Abstract
Accurate confidence judgments are important for successful decision-making. Previous work has studied to what degree the computations underlying confidence judgements, a type of metacognition, are sensitive to different noise sources (Zylberberg et al. 2016, Castanon et al. 2019). In this work, we asked whether metacognitive computations are sensitive to, and can compensate for, temporal biases during evidence integration tasks. Our experimental design leverages recent work that showed how the temporal bias of human observers could be reliably manipulated by changing the stimulus statistics (Lange et al. 2021). We compared the data from 10 human observers in a dual report orientation discrimination and confidence judgment task across two stimulus conditions. In the first condition, observers showed a primacy bias, overweighting evidence early in the trial, while in the second condition, observers weighed evidence approximately equally (like the ideal observer for our task). Observers’ confidence ratings were similarly well calibrated in both conditions, with the ratings of individual observers predictive of accuracy even when conditioned on stimulus strength, a strong demonstration of metacognition. However, observers were substantially more confident (8/10 individually statistically significant) in the first stimulus condition, where observers showed a primacy bias, compared to the second condition where they showed no such bias. Fitting a generalized drift-diffusion-model to the data from the first condition demonstrated that both temporal bias and overconfidence were driven by the same confirmation bias dynamics which favor the integration of new sensory evidence that agrees with observations earlier in the trial. The data from both conditions can be explained by as the result of approximate hierarchical inference (Lange et al. 2021) and provides further evidence against an integration-to-bound process as an alternative explanation for primacy biases during perceptual decision-making (Kiani et al. 2008).