Bryan Blette, Ph.D., Assistant Professor of Biostatistics, Vanderbilt University Medical Center

"Practical Issues in Cluster-Randomized Trials: Missing Data and Assessing Assumptions"

Cluster-randomized trials (CRTs) are clinical trials in which groups are the unit of randomization rather than individuals. These are often used when it is impractical or not possible to randomize individuals and when natural groupings can be enrolled in the trial (e.g., hospitals, schools, villages, etc.). However, this study design induces a correlation structure that typically prevents use of methods which assume all units are independent. In the first half of the talk, we consider evaluating treatment effect heterogeneity in CRTs when effect modifiers have missing data. We compare several off-the-shelf methods in a variety of simulations that reflect clustered data and generate recommendations for practitioners on which methods should be used in analyses. We then apply these methods to the Work, Family, and Health Study, a CRT which evaluated a workplace intervention intended to improve employee work-life balance and other outcomes. In the second half of the talk, we consider a nuanced assumption made in many CRT analyses, that of non-informative cluster size (i.e., that the outcomes or treatment effects are not impacted by cluster size). We show how this assumption is directly related to estimator selection and then develop formal hypothesis tests to assess this assumption. The operating characteristics of the proposed tests are explored, and the tests are applied to the Work, Family, and Health Study data.

Stay connected TwitterFacebook LinkedIn YouTubeInstagram