Accurate estimation of rare vaccine adverse events requires careful choice of epidemiologic design. The self-controlled case series (SCCS) design is widely used for short-term vaccine safety because it controls for time-invariant confounding by comparing risk periods with baseline periods within the same individuals. However, SCCS provides only indirect estimates of absolute risk differences. Cohort designs can estimate absolute incidence and risk differences directly through between-person comparisons but remain vulnerable to confounding that differs between vaccinated and unvaccinated groups.
This study evaluated whether absolute measures should be estimated exclusively using cohort designs by comparing excess stroke incidence following COVID-19 vaccination in a nationwide matched cohort with prior estimates from SCCS analyses.
A retrospective nationwide cohort study was conducted in Qatar over the period 1 December 2020 to 11 April 2023. The vaccinated cohort comprised 1,111,939 individuals who received COVID-19 vaccination. Each vaccinated individual was individually matched 1:1 to an unvaccinated person, yielding an equally sized matched unvaccinated cohort.
The predefined outcome window for assessing post-vaccination stroke incidence was the 42 days following vaccination. Stroke events occurring during follow-up in both vaccinated and matched unvaccinated cohorts were identified and compared. The analysis produced incidence rates expressed per 100,000 person-weeks and estimated an adjusted hazard ratio (aHR) comparing vaccinated with unvaccinated individuals.
Note: the abstract does not report the precise stroke case definition used, the diagnostic ascertainment method, the full list of covariates used for adjustment, nor the exact matching criteria beyond individual matching; those details were not reported in the provided source abstract.
During the follow-up window, there were 8 stroke events in the vaccinated cohort and 19 stroke events in the matched unvaccinated cohort. These counts correspond to incidence rates of 0.2 (95% CI, 0.1–0.3) per 100,000 person-weeks in the vaccinated cohort and 0.4 (95% CI, 0.2–0.6) per 100,000 person-weeks in the unvaccinated cohort.
After adjustment, the hazard ratio comparing vaccinated with unvaccinated individuals was 0.42 (95% CI, 0.18–0.96), indicating a lower observed risk of stroke in the vaccinated group in this between-person cohort comparison.
These cohort findings, which suggest a protective association of vaccination against stroke, contrast with results from SCCS analyses conducted in the same and other populations that have reported a small increase in stroke risk following vaccination. The authors interpret this divergence as potentially reflecting healthy vaccinee bias — a form of between-person confounding where individuals who receive vaccination differ systematically (for example, in baseline health, healthcare-seeking behavior, or recent illness status) from those who remain unvaccinated.
Because cohort comparisons rely on vaccinated and unvaccinated individuals who may differ in ways that affect outcome risk, cohort designs can underestimate adverse event incidence when healthy vaccinee bias operates. In contrast, SCCS reduces confounding by using within-person comparisons across different time periods, but it does not directly estimate absolute risk differences without additional assumptions or calculations.
Taken together, the contrasting results underscore the complementary strengths and weaknesses of the two designs: cohort studies directly estimate absolute incidence and risk differences but are vulnerable to between-person bias, while SCCS mitigates time-invariant confounding but does not yield straightforward absolute risk estimates.
The abstract does not provide several methodological details that would be relevant for interpreting the findings and the potential for residual confounding. Specifically, the abstract omits the precise clinical definition and ascertainment procedures for stroke events, the full set of covariates and methods used for adjustment, the matching variables and success of matching on potential confounders, and any sensitivity or subgroup analyses. These omissions limit the ability to fully assess the magnitude and sources of bias in the cohort comparison based on the abstract alone.
Because rare event counts were small (8 vs 19 events), estimates are imprecise and confidence intervals are wide; the abstract-reported 95% CI for the aHR just excludes 1.0 at the upper bound (0.96), indicating statistical uncertainty around the observed association.
In this nationwide matched cohort of more than one million vaccinated persons and an equal number of matched unvaccinated persons, COVID-19 vaccination was associated with a lower observed incidence of stroke in the 42-day risk window, yielding an adjusted hazard ratio of 0.42 (95% CI, 0.18–0.96). These cohort results conflict with prior SCCS findings that suggested a small increase in post-vaccination stroke risk.
The discordance likely reflects methodological differences, with healthy vaccinee bias as a plausible explanation for an apparent protective effect seen in the between-person cohort comparison. The authors conclude that reliance exclusively on cohort designs to estimate absolute risk differences may lead to underestimation of rare vaccine adverse events. For comprehensive vaccine safety assessment, combining within-person and between-person designs and reporting both relative and absolute measures — along with transparent reporting of case definitions, covariates, and sensitivity analyses — is important to mitigate bias and improve interpretability.
Clinicians and public health practitioners should interpret absolute incidence estimates from cohort comparisons with caution when healthy vaccinee bias is possible, and consider complementary SCCS or other within-person methods to assess temporal associations between vaccination and rare adverse events.