A1664
Title: Regression-based and design-based causal inference in two-sided experiments
Authors: Pengfei Tian - Tsinghua University (China) [presenting]
Peng Ding - University of California, Berkeley (United States)
Jizhou Liu - Peking University (China)
Abstract: Regression-based and design-based causal inference in two-sided experiments under the simple multiple randomization design is studied, in which buyers and sellers are independently and completely randomized, and pair-level treatment is induced by their joint assignments. Under a local interference assumption, a broad class of linear estimands including total effects, buyer-side and seller-side spillover effects, and direct effects is considered. The natural plug-in estimator is shown to be equivalent to a saturated regression estimator, providing a unified regression representation for causal estimation in this setting. A design-based asymptotic theory is developed using a Hoeffding decomposition into buyer-side and seller-side first-order projections and a second-order interaction remainder, yielding a central limit theorem with a transparent leading variance. For inference, the conventional two-way clustered OLS variance estimator is shown to be asymptotically conservative. A weighted class of conservative variance estimators is further introduced, asymptotically containing the two-way clustered variance as a special case, with a closed-form optimal member that is sharper than any other estimator in the class. The framework also extends to regression adjustment.
