that were correctly predicted by this model at (1) a depth of 1 on six-choice trials, (2) a depth of 2 on seven-choice trials and (3) a depth of 3 on eight-choice trials, and so on. Grey lines depict the full ‘Lookahead’ model. The blue dashed lines depict chance (i.e. 50%). The winning model correctly predicts choices of both depressed participants and healthy controls to roughly the same extent. Further, the full Lookahead model is only able to correctly predict decisions that are eight choices away in the sequence on roughly 50% of trials (i.e. at chance level). The winning model correctly predicts all choices to roughly the same extent, no matter how many choices are remaining. Note that these models include data from, and disregard differences between, trials in which transitions were displayed immediately after each button press and trials in which participants had to enter the entire sequence of transitions at once (i.e. so-called ‘plan-ahead’ trials.). Bottom: Parameters of the winning Pruning ‘rho’ model. (d) Specific and general pruning parameters. (e) Reinforcement sensitivity to each transition type. (f) Absolute ratio of reward (+140) to loss (−140) sensitivity. Red denotes depressed participants, green denotes healthy participants. Error bars denote 1 standard deviation above/below the mean (red) and 95% confidence intervals (green).