Primary Submission Category: Randomized Studies
Efficient Design-based Inference for Stepped Wedge Designs
Authors: Fan Xia, James Hughes, Patrick Heagerty, Gary Chan, Emily Voldal, Avi Kenny,
Presenting Author: Fan Xia*
Stepped wedge designs (SWD) are a type of cluster randomized trial used to evaluate new interventions on clusters like clinics and communities. In typical SWDs, all clusters start in the control group, crossover to the intervention group at randomized times, and remain until the trial ends. The unique trial structure introduces challenges due to temporal confounding and opportunities through randomized crossovers, demanding a balanced approach to optimize data use for both robustness and efficiency. Linear mixed models (LMMs), common in SWDs, require restrictive modeling assumptions. Nonparametric methods are less assumption-heavy but often lack efficiency. Following the intuition that vertical comparisons within time points can help establish robustness, and horizontal comparisons between time points can increase power, we propose an estimator based on the semiparametric efficiency theory for clustered data. The proposed estimator is robust to misspecification in the outcome model (e.g. mis-specified temporal trend) and the covariance matrix. Moreover, it is semiparametrically efficient when both are correctly specified. For inference, we propose both a sandwich-type variance estimator and a conservative plug-in estimator from exact variance enhanced with a leave-one-out correction for finite sample bias, given the typically small number of clusters in SWDs. Additionally, an unbiased estimator is introduced to further correct bias in the conservative estimator.