Details
Pharmacoepidemiologic studies based on administrative claims data are sometimes limited in their ability to control for confounding, either due to unavailability of certain clinical information, poor accuracy, and/or incompleteness of coded information. The capture of confounding covariates in U.S. administrative claims data may be improving over time, but little empirical data are available to support this statement. Electronic health records (EHRs) include more detailed information on patients’ care that is not routinely available in administrative claim databases.
The U.S. Food and Drug Administration (FDA) initiated these studies in the Sentinel System to examine the capture of key confounding covariates in electronic healthcare databases (administrative claims and EHR) to support regulatory drug safety evaluation. Covariates of interest from the Sentinel Distributed Database (SDD) included in the analyses were smoking, obesity, overweight, body mass index (BMI), alcohol abuse/dependence, drug abuse/dependence, systolic/diastolic blood pressure (BP), reported race, history of Coronary Artery Bypass Graft (CABG) or Percutaneous Transluminal Coronary Angioplasty (PTCA), and history of sudden cardiac arrest. Covariates included from the TriNetX U.S.A. Network (EHR data) were obesity, smoking, BMI, and systolic/diastolic BP.
The findings from SDD analyses using administrative claims data spanning 2006 to 2023 showed that the capture of most of these common confounding covariates improved over time with the lowest to highest prevalent values during this period being 1.9% to 18.4% for obesity, 0.2% to 9.2% for overweight, 2.3% to 14.8% for smoking, 0.4% to 2.0% for alcohol abuse/dependence, 0.2% to 1.8% for drug abuse/dependence, 0.6% to 3.8% for history of CABG or PTCA, and 0% to 0.4% for history of sudden cardiac arrest. However, a low prevalence of some of the covariates compared to national estimates in 2023 suggests that these covariates remain inadequately documented in U.S. claims data (obesity 18.4% vs 34.3%, overweight 9.2% vs 34.4%, alcohol abuse/dependence 2.0% vs 6.1%, and drug abuse/dependence 1.8% vs 17.1%, respectively).1
The analysis conducted in aggregated EHR data from TriNetX also showed a lower prevalence of obesity (14.0%) when compared with national estimates (34.3%), while the prevalence of smoking captured in both SDD (14.8%) and TriNetX (12.0%) was higher than or similar to the national estimate (12.1%). Systolic/diastolic BP and BMI were captured in about 50-60% of eligible members in TriNetX.
These exploratory Sentinel studies provided information to FDA on the availability of common confounders in electronic healthcare databases to aid decisions on the use of Sentinel data in future drug safety studies. This work highlighted the importance of carefully evaluating each data source to determine whether it can adequately address the specific study question by capturing the key confounding covariates required for the analysis. It also suggests a need to consider alternative approaches such as quantitative bias analysis in the early phase of study design when confounders are poorly captured, to assess potential bias from unmeasured confounding.
- 1National estimates from CDC Behavioral Risk Factor Surveillance System (BRFSS) and Substance Abuse and Mental Health Services Administration (SAMHSA) National Survey on Drug Use and Health.