Asymmetric, Closed-Form, Finite-Parameter Models of Multinomial Choice

TL;DR

This paper introduces a new class of asymmetric, finite-parameter multinomial choice models called logit-type models, which outperform traditional symmetric models like MNL in transportation data analysis.

Contribution

The paper develops a novel framework for constructing asymmetric multinomial choice models from binary models, filling a gap in transportation modeling.

Findings

New asymmetric models outperform MNL in likelihood tests.

Asymmetric models show practical differences in empirical transportation data.

The framework enables creation of custom binary and multinomial choice models.

Abstract

In transportation, the number of observations associated with one discrete outcome is often greatly different from the number of observations associated with another discrete outcome. This situation is known as class-imbalance. In statistics, one hypothesized explanation for class imbalance is the existence of data generating processes that are characterized by asymmetric (as opposed to typically symmetric) probability functions. Despite being a valid hypothesis for class-imbalanced choice situations, few simple models exist for testing this explanation in transportation settings---settings that are inherently multinomial. Our paper fills this gap. As such, it should be of interest to transportation scholars and practitioners alike. Overall, we addressed the following questions: "how can one construct asymmetric, closed-form, finite-parameter models of multinomial choice" and "how do such models compare against commonly used symmetric models?" To do so, we (1) introduced a new class of closed-form, finite-parameter, multinomial choice models that we call "logit-type models," (2) introduced a procedure for using our logit-type models to extend existing binary choice models to the multinomial setting, and (3) introduced a procedure for creating new binary choice models (both symmetric and asymmetric). Together, our contributions allow us to create new asymmetric, multinomial choice models by creating multinomial extensions of asymmetric, binary choice models that already exist or that we create ourselves. We demonstrated our methods by developing four new asymmetric, multinomial choice models. We found that most of our asymmetric models dominated the multinomial logit (MNL) model in terms of in-sample and out-of-sample log-likelihoods. Moreover, on our two empirical applications, we also found practical differences between the MNL model and our new asymmetric models.