A1584
Title: Inference for bipartite experiments: A superpopulation perspective
Authors: Yuehao Bai - University of Southern California (United States)
Hongchang Guo - Northwestern University (United States) [presenting]
Xun Huang - The University of Chicago (United States)
Jizhou Liu - Peking University (China)
Abstract: The analysis studies bipartite experiments in which outcome units may be linked to many intervention units and outcomes depend on the numbers of treated and untreated neighbors. Causal comparisons between all linked intervention units treated and all linked intervention units untreated are considered, with the target parameter being a degree-weighted average of the corresponding exposure contrasts. A covariate-driven sparse random graph model is introduced in which degrees are approximately Poisson with bounded mean. Under Bernoulli assignment or complete randomization, the estimator is shown to be asymptotically normal as the numbers of intervention and outcome units grow. The asymptotic variance has two components: sampling variation from the outcome side and sampling variation induced by the intervention side.
