2E-06 - Impacts of RWD-related biases in Bayesian borrowing through the use of simulation study

Background: With growing interest in the use of external controls, real world data (RWD) has moved into sharper focus in trial submission as a potential source of relevant information as shown by creation of CEDAR at the FDA for integrating such data into regulatory processes. Methods for leveraging RWD have been proposed, but concerns for biases due to misclassification and missingness remain an ongoing challenge.

Objectives: To understand the potential bias in estimation, we investigate the performance of propensity score weighted robust mixture priors for integrating RWD. Our quantitative assessments help demonstrate the types of scenarios in which inference based on Bayesian borrowing may be valid.

Methods: We conduct a simulation study to assess how the realistic RWD quality issues of misclassification of the covariates and misclassification of the outcome affect estimates obtained from a two-stage approach for augmenting the control arm from a randomized controlled trial with controls from RWD. The first stage estimates propensity score weights to uphold comparability of external subjects with trial participants. The second step leverages the external information via a robust mixture prior combined with generalized linear model for capturing the prognostic impact of covariates. The performance of this method is evaluated with respect to bias and variability.

Results: Simulation results demonstrated that our proposed approach provided low bias when there is only misclassification of the covariates and no misclassification of the outcome. As the extent of misclassification of covariates increases with no misclassification of the outcome, variability reductions are proportional to the external sample sizes. On the other hand, with increasing misclassification of the outcome, the estimation bias generally increases initially and then falls, due to lack of borrowing once a certain level of prior-data conflict is reached. With less information from RWD being leveraged, this translates to reduced and even an absence of efficiency gains.

Conclusions: Propensity score weighting with Bayesian borrowing approaches have the potential to reduce the number of trial participants in the control arm by making use of relevant information from RWD and provide valid estimates of treatment effect in certain misclassification cases. Simulations provide a quantitative look at the extent that sources of bias may result in inaccurate estimates. Aligning with regulatory agencies on the utility and reliability of potential real world data sources for supporting inference of trial data is an important initial step.