How do humans give confidence? Comparing popular process models of confidence generation

Humans have the metacognitive ability to assess the likelihood of their decisions being correct via estimates of confidence. Several theories have attempted to model the computational mechanisms that generate confidence. Yet, due to little work directly comparing these models using the same data, there is no consensus among these theories. Here, we compare twelve popular process models by fitting them to large datasets from two experiments in which participants completed a perceptual task with confidence ratings. Quantitative comparisons, validated by model recovery analysis, selected the best fitting model as one that postulates a single system for generating both choice and confidence judgments, where confidence is additionally corrupted by signal-dependent noise. These results contradict dual processing theories – according to which confidence and choice arise from coupled or independent systems. Model evidence from these data also failed to support popular notions that confidence is derived from post-decisional evidence, strictly decision-congruent evidence, or posterior probability computations. Further, we explored the link between model performance and the model’s ability to predict different qualitative patterns in the data, in order to determine the reasons why some models fail. These analyses showed that the models that consistently perform the worst fail to capture individual variations in either primary task performance or metacognitive ability. Together, these analyses establish a general framework for model evaluation that also provides qualitative insights into the successes and failures of these models. Most importantly, these results begin to reveal the nature of metacognitive computations.Ph.D