A simulation study of statistical approaches to data analysis in the stepped wedge design

Area of Science:

• Biostatistics
• Clinical Trial Design
• Statistical Modeling

Background:

• Stepped wedge randomized designs are increasingly used in health research.
• Analyzing data from these designs requires careful consideration of statistical approaches.
• Model-based and design-based methods offer different analytical frameworks.

Purpose of the Study:

• To compare the performance of various statistical approaches for stepped wedge randomized designs.
• To evaluate robustness, efficiency, Type I error rate, and power of different analytical options.
• To identify optimal strategies for analyzing stepped wedge trial data.

Main Methods:

• Comparison of generalized estimating equations (GEE) and linear mixed models (LMM).
• Evaluation of model-based versus design-based analytical approaches.
• Assessment across different scenarios, including varying cluster numbers and treatment effect assumptions.

Main Results:

• GEE models with exchangeable correlation structures are more efficient than those with independent structures.
• Model-based GEE Type I error rates can be inflated with few clusters but are improved by design-based methods.
• Linear mixed models (LMM) show Type I error inflation in design-based analysis under random treatment effects, even with correct model specification.

Conclusions:

• GEE models offer greater robustness, especially when combined with permutation testing strategies.
• Design-based approaches can mitigate Type I error inflation issues in GEE.
• Careful selection of statistical methods is crucial for accurate analysis of stepped wedge trial data.