Adaptive Treatment Assignment in Experiments with Clustered Interference

• In Event: Methods Student Research Panel

Abstract

This paper explores how adaptive treatment assignment can improve experimental design in settings where interventions target clusters of units and outcomes exhibit within-cluster interference. In such environments, a unit’s outcome depends not only on its own treatment status but also on the treatment saturation within its cluster. These conditions arise in many large-scale policy experiments, including but not limited to cluster-randomized trials (CRTs). I develop an adaptive decision framework in which the algorithm selects the treatment saturation level (action) for each cluster, observes the resulting outcomes, and updates beliefs to inform future decisions. The policy I propose here is based on Thompson Sampling and incorporates hierarchical Bayesian priors to capture underlying dependencies across treatment effects. Using simulations under both theoretical and empirically calibrated scenarios, I find that adaptive assignment improves in-sample welfare and policy outcomes relative to non-adaptive assignment policies. Hierarchical priors are especially effective in small-sample settings, though performance is sensitive to prior specification. These findings highlight the potential of adaptive, data-driven design strategies in improving the welfare outcomes of large-scale experiments.

Author

• Lixi Xu, University of California, Santa Barbara
  Presenting Author