Simulation evaluation of emerging estimation techniques for multinomial probit models. (June 2017)

• Record Type: Journal Article
• Authors: Patil, Priyadarshan N.; Dubey, Subodh K.; Pinjari, Abdul R.; Cherchi, Elisabetta; Daziano, Ricardo; Bhat, Chandra R.
• Is Part Of: Journal of choice modelling. Volume 23(2017)
• Page Start: 9
• Page End: 20
• Publication Date: 2017-06
• Subjects: Discrete choice -- GHK simulator -- Sparse grid integration -- Composite marginal likelihood (CML) method -- MACML estimation -- Bayesian Markov Chain Monte Carlo (MCMC); Decision making -- Periodicals; Social choice -- Periodicals; Decision making; Social choice; Periodicals; 302.13
• DOI: 10.1016/j.jocm.2017.01.007
• Languages: English
• ISSNs: 1755-5345

Abstract

Abstract: A simulation evaluation is presented to compare alternative estimation techniques for a five-alternative multinomial probit (MNP) model with random parameters, including cross-sectional and panel datasets and for scenarios with and without correlation among random parameters. The different estimation techniques assessed are: (1) The maximum approximate composite marginal likelihood (MACML) approach; (2) The Geweke-Hajivassiliou-Keane (GHK) simulator with Halton sequences, implemented in conjunction with the composite marginal likelihood (CML) estimation approach; (3) The GHK approach with sparse grid nodes and weights, implemented in conjunction with the composite marginal likelihood (CML) estimation approach; and (4) a Bayesian Markov Chain Monte Carlo (MCMC) approach. In addition, for comparison purposes, the GHK simulator with Halton sequences was implemented in conjunction with the traditional, full information maximum likelihood approach as well. The results indicate that the MACML approach provided the best performance in terms of the accuracy and precision of parameter recovery and estimation time for all data generation settings considered in this study. For panel data settings, the GHK approach with Halton sequences, when combined with the CML approach, provided better performance than when implemented with the full information maximum likelihood approach, albeit not better than the MACML approach. The sparse grid approach did not perform well in recovering the parameters as the dimension of integration increased, particularly so with the panel datasets. The Bayesian MCMC approach performed well in datasets without correlations among random parameters, but exhibited limitations in datasets with correlated parameters.