Skip to main content
. Author manuscript; available in PMC: 2018 Sep 12.
Published in final edited form as: Nat Neurosci. 2018 Mar 12;21(4):617–624. doi: 10.1038/s41593-018-0104-6

Figure 1. Post-decision evidence task and computational framework.

Figure 1

A) Task design. Participants made an initial left/right motion discrimination judgment, after which they saw additional post-decision motion of variable coherence moving in the same direction as pre-decision motion. They were asked to rate their confidence in their initial choice on a scale from 0% (certainly wrong) – 100% (certainly correct). Confidence scale steps were additionally labeled with the words “certainly wrong”, “probably wrong”, “maybe wrong”, “maybe correct”, “probably correct”, “certainly correct” (not shown). B) Bayesian graphical model indicating how pre- and post-decision motion samples are combined with the chosen action to update an estimate of decision confidence. C) Simulated decision variables from the model in (B) showing a distinction between updating evidence in the coordinate frame of motion direction (left panel) and choice accuracy (middle panel) as a function of post-decision motion strength and choice. A change in log-odds correct (“post-decision evidence”; PDE) is revealed by a qualitative interaction between post-decision motion strength and choice accuracy (middle panel). The right panel indicates the expected mapping between log-odds correct and both final confidence/decision value. Confidence and value are dissociated on change-of-mind trials (confidence < 0.5) through use of a quadratic scoring rule, which rewards subjects for both being confident and right, and unconfident and wrong.