Optimal Treatment Assignment Rules on Networked Populations

I analyze decision problems where a planner must choose, based on data from an experiment in a pilot population, how to simultaneously distribute a limited number of treatments that generate network spillovers across individuals in a target population. A global treatment rule maps the target population's observed covariates and network features into a jointly feasible allocation. Under local-response, stability, support, and policy-relevant overlap conditions, I identify the resulting welfare using a transported inverse propensity weighted criterion and construct an empirical welfare maximizer. I derive a finite-sample regret bound under restrictions on the complexity of network spillovers and the complexity of the global rule class, measured by its Natarajan dimension. The upper bound uses independent node-level variation, dyadic network formation without imposing a bounded maximum degree. A minimax lower bound establishes a corresponding rate comparison under additional welfare-relevant richness. Finally, I compare oracle optimal global and local rules (which allocate treatments on an individual-by-individual basis) and bound the welfare gap between them. An empirically calibrated oracle comparison illustrates the magnitude of this welfare gap.