Discrete Choice Models with Alternate Kernel Error Distributions

Abstract

The multinomial logit (MNL) and probit (MNP) models dominated
the literature on consumer behavior analysis, particularly in
the transportation planning context where the focus is on future travel
demand prediction as an aggregated outcome of individual traveler
choices. While Gumbel kernel errors in the MNL model are unbounded
and positively skewed, normal kernel errors in the MNP model are symmetric and unbounded. However, choice models with alternative kernel errors (beyond Gumbel and normal distributions) have piqued the interest of choice modelers for behavioral and prediction accuracy reasons. In addition, researchers found evidence in support of these alternate kernel errors in a wide variety of empirical contexts. This paper compiles a synthesis of the past literature that developed choices models with flexible kernel errors, including both parametric and semi-parametric methods and concludes with possible avenues for further research.