Coarse Bayesian Updating

Abstract Studies have shown that the standard law of belief updating—Bayes' rule—is descriptively invalid in various settings. In this paper, I introduce and analyze a generalization of Bayes' rule—Coarse Bayesian updating—accommodating much of the empirical evidence. I characterize the model axiomatically, show how it generates several well-known biases, and derive its main implications in static and dynamic settings. Each axiom expresses a property of Bayes' rule but, conditional on the others, stops just short of making the agent fully Bayesian. The model employs standard primitives, making it suitable for applications; I demonstrate this by applying it to a standard setting of decision under risk, leading to a close relationship with the Blackwell information ordering and comparative measures of cognitive sophistication and bias.