Capturing consumer heterogeneity in metric conjoint analysis using Bayesian mixture models

摘要

We address unobserved preference heterogeneity within an omitted variable framework which provides a theoretical rationale for more continuous preference distributions, multivariate normal in the limit. A comparison of the random coefficients model (RCM) and the latent class model (LCM) using simulated data illustrates that the RCM dominates the LCM if the underlying distribution is strictly continuous. The LCM dominates the RCM if the underlying distribution is strictly discrete once the sample is informative enough to support the true number of classes. The simulation further documents that the optimal number of classes in an LCM is an unrestricted function of the sample size if the underlying distribution is continuous. Finally, we present an application to the mineral water market, where a finite mixture with random effects model with two components performs best. All models are estimated fully Bayesian, and model comparisons are based on model likelihoods and analyses of holdout data.