Loading...
The URL can be used to link to this page
Your browser does not support the video tag.
Home
My WebLink
About
09.19.26 Board Correspondence - FW_ FDA Knew Its COVID Vaccine Safety Algorithm Could Hide Signals. It Used It Anyway.
.ATTENTION: This message originated from outside Butte County. Please exercise judgment before opening attachments, clicking on links, or replying.. From:Clerk of the Board To:Mutony, Heather Cc:Lee, Lewis Subject:Board Correspondence - FW: FDA Knew Its COVID Vaccine Safety Algorithm Could Hide Signals. It Used It Anyway. Date:Monday, September 21, 2026 8:32:30 AM Attachments:https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F78554d6e-3bc1-40fb-b2a2-35a095e70a94_1100x201.png https%3A%2F%2Fsubstack.com%2Ficon%2FLucideHeart%3Fv%3D4%26height%3D36%26fill%3Dnone%26stroke%3D%2523808080%26strokeWidth%3D2 https%3A%2F%2Fsubstack.com%2Ficon%2FLucideComments%3Fv%3D4%26height%3D36%26fill%3Dnone%26stroke%3D%2523808080%26strokeWidth%3D2 https%3A%2F%2Fsubstack.com%2Ficon%2FLucideShare2%3Fv%3D4%26height%3D36%26fill%3Dnone%26stroke%3D%2523808080%26strokeWidth%3D2 https%3A%2F%2Fsubstack.com%2Ficon%2Fnotes__NoteRestackIcon%3Fv%3D4%26height%3D36%26fill%3Dnone%26stroke%3D%2523808080%26strokeWidth%3D2%26strokeLinecap%3Dround%26strokeLinejoin%3Dround https%3A%2F%2Fsubstack.com%2Ficon%2FLucideArrowUpRight%3Fv%3D4%26height%3D36%26fill%3Dnone%26stroke%3D%2523808080%26strokeWidth%3D2 image001.png Please see Board Correspondence - Lewis LeeAdministrative Technician - ConfidentialButte County Administration25 County Center Drive, Suite 200 • Oroville, CA 95965T: 530.552.3326www.buttecounty.ca.gov | lelee@buttecounty.ca.gov From: lance dreiss <lancedreiss@att.net> Sent: Saturday, September 19, 2026 6:47 AM To: Shared Mailbox Clerk of the Board <pcbs@countyofplumas.com>; Assemblymember.Gallagher@assembly.ca.gov; Senator.Dahle@senate.ca.gov; danpargee@countyofplumas.com; davidhollister@countyofplumas.com; District Attorney <District_Attorney@buttecounty.ca.gov>; Nicolereinert@countyofplumas.com; Kitts, Melissa <mkitts@buttecounty.ca.gov>; Soderstrom, Monica <msoderstrom@buttecounty.ca.gov>; Durfee, Peter <PDurfee@buttecounty.ca.gov>; Ronald Owens <ronald@muzzledtruth.com>; Kimmelshue, Tod <TKimmelshue@buttecounty.ca.gov>; Pickett, Andy <APickett@buttecounty.ca.gov>; Connelly, Bill <BConnelly@buttecounty.ca.gov>; Teeter, Doug <DTeeter@buttecounty.ca.gov>; Beaudoin, Jarett <JBeaudoin@buttecounty.ca.gov>; Julie Threet <julie4butte5@gmail.com>; Waugh, Melanie <mwaugh@buttecounty.ca.gov>; Ritter, Tami <TRitter@buttecounty.ca.gov>; Teri DuBose <Teri.DuBose@mail.house.gov>; Clerk of the Board <clerkoftheboard@buttecounty.ca.gov>; Stephens, Brad J. <BStephens@buttecounty.ca.gov> Subject: Fwd: FDA Knew Its COVID Vaccine Safety Algorithm Could Hide Signals. It Used It Anyway. Public Record Forwarded this email? Subscribe here for more #PopularRationalism #TheNEWMEDIACHANNEL FDA Knew Its COVID Vaccine Safety Algorithm Could Hide Signals. It Used It Anyway. Perhaps it’s time for IPAK to design The People’s Vaccine Safety Tracking System. Thoughts? JAMES LYONS-WEILER, PHD SEP 17 READ IN APP A new BMJ investigation documents a preventable pharmacovigilance failure: a known masking defect, an available statistical remedy, andpublic reassurance built partly on the absence of alerts that the surveillance method was structurally capable of missing. There is a profound difference between saying, “We did not detect a safety signal” and saying, “We used a system capable of reliably detecting a safety signal, and none appeared.” diana dreiss Begin forwarded message: From: "Dr. James Lyons-Weiler from Popular Rationalism" <popularrationalism@substack.com> Date: September 17, 2026 at 6:13:44 AM PDT To: lancedreiss@att.net Subject: FDA Knew Its COVID Vaccine Safety Algorithm Could Hide Signals. It Used It Anyway. Reply-To: "Dr. James Lyons-Weiler from Popular Rationalism" <reply+3koho8&kcryl&&5c3f2addcf368e04ed9d007156beffe126028aca84c08d0a164a17aeadf2a120@mg1.substack.com> Perhaps it’s time for IPAK to design The People’s Vaccine Safety Tracking System. Thoughts?͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ Forwarded this email? Subscribe here for more #PopularRationalism #TheNEWMEDIACHANNEL FDA Knew Its COVID Vaccine Safety Algorithm Could Hide Signals. It Used It Anyway. Perhaps it’s time for IPAK to design The People’s Vaccine Safety Tracking System. Thoughts? JAMES LYONS-WEILER, PHD SEP 17 READ IN APP A new BMJ investigation documents a preventable pharmacovigilance failure: a known masking defect, an available statisticalremedy, and public reassurance built partly on the absence of alerts that the surveillance method was structurally capable ofmissing. There is a profound difference between saying, “We did not detect a safety signal” and saying, “We used a system capable of reliably detecting a safety signal, and none appeared.” Those statements are not interchangeable. According to an investigation published September 16, 2026, in The BMJ, officials at the US Food and Drug Administration knew during the initial COVID-19 vaccine rollout that their principal statistical data-mining method could fail to detect adverse-event signals under the unusual reporting conditions created by mass COVID vaccination. An FDA physician and an internationally recognized statistician warned them. A more sophisticated method existed. Analyses using it produced signals that the routine method did not. Yet the older method remained in use. The BMJ investigation, by investigative journalist David Willman, relies on internal government correspondence obtained through the Freedom of Information Act and Senate investigations, along with interviews with scientists and public-health officials. The journal reports that FDA medical officer Ana Szarfman and statistician William DuMouchel warned senior officials in early 2021 that FDA’s statistical system suffered from a potentially consequential problem known as masking. The BMJ reports that Szarfman’s efforts to pursue the issue eventually drew an instruction to “cease and desist.” Upgrade to paid This is not a story about whether every report submitted to the Vaccine Adverse Event Reporting System, VAERS, represents an injury caused by a vaccine. VAERS cannot establish that. This is a story about whether the surveillance machinery responsible for detecting hypotheses worth investigating was statistically fit for that purpose—and what federal officials did after experts inside the system told them that it was not. What VAERS Is Supposed to Do VAERS is a passive adverse-event reporting system jointly operated by FDA and CDC. Reports can come from clinicians, manufacturers, patients, caregivers, and others. Reports describe events occurring after vaccination; temporal association does not establish causation. FDA states this explicitly. VAERS lacks the denominator structure necessary to calculate incidence rates directly, and reports can suffer from stimulated reporting, underreporting, incomplete information, duplication, and other biases. But pointing out those limitations does not make statistical signal detection irrelevant. It makes signal detection essential. The purpose of a pharmacovigilance system is not to declare every report causal. Its first task is to identify unusual patterns quickly enough that better data sources, case adjudication, medical records, epidemiological studies, and active-surveillance systems can investigate them. A signal is an alarm. It is not the fire investigation. Confusing those two ideas produces errors in both directions. Treating every report as causation exaggerates risk. Treating failure of an imperfect algorithm to sound an alarm as evidence that no problem exists can conceal risk. The latter is what makes the BMJ investigation so important. Upgrade to paid The Statistical Trap: When the Comparator Becomes Contaminated FDA has long used an empirical-Bayesian approach known as the Multi-Item Gamma Poisson Shrinker, or MGPS, for data mining. The method generates an Empirical Bayes Geometric Mean, or EBGM, representing disproportional reporting of particular product- event combinations. FDA commonly considers the lower fifth percentile of the EBGM distribution—EB05—and has historically used EB05 > 2 as an alert criterion. Bayesian shrinkage has a legitimate purpose. Sparse spontaneous-report data contain enormous random variability. Shrinking unstable estimates toward the null can prevent small numbers of reports from generating endless false alarms. But every statistical correction has assumptions. And every correction can fail when the data-generating process changes. FDA itself explains a critical property of EBGM: adverse events associated with the product under study contribute to the calculation of the expected frequency. FDA contrasts that feature with proportional reporting ratios, where reports involving the target product do not contribute in the same fashion to the expected count. That distinction becomes extremely important when one new class of products suddenly dominates the surveillance database. According to The BMJ, more than 90% of the early VAERS reports were associated with the Pfizer and Moderna mRNA COVID-19 vaccines. Imagine that two products introduced almost simultaneously both increase reports of the same adverse event. The signal-detection system asks, in essence: Is this event reported disproportionately often with this product compared with the reporting background? But what happens when the background itself has become saturated with a second product producing the same event? The expected frequency rises. Observed and expected move toward one another. The signal shrinks. The surveillance system begins using part of the phenomenon it is trying to detect as its own reference distribution. That is masking. The signal has not necessarily disappeared biologically. It has disappeared statistically. This Was Not Discovered in 2026 This point matters enormously. Masking is not an explanation devised retrospectively by critics after the vaccination campaign. It is an established pharmacovigilance problem. A 2010 investigation of masking in FDA’s adverse-event database explicitly studied whether one product could suppress the detection of signals associated with another. PMID 21077702, DOI 10.2165/11584390-000000000-00000. Research using European pharmacovigilance databases has likewise demonstrated that removing masking products can reveal disproportionality signals that were previously hidden. PMID 24243665. Most importantly, Szarfman and DuMouchel did not merely complain internally and disappear. They and colleagues published the problem. In June 2022, Drug Safety published “Signaling COVID-19 Vaccine Adverse Events,” authored by Rave Harpaz, William DuMouchel, Robbert Van Manen, Alexander Nip, Steve Bright, Ana Szarfman, Joseph Tonning, and Magnus Lerch. PMID: 35737293 DOI: 10.1007/s40264-022-01186-z The investigators examined VAERS data using three conventional signal-detection approaches and a regression-based method designed to deal with masking. Their conclusion was direct: COVID-19 vaccine adverse-event signals could remain undetected or delayed because of masking, while more advanced regression-based approaches could partially correct the problem. The paper explains the mechanism in unusually clear statistical terms. Conventional disproportionality methods compare the target product-event combination against a background rate. When another product contributes heavily to reports of the same event, the background becomes elevated and the target product’s apparent disproportionality can collapse. The authors provided a simple numerical example. With the masking product included, their hypothetical relative reporting ratio was approximately 0.98, suggesting essentially no association. Removing the product responsible for masking raised the same measure to approximately 4.14. Nothing about the underlying target reports changed. The comparison changed. And the signal appeared. The paper identified Regression-Adjusted Gamma Poisson Shrinker, or RGPS, as one approach capable of adjusting statistical associations for the presence of other products rather than treating a contaminated aggregate background as though it were an unbiased reference. The authors also emphasized the necessary scientific restraint: statistical signals generate hypotheses. They do not establish causation. That distinction is indispensable. It also cuts both ways. A positive signal does not prove causation. A negative output from an insensitive method does not prove safety. FDA Had the Warning During the Rollout The chronology reported by The BMJ changes this from an abstract methodological dispute into an institutional accountability problem. CDC had planned to use two approaches to examine VAERS data: proportional reporting ratios, or PRRs, and FDA’s empirical-Bayesian method. According to The BMJ, a later letter from then-CDC director Rochelle Walensky said CDC did not actually perform the planned PRR analyses until 2022. During the critical early period, CDC and FDA instead relied on FDA’s Bayesian method. Meanwhile, Szarfman and DuMouchel were warning that the method could mask signals in precisely the abnormal database conditions created by the COVID vaccination campaign. The BMJ investigation reports that their alternative analyses generated alerts the routine FDA method had failed to surface, including signals involving myocarditis, Bell’s palsy, and other outcomes. It further reports that when CDC eventually conducted PRR analyses, hundreds of product-event combinations met the agency’s statistical alert criteria even though Walensky characterized the analyses as finding no additional unexpected safety signals. The word signals has to remain attached to that statement. “Hundreds of signals” does not mean “hundreds of proven vaccine injuries.” It means hundreds of statistical patterns met an agency-defined threshold indicating that further examination could be warranted. That distinction does not weaken the criticism. It defines it correctly. The purpose of safety surveillance is to find those patterns. Myocarditis Shows Why Detection Matters Myocarditis provides an especially instructive example because its relationship with mRNA vaccination did not remain merely a VAERS hypothesis. CDC now states that evidence from multiple safety-monitoring systems in the United States and internationally supports a causal association between mRNA COVID-19 vaccination and myocarditis and pericarditis. The events occur rarely and have been concentrated particularly among adolescent and young adult males. FDA subsequently required labeling changes for Pfizer’s Comirnaty and Moderna’s Spikevax describing the myocarditis and pericarditis risks, and in June 2025 required updated warnings containing additional risk information. That does not retroactively establish the validity of every VAERS cardiovascular signal. It establishes something much narrower and scientifically important: At least one adverse event that the early surveillance infrastructure needed to detect promptly was real enough eventually to produce a recognized causal association and formal FDA labeling. A signal-detection architecture capable of masking such an event deserved immediate methodological scrutiny. FDA’s Own Current Language Undermines the Old Reassurance There is another striking aspect of this story. FDA’s current explanation of its data-mining program states: “the absence of disproportionality does not confirm the absence of a safety signal” That is precisely correct. Read that principle carefully. If the absence of a disproportionality signal does not establish absence of a safety problem, then officials cannot logically transform “the algorithm did not alert” into “the evidence shows there is no safety concern.” The inferential error becomes even more serious once officials know that the method has a specific false-negative mechanism operating under the actual conditions of use. At that point, a negative result is no longer merely negative. Its sensitivity is in question. This is elementary measurement theory. If an instrument has a known tendency to flatten high measurements, the absence of a high reading cannot reassure you that high values are absent. You first fix the instrument. The 2022 Paper Also Requires Transparency The 2022 Drug Safety article reported no outside research funding. Its disclosures nevertheless deserve attention. Six authors, including DuMouchel, worked for Oracle Health Sciences, which provided the signal-detection and management software used in the research. The article disclosed that employment relationship. Szarfman was affiliated with FDA, and Joseph Tonning was retired from FDA and the US Public Health Service. That is a disclosed relationship, not a hidden one. It does mean independent replication matters. And replication is beginning to accumulate. A 2026 study examining Dutch and Spanish national pharmacovigilance databases specifically evaluated COVID-19 vaccine masking and alternative unmasking methods. PMID 41501319, DOI 10.1007/s40264-025- 01644-4. The appropriate scientific response was therefore never to accept RGPS uncritically. It was to compare methods openly, prospectively, and reproducibly. That is what should have happened in 2021. The Failure Was Epistemic Before It Was Political The most consequential failure documented here is not that one algorithm produced different numbers than another. Algorithms do that. The failure arises from the sequence. A surveillance method had a known weakness. The weakness was particularly relevant to the structure of the incoming data. Experts warned officials about it. An alternative method existed. Alternative analyses produced signals that the routine method missed. And according to The BMJ, efforts to pursue that discrepancy met resistance while the absence of alerts continued to contribute to public reassurance. That violates a basic principle of scientific quality control. When two reasonable analytical methods disagree, the disagreement is the finding. You investigate it. You do not select the method producing the institutionally convenient result and characterize the alternative analysis as a distraction. FDA officials had been warned that the agency’s data-mining method could miss safety signals and that later analyses exposed alerts involving myocarditis, Bell’s palsy, and other outcomes. The absence of a signal is not proof of safety. And once a surveillance authority knows its method can systematically suppress signals under the exact conditions confronting it, continuing to cite silence from that method becomes scientifically indefensible unless the limitation is disclosed and independent methods are simultaneously applied. What Should Happen Now The appropriate response is not another round of argument over whether VAERS “proves” vaccine injury. It does not. The appropriate response is a complete reconstruction of what the surveillance systems actually showed, using methods selected before looking at the desired answer. FDA and CDC should release the complete weekly COVID-19 vaccine data-mining outputs from the beginning of the rollout, including MGPS/EBGM results, PRR analyses, regression-adjusted analyses, database restrictions, stratification rules, alert thresholds, software versions, methodological changes, and internal analytical runs. Independent investigators should then rerun the surveillance chronologically—week by week—as though they were standing in 2021 without knowledge of what would subsequently become recognized adverse reactions. That distinction matters. Retrospective analysis can unconsciously optimize methods around known outcomes. A proper audit should therefore ask a prospective question: At what date would each reasonable surveillance method have generated each signal using only the data available on that date? For every event, investigators should report the first crossing of the prespecified threshold, whether the signal persisted, the number and quality of underlying reports, sensitivity to reporting stimulation, sensitivity to database composition, age and sex effects, alternative comparator structures, and the result after correction for masking. Then those signals should be compared with active-surveillance systems, claims data, electronic health records, formal epidemiological studies, and clinical adjudication. That would tell the public something scientifically useful. Not whether VAERS “proved” everything. Not whether VAERS “proved” nothing. It would tell us whether the country’s early-warning system performed as represented when millions of people depended on it. “Science” Cannot Be Allowed to Grade Its Own Mistakes Away A safety system should be designed to fail loudly. When there is uncertainty, competing methods should broaden investigation rather than narrow it. When an analyst discovers a false- negative mechanism, the response should be replication. When an improved method produces signals absent from the established method, those discrepancies should trigger scrutiny. The standard cannot be that adverse-event signals count only when the preferred algorithm sees them. The standard must be whether the analytic system can withstand adversarial methodological testing. What The BMJ has documented is therefore larger than another argument over COVID-19 vaccines. It is a lesson in how institutional science can become epistemically fragile. A regulator may possess enormous amounts of data and still fail to see what is in those data if its measurement system defines the phenomenon away. And when the regulator has already been warned that this can happen, the phrase “we found no signal” carries an obligation: Perhaps it’s time for IPAK to design The Public’s Vaccine Safety Tracking System - with HIPAA-compliant access to medical records and other features that are lacking in VAERS. Do you think we should? Would you donate to support it? Leave a comment. Leave a comment References 1. Willman D. US officials knew covid vaccine safety system was flawed—but suppressed the FDA doctor who alerted them. BMJ. 2026;394:e100806. DOI: 10.1136/bmj-2026-100806. 2. Harpaz R, DuMouchel W, Van Manen R, et al. Signaling COVID-19 Vaccine Adverse Events. Drug Safety. 2022;45:765-780. PMID: 35737293. DOI: 10.1007/s40264-022-01186-z. 3. Maignen F, Hauben M, Hung E, Van Holle L, Dogne JM. Assessing the extent and impact of the masking effect of disproportionality analyses on two spontaneous reporting systems databases. Pharmacoepidemiology and Drug Safety. 2014;23:195-207. PMID: 24243665. 4. Rachwal O, Gutiérrez-Lobón M, Sols Cueto N, et al. Evaluating COVID-19 Vaccine Masking and Unmasking Methods in Two National Pharmacovigilance Databases. Drug Safety. 2026;49:581-590. PMID: 41501319. DOI: 10.1007/s40264-025-01644-4. Thank you for being a subscriber to Popular Rationalism. For the full experience, become a paying subscriber. And check out our awesome, in- depth, live full semester courses at IPAK-EDU. Hope to see you in class! Upgrade to paid LIKE COMMENT RESTACK © 2026 James Lyons-Weiler, PhD 20714 Shady Lane Ave, Shady Lane, St. Clair Shores, MI 48080 Unsubscribe