Re-analysis of data from cluster randomised trials to explore the impact of model choice on estimates of odds ratios: study protocol.

Abstract

Background: There are numerous approaches available to analyse data from cluster randomised trials. These include cluster-level summary methods and individual-level methods accounting for clustering, such as generalised estimating equations and generalised linear mixed models. There has been much methodological work showing that estimates of treatment effects can vary depending on the choice of approach, particularly when estimating odds ratios, essentially because the different approaches target different estimands.

Methods: In this manuscript, we describe the protocol for a planned re-analysis of data from a large number of cluster randomised trials. Our main objective is to examine empirically whether and how odds ratios estimated using different approaches (for both primary and secondary binary outcomes) vary in cluster randomised trials. We describe the methods that will be used to identify the datasets for inclusion and how they will be analysed and reported.

Discussion: There have been a number of small comparisons of empirical differences between the different approaches to analysis for CRTs. The systematic approach outlined in this protocol will allow a much deeper understanding of when there are important choices around the model approach and in which settings. This will be of importance given the heightened awareness of the importance of estimands and the specification of statistical analysis plans.