Abstract.  The formal stochastic model of voting should be the theoretical benchmark against which empirical models can be gauged. A standard result in the formal model is the ‘mean voter theorem’ stating that parties converge to the electoral center. Empirical analysis based on the vote-maximizing premise, however, invalidates this convergence result. We consider both empirical and formal models that incorporate exogeneous valence terms for the parties. Valence can be regarded as an electorally perceived attribute of each party leader that is independent of the policy position of the party. We show that the mean voter theorem is valid for empirical multinomial logit and probit models of a number of elections in the Netherlands and Britain. To account for the non-centrist policy positions of parties, we consider a more general formal model where valence is also affected by the behavior of party activists. The results suggest that non-convergent policy choice by party leaders can be understood as rational, vote-maximizing calculation by leaders in response to electoral and activist motivations.