Accounting for intraclass correlations and controlling for baseline differences in a cluster-randomised evidence-based practice intervention study

Abstract

Background: In health care and community-based intervention studies, cluster-randomised designs have been increasingly used because of administrative convenience, a desire to decrease treatment contamination, and the need to avoid ethical issues that might arise. While useful, cluster-randomised designs present challenges for data analysis. First, because of dependencies that exist among subjects within a cluster, methods that account for intra-class correlations have to be used. Second, on many occasions, because of unavailability of large numbers of clusters, lack of balance on baseline measures has to be carefully examined and appropriately controlled for.

Aim/methodology: Two strategies are presented that can be used when analysing data from a cluster-randomised design; both account for baseline differences. Examples of these challenges are provided by a pain management intervention study designed to promote the adoption of evidence-based pain management practices. One approach involves use of a mixed model via SAS PROC MIXED. The other approach involves use of a marginal model: Generalised estimating equations using SAS PROC GENMOD.

Implications: In cluster-randomised design, one must adjust for intra-class correlation when evaluating the intervention effect. Although the parameter estimates and their standard errors might be comparable with both random effect and marginal strategies for certain link functions (identity link or log link only), the interpretations are quite different and the two approaches are suitable for indicating answers to different questions. If differences are present concerning baseline measures between experimental and control groups, accounting for baseline measures is important. The choice between a mixed model or marginal approach should be dictated by whether the primary interest is a population or individual.