Speaker
Brennan Kahan, PhD
MRC Clinical Trials Unit
University College London (UCL)
Keywords
Cluster-Randomized Trial, Estimands, Cluster-Average Treatment Effect, Participant-Average Cluster Size
Key Points
- The TRIGGER trial inspired statistician Brennan Kahan to ask in the questions, “But what if we’d chosen a different analysis? “and “How much would standard errors really change?”
- An estimand is a precise description of the treatment effect that a researcher aims to estimate from the trial. This concept is especially important in the context of cluster-randomized trials, but it is important to determine if the estimand will reflect participant-average treatment effect versus a cluster-average treatment effect or a marginal versus cluster-specific. The difference relates to how data is weighed.
- Two estimands will differ when there is informative cluster size in a trial. Informative cluster size refers to situations in which outcomes or treatment effects differ in large clusters versus smaller clusters. An example of this is when patients experience better outcomes in a large hospital compared to a small hospital when given the same medication.
- Which estimand (participant-average or cluster-average) to use depends on the study question. Participant-average treatment effect will demonstrate the population-effect when going from one intervention to another. A cluster-average treatment effect better enables the evaluation of an intervention or treatment’s impact directly on clusters.
- Mixed-effects models or Generalized Estimating Equations (GEE) are the most common analysis methods for clustered randomized trials, but when informative cluster size is present, both are bias. To avoid this bias, Independence Estimating Equations (IEEs) and cluster-level summaries can be used to estimate either cluster- or participant-average effects.
- The occurrence of informative cluster size has never formally been evaluated to the presenter’s knowledge. Statisticians working on pragmatic trials should consider estimand and tailor analysis around the chosen estimand.
Learn more
Many of the concepts described in the Grand Rounds presentation are outlined the article published in the International Journal of Epidemiology: Estimands in cluster-randomized trials: choosing analyses that answer the right question
Discussion Themes
– The summary measure scale is very important in determining the best estimand for a specific cluster-randomized trial.
– Equal cluster size does not necessarily mean that estimand is irrelevant because of differences in sample size. Reweighting may be necessary, and there may be additional assumptions.
– It’s difficult to determine definitively if there is informative cluster size, but it would be interesting to evaluate in pilot data clusters.
– Stratification would not be a solution to solve the issue of informative cluster size in analysis.
Tags
#pctGR, @Collaboratory1