Covariate Adjustment in Randomized Clinical Trials
Chair: Rakhi Kilaru (PPD)
Speaker: Devan Mehrotra, PhD (Merck)
Title: Covariate Adjustment Using Treatment-Blinded Covariate Selection in Randomized Clinical Trials
Abstract: When estimating treatment effects in randomized clinical trials, the primary goal of covariate adjustment is to improve precision (e.g., for linear models) and/or reduce bias (e.g., for non-linear models). In current practice, the covariates for the analysis model are pre-selected, but knowledge gaps at the trial design stage can unwittingly make that selection suboptimal. Using simulations and real data examples involving time-to-event and continuous endpoints, we demonstrate the benefits of a proposed alternative approach in which a pre-specified treatment-blinded algorithm is used to identify covariates that are jointly strongly associated with the observed trial endpoint(s) of interest, followed by a corresponding covariate-adjusted analysis after treatment unblinding. The proposed approach is tenably aligned with FDA guidance on covariate-adjustment in clinical trials and with publications by Tukey, Pocock and others that support using the same dataset for treatment-blinded covariate selection followed by a corresponding covariate-adjusted analysis for treatment effect estimation and inference.
Speaker: Yanyao Yi, PhD (Eli Lilly)
Title: A General Form of Covariate Adjustment in Randomized Clinical Trials
Abstract: In randomized clinical trials, adjusting for baseline covariates has been advocated as a way to improve credibility and efficiency for demonstrating and quantifying treatment effects. This talk will discuss a general form of covariate adjustment, the augmented inverse propensity weighting (AIPW) estimator, that includes approaches using linear models, logistic models and other non-linear models, and machine learning methods. Under covariate-adaptive randomization, including simple randomization, I will present a general theorem that shows a complete picture about the asymptotic normality, guarantee efficiency gain, and universal applicability for the general form of covariate adjustment. The general theorem will offer insights into the conditions for guarantee efficiency gain and universal applicability, which also motivate a joint calibration strategy using some constructed covariates after applying AIPW. Finally, I will illustrate the general applicability of the theorem with two examples, the parametric generalized linear models and highly adaptive machine learning models. All aforementioned methods are available in the R package RobinCar.
Speaker: Daniel Rubin, PhD (FDA)
Title: Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products
Abstract: This talk will discuss the May 2023 finalized FDA guidance document on adjusting for covariates in randomized clinical trials. Covariate adjustment is recommended because it can improve precision and power for estimating and quantifying treatment effects. When properly prespecified and implemented a covariate-adjusted analysis is a statistically reliable method that does not depend on strong model assumptions. Considerations related to covariate adjustment with linear and nonlinear models for different types of outcomes will be discussed.
Speaker: Christopher Olson, PhD (Premier-Research)
Title: Beyond Baseline; Understanding the Importance of Additional Covariate Adjustment Inclusion
Abstract: In the realm of clinical data, the importance of known and unknown covariates present a common problem and potential pitfall for many statisticians. This spans from the very start with writing the statistical analysis plan and what will be fit in the analysis models, all the way to post-hoc analyses looking at trends that were unexpected. This talk will briefly explore the impact of commonly seen covariate-related analysis issues, ways of identifying possible covariate candidates, and techniques for the inclusion of these factors within analyses. We will also cover several past anonymized examples with their respective outcomes and lessons learned. This presentation will challenge the listener to take a closer look at the data they have and think if including more than baseline as a covariate makes sense, why they should apply the lessons here, and how to do properly account for it.
