A Comparison of Statistical Models for Non-ordered Discrete Dependent Variables

Abstract

How do voters choose their most preferred candidate or political party when they have more than two alternatives? A growing number of Japanese scholars have recently examined this question using a statistical medel called the multinomial logit model. In this paper, I first theoretically introduce statistical models useful for examining how political actors make a choice among more than two non-ordered discrete alternatives. Those models include not only the multinomial logit model but also the conditional logit model and the multinomial probit model. Second, using existing survey data, I empirically examine the differences between these models, and show that the multinomial probit model could be more desirable than the others if the “Independence from Irrelevant Alternatives” condition is not satisfied. Finally, I discuss the merits and limitations of the multinomial probit model for better understanding of political behaviors and outcomes.