SW-2G - Addressing the threat of confounding in pharmacoepidemiology with innovative methods: Lessons from the FDA Sentinel Innovation Center

Background: Unmeasured confounding is a central threat in generating valid inference from pharmacoepidemiologic studies conducted using healthcare claims data. The availability of granular information achieved through linking claims to electronic health records (EHRs) may offer opportunities to address this threat by using advanced statistical and machine learning approaches.

Objectives: To describe various methodological alternatives for addressing confounding using empirical examples from a variety of projects conducted using the FDA Sentinel Real-World Evidence Data Enterprise (RWE-DE)- a data infrastructure consisting of linked claims-EHRs for >25 million individuals across six sites

Description: Members of the panel will share learnings related to the following domains: 1) the need for robust methods for confounding adjustment (Rishi Desai): we will set the stage by describing the importance of advanced methodology and newer data sources in generating regulatory grade real-world evidence for the Sentinel program , 2) Solution 1 (Richard Wyss): targeted learning with undersmoothed least absolute shrinkage and selection operator (LASSO) propensity scores for large scale covariate adjustment to accommodate thousands of natural language proceccesing (NLP) derived features from EHRs as proxy variables, 3) Solution 2 (Pamela Shaw): doubly-robust approaches including raking weights and targeted maximum likelihood estimation (TMLE) leveraging rich information available on a subset through EHRs, 4) Solution 3 (Rishi Desai): descriptive evaluation of the threat of unmeasured confounding in plasmode-based quantitative bias analysis leveraging rich information available on a subset through EHRs, 5) Solution 4 (Xu Shi): Data-driven automated negative control estimation (DANCE) as a scalable, semi-automated way to detect bias due to unmeasured confounding. The symposium will conclude with a session dedicated to address questions and comments from the attendees to facilitate discussion of all pertinent issues.