|
Abstract:
|
In a partially clustered design (PCD), individuals are randomized to one of two treatment conditions: an intervention delivered to groups created by study investigators, or a control condition where individuals remain independent. Prior research has focused on analyzing continuous outcome data from a PCD, but researchers have not evaluated approaches to analyzing binary outcome data from a PCD. We conducted a large simulation study comparing four common approaches: (1) an ordinary logistic regression model that ignores clustering; (2) a marginal model using a generalized estimating equations approach to indirectly account for clustering (3) a multilevel logistic regression model that assumes clustering occurs in both treatment conditions; (4) a multilevel logistic regression model that correctly models clustering in only one treatment condition. We evaluated the performance of each model based on the amount of bias in the estimates of the intervention effect, intracluster correlation coefficient, and type I error rate, and we made recommendations for modeling approaches based on trial size and study goals.
|
