Learning Together Slowly: Bayesian Updating About Political Facts

Although many studies suggest that voters learn about politically-relevant facts with prejudice towards their pre-existing beliefs, none have fully characterized all inputs to Bayes' Rule, leaving uncertainty about the magnitude of bias. I first show the importance of careful measures of each input and then present results from a novel experiment that measures how learning of political information departs from perfect Bayesian. Subjects learn as cautious Bayesians, updating their beliefs at about 70 percent of perfect application of Bayes' Rule. They are also modestly biased. For information consistent with prior beliefs, subject learning is not statistically distinguishable from perfect Bayesian. Inconsistent information, however, corresponds to learning less than perfect. Despite bias, I find no evidence of polarization in beliefs. With small monetary incentives for accuracy, aggregate beliefs always converge towards common truth. This suggests that cautious Bayesian learning is a reasonable model of how citizens process political information.