A0699
Title: Efficient design-based inference for the stepped wedge design
Authors: Fan Xia - UCSF (United States) [presenting]
Abstract: Stepped wedge designs (SWDs) are increasingly used to evaluate interventions delivered at the cluster level, but they present a range of analytical challenges. These include confounding by time due to staggered rollout, heterogeneous time trends across clusters, incidental imbalances from small sample sizes, complex correlation structures, and unreliable standard error estimation in finite samples. Conventional approaches, such as linear mixed models, are sensitive to the misspecification of fixed and random effects, while existing semiparametric methods often require strong assumptions on time trends or sacrifice efficiency. A unified semiparametric framework is proposed for effect estimation and inference in SWDs that directly addresses these challenges. The proposed estimator is robust to the misspecification of nuisance components, consistent and asymptotically normal, and achieves the semiparametric efficiency bound under correct covariance specification. The theory accommodates arbitrary treatment effect models and is derived using a nonstandard extension of semiparametric efficiency theory tailored to SWDs with varying cluster-period sizes. To support inference in trials with few clusters, a permutation-based standard error estimator and a leave-one-out correction are introduced to reduce plug-in bias. Through simulation studies and application to a public health trial, the method's robustness and efficiency are demonstrated relative to standard approaches.
