Tashfeen Saeed

While a large empirical literature documents systematic biases in belief updating, previous work elicits posterior beliefs directly, leaving unresolved at which stage these biases arise. Using an experimental framework, I separate and examine the two stages of belief updating: evaluating evidence and forming posteriors with that evidence. I show that systematic distortions in the form of overestimation begin during evidence evaluation before it is combined with priors. These distortions account for much of posterior under-inference: reported posteriors closely track those obtained by applying Bayes' rule to participants' own likelihood reports, while independently elicited likelihoods reproduce much of the posterior bias in a posterior-only baseline. Intervening at the evidence evaluation stage -- either through elicitation or information provision -- additionally improves posterior beliefs. In a second experiment, I show that absolute likelihood levels systematically shape likelihood judgments and posterior beliefs even when the objective likelihood ratio remains fixed, violating a key ratio-invariance restriction shared by Bayesian updating and a wide class of behavioral models. These findings establish the importance of evidence evaluation in addressing systematic belief biases and in informing richer models of belief formation.

How people respond to new information depends on how they represent that information. While standard models of inference assume that information is represented precisely, a growing literature shows that individuals represent stimuli with noise. In this paper, I examine whether exposure to different information environments shapes internal representations of information and belief updating. Drawing on insights from cognitive science, I first develop a model in which signal proportions are represented internally with noise before being used in inference. The model predicts that information is encoded more precisely in regions of the signal space encountered more frequently, leading to more responsive beliefs. I test this prediction experimentally by varying properties of the information environment -- including signal proportions and prior probabilities over states -- to examine how adaptation affects beliefs over the signal space. Participants exhibit both greater responsiveness and large level shifts to information when the information lies near their adaptation range than when it lies farther away. These results underscore the importance of understanding how individuals represent and use information in explaining systematic belief biases.