Use Case 2 (UC2) Aim 5: Plasmode Simulation Based Quantitative Bias Analysis (QBA) for Unmeasured Confounding

This project was Aim 5 of the overall UC2 project “Empirical Application of the Sentinel EHR and Claims Data Partner Network to Enhance ARIA Sufficient Inferential requests and Atypical Descriptive Requests”. 

Unmeasured confounders pose a central threat to the internal validity of non-randomized studies, but evaluation of such concerns quantitatively through bias analyses remains infrequent. Most studies that attempt to quantify the threat of confounding by unmeasured variables rely on simplistic methods that make unrealistic assumptions. We recently proposed a flexible methodology based on individual-level data simulations that can allow researchers to relax many assumptions typically needed for bias analysis and characterize the bias arising from unmeasured confounders with a specified but modifiable structure. Prior work was designed to generate simplified scenarios with fully synthetic data involving a handful of measured confounders and did not simulate all correlations between measured confounders and unmeasured confounder(s). However, in real world data, many more confounders may exist with correlations among them and with unmeasured confounders. This, was an extension of the prior work to address this limitation. Therefore, we   focused on the implementation of more complex simulations that represented a realistic description of the relationships among multiple confounders and outcomes in a ‘plasmode’ framework that built on actual individual-level data as the basis of simulations to preserve the naturally occurring correlation patterns among study variables. A software tool in the form of reusable R codes was developed to help with ease of implementation of this methodology for use by FDA staff.