Empirical application of missing data methods to address confounding in health insurance claims-based analyses using electronic health records available on a subset: an example from the Food and Drug Administration’s sentinel system.

Health insurance claims-based analyses inform many large epidemiologic studies but may have unmeasured confounding. Electronic health records (EHRs) have more detailed health information, but data may be missing on some individuals. Informed by recent work comparing approaches to address missing data, we selected and applied generalized raking (GR) and multiple imputation (MI) to illustrate how to efficiently combine data from these two sources. We considered a previously conducted claims-based study comparing 90-day arterial thromboembolism (ATE) risk among patients hospitalized with COVID-19 versus influenza, for which body mass index (BMI), a potential confounder, was unavailable. Using linked claims-EHR data, we adjusted for the same covariates as the original study and used GR and MI to additionally control for EHR-ascertained BMI, available for a subset. We included 912 hospitalized Kaiser Permanente Washington patients, 449 with COVID-19 and 463 with influenza (31.0% and 38.5% had EHR-ascertained BMI in the prior 90 days, respectively). Adjusted hazard ratios of ATE were similar with and without adjustment for BMI, in GR and MI analyses. In the setting of a claims-EHR-based study with missing data on a potential confounder, we describe our selection of the missing data approach, and its implementation, to demonstrate a broadly applicable process.