Model-robust and efficient covariate adjustment for cluster-randomized experiments.

Area of Science:

• Biostatistics
• Epidemiology
• Clinical Trials

Background:

• Cluster-randomized experiments are common for evaluating interventions in real-world settings.
• Model-based covariate adjustment is frequently used but risks bias if models are misspecified.

Purpose of the Study:

• To develop robust statistical methods for covariate adjustment in cluster-randomized experiments.
• To address limitations of existing model-based methods, including potential bias from misspecification and cluster size variation.

Main Methods:

• Adapted generalized estimating equations and linear mixed models with weighted g-computation.
• Proposed efficient, triply-robust estimators allowing flexible covariate adjustment and accounting for post-randomization cluster size variation.
• Validated methods using machine learning and parametric models for nuisance function estimation.

Main Results:

• The proposed estimators are consistent, asymptotically normal, and efficient when using machine learning.
• When using parametric models, the estimators are triply-robust, offering protection against different model misspecifications.
• Demonstrated superiority over existing methods through simulations and real-world data analyses.

Conclusions:

• The developed methods provide more reliable and less biased estimates of treatment effects in cluster-randomized trials.
• These advancements are crucial for accurate intervention evaluation in complex, real-world healthcare and public health settings.