2062. Sample Size of Studies With Patients Clustered Within Physicians
Julianne Souchek, PhD, Houston Center for Quality of Care and Utilization Studies, Houston VAMC and Baylor College of Medicine, PA Kelley, Houston Center for Quality of Care and Utilization Studies, Houston VAMC and Baylor College of Medicine, HS Gordon,
Houston Center for Quality of Care and Utilization Studies, Houston VAMC and Baylor College of Medicine
Objectives: (1) Estimate the power or sample size of a study of the effect of a characteristic of patients when patients are clustered within physicians. (2) Demonstrate how the sample size calculated by use of the usual cluster inflation factor is unnecessarily large if there is any effect of the characteristic.
Methods: We wished to calculate the power of a test of whether physicians make more information-giving statements to white than to African American patients in the operating room after heart catheterization? Patients were clustered within physician; therefore the standard deviation of the number of statements per patient was multiplied by the inflation factor for cluster sampling, [1+(n-1)r ]1/2, where n is the average number of patients per physician, and r is the intraclass correlation coefficient.
Results: This approach led to a serious underestimate of the power of the tests. The estimate of the intracluster correlation coefficient was inflated by the intervention effect. We estimated the power of the tests with an ordinary least squares procedure assuming an unbalanced split-plot design with physician the whole plot and patient the subplot.
Conclusions: The usual adjustment for clustering should not be used when the intervention is within cluster.
Impact: The study design described above is prevalent in studies of quality of patient care in the VA. An over-estimate of sample size increases the cost to the VA and the time to complete the study.