The Three Most Important Words Uttered By Secretary Kennedy in the Senate Finance Committee Hearing

CDC indeed has made chaos of data. Here's how, and why Secretary Kennedy is making history when he says "I don't know".

In last week’s Senate Finance Committee hearing (September 2025), Health and Human Services Secretary Robert F. Kennedy Jr. faced repeated questions about how many lives the COVID‑19 vaccines saved and about vaccine effectiveness. He made it clear he was not fooling around. He replied, with gravitas worthy of a polar bear: “I don’t know”. He attributed the uncertainty to “data chaos” at CDC (coverage from AP, Washington Post, PBS; see also Nature). The AP summarized his line this way: “The only confusion I expressed is exactly how many lives were saved. I don’t think anybody knows that.” AP News The Washington Post PBS Nature

Those three words—I don’t know—are not an evasion. In medical education and patient safety, acknowledging uncertainty is a feature, not a bug. WHO’s training materials literally teach that “the three most important words in medical education are ‘I don’t know.’” The difference between rigor and rhetoric is what follows after “I don’t know”: an action plan to resolve uncertainty. Iris

What the literature actually contains on “lives saved” and effectiveness

In the pre-Kennedy HHS, bias in results was second only to the hubris and exuberance with which biased results were reported.

Deaths averted. The most‑cited global estimate is merely a Lancet Infectious Diseases modeling study (Watson et al., 2022) which used reported and excess‑mortality data to calibrate counterfactuals. It estimated ~19.8 million deaths averted globally in the first vaccine year (Dec 8 2020–Dec 8 2021). That estimate is completely model‑dependent, with wide assumptions about reporting, variant waves, and non‑pharmaceutical interventions; the result is still the field’s anchor. (DOI: 10.1016/S1473‑3099(22)00320‑6, PMID: 35753318). PubMed.

Watson et al.’s 2022 estimate that COVID-19 vaccines prevented 14–20 million deaths has drawn sustained criticism on methodological and logical grounds. Klement and Walach argued that the team’s Bayesian SEIR framework assumed fixed reproduction numbers in a no-vaccine counterfactual, neglecting the adaptive ways human behavior, viral evolution, and policy would have shifted absent vaccines. In his view, this structural rigidity stripped the model of causal realism and rendered its results more an exercise in mathematics than a reflection of epidemiological dynamics. He further suggested that SEIR-style models cannot, by design, capture the full range of generative mechanisms that shape epidemic trajectories, meaning their output risks over-stating certainty while under-representing complexity.

Other critics have targeted the gap between Watson’s projections and empirical data. Amrit Šorli contended that the claim of millions of lives “saved” is logically inconsistent with the observed global mortality increase from 2020 to 2021, since such an effect would have implied fewer recorded deaths, not more. He characterized the “lives saved” construct as a theoretical fiction with no bijective mapping to real mortality outcomes. Meanwhile, Raphael Lataster situated Watson et al. within a broader metacritique of high-profile vaccine modeling, highlighting unexamined assumptions, endpoint manipulation, and potential conflicts of interest that, he argued, exaggerated benefits and muted uncertainties. Together these criticisms converge on the charge that the Watson model elevated counterfactual speculation to fact, while offering insufficient transparency about its limitations.

Regional analyses (e.g., Meslé et al., Lancet Respir Med, 2024) have produced national tallies (e.g., ~1.56 million lives saved across 34 European/North American countries), again model‑based and assumption‑sensitive. (DOI: 10.1016/S2213‑2600(24)00179‑6). The Lancet

U.S. toll and data uncertainty. CDC’s trackers and NCHS provisional death statistics remain the source of record but are subject to lags, backfills, and definitional updates—part of the “data chaos” Kennedy referenced. Multiple official dashboards exist (COVID Data Tracker, NCHS mortality, wastewater, COVID‑NET), which complicates single‑number “lives saved” claims without model choices with no assurance of accuracy or lack of bias. CDC COVID Data Tracker CDC

The bottom line is the peer‑reviewed literature does contain numeric estimates for “lives saved,” but every number is model‑contingent. Saying “I don’t know” is defensible only if accompanied by a precise plan to (1) define measurement, (2) state model structure, and (3) disclose uncertainty bands.

Methods audit — the bias‑inducing steps that distort Vaccine Efficacy and “deaths averted”

Below are the operational choices (design, coding, or reporting steps) that systematically bias VE estimates. For each, we provide the primary methodological sources and, where available, COVID‑specific demonstrations.

1. Case‑counting window bias (misaligned risk windows).
   Counting cases for unvaccinated person‑time from Day 0 while counting vaccinated cases only from ≥7–14 days after dose creates artificial VE by ignoring infections in the biologically “pre‑protection” period of vaccinees. It also biases counts of hospitalizations and deaths. This “window asymmetry” was formalized for COVID VE by Fung, Jones, Doshi and others; even small window shifts can produce large apparent VE from near‑zero efficacy. (J Eval Clin Pract, 10.1111/jep.13839; commentaries 10.1111/jep.13892, 10.1111/jep.13900; review with explicit “case counting window bias”: PMCID: PMC11557388). PubMed PubMed PMC

2. Immortal‑time bias / left truncation.
   If time between cohort study entry and vaccination is misclassified as “vaccinated,” outcomes that occurred before the shot are impossible in that interval even if the shot occurs the day before the end of the study—creating immortal person‑time for vaccinees and inflating VE. This bias is canonical in pharmacoepidemiology, especially in retrospective studies (Suissa, 2020 review; PMID 32783283), and repeatedly documented in VE analyses. PubMed

3. Healthy‑vaccinee effect (and related selection).
   People seeking vaccination are often healthier, wealthier, or more health‑seeking, leading to lower baseline risk regardless of vaccination. Classic influenza work showed implausibly large mortality benefits and identified pre‑season risk differentials as proof of bias (Jackson et al.; IJE 2006; Remschmidt et al., BMC Infect Dis 2015). COVID‑19 studies and commentaries replicate the concern. (PMIDs 16368725, 26451634, 38415513; see also Chemaitelly et al., eLife 2025). PubMed PMCeLife

4. Depletion‑of‑susceptibles (time‑varying risk set distortion).
   As high‑risk individuals get infected earlier, later comparisons involve selectively “hardier” survivors, biasing waning estimates and sometimes creating spurious negative VE. (Lipsitch, Clin Infect Dis 2019, PMCID: PMC7003633; Ray et al. 2020). PMC PubMed

5. Differential testing and care‑seeking; collider bias in Test‑Negative Design (TND).
   If vaccinated and unvaccinated differ in testing propensity or clinical thresholds, conditioning on “tested” induces bias. Use negative controls to probe residual confounding; recent simulation work quantifies distortion, especially under unequal testing. (Lipsitch Epidemiology 2010, PMID 20335814; Hitchings Epidemiology 2022 PMCID: PMC9148635; Bodner Nat Commun 2025). PubMed PMC Nature

6. Outcome misclassification (“with” vs “for” hospitalization; billing codes).
   Hospital registers frequently lack reason‑for‑admission fields; including incidental positives dilutes or inflates VE depending on directionality. A 2024 methods review documents 42% missing admission reasons in a national register and shows bidirectional distortion. (PMCID: PMC11557388). PMC

7. Prior infection and vaccination history confounding (especially in TND).
   Unobserved infection history and hybrid immunity status bias VE—sometimes to negative values in simulations—even with TND. (Wiegand et al., Nat Commun 2024 PMCID: PMC11579392). PMC

8. Endpoint selection and bundling; per‑protocol analysis starting after full dosing.
   Pivotal RCTs set symptomatic, PCR‑”confirmed” COVID‑19 (not all‑cause mortality) as the primary endpoint and conducted per‑protocol analyses ≥7 days after dose 2 (Pfizer) or ≥14 days after dose 2 (Moderna). That is legitimate but excludes early post‑dose risk windows and prioritizes disease‑specific endpoints over all‑cause severe endpoints. (Pfizer NEJM 2020, DOI 10.1056/NEJMoa2034577; Moderna NEJM 2021 final blinded phase PMCID: PMC8482810). PubMed PMC

9. Early unblinding / open‑label crossover.
   Once EUAs issued, placebo participants were offered vaccine, truncating long‑term blinded safety/efficacy. The Pfizer paper explicitly notes long‑term assessment cannot happen without a maintained placebo group; Moderna documents Part B open‑label follow‑up. (Pfizer PMCID: PMC7745181; Moderna PMCID: PMC8482810, PMCID: PMC11362294). PMC PMC

10. Target‑trial misalignment (time‑zero, eligibility, treatment assignment).
    If observational VE or effectiveness studies don’t emulate a specific “target trial,” they accumulate immortal time biases, left truncation, and time‑varying confounding. The fix is explicit target‑trial emulation. (Hernán & Robins 2016 PMCID: PMC4832051; Hernán 2016 PMCID: PMC5124536; Català et al. 2024 PMID: 37833846). PMC PubMed

11. Unequal follow‑up and attrition (socioeconomic gradients).
    Differential loss to follow‑up (by SES, occupation, access) re‑weights risk sets and biases VE; documented broadly and in COVID‑19 contexts. (PMCID: PMC11557388). PMC

12. Analytic flexibility and sponsor control.
    Not a smoking gun by itself, but pivotal trials were industry‑funded and set endpoints accordingly; interpretation must separate what was measured from broader policy claims. (Funding and endpoint disclosures in NEJM RCTs noted above). PubMed PMC

In-Depth and In Plain English: How Former CDC’s Tilted Vaccine Studies Without Telling a Lie

A Popular‑Rationalism Guide to Bias in Vaccine Research

If you control the stopwatch, you control the score. Vaccine studies often use technical moves that change when events “count,” who gets counted, and what the events mean. None of these moves needs falsification; they just tilt the table. Below, we unpack the most common bias‑inducing steps, in plain language, and spell out what each step predictably does to results.

1) The case‑counting window trick

Problem. Trials and many observational studies start counting “cases” for vaccinated people only after an arbitrary grace period (e.g., 7–14 days after dose 2), while counting cases for unvaccinated people from time zero.
Plain English. Imagine a 5K race wherein runners have to pick up marbles along the way. Team V’s clock starts at the 1‑mile mark, but Team U’s clock starts at the starting line. Team V will “run faster” on paper and pick up fewer marbles even if both teams run and pick up at the same pace. Early infections, hospitalizations, and deaths after vaccination are either ignored or reassigned to “unvaccinated,” creating a head start.
Likely Effect. Overstates vaccine effectiveness (VE), sometimes by a large margin, and can even flip a null or negative effect into a positive one (this can also be called the Lyons-Weiler/Fenton effect).

2) Immortal‑time bias (and its cousins: left truncation, immeasurable time)

Problem. People must survive long enough to become “vaccinated.” Any time before they get their shot but after study entry is “immortal” if counted toward the vaccinated group—no events can happen there by definition.
Plain English. You can’t die as a “vaccinated” person before you’ve been “vaccinated”. If analysts misallocate that survival time to the vaccinated group, the vaccinated group gets free, event‑free time.
Likely Effect. Inflates apparent protection; vaccinated groups look safer partly because they were required to be alive and at risk long enough to qualify.

3) The healthy‑vaccinee effect

Problem. People who choose or are cleared to be vaccinated differ in health status, care access, and behavior from those who don’t—often in ways that standard adjustment doesn’t capture (functional status, frailty, care‑seeking habits).
Plain English. Health‑conscious, mobile elders who see doctors and take medications get vaccinated more. Those same traits reduce hospitalization and death regardless of vaccination.
Likely Effect. Overstates benefit against severe outcomes (especially all‑cause outcomes) unless you measure and adjust for hard‑to‑see health differences.

4) Depletion‑of‑susceptibles

Problem. In any epidemic, the most exposure‑prone or biologically susceptible people get infected earlier. If a vaccine is imperfect, or impairs immunity, part of the vaccinated pool gets “picked over” faster.
Plain English. Think of two orchards. In Orchard V (vaccinated), the ripest fruit gets picked quickly; in Orchard U (unvaccinated), picking is slower (no immune suppression). As the season goes on, Orchard V has fewer “ripe” targets left, so Orchard V’s risk apparent risk falls—making protection look like it’s fading even if it isn’t.
Likely Effect. Can fabricate “efficacy” especially in the two-dose or booster setting, and distort its size and direction. Without design safeguards, within‑season VE will drift negative and then positive if true protection is stable after vaccine-induced immunosuppression culls the vaccinated.

5) Differential testing and care‑seeking (test‑negative collider problems)

Problem. In test‑negative designs (TND), people must seek testing to enter the study. If vaccinated and unvaccinated groups differ in how often, when, and why they test, the design can condition on those differences.
Plain English. If vaccinated people test more (e.g., for travel, work, or caution) and unvaccinated people test mainly when very sick, you’ve selected unequal groups at the clinic door.
Likely Effect. Direction depends on who tests more and when, but mismatched testing intensity commonly biases VE estimates—often downward for mild outcomes and unpredictably for severe ones.

6) Outcome misclassification (died “with” vs died “from”)

Problem. Many studies label hospitalizations or deaths as “COVID‑19” if a swab was positive, regardless of whether respiratory disease drove the admission.
Plain English. A person admitted for a broken hip but testing positive on screening is not the same as a pneumonia admission on oxygen. Lumping them together muddies the outcome.
Likely Effect. If incidental positives cluster differently by vaccination status (because of screening patterns, background infection, or admission reasons), VE against hospitalization and deaths gets biased—usually toward the null when incidental cases are common.

7) Prior infection and hybrid‑immunity confounding

Problem. Past infection lowers risk of future severe disease. If prior infection is uneven or poorly measured, comparisons of vaccinated and unvaccinated groups are confounded.
Plain English. Give one group quieter firewood and claim their house is more fireproof. If the “unvaccinated” group already had lots of recent infections, they bring natural protection the analysis didn’t credit.
Likely Effect. Underestimates VE against severe disease in some settings; overestimates protection against infection if prior infection is more common in the vaccinated.

8) Endpoint selection and bundling

Problem. Trials often choose endpoints like “symptomatic PCR‑confirmed COVID‑19” that start counting after a per‑protocol window; severe outcomes are secondary, rarer, and less powered.
Plain English. If you choose the easiest target (mild disease in a carefully timed window), you’ll hit it. That tells you less about hospitalization and death.
Likely Effect. Large VE against mild endpoints; uncertain or less precise VE against severe endpoints; optimistic headlines.

9) Early unblinding / open‑label crossover

Problem. After early positive results or authorization, placebo recipients are offered vaccine, and blinding dissolves.
Plain English. You changed the rules mid‑game and let teams switch jerseys. Long‑term, blinded comparison disappears.
Likely Effect. Limits robust estimates of durability, late adverse events, and true long‑term VE; invites informative censoring and behavior changes.

10) Target‑trial misalignment (“time zero” eligibility, follow‑up)

Problem. Observational emulations must define who’s eligible, when follow‑up starts, and what treatment strategies are compared. Misalign any of these and bias rushes in.
Plain English. If you compare “people who ever got vaccinated” to “people who haven’t yet,” you gave the vaccinated a head start and counted different calendars. (We saw this recently in a study of “ever used Acetaminophen” instead of “number of doses” or “doses administered prior to or after vaccination”).
Likely Effect. Immortal‑time bias and confounding masquerading as causation; VE estimates that depend more on design choices than biology.

11) Unequal follow‑up / informative censoring

Problem. Loss to follow‑up (moving, death from other causes, switching health systems) differs by vaccination status, age, deprivation, or comorbidity. If censoring relates to outcome risk, standard analyses skew.
Plain English. If sicker, unvaccinated people exit your dataset earlier, you’ll keep measuring the healthier remainder.
Likely Effect. Typically overstates protection unless you correct with methods like inverse‑probability‑of‑censoring weighting.

12) Flexible analytics under sponsor or operational control

Problem. Pre‑specification can be loose; analyses can shift windows, covariate sets, or definitions. Funding entities may set or influence endpoints and analysis timing.
Plain English. If you can move the goalposts after seeing the field, you won’t miss many kicks.
Likely Effect. Endpoint and window choices that maximize early efficacy against mild outcomes, with less clarity on hard endpoints and durability.

Why this matters

Every bias-inducing step pushes estimates in a predictable direction, making data interpretation impossible. Thus, Kennedy’s “I Don’t Know”.

Stack enough of these biases—post‑dose “grace” windows, immortal time, healthy‑vaccinee selection, outcome bundling—and you can manufacture false certainty. None of this says vaccines don’t work. It says you must read the design before you read the results.

How to read a vaccine study like a hawk

1. Ask when clocks start. Do cases “count” for vaccinated people only after a delay?

2. Check time zero. Does case, hospitalization and death counting and follow‑up begin at the same calendar time and biologic state for both groups?

3. Inspect testing behavior. Are inclusion and testing rules symmetric across exposure groups?

4. Demand clear outcomes. “On oxygen” or “ICU” beats “any admission with a positive swab.”

5. Look for prior infection control. Serology, documented positives, or proxy adjustment?

6. Check blinding and crossover. Was long‑term comparison preserved?

7. Read the statistical appendix. Are censoring, attrition, and sensitivity analyses (negative controls, recently vaccinated bias checks) done and reported?

Takeaway: These steps—misaligned windows, immortal time, selection, depletion, misclassification, endpoint choice, unblinding, and target‑trial drift—can convert modest or context‑dependent effects into confident headlines. That is exactly why “I don’t know (yet)” must be followed by design and code that neutralize the biases.

IV. Endpoint ecology example: HPV “type replacement” is contested—and methodological

Because you asked us to keep every bias lens we use across vaccines, a brief, tightly sourced note:

• Signals for non‑vaccine type increases after HPV programs have been reported across the literature in reviews and regional studies, including observed rises in certain high‑risk non‑vaccine types and in CIN2/3 attributable to non‑vaccine types in some cohorts. (Hampson 2024, PMCID: PMC11359832, summarizing regional studies in Spain, Italy, New York, Stockholm; also reports from Scotland on CIN type shifts). PMC

• Signals against: England’s high‑coverage program reports no indication of type replacement to date in surveillance of vaccine‑eligible cohorts (Checchi 2023, PMID: 37821315). Cervical cancer incidence drops by deprivation stratum continue (Falcaro 2024, BMJ 2024). PubMed BMJ

• Methodological hinge: Shifts can reflect true ecological replacement, cross‑protection, or clinical unmasking (fewer excisions for 16/18 leaves tissue at risk for other types). The same causal‑inference cautions—case definitions, risk windows, and surveillance drift—apply here. PMC

This “type replacement” debate illustrates why endpoint ecology and ascertainment rules must be pre‑registered and audited—precisely the kind of discipline CDC needs to impose on respiratory‑virus VE reporting.

V. What “I don’t know” should trigger—concrete next steps to end the chaos

1. Publish the schemas. CDC should release (a) the exact case definitions used in VE denominators and numerators across COVID‑NET, VISION/IVY, and death tabulations; (b) the windowing rules for case accrual in vaccinated vs unvaccinated person‑time; (c) how prior infection and hybrid immunity are measured or proxied. These are the knobs that move VE the most. (See CDC surveillance pages for current fragmented dashboards). CDC COVID Data Tracker CDC

2. Enforce target‑trial emulation in all VE updates: align time zero, eligibility, and treatment assignment; make immortal‑time elimination and left‑truncation handling visible in code supplements. (Hernán & Robins framework). PMC

3. Mandate negative controls (recently vaccinated as exposure negative control; non‑COVID outcomes as outcome negative controls) and publish the results alongside VE. PubMed PMC

4. Dual‑reporting: present (i) design‑based VE with all censoring rules and (ii) model‑based “deaths averted,” with structural diagrams and sensitivity bands. Put the Watson‑style counterfactuals next to raw counts to prevent conflation. PubMed

5. Prospectively lock windowing rules (e.g., count infections from Day 0 and Day 14 in parallel) to quantify window bias directly, as recommended in the methodological literature. PubMed

If HHS/CDC produces this level of pre‑specification and open code, “I don’t know” is seen as an honest interim status, and a call to science, not an evasion strategy.

Selected examples & sources (one to three strong exemplars per bias)

Case‑counting window bias; endpoint timing
Doshi P, Fung K. How the case counting window affected vaccine efficacy calculations in randomized trials of COVID‑19 vaccines. J Eval Clin Pract. 2024;30(1):105‑106. DOI:10.1111/jep.13900; PMID:37452751. Lataster R. Reply to Fung et al. on COVID‑19 vaccine case‑counting window biases overstating vaccine effectiveness. J Eval Clin Pract. 2024;30(1):82‑85. DOI:10.1111/jep.13892; PMID:37403424. Polack FP et al. Safety and Efficacy of the BNT162b2 mRNA Covid‑19 Vaccine. N Engl J Med. 2020;383:2603‑2615. DOI:10.1056/NEJMoa2034577; PMID:33301246. El Sahly HM et al. Efficacy of the mRNA‑1273 SARS‑CoV‑2 Vaccine at Completion of Blinded Phase. N Engl J Med. 2021; doi:10.1056/NEJMoa2113017; PMCID:PMC8482810. PubMed PubMed PMC

Immortal‑time / time‑related biases
Suissa S. Immortal time bias in observational studies of drug effects. Pharmacoepidemiol Drug Saf. 2007;16(3):241‑249. DOI:10.1002/pds.1357; PMID:17252614. Suissa S. Immortal time bias in pharmacoepidemiology. Am J Epidemiol. 2008;167(4):492‑499. DOI:10.1093/aje/kwm324; PMID:18056625. Suissa S. Time‑related biases in pharmacoepidemiology. Pharmacoepidemiol Drug Saf. 2020;29(10):1101‑1110. DOI:10.1002/pds.5083. PubMed Oxford AcademicWiley Online Library

Healthy‑vaccinee effect
Jackson LA et al. Evidence of bias in estimates of influenza vaccine effectiveness in seniors. Int J Epidemiol. 2006;35:337‑344. PMID:16368725. Jackson LA et al. Functional status is a confounder… Int J Epidemiol. 2006;35:345‑352. PMID:16368724. Remschmidt C et al. Frequency and impact of confounding by indication and healthy vaccinee bias… BMC Infect Dis. 2015;15:429. DOI:10.1186/s12879‑015‑1154‑y. Furst T et al. Healthy vaccinee effect: a bias not to be forgotten… Pol Arch Intern Med. 2024;134(2):16634. DOI:10.20452/pamw.16634; PMID:38415516. PubMed+2PubMed+2BioMed Central

Depletion‑of‑susceptibles (spurious waning)
Lipsitch M, Goldstein E, Ray GT, Fireman B. Depletion‑of‑susceptibles bias in influenza vaccine waning studies. Epidemiol Infect. 2019. PMID:31774051; PMCID:PMC7003633. Ray GT et al. Depletion‑of‑susceptibles Bias in Analyses of Intra‑season Waning of Influenza VE. Clin Infect Dis. 2020;70(7):1484‑1491. PMID:31351439; PMCID:PMC7318775. PMC

Differential testing / TND collider issues; negative‑control checks
Hitchings MDT et al. Use of Recently Vaccinated Individuals to Detect Bias in Test‑Negative Case‑Control Studies of COVID‑19 VE. Epidemiology. 2022;33(4):577‑584. PMID:35384900; PMCID:PMC9148635. Jackson ML, Nelson JC. The test‑negative design for estimating influenza VE. Vaccine. 2013;31(17):2165‑2168. DOI:10.1016/j.vaccine.2013.02.053. Lipsitch M, Tchetgen Tchetgen E, Cohen T. Negative controls: a tool for detecting confounding and bias in observational studies. Epidemiology. 2010;21(3):383‑388. PMID:20335814. PubMedPMClibrarysearch.mtroyal.caLippincott Journals

Outcome misclassification (“with” vs “for” hospitalization)
Klann JG et al. Distinguishing admissions specifically for COVID‑19 from incidental SARS‑CoV‑2 admissions. J Med Internet Res. 2022;24(5):e37931. PMID:35476727; PMCID:PMC9119395. (Also see cohort validations comparing “oxygen‑on‑admission” definitions.) PubMedPMC

Prior infection / hybrid immunity confounding
Wiegand RE et al. Bias and negative values of COVID‑19 vaccine effectiveness from the test‑negative design without prior infection information. Nat Commun. 2024;15:8561. DOI:10.1038/s41467‑024‑54404‑w. Nature

Target‑trial specification and time‑zero alignment
Hernán MA, Robins JM. Using big data to emulate a target trial when a randomized trial is not available. Ann Intern Med. 2016;165(9):S1‑S12. DOI:10.7326/M16‑0829. BioMed Central

Unequal follow‑up / informative censoring
Lin HM et al. Informative Censoring—A Cause of Bias in Estimating COVID‑19 Outcomes. Biomedicines. 2023;11(1):250. PMID:36676159; PMCID:PMC9865049. Agampodi S et al. Biases in COVID‑19 vaccine effectiveness studies using observational designs: a narrative review. Front Med. 2024;11:1474045. PMCID:PMC11557388. PubMed PMC

Registry linkage and exposure misclassification
Brookmeyer R et al. Estimating Vaccine Effectiveness by Linking Population Registries. J Infect Dis. 2022;226(7):1165‑1168. PMID:35938874. Ioannidis JPA. Factors influencing estimated effectiveness of COVID‑19 vaccines in observational studies. BMJ Evid Based Med. 2022;27(6):324‑327. DOI:10.1136/bmjebm‑2021‑111735. PubMed BMJ Evidence-Based Medicine

Endpoint choice and trial focus on mild disease
Polack FP et al. BNT162b2 trial (primary endpoint: symptomatic, PCR‑positive Covid‑19 counted ≥7 days after dose 2). N Engl J Med. 2020;383:2603‑2615. DOI:10.1056/NEJMoa2034577; PMID:33301246. Doshi P. Will covid‑19 vaccines save lives? Current trials aren’t designed to tell us. BMJ. 2020;371:m4037. DOI:10.1136/bmj.m4037; PMID:33087398. PubMed

Early unblinding / open‑label crossover
Thomas SJ et al. Safety and Efficacy of the BNT162b2 mRNA Covid‑19 Vaccine through 6 Months. N Engl J Med. 2021;385:1761‑1773. DOI:10.1056/NEJMoa2110345 (includes blinded + open‑label periods). El Sahly H et al. mRNA‑1273 trial (open‑label phase with crossover). N Engl J Med. 2021; doi:10.1056/NEJMoa2113017. PubMed New England Journal of Medicine