In a correspondence published in the European Journal of Epidemiology, researcher Anton Theunissen evaluates findings on infection, vaccination, and excess mortality among older Dutch adults. The underlying study by Koster and colleagues integrated relative survival into a multi-state model to assess pandemic outcomes. While Theunissen describes this modeling framework as a genuine methodological advance and commends the authors’ open-access code and candid limitations, he questions their primary causal conclusions. The original authors reported a strong negative association between vaccination and excess mortality across all groups, asserting an essential role for vaccination. However, Theunissen argues that these protective claims cannot be reliably identified from the underlying registry data due to structural exposure misclassification and unadjusted selection mechanisms.
The central problem identified in the correspondence concerns how vaccination status was classified using administrative data. The original study relied on the Dutch national registry, CIMS, where a substantial and unevenly distributed portion of administered doses was never recorded. Individuals lacking official records were automatically categorized as unvaccinated, directly contaminating the comparison group. Koster and colleagues recognized this limitation, estimating that 8 to 25 percent of birth cohorts were misclassified. However, Theunissen notes that this estimate accounted solely for individuals who refused data-sharing consent. It represented an absolute lower bound that omitted major institutional gaps, such as unrecorded nursing home vaccinations and doses given to patients near death.
Citing independent Dutch sources, Theunissen highlights the scale of these institutional registration gaps. Research by Rijpkema and colleagues estimated that roughly 25 percent of doses were missing overall, rising above 75 percent in the most highly vaccinated elderly strata where genuinely unvaccinated individuals were scarce. Additionally, national Outbreak Management Team reporting indicated approximately 30 percent missing registrations among agencies serving vulnerable groups in spring 2021. Because administrative omissions concentrated among nursing home residents and patients near the end of life, misclassified individuals did not represent a random sample of vaccinees. Instead, they were disproportionately frail individuals with substantially higher baseline mortality.
This non-random misclassification creates opposing biases. While healthy non-consenting vaccinees diluted the unvaccinated group and attenuated contrasts—as described by van Werkhoven and colleagues—unrecorded frail or dying individuals added high-mortality person-time directly to the comparator, falsely inflating apparent vaccine protection. Theunissen observes that Koster and colleagues addressed only the dilution effect from non-consent, dismissing institutional gaps as having probably no large impact. That dismissal rested on an argument about population head-counts rather than death counts, overlooking how a small misclassified subgroup with elevated mortality exerts decisive leverage in older cohorts where the truly unvaccinated group was vanishingly small.
Evidence of strong baseline confounding also appears in the study’s excess mortality calculations. The original authors observed a negative excess-mortality probability from the vaccinated state across every subgroup, which they attributed to healthy-vaccinee selection. While Theunissen notes that a negative all-cause excess is not biologically impossible, it cannot be produced by an intervention operating solely by preventing COVID-19 deaths. Instead, it demonstrates that the vaccinated group was inherently healthier than the reference population. This confirms that the vaccinated state was populated by a favorably selected cohort, introducing baseline differences that fundamentally conditioned the comparative mortality contrasts.
This selection bias was neither negligible nor speculative. At the 1 June landmark, the authors reported a negative excess-mortality probability of −0.14 percent, with a narrow 95 percent confidence interval from −0.15 to −0.12 percent. Theunissen argues that this precise and consistent estimate indicates selection bias of a magnitude the data can directly resolve. Although the negative sign of this selection is established, its quantitative contribution to the vaccination contrasts was left uncorrected. With the net direction of misclassification unresolved and selection bias conceded, the dataset cannot distinguish true vaccine protection from underlying health disparities between the compared groups.
The critique also addresses the authors’ defense that survival differences persisted across landmark times when comparing recent infection with and without prior vaccination. Theunissen points out that this comparison was not independent but remained differentially exposed to registration gaps. Entry into the unvaccinated, infected group required both documented infection and absence of a vaccination record—a combination enriched among institutionalized and end-of-life patients where both infection ascertainment and recording failures clustered. Because registration gaps contaminated only one arm of the contrast, Theunissen suggests describing the findings as associations between registry-defined groups rather than proven protective effects until quantitative bias analysis incorporating vulnerability-weighted misclassification is conducted.
Subject of Research: Methodological critique of exposure misclassification and selection bias in Dutch registry-based COVID-19 mortality analyses
Article Title: Exposure misclassification and residual selection limit the vaccination-related inference of Koster et al.
Article References: Theunissen, A. (2026). Exposure misclassification and residual selection limit the vaccination-related inference of Koster et al.. European Journal of Epidemiology. https://doi.org/10.1007/s10654-026-01472-5
Image Credits: AI Generated
DOI: 10.1007/s10654-026-01472-5
Keywords: COVID-19 vaccination, excess mortality, exposure misclassification, selection bias, healthy-vaccinee effect, multi-state model, relative survival, Dutch registry data, CIMS registry, observational epidemiology, quantitative bias analysis, landmark analysis, nursing home vaccination, older adults, Outbreak Management Team, data linkage, mortality contrast
News Source: Ophelia Keating. (November 21, 2026). Exposure Misclassification Limits Vaccination Inferences in Dutch Excess Mortality Study. Scienmag.



