Batch Sampling To Improve Power in a Community Trial

Abstract

Experiments involving large social units, such as schools, work sites, or whole cities, are commonly limited in statistical power because the number of randomized units is small, leaving few degrees of freedom for residual (between-unit) error. The authors describe a method for increasing residual degrees of freedom in a community experiment without substantially increasing cost or difficulty. In brief, they propose that the experimental units should be divided into random subsamples (batches). Batch sampling can improve statistical power if the community endpoint means are stable over time or if their temporal variation is comparable in period to the batch-sampling schedule. The authors demonstrate the theoretical advantages of the batch system and illustrate its use with data from the Pawtucket Heart Health Program, in which such a design was implemented.