Randomization Tests for Bipartite Experiments

Randomized experiments with bipartite structures are increasingly used to study interventions where treatments and outcomes are defined on distinct but interdependent populations. This paper develops Fisherian-style randomization tests tailored for bipartite experiments. Our framework formalizes randomization tests on a subset of units over a restricted assignment space defined by the bipartite structure. The proposed procedures are finite-sample valid under commonly used experimental designs, including complete randomization and Bernoulli trials, and extend naturally to more general designs through conditional sampling. We further demonstrate how general network experiments can be represented within the bipartite framework, enabling the analysis of total and spillover effects on networks. Simulation and empirical studies illustrate the validity and practical relevance of our approach.