YSPH Biostatistics Seminar: “Multiply robust generalized estimating questions for cluster randomized trials with missing outcomes”

Title: “Multiply robust generalized estimating questions for cluster randomized trials with missing outcomes”

ABSTRACT

In a cluster randomized trial (CRT), groups of people (rather than individuals) are randomly assigned to different interventions. Regression models fit using generalized estimating equations (GEEs) can be used to estimate treatment effects from CRTs but can give inaccurate estimates if some outcomes are missing. There are existing methods to handle missing data in CRTs, but they rely on specifying a single correct statistical model for missingness or the outcome—as we know, this can be difficult to do in practice!

In this paper, we developed a new method called “multiply robust GEE”, which provides extra protection against model misspecification when analyzing CRTs with missing outcomes. This method allows us to simultaneously specify multiple missingness models and multiple outcome models, providing accurate estimates if even one of these multiple models is correct. In addition to working out some of the mathematical details and evaluating the method’s performance in simulations, we also applied the multiply robust estimator to analyze the Botswana Combination Prevention Project, a large HIV prevention CRT designed to evaluate whether a combination of HIV-prevention measures could reduce HIV incidence. I also developed an open-source R package on GitHub that can be used to fit a multiply robust GEE.