Defining and estimating effects in cluster randomized trials: A methods comparison.

Area of Science:

• Epidemiology and Biostatistics
• Causal Inference

Background:

• Cluster randomized trials (CRTs) are widely used to evaluate group-level interventions but face challenges in defining causal effects and understanding analysis methods.
• The theoretical and practical performance of common CRT analysis methods, especially with limited or varying cluster sizes, requires further clarification.

Approach:

• A general framework is presented to formally define various causal effects using counterfactual outcomes.
• The study provides an overview of CRT estimators, including t-test, GEE, augmented-GEE, and TMLE.
• Finite sample simulations and a real-world application (Preterm Birth Initiative study) assess estimator performance with varying cluster sizes and effect definitions.

Key Points:

• TMLE demonstrates flexibility in estimating user-specified causal effects, adapting for covariates to improve precision while controlling Type I error.
• The study illustrates how effect estimates differ when targeting individual-level versus cluster-level outcomes.
• For the PTBi data, cluster-level effects showed a 19% reduction, while individual-level effects indicated a 34% reduction.

Conclusions:

• Targeted Maximum Likelihood Estimation (TMLE) is a robust and adaptable method for cluster randomized trial analysis.
• The choice of causal effect definition (individual vs. cluster level) significantly impacts intervention effect estimation in CRTs.
• Understanding the performance of different estimators under realistic conditions, such as limited and heterogeneous cluster sizes, is crucial for reliable CRT analysis.