Reducing Bias Due to Outcome Misclassification for Epidemiologic Studies Using EHR-derived Probabilistic Phenotypes

Abstract

Epidemiologic studies using electronic health record (EHR)-derived phenotypes as outcomes are subject to bias due to phenotyping error. In the case of dichotomous phenotypes, existing methods for misclassified outcomes can be used to reduce bias. In this article, we present a bias correction approach for EHR-derived probabilistic phenotypes: continuous predicted probabilities of the outcome of interest. This approach makes use of correction factors that can be computed by hand and do not require specialized software. We used simulation studies to investigate the performance of the proposed approach under a variety of scenarios for accuracy of the probabilistic phenotype, strength of the outcome/exposure association, and prevalence of the outcome of interest. Across all scenarios investigated, the proposed approach substantially reduced bias in association parameter estimates relative to a naive approach. We demonstrate the application of this approach to a study of pediatric type 2 diabetes using data from the PEDSnet network of children’s hospitals. This straightforward correction factor can substantially reduce bias and improve the validity of EHR-based epidemiology.