Design and Inference for Conjoint Analysis

• In Event: Causal Inference: Post-treatment Bias and Conjoint Experiments

Abstract

Conjoint analysis, which is a type of factorial design, has become popular among social scientists as a tool for measuring multidimensional preferences across several attributes. Because such experiments are based on multiple factors, each of which has several levels, researchers often focus on the average causal effect of a single attribute while marginalizing over the other attributes. What has been overlooked, however, is the fact that this so-called average marginal component effect (AMCE) critically relies upon the distribution of other attributes. This is problematic because most researchers use the uniform distribution when randomizing factors, and yet the population distribution of attributes in the real world is often far from uniform. Using an existing conjoint experiment and a simulation study, we demonstrate that the standard estimate of the AMCE can suffer from a substantial bias when the population distribution of attributes differs from the randomization distribution used in experiments. We address this problem by proposing a new experimental design and estimation method. The proposed methodologies are implemented through an open-source software package.

Authors

• Brandon de la Cuesta, Princeton University

• Naoki Egami, Princeton University

• Kosuke Imai, Harvard University