In many clinical trials, data is collected on different patient characteristics at the time of entry (e.g., age, baseline severity and comorbidities). Covariate adjusted estimation methods that can both be more efficient than unadjusted estimators whilst also remaining robust to model misspecification (i.e., we require consistent estimators user arbitrary model misspecification) are available. The resulting sample size reductions can lead to substantial cost savings, and also can lead to more ethical trials since they avoid exposing more participants than necessary to experimental treatments. This was also emphasized in the recent guidance released by the U.S. Food and Drug Administration (FDA) for industry on “Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products”.

In this talk, we explain what covariate adjustment is, how it works, when it may be useful to apply, and how to implement it. We will then discuss recent contributions to the field of covariate adjustment. In particular, we will touch on the role of data-adaptive methods (i.e., any data analysis method that adapts the data to learn structure). This includes sophisticated methods such as machine learning methods, with much flexibility to `adapt’ to the data, but also flexible parametric models with variable selection (e.g., stepwise variable selection, lasso, …), with or without the inclusion of splines.

Covariate adjustment in randomized trials is a currently a topic of interest for health authorities, highlighted by a recent FDA guidance document and EMA qualification opinion on a particular type of adjustment.

How should pharmaceutical companies react to these developments? Are new analysis techniques required? How can awareness and comprehension of new methods be spread across a large organization? To help address these questions, an internal “Covariate Adjustment Challenge” was run within the Analytics department at Novartis. 23 participating teams were given access to data from five prior studies in a particular indication. The aim was to propose either a single "super" covariate or a pre-specified set of baseline covariates that could be used in covariate-adjusted analyses of key endpoints in a subsequent study in the same indication. Upon the “test data” becoming available, teams were scored according to the gain in precision from their proposed adjustment compared to an unadjusted analysis.

This talk will cover the background to the project, the scoring metrics, as well as a high-level overview of the results.

Covariate adjustment in randomized trials is a currently a topic of interest for health authorities, highlighted by a recent FDA guidance document and EMA qualification opinion on a particular type of adjustment.

An internal “Covariate Adjustment Challenge” was run within the Analytics department at Novartis. 23 participating teams were given access to data from five prior studies in a particular indication. The aim was to propose either a single "super" covariate or a pre-specified set of baseline covariates that could be used in covariate-adjusted analyses of key endpoints in a subsequent study in the same indication. Upon the “test data” becoming available, teams were scored according to the gain in precision from their proposed adjustment compared to an unadjusted analysis.

This talk will discuss the challenge from a participant’s perspective, including the overall strategy taken, the specific methods used, implementation, as well as how the proposal could be communicated to clinical colleagues.