Ratings-Informed Probit for Predicting Substitution

Abstract

Choice-based conjoint analysis is widely used in marketing research to predict consumer preferences. However, some product attributes, like brands, or product styling, are too complex to decompose into a small number of distinct components that can be included as variables in a linear utility function. We propose an approach to handle any such "complex attribute" within a conjoint study. Specifically, we supplement the conjoint questions with appeal ratings for each potential level of the complex attribute. We then estimate a heterogeneous multinomial probit model where the covariance structure is a function of the correlations in the appeal ratings. The proposed ratings-informed probit model is theoretically-motivated and requires only one additional parameter per respondent. The structure allows us to collect the choice and ratings data for different groups of respondents, and so scales easily to settings with a large number of concepts. Two empirical applications demonstrate that the ratings-informed probit makes substantially better predictions about which concepts will compete against each other compared to benchmarks. The approach provides a parsimonious way to incorporate complex attributes into choice models early in the product development process when evaluating many concepts.

Keywords: hoice Models, Conjoint, Product Line Design, Brands, Substitution, Styling, Hierarchical Bayes, Multinomial Probit Model