Why This Matters
In April 2025 Florida’s Department of Health posted a medRxiv preprint claiming adults who received the Pfizer‑BioNTech COVID‑19 vaccine had 38 % higher 12‑month all‑cause mortality than adults matched to receive Moderna. The first author is Joseph A. Ladapo, Florida’s Surgeon General and a tenured professor at the University of Florida - an unusual dual role that journalists and UF faculty have flagged for potential conflicts of interest and irregular research oversight.
Having a political appointment or prior views does not automatically invalidate anyone’s science, especially when conflicts are properly disclosed, as they were here. Still, such dual roles mean readers should examine the work with extra care. Let’s do that.
1 · What the Study Tried to Show
The authors extracted Florida immunization and death-certificate data for 9.16 million adults who completed a two-dose mRNA series between December 2020 and August 2021. They then created one-to-one "exact match" pairs between Pfizer and Moderna recipients based on seven characteristics: age band, sex, race, ethnicity, month of vaccination, vaccination-site type, and census tract. After matching, only 735,050 people in each group remained - just 16 percent of the total dataset. They then counted any death within 12 months of the second dose. Result: 1.0% of the Pfizer group died, versus 0.7% of the Moderna group - an odds ratio of 1.38. The authors used homicide and suicide deaths as a "negative control" and, finding no difference, claimed minimal residual bias.
At first glance, this setup looks standard - no major methodological red flags in terms of how the data is matched or how outcomes are defined. The authors clearly know how to conduct this type of analysis and structure a health data study. However, the implications of a claim that one vaccine brand is associated with significantly higher mortality are extremely serious. This is where the principle applies: extraordinary claims require extraordinary evidence. When the consequences of a finding could affect vaccine confidence or public health decisions at scale, even subtle limitations or untested assumptions need to be evaluated with care.
2 · Confounding, Matching, and What Was Left Out
When comparing vaccine outcomes, it’s essential to consider confounding factors - hidden differences between groups that might affect the outcome. In this case, important potential confounders include chronic illness, frailty, history of COVID infection, the timing and uptake of booster doses, and socioeconomic status differences that are not fully captured by census-tract averages. If one group had more people with chronic conditions or was vaccinated earlier when COVID risk was higher, their baseline risk of death would naturally be higher, regardless of which vaccine they received.
The authors addressed this by using exact matching on seven factors: age band, sex, race, ethnicity, month of vaccination, type of vaccination site, and census tract. This is a reasonable approach, and the list is not short. However, they did not include many medically relevant variables such as comorbidities, prior infections, disability status, or whether someone later received a booster. These omissions matter, especially when studying all-cause mortality, because they leave the door open for major hidden biases.
Moreover, the matching procedure led to the exclusion of 84 % of vaccinated individuals. This extreme data loss suggests that the original Pfizer and Moderna groups were not very similar - if they had been, most people could have been matched. Within that subset, propensity matching has limitations. It balances what’s measured, but not what’s unmeasured. If unmeasured differences (like health status) strongly affect mortality, then even perfectly matched groups can still produce misleading comparisons.
A practical analogy: imagine comparing student performance between public and private schools. You only keep the students who match exactly on age, parental income, and neighborhood. You discard 84 % of your data. The students who remain might be comparable on paper, but if you didn’t account for factors like parental education, special needs, or prior academic history, your results can still be skewed. The same holds here.
While the authors employed an accepted statistical tool, the way it was applied, combined with what it left out, raises genuine concerns about whether the groups were truly comparable in the ways that matter for mortality.
4 · Weak “Negative Controls”
The authors likely understood that their matching approach might not fully account for all relevant differences between the Pfizer and Moderna groups. To address this, they included a so-called "negative control" outcome - deaths from homicide and suicide - under the logic that these types of deaths should not be affected by vaccine brand. If the rates of homicide and suicide were similar across groups, the thinking goes, then the groups must have been sufficiently balanced.
But this control is weak, for two main reasons. First, these are rare events, in a middle-aged-plus cohort, violent deaths are infrequent, and random variation easily hides moderate imbalances. Second, and more importantly, the risk factors that drive all-cause mortality, like heart disease or immunosuppression, have almost nothing to do with whether someone is murdered or dies by suicide. Equal homicide rates, therefore, tell us very little about whether the two groups were truly similar in terms of underlying health.
A better negative control would have been short-term hospitalisations - events that occur more frequently and correlate more closely with general health status. As it stands, the chosen control offers limited reassurance that bias was adequately addressed.
5 · Ignoring Florida’s Roll‑Out Strategy
Florida’s “Seniors First” policy (December 2020) sent Pfizer to hospitals and nursing homes first, while Moderna scaled up in county clinics weeks later. This policy led to a significant practical difference: older, frailer adults were more likely to receive Pfizer during the deadliest phase of the pandemic, while relatively healthier individuals received Moderna later under different conditions.
This real-world vaccine deployment context fits precisely with the problem described earlier: if the original populations who received Pfizer and Moderna were very different - as suggested by the need to exclude 84% of people during matching - then statistical comparisons become vulnerable to bias. Propensity matching cannot fully correct for systematic differences in how vaccines were offered to different groups. When the rollout strategy and clinical context have already selected for different risk profiles, observed outcome differences may reflect those selection effects more than any intrinsic property of the vaccines themselves.
6 · No Unvaccinated Baseline
One further omission stands out: the complete exclusion of an unvaccinated comparison group. The authors, who clearly understand how to conduct complex observational research, made a deliberate choice not to include what would seem like the most obvious and relevant baseline - how mortality in vaccinated individuals compares to those who received no vaccine at all. That decision is striking. Including an unvaccinated group would have provided context for whether the mortality observed in either the Pfizer or Moderna cohorts was elevated or suppressed relative to background risk. Without it, readers are left with a relative comparison that may overemphasise internal differences without showing the net benefit, or lack thereof, compared to being unvaccinated. In studies where public messaging and policy may be influenced, that missing reference point matters.
7 · Key Points for Readers
Matching is useful, but losing most data is a warning sign. Unmeasured confounders can fully explain observed gaps. Negative controls must be relevant and common enough to be informative. Findings should fit the broader evidence landscape. Extraordinary claims need robust, transparent analysis.
Public‑health guidance should continue to rest on the convergence of multiple rigorous studies, not on a single un‑reviewed analysis that slices away most of the available data.
No comments:
Post a Comment