In this paper, I develop a model of confirmation bias, which arises from a cognitive mechanism conceptually linked to the Bayesian brain paradigm and the efficient coding hypothesis in neuroscience. Specifically, I hypothesize that cognitive effort - quantified as the statistical divergence between prior and posterior beliefs -influences the precision of the sampling process, thereby allowing the probability of misperception to emerge endogenously. Consequently, disconfirming evidence, which requires greater cognitive effort to process, attenuates the impact of new information during Bayesian updating, resulting in posterior mean beliefs that are disproportionately influenced by prior mean beliefs. To test the model's predictions, I conduct an experiment, and I find empirical support for the hypothesis: noise in perceived probabilities increases proportionally with the divergence between the decision-maker's initial expectations and the sample.

We develop a novel method that allows us to econometrically recover belief distributions from binary choices between bets on different events. We delpoy the method in a strategic context by studying the predictive power of the recovered beliefs about reciprocity on trust. We econometrically recover two measures: a measure of the mean belief of the decision-maker; and a measure of the subjective uncertainty surrounding that mean belief. We show that mean beliefs are a significant predictor of trusting behaviour. Belief uncertainty, however, plays an even more important role in explaining behaviour. We illustrate this by estimating the generalized Arrow-Pratt approximation of our `trust equivalents' under model uncertainty proposed by Maccheroni, Marinacci, and Ruffino (2014). We find aversion to model uncertainty (i.e., uncertainty characterizing the trustor's own belief distribution) to be the single most important driver of trust. This showcases the role of the dispersion of the belief distribution when it comes to explaining behaviour in strategic interactions, and supports multiple-prior models of decision-making under ambiguity.