ESMARConf2023: {metaUI} tutorial

Presenter: Martijn Schuemie
Authors: Schuemie , Martijn; Chen, Yong; Madigan, David; Suchard, Marc
Session: Tutorials 8
Title: The EvidenceSynthesis package: combining estimates across a heterogeneous distributed research network facing small and zero counts using likelihood profiles
Abstract: Studies of the effects of medical interventions increasingly take place in distributed research settings using data from multiple clinical data sources including electronic health records and administrative claims. In such settings, privacy concerns typically prohibit sharing of individual patient data, and instead, cross-network analyses can only utilize summary statistics from the individual databases. Typical models used in such studies, such as conditional Cox proportional hazards models or conditional Poisson regressions cannot be pooled by communicating 2-by-2 tables, so often the point estimate and confidence interval of the effect size estimate are shared instead, requiring an assumption of normality on the per-database likelihood distributions. In our paper we show that often this assumption is violated due to low counts, leading to strongly biased summary estimates. The EvidenceSynthesis R package solves this problem by allowing sites to share the shape of the likelihood curve, for example using a grid approximation. The package provides functions for creating these approximations, and for combining them at the study-coordinating site using either a fixed-effects or (Bayesian) random-effects model. We’ve shown this approach avoid the aforementioned bias due to low counts, without sharing patient-level data. This package is available in CRAN, and is actively used by the Observational Health Data Science and Informatics (OHDSI) in its global studies.
GitHub repository: https://github.com/OHDSI/EvidenceSynt...

Links: https://esmarconf.org/ESMARConf2023Ad...
https://osf.io/ubt9a