An empirical study of cluster randomization

Abstract

Increasing attention has been given recently to the methodological issues associated with randomization of clusters rather than individuals in lifestyle intervention trials. These issues are explored through an empirical study of the 'effective sample size' imposed by randomization of three experimental units frequently considered in epidemiological research: the spouse pair, the general practice, and the large geographic area (county). The measurement of within-cluster dependence for a dichotomous outcome variable is also discussed, and a relationship shown between Fleiss's kappa and Cornfield's inflation factor.