A Bayesian Learning Model with Applications to Party Identification

Previous research has suggested that party labels operate like brand names that citizens use to inform their votes. The article argues that this earlier work has focused too much on the content of party messages, while the informational value of party labels also depends on voter uncertainty about the party's behavior in office. These intuitions are developed in a Bayesian learning model in which voters update their beliefs about the mean and variance in the distribution of parties' ideologies, and apply these beliefs to a spatial model of partisan choice with risk-averse voters. The model predicts that if party unity is high, then party labels will provide a useful signal to voters about candidate characteristics and identifications with the parties will be strong, but if party unity is low, then party attachments will be weak. This approach seems to explain both the stability in respondents' political preferences over the life-cycle and the decline and resurgence in the strength of party identifications in the American electorate over the last half-century.