Primary Submission Category: Causal Inference in Networks
AIPW Estimation in Network Experiments
Authors: Christopher Harshaw, Gernot Zöcklein,
Presenting Author: Gernot Zöcklein*
In classical experiments without interference, AIPW estimators use covariate information to achieve improved efficiency over unadjusted estimators such as Horvitz—Thompson. Considerably less is known about the efficiency of similarly adjusted estimators for network experiments, where treatment given to one unit can affect the outcomes of neighboring units. While AIPW-style estimators for network experiments have been proposed and analyzed throughout the literature, they may — in some cases — incur a larger variance than their unadjusted counterparts. A central challenge which prevents efficiency gains is the irregular correlations of exposures which are necessarily induced by the underlying network.
In this paper, we propose an AIPW estimator for arbitrary contrastive effects in network experiments. Our AIPW estimator is constructed to perform well under the recently introduced Conflict Graph Design of Kandiros et al (2024), which achieves the currently best known rates of estimation. We show that the AIPW estimator achieves improved efficiency over the Horvitz–Thompson estimator when the dimension of the covariates is appropriately asymptotically bounded. An important aspect of our AIPW estimator is a “pruning away” of high degree nodes as a means to resolve a bias-variance trade-off, which may be of independent interest. We provide conservative variance estimators which facilitate asymptotically valid inference.