Primary Submission Category: Causal Inference in Networks
Adaptive Experiments for Exposure Mapping Estimation
Authors: Ravi Sojitra, Ruohan Zhan,
Presenting Author: Ravi Sojitra*
During experimentation, treating one unit can interfere with outcomes of other units. Our goal is to estimate the proximity within which such interference can be induced in networks. For example, if some people are vaccinated against the flu, we expect unvaccinated people within the same households to benefit from reduced household level risk. Such information about treatment-control spillover is valuable for both experiment design and welfare maximization under resource constraints. First, we formalize estimands to quantify how far spillovers propagate in networks and their identifying assumptions. For instance, the (average) minimum numbers of vertices that need to be treated at varying proximities to induce a minimum spillover effect. Then, we propose algorithms for sequential experiments to simultaneously estimate these estimands and improve outcomes of experimental units during the experiment. We show how one can straightforwardly apply regret minimization and pure exploration algorithms under these assumptions. Moreover, we extend these algorithms to combine randomized treatments and observational exposures to mitigate the impact of treatment dimensionality on learning.