Primary Submission Category: Randomized Studies
Design-based inference for paired cluster-randomized experiments
Authors: Charlotte Mann, Adam Sales, Johann Gaganon-Bartsch,
Presenting Author: Charlotte Mann*
Paired cluster-randomized experiments (pCRTs) are common in education and medicine because there is a natural clustering of students/patients within classrooms/hospitals and within schools/cities. Additionally, paired randomization can help balance baseline covariates to improve experimental precision. Although paired cluster-randomization is very common, there is surprisingly no obvious way to analyze this randomization design if an individual-level (rather than cluster-level) treatment effect is of interest. For example, the most basic and common estimator, the difference in mean outcomes, is biased in this setting. Variance estimation is also complicated due to the dependency created through pairing clusters in pCRTs. In order to provide guidance for practitioners analyzing pCTS, first, we review point estimators and associated variance estimators for an individual-level sample average treatment effect for pCRTs, unifying the notation. We propose a design-based point estimator and variance estimator, and show how this estimator relates and, in fact, provides a unifying framework for previous estimators in the literature. Through analysis based on this framework and extensive simulation studies, we illustrate the trade-offs between the reviewed point and variance estimators in practice.