Balancing Centers and Observations in Multicenter Clinical Trials

Dalene Stangl, Saurabh Mukhopadhyay
Duke University

Apr 10 1993

Multicenter clinical trials provide a means by which to asses treatment effects across variations in patient characteristics, protocol adherence, and treatment milieu. In multicenter trials, one design problem is choosing the number of centers. Intuitively, choosing too few centers, we will fail to represent the population of centers, choosing too many centers will prevent us from learning about the individual centers. In this article we present a general design procedure for determining the optimal number of centers. constraints of the optimization include monetary resources and the desire posterior precision of the population and center-specific treatment effect parameters. The methodology uses a Bayesian hierarchical model framework and is implemented via the Gibbs sampling algorithm. We obtain a design that is optimal for a particular elicitation of the prior hyper-parameters and that is reasonably robust against prior misspecification. While the methodology is described in the context of multicenter clicical trials, similar arguments apply for choosing number of clusters in survey sampling, number of observation points in longitudinal studies, and number of families in genetic studies.


hierarchical Bayes method, Gibbs sampling, multicenter clinical trial, optimal designs


PDF icon 1993-13.pdf