Putting the Bayesian confidence hypothesis to rest

Influential models of confidence generation postulate that confidence computation is Bayesian, that is, confidence reflects the posterior probability that a decision is correct. Although several recent studies have found evidence against a strictly Bayesian account in the context of relatively complex tasks, the Bayesian confidence model remains dominant for simpler tasks. Here, we performed extensive comparison of 32 models that implemented either Bayesian or distance-to-criterion confidence computations, and systematically differed in their auxiliary assumptions. We fit all models to three different datasets, including two that have previously been used to support the Bayesian confidence model. We found overwhelming support for the distance-to-criterion models over their Bayesian counterparts across all model variants and across all three datasets. We further traced the discrepancy with previous results to the details of the fitting procedure and uncovered a previously unappreciated mimicry between metacognitive noise and lapse rate parameters. These observations provide strong evidence against the notion that confidence is formed on the basis of Bayesian computations and bring important insights into many components of confidence computations.