Primary Submission Category: Weighting
Balancing weights in factorial observational studies
Authors: Ruoqi Yu, Peng Ding,
Presenting Author: Ruoqi Yu*
Factorial design is a common and easy-to-use tool to evaluate causal effects with multiple treatments. The main literature focuses on randomized experiments, but it remains challenging to draw reliable causal inferences in observational studies. In recent years, several methods have been proposed to deal with observational data, ignoring the factorial structures and treating the treatment combinations as a multi-leveled treatment. However, as the number of treatment combinations grows exponentially as the number of treatments, some treatment combinations can be rare or unobserved, raising new challenges in the definition of causal estimands and the downstream inference. To overcome the limitations, we propose a new weighting framework that (i) adjusts the confounding effects of observed covariates for all contrasts of interest, (ii) takes care of the factorial structures of any number of treatments, (iii) can be easily generalized to fractional factorial design and incomplete factorial design, (iv) is computationally efficient. We also conduct numerical studies to evaluate the performance of the newly proposed method in simulations and an empirical application.