The Generalised Multinomial Logit Model: Misinterpreting Scale and Preference Heterogeneity in Discrete Choice Models or Untangling the Un-Untanglable?

Recently published papers dealing with issues of scale and preference heterogeneity are having a significant impact on the choice modeling community. In particular, the generalized multinomial logit model (GMNL) is now being widely promoted as the model of choice in many discipline areas given its purported ability to separately identify scale and preference heterogeneity. The purpose of this paper is to firstly discuss a number of issues related to the estimation of the GMNL model. The second objective of the paper is to argue that the GMNL model does not in fact untangle scale and preference heterogeneity as has been reported and that the outputs derived from the model have been misinterpreted. The authors further argue that the model is not a generalized version of the mixed multinomial logit model, but in fact is a mixed multinomial logit with more flexible, but still restrictive, mixtures of distributions that do not necessarily equate to an ability to capture scale heterogeneity. They finally discuss the only theoretical circumstance under which random scale and preference can be separately identified within the logit family of models.