Nearly two million matched adults showed a substantial difference in deaths over one year. The finding deserves an explanation, and the methods determine which explanations remain possible.
James Lyons-Weiler, PhD
For every 100,000 adults in a matched Florida cohort, approximately 784 people who began vaccination with Pfizer’s COVID-19 vaccine died within a year, compared with 562 who began with Moderna’s. The difference was 221 deaths per 100,000 people, or about 2.21 per 1,000. Most of that difference appeared among deaths classified as unrelated to COVID-19. These findings make the choice of comparison, the health of the recipients, and the completeness of their medical records consequential questions for anyone trying to understand what the vaccines did to overall health.
The study, accepted by Global Epidemiology, was conducted by Retsef Levi, Fahad Mansuri, Melissa Jordan, and Joseph Ladapo. It examined Florida residents who received their first Pfizer or Moderna dose between December 18, 2020, and August 31, 2021, following deaths for 12 months after that first dose. The version reviewed here is an accepted journal pre-proof, still awaiting final production changes (Levi et al., 2026).
The result warrants serious investigation. Whether the vaccines caused the difference is a separate question, and the study provides useful but incomplete evidence toward answering it. A fair reading must give the investigators credit for the problems they addressed while examining the problems their data could not resolve.
The investigators started with more than 11.2 million eligible adults and constructed two groups of 981,411 people each. Every Pfizer recipient was matched to a Moderna recipient on seven characteristics: five-year age group, sex, race, ethnicity, vaccination month, type of vaccination site, and residential census tract. The geographic matching covered 4,832 census tracts. The resulting comparison was considerably more disciplined than a simple count of deaths following vaccination.
Comparing two vaccinated groups also addresses part of a familiar epidemiological difficulty. People who seek vaccination may differ from people who decline or postpone it in health, access to care, and behavior. Having both groups initiate vaccination reduces some of that contrast. Starting follow-up at dose one is another strength. Requiring people to complete a second dose before admitting them into the analysis would exclude anyone who died before completion and could select out people whose health deteriorated after their first exposure.
Across the matched groups, Pfizer recipients had a 39% higher estimated cumulative risk of death over the year. The adjusted hazard ratio, a related measure comparing death hazards during follow-up, was 1.40, with a 95% confidence interval of 1.36 to 1.45. The absolute risks were approximately 0.784% and 0.562%. Reporting both absolute and relative differences gives readers the scale of the finding without either magnifying or minimizing it.
Source: Levi et al., Table 3. Risks are the authors’ Kaplan–Meier estimates. Cardiovascular deaths overlap the broader categories; the rows must not be added together. Cause-specific estimates require the competing-risk qualification discussed below.
Approximately 87% of the reported all-cause difference falls within the non-COVID-19 category. The authors also report that the cardiovascular and COVID-19 mortality differences were primarily driven by adults aged 60 and older, although all-cause mortality was higher among Pfizer recipients across their age-group analyses. Those patterns make age-specific clinical histories particularly important to the interpretation.
All-cause mortality is valuable because it does not depend on deciding which disease belongs on a death certificate. Provided deaths are captured and linked accurately, a death remains part of the primary outcome whether it is attributed to COVID-19, cardiovascular disease, cancer, or another cause. This makes all-cause mortality resistant to one important form of classification error. It does not make it resistant to differences in the health of the people being compared.
That is where the study’s central uncertainty lies. Two people can live in the same census tract, fall within the same five-year age band, and receive a vaccine during the same month at the same category of provider, while having very different probabilities of surviving the next year. One may have advanced renal disease, poorly controlled diabetes, or recent functional decline. The other may be medically stable. Matching works on the characteristics supplied to it; it cannot certify equivalence on histories that were never measured.
The authors acknowledge that smoking, obesity, and comprehensive comorbidity information were unavailable. They added an analysis adjusting for previous cardiovascular hospitalization and report that the results changed little. That is a useful test of one potential source of imbalance. It leaves substantial room for differences in outpatient disease, severity, medications, frailty, and health-care use. A person need not have been hospitalized for cardiovascular disease to have a substantially elevated risk of death.
But then again, there are no causal reasons to suspect that smoking or obesity will have caused on to prefer Pfizer or Moderna. They are variables that only rate as speculative, not causal, confounders.
These omissions therefore do not establish that Pfizer recipients were sicker. They establish that the study cannot directly demonstrate that the groups were equally healthy. Treating possible confounding as a proven explanation would be as unwarranted as treating the matching process as proof that confounding was eliminated.
Some residual differences could also arise within the categories used for matching. A calendar month is a broad interval during a rapidly changing epidemic and vaccination rollout. A hospital or pharmacy category can contain facilities serving different patient populations. More precise matching on dates and individual facilities would help test these possibilities. The claim that similar matching produced clinical balance in earlier Veterans Affairs studies is reassuring context, but balance in veterans cannot substitute for measuring balance in Florida residents.
The investigators did more than acknowledge this problem. They tested outcomes that, under their assumptions, should not differ because of the vaccine product but might reveal differences between recipients. These “negative controls” included cancer deaths and deaths from suicide or homicide. Cancer mortality was close to equal, with an adjusted hazard ratio of 1.05 and a confidence interval of 0.98 to 1.13. That provides evidence against a large cancer-mortality imbalance during the study period.
The suicide/homicide comparison was less conclusive. Its hazard ratio was 1.22, with a confidence interval of 0.94 to 1.59. Although that interval includes no difference, it also includes a substantial relative difference. Calling the result statistically nonsignificant does not establish equivalence. Nor would similar suicide and homicide rates establish similar cardiovascular vulnerability: the unmeasured conditions that predispose someone to heart failure need not affect the negative-control outcomes in the same way.
The controls therefore narrow the concern. They suggest that the mortality difference was not accompanied by equally large differences in every measured cause of death. They cannot establish that the particular medical characteristics capable of producing the cardiovascular and all-cause findings were balanced.
The paper’s sensitivity analysis also deserves attention. The authors report that, under their specified binary-confounder scenario, an unmeasured factor twice as prevalent among Pfizer recipients would need an approximately 2.8-fold association with mortality odds to bring the confidence interval across the null. Such calculations ask how strong a hypothetical confounder would have to be. They do not tell us whether that confounder, or an equivalent combination of unmeasured health differences, actually existed. Their evidentiary value depends on the parameter definitions and assumptions, and the final supplementary analysis is needed to inspect those details.
The distinction between a non-COVID-19 death and a death unrelated to infection is equally important. The authors examined a broader infection-related category that included deaths following documented SARS-CoV-2 infection. A mortality difference persisted among deaths without a recorded prior infection. This weakens an explanation based solely on infections captured by their records. It cannot exclude infections that were never tested, diagnosed elsewhere, or otherwise absent from the dataset.
Undocumented infection is not synonymous with asymptomatic infection. A symptomatic illness can also go unrecorded. Consequently, the remaining difference cannot be assigned to an infection-independent vaccine effect simply because no positive test was found. The paper discusses product differences involving formulation, delivery, and other biological features, but this study does not measure a mechanism that connects any of those features to the mortality gap.
The time course offers questions worth testing. The authors describe separation of several mortality curves around 30 to 40 days after the first dose, near the period when many participants received dose two. A formal analysis aligned to actual second-dose dates would be more informative than interpreting that visual proximity alone. Likewise, a divergence around day 200 cannot be mapped cleanly onto a single epidemic wave when participants entered the study across more than eight calendar months. Calendar-time analyses would help distinguish those influences.
One technical issue affects the cause-specific risk estimates. The investigators used Kaplan–Meier curves and censored deaths from other causes. When estimating the actual probability of dying from a particular cause, those other deaths are competing events: someone who has died cannot later die from the cause being studied. A competing-risk cumulative-incidence analysis should be reported for that question. With total mortality below 1%, the numerical difference may be modest, but it needs to be calculated. This issue does not invalidate the all-cause mortality comparison, for which every death is counted as the outcome.
The size of the matched sample also needs the right interpretation. Nearly two million people provide considerable statistical precision, yet they constitute only 17.4% of the eligible population. Matching removed people without suitable counterparts. That can improve the comparison while narrowing the population to which it applies. The study also excluded institutionalized and homeless individuals. It would be inappropriate to multiply its risk difference across all vaccinated Floridians, much less the national population, and call the result a count of vaccine-caused deaths.
There is relevant evidence beyond Florida. In a nationwide Veterans Affairs study, Dickerman and colleagues found higher risks of several adverse outcomes among Pfizer recipients than among Moderna recipients over 38 weeks, including myocardial infarction, ischemic stroke, and other thromboembolic events. That gives the Florida cardiovascular findings a precedent. The VA investigators also acknowledged that better protection against infection could partly explain the apparent safety differences (Dickerman et al., 2022b).
The earlier VA effectiveness study found advantages for Moderna on several COVID-19 outcomes, while the estimated COVID-19 mortality difference was small and uncertain. As the Florida authors acknowledge, that earlier study did not find a difference in non-COVID-19 mortality. Differences in populations and follow-up matter, but they do not justify describing the literature as uniformly confirming the Florida result. Agreement on some cardiovascular outcomes and disagreement on non-COVID-19 mortality should both shape the next investigation (Dickerman et al., 2022a; Levi et al., 2026).
The absence of an unvaccinated comparison group defines another boundary. A difference between two products cannot establish the effect of either product relative to receiving neither. The observed pattern could reflect greater benefit from Moderna, greater harm from Pfizer, different combinations of benefit and harm, residual differences between recipients, or several of these at once. Both products could also share an effect that their comparison would fail to reveal. Adding unvaccinated people would introduce its own confounding problems, so the answer is a carefully designed additional comparison, not an assumption that an unvaccinated control automatically resolves the question.
Nor is this an isolated test of the first injection. Most recipients completed a second dose, and the one-year period allowed subsequent vaccination and changes in health care. The analysis follows people according to the product with which they began. That can be a useful practical comparison, but assigning the entire difference to the biological effect of dose one would misstate the exposure being studied. The authors themselves caution that these findings do not establish the comparative safety of newer formulations.
The disclosures should be read with the same discipline as the methods. Mansuri, Jordan, and Ladapo report Florida Department of Health support; Ladapo serves as State Surgeon General. Levi discloses an ACIP role and a financial relationship with Bluebell Foundation/Chicago Community Trust. The authors report no external study funding. These relationships belong in an assessment of the work, just as manufacturer relationships belong in an assessment of industry-associated research. None determines whether the reported association is correct or whether it is causal.
A credible next step is independent replication using vaccination records linked to detailed clinical histories and validated mortality data. Investigators should measure frailty, comorbidities, recent health-care use, and medications; test tighter date and facility matching; examine infection capture; and account explicitly for subsequent doses. They should publish cause-specific competing-risk estimates, explain the handling of migration and out-of-state deaths, and provide the code and definitions needed to reproduce their results. These steps address identifiable weaknesses in the present comparison.
The public-health implication already supported by this study is that comparative vaccine evaluation should examine overall survival and serious outcomes beyond the target disease. The unresolved mortality difference makes that work necessary even before its cause is settled. An observational association can be sufficient to justify investigation without being sufficient to assign injury to an individual patient or change a recommendation for a different population and formulation.
The Florida study leaves us with a substantial finding and an unfinished explanation. Its detailed matching, first-dose follow-up, and all-cause endpoint deserve credit. Its missing clinical histories prevent the confidence intervals from carrying the full weight of a causal conclusion. The responsibility now is to determine how much of the mortality gap survives a comparison that measures those histories. If the difference persists, product-specific explanations gain force. If it contracts, the sources of that contraction will tell us why the original groups differed. Either outcome would give patients information they do not yet have.
References
Levi R, Mansuri F, Jordan MM, Ladapo JA. (2026). Twelve-month all-cause mortality after initial COVID-19 vaccination with Pfizer-BioNTech or mRNA-1273 among adults living in Florida: Statewide cohort study. Global Epidemiology. Accepted journal pre-proof. https://doi.org/10.1016/j.gloepi.2026.100296
Dickerman BA, Gerlovin H, Madenci AL, et al. (2022a). Comparative Effectiveness of BNT162b2 and mRNA-1273 Vaccines in U.S. Veterans. New England Journal of Medicine, 386, 105–115. Published online December 1, 2021. https://doi.org/10.1056/NEJMoa2115463
Dickerman BA, Madenci AL, Gerlovin H, et al. (2022b). Comparative Safety of BNT162b2 and mRNA-1273 Vaccines in a Nationwide Cohort of US Veterans. JAMA Internal Medicine, 182(7), 739–746. https://doi.org/10.1001/jamainternmed.2022.2109
IPAK-EDU is grateful to Popular Rationalism as this piece was originally published there and is included in this news feed with mutual agreement. Read More
Leave a Reply