Heteroscedastic Control for Random Coefficients and Error Components in Mixed Logit

Abstract

Developments in simulation methods, and the computational power that is now available, have enabled open-form discrete choice models such as mixed logit to be estimated with relative ease. The random parameter (RP) form has been used to identify preference heterogeneity, which can be mapped to specific individuals through re-parameterisation of the mean and/or variance of each RP’s distribution. However this formulation depends on the selection of random parameters to reveal such heterogeneity, with any residual heterogeneity forced into the constant variance condition of the extreme value type 1 distribution of the classical multinomial logit model. In this paper we enhance the mixed logit model to capture additional alternative-specific unobserved variation not subject to the constant variance condition, which is independent of sources revealed through random parameters. An empirical example is presented to illustrate the additional information obtained from this model.