The Reach of Treatment Effects: Randomization Inference for Zero-Effect Regions

Abstract

A treatment may affect some individuals while leaving others unaffected. We study the reach of treatment effects by asking how large an unaffected covariate region could be. We develop efficient local randomization tests over nearest-neighbour regions to construct a finite-sample valid upper confidence bound for the size of the largest zero-effect region. A small bound provides evidence that treatment effects extend broadly across covariate values by ruling out large zero-effect regions. Under a general outcome model, we characterize how the size of the bound depends on the number of affected units, the size of their treatment effects, and the geometry of the true zero-effect set. We show that, when region size is measured as a sample fraction, the excess of the upper confidence bound over the true size decreases to zero at a minimax optimal rate up to a logarithmic factor. Numerical simulations assess the coverage, power and computational efficiency of our method. Applications to voter-turnout, HIV clinical-trial, and AI-assisted writing data illustrate how our method rules out large zero-effect regions across the covariates of interest.