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

摘要: Variational Bayes (VB) methods have emerged as a fast and computationally-efficient alternative to Markov chain Monte Carlo (MCMC) for scalable Bayesian estimation of mixed multinomial logit (MMNL) models. It has been established that VB is substantially faster than MCMC at practically no compromises in predictive accuracy. In this paper, we address two critical gaps concerning the usage understanding MMNL. First, extant are limited utility specifications involving only individual-specific taste parameters. Second, finite-sample properties estimators relative performance VB, maximum simulated likelihood (MSLE) not known. To former, study extends several MMNL admit including both fixed random latter, conduct an extensive simulation-based evaluation benchmark extended against MSLE terms times, parameter recovery The results suggest all variants with exception ones relying on variational lower bound constructed help modified Jensen's inequality perform well prediction recovery. particular, nonconjugate message passing delta-method (VB-NCVMP-Delta) up 16 times MSLE. Thus, VB-NCVMP-Delta can be attractive fast, accurate