Different computational signatures of rule- and stimulus-based biases.

Each panel shows best-fitting evidence-accumulation (ordinate) and starting-point biases (abscissa), plotted per subject (points). A Data from the rule-based condition, for each of the three prior cues (blue: low, black: neutral; red: high). Lines are linear fits (p B Data from in the stimulus-based condition (no linear relationship). C Data from the rule-based biases of the mixed condition, plotted as in A. D Data from the stimulus-based biases of the mixed condition, plotted as in B (no linear relationship). C and D show results for mixed-condition 1, which included block-level prior cues; results were similar in mixed-condition 2, which had trial-wise priors. E Data from A (the low- and high-prior conditions) plotted with respect to optimal performance contours from the rule-based condition. The average 97% contours (curved lines) demarcate the area of the parameter space in which choice accuracy is at least 97% of the maximum possible accuracy. Accuracy was predicted by the DDM applied to the actual task sequences. Closed/open circles indicate that the data fell within/outside of the subject-specific 97% contour. Lines are linear fits as in A, and shaded regions are 95% confidence intervals. F Same as E, but for the stimulus-based condition. The 97% contour is for the context of two low tones before the test tone (similar results were obtained for two high tones but not shown).