Convergence Analysis of Markov Chain Monte Carlo Estimators of Discrete Choice Models in Transportation

Ricardo A. Daziano,Chen Wang

2013-01-01

Abstract:Although Bayes estimators are attractive for problems involving non-convex optimization and for weakly identified discrete choice models, researchers in travel behavior modeling seem somewhat reluctant to adopt the Bayesian framework. A common argument against simulation-based Bayes estimators is that there are no general rules for assessing convergence. In this paper, the authors study convergence of the Markov chain Monte Carlo (MCMC) estimator of a discrete choice model via the empirical analysis of travel mode choice using a series of commonly used diagnostics. The authors also compare the behavior of the posterior first and second moments with the point estimates of maximum simulated likelihood. Some of the tests are more conservative than others, suggesting the use of a longer chain (500,000 iterations for the specific empirical problem being analyzed), while other test provide evidence in favor of a shorter chain (50,000 iterations). The comparison with the maximum simulated likelihood estimator shows that the only problem with 50,000 iterations is the loss of efficiency in the determination of the nuisance parameters.