Primary Submission Category: Causal Inference in Networks
Adaptive Experimental Design for Efficient Causal Estimators under Neighborhood and Temporal Interference
Authors: Fei Fang, Laura Forastiere,
Presenting Author: Fei Fang*
Network interference poses fundamental challenges for experimental design, as incompatibilities among nearby units in simultaneously attaining the exposure conditions defining the target estimand can lead to low estimation efficiency. We develop an adaptive design for estimating causal effects under neighborhood interference to improve efficiency. We first consider experiments conducted on multiple networks, where network structure, potential outcomes, and covariates are sampled from a common distribution over time. In this regime, we build on the conflict graph design of Kandiros et al. (2025), which reduces conflicts in realizing target exposure conditions by assigning treatments according to an importance ordering of neighboring units. In the proposed adaptive conflict graph design, exposure sampling probabilities and importance orderings are updated based on observed history and chosen to minimize estimator variance. We then study adaptive design on a single fixed network observed repeatedly over time, where temporal carryover and dependence arise. To address this setting, we propose a block-adaptive design that applies the adaptive conflict graph design at the beginning of each block, with estimation leveraging observations from the carryover period. We jointly minimize estimator variance over block length, exposure sampling probabilities, and importance orderings in an adaptive manner. Synthetic data applications show efficiency gains of our designs.