The detection bar was set higher than anywhere else
https://www.medrxiv.org/content/10.1101/2025.01.03.25319975v1.full
inga314 Analysis
In January 2025, scientists at the U.S. Food and Drug Administration published a paper on medRxiv. They had analyzed health records from 7.6 million elderly Americans who received the 2023–2024 COVID-19 vaccines. They were looking for serious health problems: heart attacks, strokes, heart inflammation, nerve damage, severe allergic reactions, blood clots.
The first sentence of the abstract read: “COVID-19 vaccines are well-established as safe.”
That sentence is not a finding. It is the conclusion, placed before the evidence. A genuine safety surveillance study would ask: what are the risks? This one opened by declaring there were none. Everything that follows is shaped by that declaration.
By October 2025, the FDA directed its own scientists to withdraw the paper from the journals that had accepted it. The official reason given by HHS: “The authors drew broad conclusions that were not supported by the underlying data.”
inga314 agrees. Here is the full case for why.
Finding 1: They found a signal and made it disappear
The paper found a statistically significant safety signal.
Among people who received the Pfizer-BioNTech vaccine, the rate of severe allergic reactions — anaphylaxis — was four times higher in the weeks after vaccination than in the comparison period. Four times. The result cleared even the paper’s own stringent statistical bar.
In pharmacovigilance — the science of drug safety — a fourfold elevation that survives statistical testing is not a footnote. It is the finding. It is what the study exists to detect.
Here is what happened next.
A second adjustment was applied — specifically to this one result, after the signal appeared. The method is called PPV-based imputation for “potential outcome misclassification.” Here is exactly what it did to the numbers:
The risk estimate moved from 4.04 to 3.90. The risk barely changed — a movement of 0.14. Nearly fourfold elevated before. Nearly fourfold elevated after.
But the confidence range — the statistical window around the estimate — expanded enormously. Before the adjustment: [1.07 to 15.30]. The entire range sat above 1.0, the line that separates signal from no signal. After the adjustment: [0.49 to 30.90]. The lower end dropped below 1.0.
Once the lower end crosses below 1.0, a result is classified as non-significant. The signal disappeared — not because the risk went away, but because the adjustment made the researchers less certain about a risk that was still nearly fourfold elevated.
The paper then concluded: “No new safety signals identified.”
The parameters of this adjustment are not disclosed in the main text. What positive predictive value was assumed for the anaphylaxis billing code? Where did that value come from? What happens to the result if a different value is used? None of this is provided. The reader cannot verify whether the adjustment was appropriate or whether it was chosen to produce a specific outcome.
This is the most important sequence in the entire paper. A signal was found. A targeted adjustment was applied. The signal disappeared. The conclusion said nothing was found.
Finding 2: The detection bar was set higher than anywhere else
To call something a safety signal in science, a result has to pass a statistical threshold. The standard threshold in drug and vaccine safety research — used by the FDA in all its prior COVID-19 safety studies, used by the European Medicines Agency, used by the World Health Organization — corresponds to 95% confidence.
This paper used 99% confidence.
The difference matters more than it sounds. At 95%, you are willing to flag a result if you are 95% certain it is real. At 99%, you need to be 99% certain. That extra 4% of required certainty means the bar is substantially higher. Results that would be called signals at 95% are not called signals at 99%. The higher the confidence requirement, the fewer signals a study will find.
The paper justifies this in a single sentence: “to reduce false positive signals.” In drug safety, this justification is backwards. A false positive means you investigate something that turns out to be harmless. A false negative means you miss a real harm and people continue to be exposed to it. The entire purpose of post-authorization safety surveillance is to catch false negatives — harms that were not detected before approval. Designing a study to minimize false positives in this context is designing it to minimize its own ability to do its job.
No explanation is given for why this specific study departed from the standard threshold. No sensitivity analysis comparing results at 95% versus 99% confidence is presented. The reader is never shown what the results would look like under the standard used by every comparable study.
This single choice determined which results were called significant and which were not. It shaped the conclusion of the paper more directly than any other decision.
Finding 3: They only watched for 42 days
The study divided post-vaccination time into two periods. Days 1 to 42 after vaccination: the risk window — any event here could be vaccine-related. Days 43 to 90: the control window — treated as normal background, unaffected by the vaccine.
This 42-day boundary was not derived from biology. It was not measured. It was imported from a general methodological guideline developed for vaccine safety research across all populations and all vaccine types. The assumption embedded in it is precise: by day 43, the vaccine’s biological effects on every condition being monitored have fully resolved.
For elderly patients receiving mRNA COVID vaccines, this assumption is not justified for several of the conditions the study was monitoring.
Guillain-Barré syndrome — a condition in which the immune system attacks the nervous system — develops through a process of molecular mimicry that can unfold over weeks to months. Cases presenting beyond 42 days after vaccine administration are documented in the medical literature. An elderly patient who begins developing nerve damage on day 50 after vaccination has their case counted as normal background in this study.
Myocarditis and pericarditis — inflammation of the heart — in elderly patients follows a different timeline than in the young adults in whom vaccine-associated cardiac inflammation was first characterized. Inflammatory processes in older bodies resolve more slowly. Late presentation beyond six weeks is real and documented.
Thrombotic events — blood clots, pulmonary embolism, stroke — can result from sustained endothelial activation and disruption of the coagulation system. Some research suggests that mRNA vaccine-related immune activation persists beyond six weeks in some individuals. A clot forming on day 50 is not counted as a potential vaccine effect.
Every event that happens after day 42 is absorbed into the control window and treated as normal. This inflates the control window event rate. An inflated control rate makes the first 42 days look less unusual by comparison. The signal is compressed.
The 99% confidence interval suppresses signals from above — by requiring more certainty. The 42-day window suppresses signals from below — by inflating the denominator. They operate on different parts of the same calculation and move in the same direction.
Finding 4: The vaccine did not prevent COVID — and COVID was everywhere
This is the finding that most directly breaks the study’s core logic.
The study works by comparing two time windows for each person: the 42 days after vaccination versus the following 48 days. The assumption is that the second window — the control — represents the person’s true normal baseline, unaffected by the vaccine.
For this assumption to hold, the two windows have to be comparable. The main difference between them should be the vaccine’s effects, which are present in the risk window and resolved by the control window.
The study was conducted from September 2023 through April 2024. The JN.1 variant of COVID-19 emerged in September 2023 and became dominant through the autumn and winter. This matters because the 2023–2024 XBB.1.5 vaccine — the vaccine being studied — was designed against an earlier variant. Against JN.1 specifically, its effectiveness against infection was reduced. Multiple independent studies found that effectiveness against infection in elderly adults was approximately 50% overall, and lower still against JN.1.
Roughly half of the 7.6 million vaccinated Medicare beneficiaries in this study were not protected against COVID infection.
COVID-19 independently causes heart attacks, strokes, blood clots, and cardiac inflammation. This is established by multiple large studies conducted before vaccines existed — when the question was clean and vaccination could not be a confound. A Swedish study of 86,742 patients published in The Lancet found that AMI risk was elevated approximately threefold in the first two weeks after COVID infection.
Now consider what this means for the study’s comparison.
COVID infections happened throughout September and October 2023 — the risk window for the earliest vaccinees. They also happened throughout November and December 2023 — the control window. The JN.1 wave intensified through the autumn and peaked in late 2023 and early 2024. This means COVID infections may have been more frequent during the control window than during the risk window.
COVID infections in the risk window add cardiac events to the risk window — events that are then attributed, in this study’s model, to the vaccine. COVID infections in the control window add cardiac events to the control window — inflating the baseline rate and making the risk window look less elevated by comparison.
Both windows are contaminated by the same pathogen whose cardiac effects are indistinguishable from the vaccine’s cardiac effects in this design. The vaccine did not prevent the contamination because the vaccine did not reliably prevent infection.
The SCCS design was built for a clean, discrete exposure with a clean, uncontaminated control period. Neither condition exists here. The entire comparison is invalid as a result.
Finding 5: Prior COVID infection was completely ignored
Beyond the ongoing infection problem during the study period, there is the prior infection problem before it began.
Many of the 7.6 million people in the study had COVID-19 in the weeks or months before they vaccinated in September and October 2023. COVID’s damage to the cardiovascular system — endothelial injury, microclots, inflammation — does not end the day the acute infection resolves. It decays over weeks. The elevated cardiac risk from a COVID infection in August is still present — at a lower level — in September and October.
A person infected in August who vaccinates in September and has a heart attack in October is carrying COVID-related cardiac risk that has not fully resolved. This study counts that heart attack as a post-vaccination event in the risk window. The prior COVID infection is invisible. The study never measured it.
The word “infection” in the context of prior exposure does not appear anywhere in the paper — not in the methods, not in the limitations, not in the discussion. The data to identify recently infected Medicare beneficiaries exists in the same claims database the study used. COVID-19 diagnosis codes are standard in Medicare billing. The adjustment was technically feasible. It was not made.
Without knowing which patients had recent prior infections, the study cannot separate three distinct causal explanations for any cardiac or neurological event:
The first explanation: the event was caused by lingering COVID damage. The vaccine had nothing to do with it.
The second explanation: the event was caused by the vaccine directly — through inflammatory responses, spike protein effects on blood vessel walls, or disruption of the clotting system.
The third explanation: the event was caused by the combination. A person whose immune system was already activated and stressed by COVID received a vaccine that triggered a second wave of immune activation. The combination pushed them past a threshold that neither exposure alone would have reached.
The third explanation is the most biologically plausible for elderly patients with recent prior infection. A weakened cardiovascular system, already dealing with the aftermath of COVID, receiving an immune stimulus that would be tolerable in a fully recovered person — this is a credible mechanism for excess cardiac events that would be entirely invisible in this study.
When none of these three explanations can be separated, “no signal detected” cannot be interpreted as “no risk.” It means the question was not answered.
Finding 6: The sickest patients were excluded before the study began
The study uses a design called the self-controlled case series. It compares each person against themselves at different time points. This is mathematically elegant because it controls for stable individual characteristics — a person’s age, genetics, baseline health — by using the person as their own control.
What it cannot control for is who is in the study at all.
In the elderly Medicare population, the decision to get vaccinated in autumn 2023 is not random. It correlates with health. People who were vaccinated were, on average, healthier, more mobile, more engaged with the healthcare system, and less frail than people who were not vaccinated. The people who chose not to vaccinate — or were told not to by their doctors — include the group with the highest burden of cardiovascular disease, the most severe comorbidities, and the highest baseline probability of experiencing a serious adverse event.
This group is entirely absent from the data.
The study is measuring vaccine safety in a pre-selected population of people healthy enough to have chosen vaccination. The risk estimates produced from this group are systematically lower than what would be found in the full elderly population — including the people most vulnerable to vaccine-related harm.
There is no adjustment for this. It is not mentioned in the limitations.
Finding 7: The records used are billing documents, not medical records
Every health event in this study — every heart attack, every case of nerve damage, every allergic reaction, every blood clot — was identified by a billing code in an insurance claim. These codes exist to justify payment to hospitals and clinics. They are generated by billing departments under pressure to process claims quickly and accurately enough for reimbursement.
For some conditions, billing codes are reliable proxies for clinical diagnoses. For others, they are not.
For myocarditis — inflammation of the heart muscle — studies comparing billing code ascertainment against clinical record review have found that claims-based methods capture roughly 40 to 60 percent of actual cases. The majority of real myocarditis cases generate billing codes for something else: chest pain, cardiac arrhythmia, unspecified heart condition. The myocarditis itself is systematically under-counted.
For anaphylaxis — the severe allergic reaction for which the paper found its fourfold signal — the problem is worse. In emergency settings, anaphylaxis frequently generates billing codes for its symptoms: airway obstruction, hypotension, allergic reaction, urticaria. The unifying diagnosis of anaphylaxis is under-coded. The paper found a fourfold signal through this filter — which means the true rate of anaphylactic reactions is likely higher than what was measured, not lower.
The study found a signal despite undercounting events. Then it applied an adjustment that cited outcome misclassification as the reason to widen the confidence interval. The same problem that made the signal smaller in the first place was used as the justification for making the uncertainty larger until the signal disappeared.
Finding 8: The regulator was marking its own homework
The FDA authorized the 2023–2024 COVID-19 vaccines. The FDA then designed a study to check whether those same vaccines were causing serious health problems. The FDA controls the Medicare claims database used for the analysis. The FDA set the statistical threshold at 99% instead of the standard 95%. The FDA defined which health outcomes qualified for analysis and which were dropped. The FDA conducted the analysis through Acumen LLC, a contractor funded by the FDA that works exclusively on FDA-directed projects. The FDA published the findings.
At no point in this chain does an independent party enter.
Independent drug and vaccine safety surveillance — the kind that produces reliable results — requires, at minimum, independent access to the data, independent statistical analysis, and independent adjudication of adverse events. The FDA’s commercial drug safety system, the Sentinel System, was designed partly to address this problem by using a distributed model. This study used none of those safeguards.
The paper contains no disclosure of this structural situation. There is no conflict of interest statement noting that the authorizing body is also the monitoring body. The authors are listed as FDA employees, but the institutional implications of this are not addressed.
This is not an accusation of individual misconduct. It is a structural observation: an institution whose public credibility is linked to having authorized a product has a systematic institutional interest in that product’s safety surveillance finding no new problems. That interest shapes decisions — about threshold choices, about which adjustments to apply, about how conclusions are framed — in ways that may not involve any deliberate act by any individual researcher.
Finding 9: Cardiac scars do not go away
Every prior finding in this analysis has involved a confound that, in principle, could be measured or adjusted for. Prior infection dates exist in the database. Vaccination records exist. Confidence interval thresholds can be recalculated. Window definitions can be extended.
This finding is different. It cannot be adjusted for. The data does not exist. It never will.
When COVID-19 causes inflammation of the heart — myocarditis, pericarditis, or subclinical cardiac injury — the resulting damage to heart muscle tissue produces scar tissue. Fibrosis. This is a permanent structural change. Unlike the acute vascular inflammation and clotting risk that peaks in the weeks after COVID infection and then decays, myocardial scarring does not resolve over time. It does not appear in billing records unless it causes a hospitalization. It cannot be identified from claims data. It is invisible to this study by design.
Scarred heart muscle is permanently different from healthy heart muscle. It is electrically unstable — more prone to dangerous arrhythmias. It is mechanically weaker — more vulnerable to heart failure under stress. The threshold for a serious cardiac event in a person with myocardial scarring is permanently lower than it was before the scarring occurred.
By autumn 2023, a substantial fraction of the 7.6 million Medicare beneficiaries in this study had experienced one or more COVID-19 infections. Research using cardiac MRI in recovered COVID patients — most of whom never had a hospitalization — has found rates of myocardial abnormality between 20 and 60 percent in various study populations. Most of this damage never generates a billing code. It is subclinical. The person feels recovered. But the scar is there.
This destroys the SCCS assumption at its foundation.
The self-controlled case series works by comparing each person against themselves across time. The elegance of the design is that each person serves as their own control, eliminating the effect of stable individual characteristics — genetics, age, baseline health. The assumption is that the person in the risk window and the person in the control window are the same person, with the same cardiac baseline.
A person with post-COVID myocardial scarring is not the same person they were before COVID. Their cardiac baseline has permanently shifted. That shift is present in October — the risk window. It is equally present in November and December — the control window. The scar does not make the risk window look more dangerous than the control window. It elevates the event rate in both windows equally.
This means the SCCS cannot see the damage. It compares elevated risk against elevated risk and records the ratio as normal. The absolute elevation above the person’s pre-COVID baseline is invisible — not because the study missed it, but because the study design cannot in principle detect it. You cannot compare against a healthy baseline that no longer exists.
If a vaccine administered to a person with post-COVID myocardial scarring causes an additional inflammatory insult to already damaged tissue — the second hit described in Finding 5 — that event is measured against a damaged cardiac substrate, not a healthy one. The threshold is lower. The apparent IRR reflects the difference between the risk window and the control window on top of permanent damage. The absolute risk above the person’s pre-COVID self cannot be computed from this data.
The 42-day window argument assumed a decaying confound. The prior infection argument assumed a confound that could be adjusted for if the data were used. This finding assumes neither. Myocardial scar does not decay. It cannot be measured in claims data. It cannot be adjusted for. The population receiving vaccines in September 2023 is, to an unknown and unknowable degree using this dataset, a population whose cardiac baselines have already been permanently altered by prior COVID infections that left no trace in the billing records.
The study is measuring cardiac event rates in people whose hearts it cannot characterize, comparing windows of time on top of invisible damage, and concluding the vaccine had no effect on cardiac outcomes.
What this means for heart attacks
The paper reported no elevated risk of heart attacks. This cannot be read as strong evidence that COVID vaccines are safe for the heart in elderly patients.
The population most at risk of cardiac events — known coronary artery disease, prior heart attacks, diabetes with vascular complications — was largely excluded by the healthy vaccinee effect.
A 20 to 30 percent elevation in heart attack risk would not register as significant with 99% confidence intervals. In 7.6 million elderly patients, that is thousands of additional heart attacks the study cannot see by design.
The 42-day window means late-onset thrombotic cardiac events are counted as background.
Prior COVID infection — which independently elevates acute cardiac risk — was never measured, making the attribution of any post-vaccination cardiac event unanswerable.
Throughout both windows, an active COVID wave contaminated the comparison in ways that cannot be separated.
And beneath all of this: an unknown fraction of the study population has permanent myocardial scarring from prior COVID infections that never generated a billing code. Their cardiac baseline is not the baseline this study assumes. The SCCS compares their risk window against their control window — both measured on top of invisible, permanent damage. A heart attack in a scarred heart triggered by a vaccine-related inflammatory insult looks identical, in this data, to a heart attack in a healthy heart from any other cause. The study cannot tell the difference because it cannot see the scar.
The null heart attack finding does not mean no elevated cardiac risk from COVID vaccination. It means: we could not detect a signal with this instrument, in this population, with these design choices, measuring event rates on top of permanent cardiac damage we cannot see, in the presence of an active COVID wave this vaccine did not reliably prevent.
These are very different statements. Only the first appears in the paper’s conclusion.
The withdrawal
In October 2025, the FDA directed its own scientists to pull this paper from the journals that had accepted it. HHS confirmed this publicly in May 2026. The stated reason: the authors drew broad conclusions not supported by the underlying data.
This is technically accurate. inga314 reached the same conclusion independently — before the withdrawal was announced, without any political interest in the outcome.
The withdrawal occurred under HHS Secretary Robert F. Kennedy Jr., as part of a documented pattern of actions affecting vaccine policy and access. Some credible researchers reviewed the paper and said the underlying methodology was not inherently flawed. That assessment addresses the methods in isolation. It does not address whether the conclusions those methods produced were warranted.
A study that found a fourfold anaphylaxis signal and concluded no signals were found is not an honest document — regardless of who finds it useful to say so, and regardless of what policy conclusions are drawn from that observation.
What inga314 found
inga314 asks one question: do the conclusions follow from the data and methodology?
Here, they do not.
The paper found a fourfold anaphylaxis signal. The methodology consumed it through a post-hoc adjustment that widened statistical uncertainty without changing the measured risk. The conclusion declared nothing was found.
For every other outcome — heart attacks, strokes, blood clots, nerve damage — the null findings rest on an instrument that cannot detect modest elevations, excludes the highest-risk patients, uses a time window that misses late-onset effects, and operates throughout a period when an active COVID wave contaminated both comparison windows because the vaccine being studied did not reliably prevent infection.
The study cannot separate vaccine-caused events from COVID-caused events from the combination of both. It treats that inability as equivalent to absence of risk. It is not.
The HHS withdrawal justification is correct on the narrow technical point. What follows from it — in policy, in clinical practice, in how we assess vaccine safety going forward — is outside inga314’s scope.
The paper failed the test of logical integrity before anyone decided it was useful to say so.
inga314.ai — They find papers. We find flaws.
Paper: Gruber JF, Ondari M, Zelaya CE, et al. “Safety Monitoring of Multiple Health Outcomes Following 2023–2024 COVID-19 Vaccination among Medicare Beneficiaries Aged 65 Years and Older in the United States.” medRxiv 2025.01.03.25319975. Withdrawn from journal publication, October 2025.
System D 