EHR and claims data pull different clinical histories into focus in N3C
A new analysis from the National Clinical Cohort Collaborative argues for a harder, less convenient standard in health-data research: if the question depends on a patient’s full trajectory, a single data stream may be too thin. By combining electronic health records with CMS claims data, investigators measured how much each source captured on its own, where they overlapped, and how those differences changed a real epidemiologic estimate. The result was not just a data-quality exercise. It was a demonstration that missingness and duplication can move an effect estimate far enough to change the apparent direction of a study.
The work focused on a large public US EHR resource, N3C, paired with Medicare and Medicaid claims. The central premise is straightforward but consequential: EHRs and claims are not interchangeable records of care. EHRs often capture medication detail and clinical context inside the health system, while claims can better preserve billed diagnoses and encounters across settings. When each source is used alone, the patient story is incomplete in a different way.
Overlap was uneven, and duplication surfaced quickly
The first part of the analysis compared information coverage across the two sources before and after a landmark COVID-19 diagnosis. CMS data identified many individuals who appeared duplicated among EHR identifiers, a reminder that multi-site data aggregation can create false fragmentation unless linkage is handled carefully. Even before endpoint modeling began, the merged dataset was telling a more coherent story about who was actually represented in the cohort.
Across variables, CMS captured conditions more completely, while EHR data captured medications more completely after COVID-19 diagnosis. This split is clinically intuitive. Claims are built from reimbursement events and diagnosis coding, so they tend to preserve coded disease history and utilization patterns. EHR systems, in contrast, are closer to the point of prescribing and medication reconciliation, which can make them better at recording treatment exposure, especially when the therapeutic signal depends on orders rather than billing.
The time component mattered. Some variables showed marked differences after the landmark COVID-19 diagnosis, suggesting that the same patient can look substantially different depending on whether the analyst is reading the chart or the claims ledger. In a post-infectious setting, where care often shifts across outpatient, inpatient, and pharmacy environments, those differences can be large enough to alter who appears exposed, who appears diseased, and when a risk window begins.
The ischemic stroke case study changed with the dataset
The investigators then tested whether the data source itself altered a concrete causal estimate: the effect of COVID-19 vaccination on ischemic stroke. They used both target trial emulation and self-controlled case series methods, which allowed them to examine the same clinical question through two analytic lenses. The point was not to prove a vaccine-stroke relationship. The point was to see whether incomplete capture would distort the estimate of that relationship.
It did.
Using EHR data alone, target trial emulation produced an odds ratio of 0.47 with a 95% CI of 0.28 to 0.79, suggesting a lower risk of ischemic stroke after vaccination. The self-controlled case series using EHR data showed a smaller reduction that was not statistically significant, with an incidence rate ratio of 0.81 and a 95% CI of 0.64 to 1.02. When the analysis relied on CMS data alone, or when both sources were combined, the estimates moved close to the null in both designs.
That divergence is the core finding. A single data source created an apparent protective association that disappeared when claims were added. In practical epidemiologic terms, the estimate was not merely noisy; it was vulnerable to differential capture of medication exposure, diagnosis history, and duplicated identity structure. When those pieces were corrected, the effect estimate settled back toward no association.
Why the biology of data matters as much as the biology of disease
The study is about methodology, but the bedside relevance is real. Stroke risk analyses after vaccination sit at the intersection of infectious disease, vascular medicine, and public trust, so even a modest analytic distortion can carry outsized interpretive consequences. If an EHR misses prior illness or incomplete enrollment history, vaccinated and unvaccinated groups can look more different than they truly are. If claims omit recent medication detail, exposure classification can drift. Both errors can bias the estimated effect of an intervention on a vascular endpoint that is already sensitive to confounding by health status and health-seeking behavior.
The merged N3C-CMS approach reduced those blind spots by filling in the gaps each system left behind. It also exposed a second-order problem: apparent findings may reflect data architecture as much as patient biology. For observational research, especially in large national datasets, that distinction is not academic. It determines whether a signal survives contact with fuller information.
A stronger record does not mean a perfect record
The investigators conclude that combining data sources produces fuller medical histories and more valid scientific results than either source alone. Their case study shows why. Single-source analyses can generate biases that are difficult to recognize until a second data stream reveals missingness, overlap, or duplication. In this setting, the combined dataset did more than add rows; it corrected the shape of the cohort itself.
The study also points to a broader lesson for researchers using national databases. Data completeness is not fixed across domains, and it can shift after a major clinical event such as COVID-19 diagnosis. Analyses that ignore those shifts risk treating partial documentation as if it were total history. That is especially hazardous when the endpoint is uncommon, the exposure is time-dependent, and the treatment assignment is entangled with healthcare use.
As a preprint, the work remains unreviewed, and the analytic framework will need scrutiny from other teams working with linked EHR-claims systems. Still, the signal is hard to miss: when the outcome is ischemic stroke and the exposure is vaccination, the answer can depend on whether the researcher sees the patient through one lens or two.
The unresolved issue is not whether linked data are useful. It is how often single-source analyses in large observational cohorts are quietly shifting estimates before anyone notices the shift at all.