Use Case 2 (UC2) Aim 2: Empirical Application of Linked Dev Network (Y5): Correcting Claims Analyses for Unmeasured Confounding Using Subset Calibration Tools

This project was aim 2 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.”

As a component of a different Sentinel Innovation Center (IC) activity “Subset Calibration Methods (CI4)”, IC compared three methods of analysis designed to address incomplete confounder information when assessing the relative safety of medical products. These methods leveraged additional data available on a subset of the study population that is not available for the whole cohort. The three approaches studied were multiple imputation, generalized raking, and targeted maximum likelihood estimation (TMLE). Under this project (UC2) aim 2, the learnings, R code, and instructional vignettes from the Subset Calibration Methods project activities were used to facilitate implementation of at least one of these methods to address an epidemiologic question. Selection of the method(s) was based on those found to be both practical to implement and successful for settings similar to the use case selected for the UC2 aim 2 project.

The use case selected by the FDA and workgroup members was a previously conducted claims-based study which compared 90-day risk of arterial thromboembolism among patients hospitalized with COVID-19 versus influenza. Data on body mass index, a potential confounder, was not available in that analysis but is available fora subset of individuals with electronic health record data (EHR). In this study conducted at Kaiser Permanente Washington, generalized raking and multiple imputation were used to control for the EHR-ascertained body mass index data. These methods demonstrated how additional data available on a subset of a population could be leveraged to bolster study validity from claims-based studies.