Analysing data from a cluster randomized trial (cRCT) in primary care: a case study

Health technology assessment often requires the evaluation of interventions which are implemented at the level of the health service organization unit (e. g. GP practice) for clusters of individuals. In a cluster randomized controlled trial (cRCT), clusters of patients are randomized; not each patient individually. The majority of statistical analyses, in individually RCT, assume that the outcomes on different patients are independent. In cRCTs there is doubt about the validity of this assumption as the outcomes of patients, in the same cluster, may be correlated. Hence, the analysis of data from cRCTs presents a number of difficulties. The aim of this paper is to describe the statistical methods of adjusting for clustering, in the context of cRCTs. There are essentially four approaches to analysing cRCTs: 1. Cluster-level analysis using aggregate summary data. 2. Regression analysis with robust standard errors. 3. Random-effects/cluster-specific approach. 4. Marginal/population-averaged approach. This paper will compare and contrast the four approaches, using example data, with binary and continuous outcomes, from a cRCT designed to evaluate the effectiveness of training Health Visitors in psychological approaches to identify post-natal depressive symptoms and support post-natal women compared with usual care. The PoNDER Trial randomized 101 clusters (GP practices) and collected data on 2659 new mothers with an 18-month follow-up.