Officials knew vaccine safety surveillance was flawed - sidelined the doctor.

“The strategy that she is using could create erroneous conflicts that feed in to anti-vaccination rhetoric.”

That was Peter Marks, FDA’s top COVID vaccine regulator, asking another center to make career safety officer Ana Szarfman cease and desist. Not “the math is wrong.” Don’t feed the critics.

React19 members were told, over and over, that there was “no signal.”

A new BMJ investigation by David Willman shows what that phrase was sitting on.

Before the shots went out in December 2020, CDC promised VAERS would be the country’s early-warning system. Two tools were supposed to run together: CDC’s proportional reporting ratios and FDA’s empirical Bayesian data mining. CDC later admitted it did not run the PRR analyses until 2022. For more than a year, officials leaned on the FDA method and called it more robust.

It was not robust for this rollout. It was easy to blind.

In the first year, more than 90 percent of 1.1 million VAERS reports were for Pfizer or Moderna. The Bayesian tool compares one vaccine with “other vaccines.” The comparison product was mostly the other mRNA shot. If both raised the same problem, observed and expected rates looked similar and the alarm never rang. Statisticians call that masking.

Szarfman and William DuMouchel, the statistician behind the original method, warned Marks in early 2021. She emailed him after the New York Times reported that other safety systems were not ready. On February 29 she showed him the problem in one slide:

PRR lit up red with signals.

FDA’s official tool stayed green and found almost nothing. She said they were not getting useful information. Marks’ center did not switch.

That is the backdrop for the public line on myocarditis. Israel and the Pentagon were already reporting clusters in young people. Walensky told reporters they had looked intentionally across hundreds of millions of doses and seen no signal. VAERS already had hundreds of heart reports. When CDC wanted to warn clinicians, Marks objected because myocarditis and pericarditis “have not actually signaled.” Gregory Poland at Mayo got severe tinnitus after dose two and was told the same thing.

Weekly FDA mining reports reviewed by BMJ never flagged myocarditis, pericarditis, or tinnitus.

CDC’s Tom Shimabukuro later showed that young males had far more myocarditis and pericarditis reports than background rates would predict. The day before that talk he received FDA’s latest mining report. Still no alert. He emailed a colleague:

“I’m perplexed that myocarditis isn’t alerting.”

He knew the siren was dead.

On May 7, 2021, FDA’s Craig Zinderman told Szarfman to stop creating and sending COVID vaccine data-mining reports. Changing the method mid-campaign, he wrote, could cause confusion and have “unintended consequences (eg, regarding vaccine confidence).” They pointed her to other systems such as VSD. VSD also missed myocarditis in April 2021 while looking for it, and still had not signaled in August, after officials had already accepted a causal link.

In July, Szarfman brought a strong signal for death and sudden death from DuMouchel’s updated method.

Richard Forshee told Marks he thought it was spurious because of who got vaccinated first and because reporting was being pushed so hard. Then came Marks’ cease-and-desist email about anti-vaccination rhetoric.

Then he sat with React19’s membership, who had prolonged neurologic symptoms after vaccination. He said FDA would look for a neurologic signal. Three months later he told them they had looked and had not found a rate higher than expected. He said that after his shop had shut down the analyst whose method was designed to unmask the signals patients were describing.

CDC’s delayed PRRs did not quietly confirm that everything was fine. In thirteen straight weekly reports from May to July 2022, comparing mRNA shots with non-COVID vaccines, CDC flagged myocarditis, pericarditis, Bell’s palsy, and tinnitus.

The FDA tool stayed quiet.

A journal reviewer later asked why anyone would run Bayesian mining at all if more than 90 percent of VAERS was the vaccines of interest. The chance of an alert was near zero. FDA coauthors agreed there was “considerable bias toward the null,” pulled the analysis, and took their names off the paper.

In 2023, FDA pharmacovigilance chief Narayan Nair wrote that they had known this limitation before and during the pandemic. DuMouchel, who received Moderna himself, was not looking for a scandal. His verdict was simple: they were wrong, and if they had listened to Szarfman they would have done better.

React19 was not crazy for noticing the gap between lived injury and official calm.

Myocarditis was real. Tinnitus reports were real. Neuro injury was real.

“No signal” often meant the detector had been left on a setting that could not hear it, then used as proof there was nothing to hear.

The investigation is here: