Testing Bayesian Updating with the AP Top 25

Most studies of Bayesian updating use experimental data. This paper uses a non-experimental data source– the voter ballots of the Associated Press (AP) college football poll, a weekly subjective ranking of the top 25 teams–to test Bayes’ rule as a descriptive model. I estimate the voters’ Bayesian posterior rankings for the first seven weeks of the 2006-08 seasons (a total of over 31,000 rankings) based on the assumption that the voters’ individual final rankings for each season are their true rankings throughout the season. I compare the estimated Bayesian rank updates to the observed updates and find disparate evidence: voter behavior is sometimes consistent with estimated Bayesian updating, while in other contexts voters systematically underreact to new information, and in others still they overreact. A parsimonious explanation for the different results is that voters are less responsive to aspects of the signals and priors that are relatively