Abstract :
Background. Large field trials that randomize naturally occurring clusters such as communities, worksites, or schools are becoming widely accepted for evaluating complex interventions. The within-cluster measurement of individuals typically uses a cohort, followed throughout the trial or cross-sectional samples selected independently at each time point. The relative costs of these approaches is of concern in designing such trials. Methods. This paper takes the unified model for analyzing large cluster unit trials developed by Feldman and McKinlay (State Med 1993) and combines the resulting expression for the variance of the treatment effect with a simple cost function into an algorithm that produces the optimal trial design in terms of the number of clusters and the number of observations per cluster. Using the unified model developed in the prior paper, this algorithm also allows direct comparison of the cost of designs with equivalent precision. In particular, designs that use cohorts in each cluster unit and observe cohort members over time are contrasted with designs that draw independent cross-sectional samples from each cluster at each time point. Results. Using the algorithm and a realistic design problem, it is demonstrated that cohort designs are more cost efficient for short trials and high (≥ 0.75) autocorrelations. Conclusions. The power of the algorithm in designing cost-efficient cluster unit trials is well demonstrated. Estimates of variance and cost components from prior trials need to be readily accessible for use in the algorithm, for planning subsequent trials.
