Advisory Committee Chair
Leslie A McClure
Advisory Committee Members
Date of Award
Degree Name by School
Doctor of Philosophy (PhD) School of Public Health
Cluster-randomized trials that randomize groups of individuals, instead of individuals themselves, run the risk of being underpowered when they are designed assuming equal cluster sizes but end up recruiting unequal clusters. The loss in power of the trial is directly related to the amount of variation present among the cluster sizes. To overcome the loss in power, researchers often employ a weighted analysis of the cluster means. Different weighting procedures have been developed that appear to maintain the power of the trial but there have been fewer attempts to describe the impact of different weights on the type I error of the study. We examined the effect of using no weights, minimum variance weights and cluster size weights under various scenarios of imbalance to recommend situations where one method performs better than the others in maintaining the type I error while achieving the trial power. We found that the minimum variance weights best controlled the type I error and provided adequate power compared to other weights. However, we also found that a weighted analysis method may not be always able to achieve the target power when the variation in the cluster size is high. Under such circumstances a more proactive approach would be to use sample size re-estimation within an internal pilot design. Using such designs, one can adjust the final number of clusters required for the trial based on the variability in clusters observed in the accrued data. An internal pilot design is also helpful in situations when there is uncertainty regarding precise estimates of certain parameters used in initial computation of sample size. Using simulations, we found that using an internal pilot design could lead to non-trivial inflation of type I error, especially, in trials that start out with small number of clusters. Currently, we recommend that such methods be used with caution if inflation of type I error is a major concern. Finally, we demonstrate the usefulness of an internal pilot design in the context of a real world cluster-randomized trial, the CAPTION trial, and discuss its practical implications.
Ranjan, Ashutosh, "Power Issues And Internal Pilot Designs For Cluster-Randomized Trials With Unequal Cluster Sizes" (2014). All ETDs from UAB. 2787.