Bayesian Estimation of Mixed Multinomial Logit Models: Advances and Simulation-Based Evaluations

  • 2019-04-12 15:48:39
  • Prateek Bansal, Rico Krueger, Michel Bierlaire, Ricardo A. Daziano, Taha H. Rashidi
  • 0


Variational Bayes (VB) methods have emerged as a fast andcomputationally-efficient alternative to Markov chain Monte Carlo (MCMC)methods for Bayesian estimation of mixed multinomial logit (MMNL) models. Ithas been established that VB is substantially faster than MCMC at practicallyno compromises in predictive accuracy. In this paper, we address two criticalgaps concerning the usage and understanding of VB for MMNL. First, extant VBmethods are limited to utility specifications involving onlyindividual-specific taste parameters. Second, the finite-sample properties ofVB estimators and the relative performance of VB, MCMC and maximum simulatedlikelihood estimation (MSLE) are not known. To address the former, this studyextends several VB methods for MMNL to admit utility specifications includingboth fixed and random utility parameters. To address the latter, we conduct anextensive simulation-based evaluation to benchmark the extended VB methodsagainst MCMC and MSLE in terms of estimation times, parameter recovery andpredictive accuracy. The results suggest that all VB variants perform as wellas MCMC and MSLE at prediction and recovery of all model parameters with theexception of the covariance matrix of the multivariate normal mixingdistribution. In particular, VB with nonconjugate variational message passingand the delta-method (VB-NCVMP-Delta) is relatively accurate and up to 15 timesfaster than MCMC and MSLE. On the whole, VB-NCVMP-Delta is most suitable forapplications in which fast predictions are paramount, while MCMC should bepreferred in applications in which accurate inferences are most important.


Quick Read (beta)

loading the full paper ...