Protocol for a review of statistical methods used to estimate risk ratios and risk differences in parallel cluster randomised trials

Abstract

Abstract Background Cluster randomised trials randomise groups of individuals, such as clinics, schools, or communities, and are used when interventions apply at the group level, when individual-level interventions risk contamination between participants, or to reflect real-world implementation. When outcomes are binary, treatment effects may be expressed as relative measures (such as odds ratios or risk ratios) or absolute measures (such as risk differences). CONSORT guidelines recommend reporting both, but risk ratios and risk differences are often underreported compared to odds ratios. Estimating these measures in cluster trials is more complex than in individually randomised trials, requiring appropriate handling of clustering, convergence issues, and small sample corrections. There is currently little empirical evidence describing which statistical methods are used to estimate these effect measures in published cluster trials. Methods This protocol describes the planned methods for a methodological review of published cluster randomised trials. We will use an existing database of 800 trials conducted in low- and middle-income countries. From this, we will identify a subset of trials with a parallel design and a binary primary outcome. Trials reporting a risk ratio or risk difference for the primary outcome will undergo further detailed data extraction. We will summarise the methods used to estimate these effects, including how clustering and small sample sizes were handled, and whether estimates were adjusted for covariates. Discussion This review will provide the first detailed description of how risk ratios and risk differences are currently estimated and reported in cluster randomised trials. The findings will inform the development of methodological guidance and help identify gaps in reporting and implementation. This is particularly important as interest grows in improving estimand specification and the clarity of statistical analysis plans.