International Choice Modelling Conference, International Choice Modelling Conference 2017

Font Size: 
Model selection and model averaging in MACML-estimated Multinomial Probit (MNP) models
Manuel Batram

Last modified: 28 March 2017

Abstract


SEE ATTACHED PDF-DOCUMENT for the full Abstract (incl. formulas)

(Very) short version:
The two most commonly used families of discrete choice models are the multinomial logit (MNL) and the MNP. Between these two the MNP offers better modeling flexibility at the expense of higher computational costs. In this respect, Bhats Maximum Approximate Composite Marginal Likelihood (MACML) estimation is a method which combines Composite Marginal Likelihood (CML) and an analytic approximation to allow for simulation-free, and fast but approximate estimation of MNP models. The literature on model selection for MNP models using the MACML estimation methodology is not extensive. Model averaging in this context has not been dealt with. Therefore this paper aims to stimulate discussion on this important topic by surveying proposals for model selection methods for MNP models estimated using MACML. Furthermore, we introduce model averaging methods for MACML-estimated MNP models. Preliminary evaluation of the corresponding merits of the various proposals is obtained using a simulation exercise on synthetic data generated using models obtained from real world data sets.

Conference registration is required in order to view papers.