Journal club: Covariate Adjustment for Two‐Sample Treatment Comparisons in Randomized Clinical Trials

Led by David Schoenfeld, PhD, of Massachusetts General Hospital, this discussion will focus on the methods outlined in the paper Covariate Adjustment for Two‐sample Treatment Comparisons in Randomized Clinical Trials: A Principled yet Flexible Approach. Practical application of these methods and the use of the software developed for them will also be discussed. Register.

This discussion is part of the Harvard Catalyst Biostatistics Journal Club.

Designers of clinical trials often obsess about whether to control for covariates in the statistical analysis. By “control” they usually mean including these covariates in a model for the outcome and estimating the treatment effect using this model. Since the covariates will be balanced across the two treatment arms, this practice is optional but may increase the power of the test for a treatment effect. Unfortunately if the model is wrong it may decrease the power of a test of for a treatment effect and the correction may change the meaning of the estimator of the treatment effect(the estimand). Tsiatis, Davidian, Zhang and Lu, have developed an ingenious solution to these problems, which they demonstrate is optimal over a wide class of methods for estimating the treatment effect. Their method is to develop a separate model for patients in each treatment group. This model can be developed in a post hoc manner and be as complicated as necessary as long as it is developed independently for each treatment group. Then one has an observed mean for each treatment group and an expected mean for each treatment group that can be calculated using either of the two models. In the setting of equal group sizes the estimate of the treatment effect is the difference in observed means corrected using the difference in expected means averaged over both models. This method has the advantage that one does not have to prespecify the model so that one can make it fit the data well. Further, the estimand is the average causal effect which is unambiguously defined. We will not discuss the methods optimality properties in this talk but rather discuss the practical application of these methods and the use of the software I have developed for themTo facilitate the upcoming discussion, please read the following paper:

Tsiatis, Anastasios A., et al. “Covariate adjustment for two‐sample treatment comparisons in randomized clinical trials: a principled yet flexible approach.” Statistics in medicine 27.23 (2008): 4658-4677. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2562926/ 

Slides from Dr. Schoenfeld’s presentation [PDF]

R files from Dr. Schoenfeld’s presentation:
• (correctedEstimate.R)
• (run.R)