| Session Assignment: 1715 | |
| IMPLICATIONS OF CLUSTERING ON SAMPLE SIZE CALCULATIONS OF RANDOMIZED CONTROLLED TRIALS | |
| Author: Kia Brown | Presenter: Kia Brown |
| Department: Graduate School of Biomedical Sciences | |
| Research Area: Other | |
 |
|
| (1) Cluster Randomized Controlled Trials, (2) Sample Size Calculations, (3) Statistics | |
| Kia S. Brown, MS*; Roberto Cardarelli, DO, MPH*; Anna Espinoza, MD*; Patricia A. Gwirtz, PhD**; John Schetz, PhD***; *Primary Care Research Institute, Department of Family Medicine; **Department of Integrative Physiology, ***Department of Pharmacology and Neuroscience University of North Texas Health Science Center, Fort Worth, TX 76107 | |
| Short Description: Primary care clinical research is a growing field of study within clinical research. The cluster randomized controlled trial design is one that is greatly employed by primary care clinical researchers. Clustering brings out a unique set of issues that the investigator must take into account before planning such a study and one of the most important issues is calculating an adequate sample size. An inadequate sample size may result in an underpowered study, thus limiting the ability to draw conclusive results. Unfortunately, cluster randomized controlled trials are only recently gaining popularity in the literature and few investigators fully understand the intricacies of designing these trials. This report seeks to educate clinical researchers on the effects of clustering on sample size calculations and the factors that need to be considered in the design and implementation of these trials. These factors include the intracluster correlation coefficient, the design effect and the effective sample size. | |
|
Purpose: The main goal of this project was to educate investigators and clinical researchers about the intricate details related to sample size calculations for cluster randomized controlled trials. These include the impact of the intracluster correlation coefficient (ICC), the design effect (DE), and the effective sample size (ESS).
Methods: A literature review on relevant topics was conducted using PubMed and print resources from the UNTHSC library. Sample size calculations were conducted with and without accounting for the clustering effect for a cluster randomized controlled trial currently being conducted by the North Texas Primary Care Practice-Based Research Network (NorTex). Sampsize Software was used to determine the unadjusted sample size without clustering effects based on a specified power, significance level, standard deviation and minimum difference. The software was also used to determine the adjusted sample size for a range of ICC values and differences in power. Separate calculations were done to determine the ESS and DE values for a range of ICC’s. Results: The sample size calculated for a clustered randomized controlled trial was found to be 4 times the sample size needed for a simple randomized trial at the same significance level and power. A 0.1 increase in the ICC almost doubled the total sample size needed for a clustered trial. Inflation of the sample size increased as the ICC increased. The ESS decreased as the ICC increased. The DE increased as the ICC increased as well. A 5% increase in power of a clustered trial required a 20% increase in the sample size and was best accomplished by keeping the cluster size the same and increasing the number of clusters. Conclusions: Due to the loss of statistical efficiency caused by clustering, a larger sample size is needed for these trials. It is important to estimate this value carefully because even a small change in the ICC can drastically change the sample size. In order for a cluster randomized trial to achieve the equivalent power of a simple randomized trial, the standard sample size must be inflated by the DE. The DE should be kept low by increasing the number of clusters and keeping the cluster size small. This will create an adequately powered study and keep the ESS large. Without knowing the implications of clustering effects on sample size calculations, studies may produce results with no true value to the investigator or to the medical community. |
|
| N/A | |
