Two-for-One Conjoint: Bayesian Cross-Category Learning for Shared-Attribute Categories

Abstract

Conjoint analysis is an ubiquitous market research tool. As normally implemented, each time you wish to understand consumer utilities for a product category, you run a conjoint study for that category. For example, if you require consumer preferences for yogurt features, you run a (choice-based) conjoint using choices between yogurts. In this research, we propose two conjoint methodologies. The first, based on transfer learning, has the researcher learn about a focal category (e.g. yogurt) by leveraging results from a conjoint study on a related category (e.g. ice cream). The second contribution is a data collection task that uses choice pairs between products in different categories. We demonstrate our approach using three category pairs (hiking jackets and sleeping bags, ice creams and yogurts, and televisions and computer monitors) where paired categories have fully or partially overlapping features. We show, with a Bayesian model, how out-of-sample prediction accuracy is improved when using out- of-category information, with both no data and sparse data from the focal category. We demonstrate the effects of out-of-category information on managerially relevant decisions such as market share predictions. Since our approach uses standard conjoint software with no modifications, it can be used immediately as a practical marketing research tool.