The exact variance of the average treatment effect estimator in cluster randomised controlled trials

Abstract

In cluster randomised controlled trials (CRCTs) with finite populations, the exact design-based variance of the Horvitz-Thompson (HT) estimator of the average treatment effect (ATE) depends on the joint distribution of unobserved, cluster aggregates of potential outcomes and is therefore not point-identified. We study a two-stage sampling design in which clusters are sampled at random, units are then sampled within selected clusters, and treatment is assigned at the cluster level. We first derive the exact design-based variance of the HT estimator of the ATE, accounting for cluster sampling, within-cluster sampling, and treatment assignment. We then extend Aronow et al. (2014) to derive sharp bounds for an estimated variance based on estimated cluster totals and propose a feasible consistent estimator of the upper bound. In simulations and an empirical application, confidence intervals based on the proposed upper-bound estimator are typically narrower than those based on conventional cluster-robust standard errors.