Abstract:

We propose a model of belief formation based on sampling from memory. The agent in our model retrieves memories and combines them with the prior to form a belief. The agent is fully Bayesian and rational but faces a constraint on memory retrieval — she can only sample observations one at a time instead of retrieving all of them at once. Retrieval is mostly random, but the agent can partially target retrieval using an index. The index splits the database of memories into two (or more) groups based on the values of one (or more) attribute. To ensure that her beliefs are as accurate as possible, the agent chooses which indexed group tosample from in each period. We show that the expert will generically oversample one group and characterize three forces that determine which group is sampled more intensely. We then show that oversampling translates directly into ex-post belief bias. We use this insight to explain well-known biases in beliefs across individuals such as the “depression babies” effect, rational stereotypes, and the dependence of beliefs on the history of previously encountered problems.