We investigate the difference between confidence in a belief distribution versus confidence over multiple priors using a lab experiment. Theory predicts that the average Bayesian posterior is affected by the former but is unaffected by the latter. We manipulate confidence over multiple priors by varying the time subjects view a black-and-white grid, of which the relative composition represents the prior. We find that when subjects view the grid for a longer duration, they have more confidence, under-update more, placing more weight on priors and less weight on signals when updating. Confidence within a belief distribution is varied by changing the prior beliefs
subjects are insensitive to this notion of confidence. Overall we find that confidence over multiple priors matters when it should not and confidence in prior beliefs does not matter when it should.