We propose a mathematical framework for modeling opinion change using large-scale longitudinal data sets. Our framework encompasses two varieties of Bayesian learning theory as well as Mannheim’s theory of generational responses to political events. The basic assumptions underlying the model are (1) that historical periods are characterized by shocks to existing political opinions, and (2) that individuals of different ages may attach different weights to those political shocks. Political generations emerge endogenously from these basic assumptions: the political views of identifiable birth cohorts differ, and evolve distinctively through time, due to the interaction of age-specific weights with period-specific shocks. We employ this model to examine generational changes in party identification using survey data from the 1952– 2008 American National Election Studies.