Who Belongs in the Denominator
The best statistical lies contain no false numbers.
In August 2026, the Senate Homeland Security Permanent Subcommittee on Investigations released a batch of Fauci-era texts alongside a rebuttal claiming that COVID-19 mRNA vaccination produced an 82% miscarriage rate in early pregnancy. The claim traces to a January 2022 open letter from OB/GYN James Thorp. Every number in that calculation is real. Every input was published in the New England Journal of Medicine.
The arithmetic is correct. The answer is wrong.
The original math
Shimabukuro et al. published preliminary V-safe registry data in April 2021, covering vaccinations from December 14, 2020 through February 28, 2021, with outcomes followed up by telephone through March 30. Of 3,958 enrollees, 827 had a pregnancy that had completed and been reported by that cutoff. Of those, 104 ended in loss before 20 weeks.
Three terms, as they are used here:
- Spontaneous abortion, or miscarriage: pregnancy loss before 20 weeks of gestation.
- Stillbirth: fetal death at 20 weeks or later.
- Completed pregnancy: a registry pregnancy whose outcome had been reported by the time of analysis.
The first two follow US reporting conventions, and other countries draw the line elsewhere. The 20-week boundary matters here because it is what makes a third-trimester vaccination incapable of producing a miscarriage.
That denominator is wrong. A footnote in Shimabukuro’s Table 4 says 700 of the 827 received their first dose of the COVID vaccine in the third trimester. A woman vaccinated at 32 weeks cannot lose a pregnancy before 20 weeks. She was never at risk of the outcome. Including her dilutes the denominator.
Hong Sun said so in a letter to the journal. The CDC authors agreed: 827 is not an appropriate denominator for a risk estimate. The journal issued a correction.
Everybody agreed the denominator was wrong.
The edit
In November 2021, Aleisha Brock and Simon Thornley published a reanalysis in Science, Public Health Policy & the Law. Same observation about the 700, then the obvious move:
Then divide the 104 losses by what is left:
That’s the 82%. That’s the whole derivation. The first operation is correct. The second is not.
Correct in direction, at least. The third trimester begins at 28 weeks, so those 127 still include women vaccinated between 20 and 27 weeks, who also cannot lose a pregnancy before 20 weeks. Brock and Thornley described their cohort as exposed before 20 weeks. The arithmetic does not produce that cohort.
Why removal is invalid
“Completed pregnancy” is doing two jobs, and Brock and Thornley only credited one.
To land in the 827, a pregnancy had to clear two filters: vaccination inside the study window, and an outcome already reported by the data cut. Removing 700 women corrects the first. It does nothing about the second, with the second filter selecting on outcome.
Follow-up ran fifteen weeks at the outside, and far less for anyone enrolling late. Gestation is forty. For a woman vaccinated at ten weeks, a loss reaches the registry almost immediately. A birth reaches it thirty weeks later, long after reporting closed. Only one of them could be recorded by March 30.
The Shimabukuro paper lets you check this. Table 4’s footnote says 700 of the 827 received a first dose in the third trimester. The Results text says 700 of the 712 live births were among third-trimester vaccinees. Same 700, so every third-trimester vaccinee in the completed set delivered a live infant. The paper never cross-tabulates outcome by trimester, but the remainder is forced:
| First dose | Completed | Live births | Live-birth rate |
|---|---|---|---|
| Third trimester | 700 | 700 | 100% |
| Before third trimester | 127 | 12 | 9.4% |
Same vaccine, same registry, same follow-up window. The distance between a 100% live-birth rate and a 9.4% one is not pharmacology. It is how many calendar days remained between the first injection and the data cutoff, which is what “completed pregnancy” measures.
Both papers counted the wrong person-time, in opposite directions.
- Including the 700 is immortal time bias. Pregnant women held in a denominator across a stretch when the outcome was impossible for them.
- Dropping the still-pregnant is informative censoring. A pregnancy leaves the count for a reason tied to how it ends. Losses reach the registry and get counted. Healthy pregnancies are still pending, so they get dropped.
This is survivorship bias inverted. Here, the survivors are the ones thrown out.
The denominator that works
You need everyone at risk during the interval, including the women who are still pregnant. That’s a life table with censoring for incomplete follow-up.
Zauche et al. ran it in September 2021, and the complete registry data landed in Vaccine in March 2026.
| Analysis | Denominator | Result |
|---|---|---|
| Shimabukuro et al. (2021) | 827 completed pregnancies | 12.6% |
| Brock & Thornley (2021, retracted) | 127 completed, dosed pre-third-trimester | 81.9% |
| Zauche et al. (2021) | 2,456 at risk, censored | 14.1% |
| Madni et al. (2026) | 12,097 at risk, censored | 10.79% |
| Pre-pandemic background | — | 11–22% |
The first two rows divide the same 104 losses out of the same snapshot. The only thing that changed between 12.6% and 81.9% is the denominator.
Zauche’s denominator includes 2,022 pregnancies still ongoing at 20 weeks. Brock and Thornley’s denominator had no room for pregnancies like these. Not because anything was wrong with them. Because “going fine” had not yet reached its conclusion.
The number outlived the paper
Thornley retracted the paper himself in November 2021, after the head of Auckland’s School of Population Health emailed the faculty and named the error: incorrect denominator. He told 1News they had made a major mathematical error and that the conclusion was false.
The 82% reached the subcommittee through Thorp’s open letter rather than the retracted paper, but it is the same arithmetic, withdrawn by its own authors nearly five years earlier.
What to take from this
None of the arithmetic shown here is difficult. The only operations are subtraction and division. Deciding who to include in a study is a challenge, and understanding what stays in but gets censored is even more so. The hard part is knowing who belongs in the denominator.
Sources:
- Shimabukuro TT, Kim SY, Myers TR, et al. Preliminary Findings of mRNA Covid-19 Vaccine Safety in Pregnant Persons. N Engl J Med. 2021;384(24):2273-2282. doi:10.1056/NEJMoa2104983
- Brock & Thornley, Science, Public Health Policy & the Law 4:130 (2021, retracted).
- Sun H. On Preliminary Findings of mRNA Covid-19 Vaccine Safety in Pregnant Persons. N Engl J Med. 2021;385(16):1535-1536. doi:10.1056/NEJMc2113516
- Zauche LH, Wallace B, Smoots AN, et al. Receipt of mRNA Covid-19 Vaccines and Risk of Spontaneous Abortion. N Engl J Med. 2021;385(16):1533-1535. doi:10.1056/NEJMc2113891
- Madni SA, Olson CK, Zauche LH, et al. Risk of spontaneous abortion after mRNA COVID-19 vaccination received just prior to or during pregnancy: Complete data from CDC COVID-19 vaccine pregnancy registry. Vaccine. 2026;76:128340. doi:10.1016/j.vaccine.2026.128340
- Goldhaber MK, Fireman BH. The fetal life table revisited: spontaneous abortion rates in three Kaiser Permanente cohorts. Epidemiology. 1991;2(1):33-39.
- Li DK, Odouli R, Wi S, et al. A population-based prospective cohort study of personal exposure to magnetic fields during pregnancy and the risk of miscarriage. Epidemiology. 2002;13(1):9-20. doi:10.1097/00001648-200201000-00004
- Mukherjee S, Velez Edwards DR, Baird DD, Savitz DA, Hartmann KE. Risk of miscarriage among black women and white women in a U.S. prospective cohort study. Am J Epidemiol. 2013;177(11):1271-1278. doi:10.1093/aje/kws393