A Heteroscedastic Generalized Extreme Value Discrete Choice Model

Abstract

Of the commonly used discrete choice models, the probit class allows flexible covariance structures for disturbances but is computationally burdensome for problems with more than a few alternatives. The generalized extreme value (GEV) class, including the widely used logit and nested logit models, has the advantage of computational ease but suffers in general from the restriction of homoscedastic disturbances. This article generalizes the GEV class to allow heteroscedastic disturbances across decision makers as well as across choice alternatives. The resulting models include the heteroscedastic extreme value model as a special case, which is a generalized logit model with heteroscedasticity across choice alternatives. Particular attention is paid to the heteroscedastic logit and nested logit models because of their widespread use in practice. An empirical application reanalyzing data from the 1980 presidential election tests the hypothesis of information-induced heteroscedasticity across voters and finds support for a heteroscedastic logit model that reveals stronger effects of voter information on the turnout decision than suggested by the original standard logit model in Ordeshook and Zeng.