Aggregation, Interpretation, and Estimation of Preferences in Conjoint Experiments

Abstract

Conjoint experiments enable the measurement of preferences in complex, multi- dimensional choice settings. But the problem of how to aggregate over multiple dimensions to make substantively meaningful statements of the form “respondents typically prefer feature A to B” has not received a concise, systematic treatment in the literature. This paper provides a set of theoretical and statistical tools for understanding the behavior of conjoint estimands that do just that. Specifically, we focus on the choice of whether to target an estimand that includes indirect comparisons between two features in addition to direct comparisons. We show that although this permits researchers to incorporate more observed tasks in estimation, it can also raise problems of interpretability when indirect and direct comparisons diverge in sign and magnitude. We develop a novel set of statisti- cal tools, which integrate easily with existing workflows, to guide practitioners in choosing the estimand that best suits their needs.