An approach to improve the predictive power of choice-based conjoint analysis

Abstract

Conjoint analysis continues to be popular with over 18,000 applications each year. Choice-based conjoint (CBC) analysis is currently the most often used method of conjoint analysis accounting for eight-tenths of all conjoint studies. The CBC employs a multinomial logit model with heterogeneous parameters across the population. The most commonly used models of heterogeneity are the Latent Class Model, the single multivariate normal distribution, or a mixture of multivariate normal distributions. A more recent approach to capture heterogeneity is the Dirichlet Process Mixture (DPM) model and its predecessor Dirichlet Process Prior (DPP) model. The alternative models are empirically tested over eleven CBC data sets with varying characteristics. The DPM model provides the best predictive validity (percent of choices correctly predicted) for each of the eleven datasets studied, and provides a significant improvement over extant models of heterogeneity.