Keith Goldfeld, DrPH, MS, MPA of the IMPACT Design and Statistics Core, recently published two blog posts examining whether it is possible to reduce the sample size requirements of a stepped wedge cluster randomized trial by collecting baseline information.

Goldfeld explains that reducing sample size requirements is possible for randomization at the individual level and in cluster randomized trails, but he and colleagues Monica Taljaard and Fan Li, also members of the IMPACT Design and Statics Core, are exploring the question for stepped wedge designs.

The first blog post focuses on work already done to derive design effects for parallel cluster randomized trials (CRTs) that collect baseline measurements, and the second addresses the questions for stepped wedge designs.

Blog post 1 of the series:  The design effect of a cluster randomized trial with baseline measurements

Posted on November 23, 2021

In the first blog Goldfeld lays the groundwork for future posts by focusing on work already done to derive “design effects” for parallel cluster randomized trials (CRTs) that collect baseline measurements.  He shares a discussion about why baseline measurements may have impact and how they can be used to reduce required sample size in cluster randomized trials.  Goldfeld cites a paper published in 2012 by Teerenstra et al, “A simple sample size formula for analysis of covariance in cluster randomized trials” as a great foundation to understand how baseline measurements can impact sample sizes in clustered designs. He provides examples and simulations, and describes the work being done by himself fellow researchers to expand this to stepped wedge cluster randomized trials in particular.

Blog post 2 of the series: Exploring design effects of stepped wedge designs with baseline measurements

Posted on December 7, 2021

In this post, Goldfeld extends the simulations provided in the previous post to stepped wedge trials in an effort to identify and define design effects on stepped wedge designs in particular.  He concludes by sharing the code to estimate statistical power of each model reviewed under a range of sample size assumptions.