A Learning-based Model for Imputing Missing Levels in Partial Conjoint Profiles

A Learning-based Model for Imputing Missing Levels in Partial Conjoint Profiles Respondents in a conjoint experiment are sometimes presented with successive partial product profiles (i.e., profiles with missing attributes). The manner in which these respondents integrate available information in the current profile, information embedded in all previously shown profiles (perhaps through memory recall), and their prior knowledge about the product category, to impute values for missing attribute levels, is both theoretically interesting and practically relevant. Theoretically, this investigation sheds light on how customers integrate different sources of information in evaluating products with incomplete attribute information. Practically, this study highlights the potential pitfalls of imputing missing attribute levels using simple rules (e.g., an averaging model) and develops a better behavioral model for describing and predicting customers’ ratings for partial conjoint profiles. This research has two goals. First, we model how respondents infer missing levels of product attributes in a partial conjoint profile by developing a learning-based imputation model that nests several extant models. The advantage of our approach over previous research is that our general class of imputation models infers missing levels of an attribute not only from prior levels of the same attribute, but also from prior levels of other attributes (especially those that match the attribute levels of the current product profile). To account for heterogeneity in learning across individuals, we estimate this class of imputation models using a hierarchical Bayesian approach. A second goal is to provide an empirical demonstration of our approach, and to test whether learning in conjoint studies occurs, to what extent, and in what manner it affects responses, partworths, and the relative importance of attributes. We show that the relative importance of attribute partworths can shift when subjects evaluate partial profiles. Such behavioral distortion suggests that consumers may “construct” rather than “retrieve” part-worths and hence consumers are sensitive to the way the profiles are presented. Finally, our results show that consumers’ imputation process can also be influenced by manipulating their “prior” information about a product category.