Primary Submission Category: Heterogeneous Treatment Effects
Efficiency Gain of Covariate-Adjusted Differential Variance Estimators in Randomized Controlled Trials
Authors: Hani Zaki, Tania Janaudis-Fereirra, Philippe Boileau, Mireille Schnitzer,
Presenting Author: Hani Zaki*
Contrasts of potential outcomes’ variances have recently been proposed to detect treatment effect heterogeneity, even when treatment effect modifiers are missing or mismeasured. Adjusted causal machine learning estimators of these causal estimands have been derived and shown to be doubly robust and asymptotically linear under mild conditions. They were successfully applied to detect treatment effect heterogeneity in the re-analysis of randomized controlled trials (RCTs). However, it was not previously demonstrated that the covariate-adjusted estimators are more efficient compared to unadjusted estimators in the context of RCTs. Best practices for inference of these new causal estimands in RCTs are therefore unclear, particularly in situations where the asymptotic guarantees of the adjusted estimators may not be achieved due to small sample sizes. To address this gap, we derive unadjusted estimators and compare their asymptotic and finite-sample behavior to that of the adjusted estimators. We show theoretically and empirically that the adjusted estimator is asymptotically more efficient, but that empirically the unadjusted estimators can have smaller variance in small-sample settings. We then apply the adjusted and unadjusted estimators to data from an RCT aimed at evaluating the effectiveness of a virtual home-based physical rehabilitation program for patients living with long COVID.
Â