Integrating External Control Subjects in the Design and Analysis of Randomized Controlled Trials

Researchers, drug developers, and regulatory agencies (as seen by recent FDA guidance documents) are increasingly interested in leveraging subjects from real world data (RWD) as external controls in clinical trials (RCTs). This has the potential to reduce the sample size or shorten trial duration while demonstrating the safety and efficacy of a treatment. However, external data can differ from RCTs in distribution of both covariates and outcomes. Therefore, borrowing may result in biased treatment effect estimates, lower power, and inflated type I error rate. To enable robust medical decision-making, careful consideration needs to be given to the design and analysis of externally controlled trials (ECT). This course will provide an overview of the statistical framework and challenges in ECTs, recommendations for designing studies with hybrid external control arms, strategies for selecting external subjects according to similarity metrics to RCT data, and popular Bayesian approaches that use data driven comparisons to determine subject-specific contributions for external controls. In the first part of this course, we will provide an overview of externally controlled designs and a broad overview of statistical considerations for designing such studies including 1) causal inference framework for estimating treatment effects when incorporating external data, 2) identifiability assumptions and types of potential bias, 3) considerations for selecting fit-for-purpose external data, and 4) strategies to select external control subjects using propensity score and entropy balancing. The second portion of this course will focus on Bayesian adaptive borrowing methods including 1) overview of common priors (power prior, commensurate prior, meta-analytic predictive prior, and variations), 2) data-driven forms of prior distributions using similarity metrics (propensity score, posterior predictive p-value, Bayes factors), 3) simulation evaluation of operating characteristics of Bayesian methods. We will demonstrate the implementation of Bayesian approaches using the R packages psborrow and rjags. Simulated and/or real data examples will be integrated throughout the course to motivate methodology and tools presented.