Abstract
Humans have the metacognitive ability to estimate the accuracy of their decisions using confidence estimates. Several theories have attempted to describe the computational mechanisms of confidence generation by instantiating process models. Yet, there has been little work on comparing these models using the same data. In this study, we aim to uncover the computational mechanisms of confidence generation by extensively comparing models on their ability to fit confidence data and predict behavior, while simultaneously establishing a multi-dimensional framework for robust model comparisons. We fit twelve popular process models to a large dataset (20 subjects, 2,800 trials/subject) in which participants completed a perceptual task with confidence ratings. Our quantitative comparisons show that the best fitting model postulates a single system for generating both choice and confidence judgments where confidence is additionally corrupted by signal-dependent noise (Shekhar & Rahnev, 2020; Psych Review). These results suggest that dual processing assumptions – according to which confidence and choice arise from coupled or independent systems – is unnecessary. Model evidence also contradicted popular notions that confidence is derived from post-decisional evidence or posterior probability computations. Further, qualitative analyses revealed that the best fitting models were those that could closely predict individual variations in metacognitive ability and zROC functions. On the other hand, the worst performing models were characterized by a failure to predict the folded X-pattern, which is considered a basic feature of confidence. Finally, through model recovery analyses, we also establish practical guidelines for designing experiments that allow us to maximally discriminate between models. Together, these analyses establish a general framework for model evaluation that also provides qualitative insights into their successes and failures. Most importantly, these results, by confirming and falsifying theories about confidence, begin to reveal the nature of metacognitive computations.