This article is a companion to When Math Outruns Methodology: A Critical Look at the Czech Vaccine Analysis. The original article examined Steve Kirsch’s comparison of Pfizer and Moderna recipients and asked whether the analysis met the standards required for causal inference.
It did not attempt to prove that Kirsch was wrong. It asked whether his evidence was sufficient to prove that he was right.
Kirsch has since pointed to another analysis, KCOR, and argued that it addresses the methodological concerns raised in the original article.
KCOR stands for Kirsch Cumulative Outcomes Ratio. Kirsch describes it as a new method that corrects for healthy vaccinee bias, latent frailty, cohort depletion, and non-proportional hazards. He claims that it can determine whether an intervention saved or cost lives using only dates of birth, intervention, and death.
He goes well beyond calling it a useful statistical tool.
Kirsch describes KCOR as a “lie detector for data,” says that it cannot be gamed, and presents it as an objective way to determine whether an intervention caused net benefit or harm. He concludes that the Czech data show the COVID vaccines probably killed more people than they saved.
That is an extraordinary claim.
It is also a claim that his own methods paper does not support.
The Problem KCOR Is Trying to Solve Is Real
KCOR begins with a legitimate problem.
Comparing vaccinated and unvaccinated people is difficult because those groups are not randomly formed. People who accept, refuse, delay, or become eligible for vaccination may differ in ways related to their risk of death.
Age is only the most obvious difference. The groups may also differ in frailty, institutional residence, healthcare use, chronic illness, previous infection, socioeconomic status, physician recommendations, willingness to seek care, and whether they were healthy enough to receive another dose.
One particularly important problem is the healthy vaccinee effect.
Someone who is acutely ill, hospitalized, rapidly declining, or approaching death is less likely to be vaccinated at that moment. This can make recently vaccinated people appear unusually healthy, even if the vaccine itself does not affect mortality.
There is also a related problem known as depletion of susceptibles.
Imagine two groups containing different proportions of relatively healthy and highly vulnerable people. The most vulnerable tend to die first. As they leave the population at risk, the surviving group becomes progressively healthier. Its mortality curve can change over time even if the underlying treatment effect remains constant.
KCOR attempts to model that change.
This concern is not trivial. Frailty depletion is a recognized problem in survival analysis, and adjustment for measured variables may not fully account for it.
The disagreement is not about whether the problem exists.
The disagreement is about whether KCOR solves enough of it to support Kirsch’s causal conclusion.
What KCOR Actually Does
KCOR begins by assigning people to fixed cohorts on a chosen enrollment date. A person may be classified as having received zero, two, or three doses by that date.
Once assigned, the person remains in that original cohort throughout the analysis. Later vaccinations do not change the classification.
KCOR then calculates mortality hazards for the cohorts over time. It assumes that the mortality curves can be represented by a Gompertz baseline mortality process combined with a gamma distribution of unobserved frailty.
In plain English, the model assumes that each cohort contains people with different underlying vulnerabilities and that those vulnerabilities follow a particular mathematical distribution.
KCOR identifies periods called “quiet windows.” These are intended to be periods relatively free from major epidemic waves or other mortality shocks. It uses the shape of the mortality curve during those windows to estimate how quickly frailer members of each cohort are being depleted.
The method then mathematically reverses, or neutralizes, that estimated depletion process. It compares the resulting adjusted cumulative hazards and produces the KCOR ratio.
A ratio above one means that one cohort accumulated more mortality than the comparison cohort after the model-based adjustment. A ratio below one means that it accumulated less.
That is the calculation.
The more difficult question is what the remaining difference means
KCOR can show: a mortality divergence remains after normalization.
KCOR cannot determine by itself whether that divergence was caused by vaccination, residual confounding, crossover, external hazards, or model misspecification.
Kirsch’s Newsletter Says “Causal.” His Paper Says “Non-Causal.”
Kirsch’s public description says KCOR can determine whether an intervention caused net benefit or harm.
His technical manuscript says something much narrower.
The paper places its estimand under the heading “Target estimand and scope (non-causal).” It describes KCOR as a diagnostic signal showing what remains after a working-model adjustment. It then states explicitly that KCOR is “not a causal effect estimator” and does not recover outcomes under hypothetical alternative interventions.
Later, the paper calls KCOR a “pre-causal diagnostic layer.”
It says the method does not adjust for general confounding, does not replace covariate adjustment, and cannot transform a remaining contrast into a causal effect. It acknowledges that a persistent difference after KCOR normalization could represent a genuine intervention effect, residual confounding, model misspecification, external hazards, or some other mechanism.
This is not a minor disagreement over terminology.
It is the central contradiction:
The newsletter presents KCOR as a causal verdict. The methods paper defines it as a non-causal diagnostic.
KCOR may indicate that a difference remains after a particular adjustment.
It cannot, by itself, tell us why that difference remains.
That final step is exactly what Kirsch’s public conclusion assumes.
A Plain-English Translation
Suppose KCOR adjusts two mortality curves and they still diverge.
That tells us:
The divergence was not fully explained by the particular depletion process KCOR modeled.
It does not automatically tell us:
The vaccine caused the divergence.
Those are different statements.
The first is a statistical observation.
The second is a causal conclusion.
KCOR’s own paper permits the first. Kirsch’s public interpretation jumps to the second.
That gap is where the entire argument lives.
KCOR Is Not Model-Free
Kirsch says KCOR “does no modeling,” has no meaningful tuning parameters, and allows the data to determine the answer.
That description is difficult to reconcile with the method itself.
KCOR depends on:
a Gompertz baseline mortality model,
a gamma distribution of latent frailty,
multiplicative frailty effects,
nonlinear parameter estimation,
selected enrollment dates,
skipped post-enrollment periods,
selected quiet windows,
baseline anchoring,
assumptions about shared external hazards,
and, in some applications, an additional mortality-wave adjustment.
Kirsch’s methods paper is clear about this. It states that normalization depends on the gamma-frailty model, the Gompertz baseline, and the method used to align quiet windows. It describes KCOR as a working-model normalization rather than a model-free correction.
Estimating parameters from data does not make a method model-free.
Every fitted statistical model uses data to estimate its parameters. The real question is whether the assumptions connecting those parameters to the real world are justified.
Kirsch also says KCOR cannot be gamed because it is deterministic.
But determinism means only that the same data and analytical choices will produce the same result.
It does not establish that those choices are correct.
An improperly calibrated bathroom scale can give the same wrong weight every morning. Reproducibility is valuable. It is not the same as validity.
Three Dates Cannot Reconstruct Every Difference Between Cohorts
Kirsch says KCOR needs only a date of birth, date of intervention, and date of outcome. He argues that age, sex, comorbidities, socioeconomic status, and other factors are handled implicitly through the shape of the mortality curves.
That claim goes beyond what the methods paper establishes.
KCOR adjusts for baseline differences only to the extent that those differences produce the type of depletion curvature represented by its working model.
That is not the same as adjusting for every cause of non-exchangeability.
Two cohorts could have similar aggregate mortality curvature while differing in:
previous infection,
institutional residence,
regional epidemic intensity,
access to treatment,
occupational exposure,
healthcare use,
subsequent vaccination,
cause-of-death composition,
and changes in health behavior.
Those differences cannot be reconstructed merely from dates of birth, vaccination, and death.
The KCOR paper acknowledges that mortality curvature may arise from frailty selection, behavior, seasonality, treatment effects, reporting artifacts, or some combination of these factors. It states that KCOR does not uniquely identify the biological, clinical, or behavioral mechanism responsible for the observed pattern.
That admission matters.
The method cannot claim to control every important difference while also acknowledging that several different processes can produce the same curve and cannot be distinguished by KCOR.
The Optional COVID-Wave Correction Was Not Reliably Identified
Kirsch says KCOR corrects three major sources of bias:
dynamic healthy vaccinee effects,
static healthy vaccinee effects,
COVID-related non-proportional hazards.
He presents these corrections as the reason KCOR succeeds where conventional methods fail.
The third correction, however, is not part of the essential KCOR framework. It is an optional module involving an exponent identified as alpha.
More importantly, in the methods paper reviewed here, alpha was not reliably identified in the pooled Czech analysis. The apparent optimum was too unstable and model-dependent to meet the paper’s prespecified identification criteria.
The paper therefore did not treat the Czech estimate as a reliable correction parameter.
That creates another gap between the promotional claim and the technical result.
The public article says KCOR corrects the COVID-wave problem.
The methods paper says that the relevant parameter could not be reliably identified in the Czech data.
The Booster Data and the “Smoking Gun”
Kirsch’s strongest response to the original article concerns his booster analysis.
He argues that the Dose 3 mortality pattern cannot be explained by ordinary selection bias. In his interpretation, the curves change in temporal alignment with booster vaccination, rise for a limited period, and then plateau.
That, he says, makes vaccine causation “crystal clear.”
An AlterAI chatbot later described the booster result as a “smoking gun” and argued:
“Selection bias doesn’t have a calendar. Frailty doesn’t know when you got your booster.”
It is a memorable line.
It does not resolve the underlying methodological problem.
The curves are not aligned to each person’s booster date
KCOR organizes cohorts according to vaccination status on a fixed enrollment date. Its time axis represents weeks since cohort enrollment, not necessarily weeks since each individual received a booster.
Kirsch’s methods paper states that the Dose 2 and Dose 3 curves were constructed for cohorts enrolled during ISO week 2022-26 and followed during 2023. It identifies the x-axis as weeks since enrollment.
It also describes the figures as illustrative and says they do not support causal inference.
People classified as Dose 3 may have received their boosters at different times before the common enrollment date.
A curve that rises after enrollment therefore does not necessarily show mortality rising at a consistent biological interval after each person’s booster.
To establish that timing relationship, the analysis would need to align individuals by their actual booster dates, define a comparable time zero for the control group, and account for the factors determining who was eligible and healthy enough to receive the dose.
The fixed-cohort graph alone does not accomplish that.
Selection can operate through calendar time
Booster status on a particular date is not random.
It may be influenced by:
age,
medical risk,
timing of previous doses,
survival until booster eligibility,
recent infection or acute illness,
healthcare access,
institutional residence,
physician recommendations,
and willingness to receive another dose.
Czech booster eligibility also changed over calendar time. In December 2021, high-risk patients and people over 60 could receive a booster five months after their second dose, while others generally waited six months. Eligibility intervals and age thresholds were subsequently changed.
Those policies helped determine who had reached Dose 3 by a later enrollment date.
Frailty does not need to read a calendar.
Eligibility and selection processes do.
Kirsch’s own paper describes Czech vaccine uptake as voluntary, rapidly changing over time, and associated with baseline health status.
That is selection operating through calendar time.
Dose 2 and Dose 3 are not identical groups separated by one random injection
Everyone in the Dose 3 cohort had to:
complete the primary series,
survive until booster eligibility,
remain eligible,
choose or be advised to receive another dose,
and receive it before the enrollment cutoff.
The Dose 2 group consisted of people who had not received that additional dose by the cutoff.
Some may not have been eligible long enough. Others may have experienced a recent illness or infection. Some may have had contraindications, difficulty accessing vaccination, different healthcare behavior, or less willingness to continue vaccination.
None of these possibilities proves that confounding caused the observed curve.
That is not the point.
The point is that Dose 2 and Dose 3 were not randomized versions of the same group. The booster was not the only difference between them.
KCOR must demonstrate that its normalization removed the outcome-related differences that mattered. The existence of a post-adjustment curve does not prove that it did.
A plateau does not identify the cause of the earlier divergence
Kirsch interprets an increase followed by a plateau as evidence of a temporary vaccine injury.
That is one possible interpretation.
But a cumulative ratio can plateau whenever the difference in the underlying rates diminishes.
The earlier divergence could reflect:
a genuine vaccine effect,
a temporary external hazard,
an infection wave affecting the cohorts differently,
differing previous-infection histories,
time-varying selection,
baseline anchoring,
quiet-window misspecification,
or a mechanism not represented by gamma frailty.
KCOR cannot distinguish among those explanations using three dates alone.
A curve that is compatible with one causal story is not proof that no other story could produce it.
Fixed Cohorts Create a Crossover Problem
KCOR freezes vaccination status at enrollment.
Someone classified as Dose 2 remains in the Dose 2 cohort after receiving a booster. Someone classified as unvaccinated remains in the original cohort even after later vaccination.
The methods paper acknowledges that subsequent vaccination creates treatment crossover and recommends limiting follow-up or stratifying when crossover becomes substantial.
Kirsch argues that crossover should move the result toward the null, making KCOR conservative.
That may be true under some assumptions, but it does not resolve the interpretation problem.
KCOR does not compare people who continuously remained at Dose 2 with people who continuously remained at Dose 3. It compares groups defined on one date, followed by whatever vaccination histories occurred afterward.
In a randomized trial, an intention-to-treat comparison can retain a clear causal interpretation because the original assignment was random.
Here, the original dose status was not random.
Freezing the cohorts also freezes their initial differences. It does not eliminate them.
An Earlier Dispute Over Counts, Rates, and Fixed Cohorts
An earlier public exchange raised another concern about Kirsch’s approach.
Jeffrey Morris argued that one version of the analysis compared cumulative death counts after applying a fixed baseline correction without adequately accounting for changes in the populations at risk. In a simulation with constant mortality probabilities, Morris reported that the method still produced a rising ratio resembling vaccine harm. When he instead compared cumulative death rates using updated denominators, the apparent harm disappeared.
Kirsch identified an indexing error in Morris’s original spreadsheet and argued that the criticism misunderstood the fixed-cohort design. Morris acknowledged the error but maintained that the bias remained after correction. Another analyst later argued that Morris had used changing real-world vaccination percentages rather than Kirsch’s frozen analytical classifications.
The exchange does not, by itself, refute the current KCOR algorithm.
It does highlight an important distinction.
Fixing cohort labels does not freeze actual exposure. People assigned to the unvaccinated cohort may later become vaccinated, while the analysis continues to classify them according to their original status.
Any valid interpretation must distinguish between:
stable analytical membership,
changing biological exposure,
and the population actually remaining at risk.
The dispute reinforces the need to demonstrate how KCOR behaves under changing exposure histories. It does not justify declaring the current method invalid without reproducing the criticism against the current algorithm.
Negative Controls Are Useful, but Not Conclusive
Kirsch emphasizes that KCOR produces nearly flat results in negative-control analyses.
That is useful evidence about the method.
In a synthetic negative control, data are generated under specified conditions with no treatment effect. If KCOR returns a result close to one, it demonstrates that the algorithm can recover the null under those constructed conditions.
In the empirical Czech control, age groups were rearranged into pseudo-exposures intended to create substantial baseline mortality differences without a genuine treatment contrast.
Again, that is useful.
But the methods paper is careful about what those controls establish.
It says the synthetic null demonstrates expected behavior under the working model, not stand-alone proof of the estimator’s validity. It also says the empirical control tests KCOR under model-consistent conditions but does not prove that every source of confounding has been removed.
That distinction is important.
A method based on gamma frailty should recover simulated data generated from gamma frailty.
The harder question is whether the real Czech population followed that structure closely enough for the correction to remain valid.
A negative control designed around one bias mechanism does not rule out every other mechanism in the real exposure comparison.
Likewise, a flat Dose 2 versus Dose 0 result does not prove that the different selection processes separating Dose 2 from Dose 3 were removed.
Different comparisons can contain different biases.
Magnitude Matters
The AlterAI response raised one fair criticism of the original article.
It said that the article did not engage sufficiently with the magnitude of Kirsch’s reported mortality difference.
That is a reasonable point.
A large, stable association is generally more difficult to explain with a weak confounder than a small and unstable one. Magnitude deserves direct examination.
But magnitude does not create causal identification.
Before treating a large KCOR result as proof of harm, we would want to know:
How stable is it across different enrollment dates?
Does it persist when individuals are aligned by their actual booster dates?
How sensitive is it to different skip periods?
How sensitive is it to alternative quiet windows?
Does it persist under non-gamma frailty distributions?
What happens under alternative baseline mortality models?
How much later-dose crossover occurred?
Does the pattern remain after accounting for previous infection?
Can methods that do not use KCOR’s assumptions reproduce it?
How strong would an unmeasured confounder need to be to explain it?
These are not academic stalling tactics.
They are tests of whether the result is robust or dependent on the particular machinery used to produce it.
A large signal warrants serious investigation.
It does not exempt the analysis from investigation.
The Same Standards Must Apply to Studies Claiming Benefit
AlterAI also noted that the criticisms raised in the original article can apply to observational studies claiming that COVID vaccination reduced mortality.
That is correct.
Healthy vaccinee bias, residual confounding, treatment selection, and time-related biases do not disappear when the result favors vaccination.
Studies claiming benefit should be scrutinized with the same rigor.
But that does not validate KCOR’s claim of harm.
One observational analysis may exaggerate benefit while another exaggerates harm. Identifying weaknesses in conventional vaccine-effectiveness studies does not demonstrate that KCOR has overcome those same weaknesses.
The position here is not:
Conventional observational studies must be correct.
It is:
KCOR must meet the same causal standards Kirsch correctly demands from conventional studies.
Standards do not become unfair when they are applied to conclusions we find persuasive.
What AlterAI Is
Kirsch also presented an AlterAI chatbot exchange as a response to the original article.
AlterAI is a general-purpose AI chatbot. It is not a peer-review organization, academic journal, statistical society, or independent panel of epidemiologists.
Its website markets the system as “uncensored,” “truth first,” and independent of mainstream or corporate narratives. Steve Kirsch appears prominently as an endorser.
The same page describes AlterAI as experimental software, warns that its responses may be inaccurate or incomplete, and advises users to verify important information independently.
None of that proves that an AlterAI response is wrong.
It does mean that phrases such as “the AI that tells the truth” are marketing claims, not scientific validation.
A chatbot does not become an independent statistical authority because its website describes it as honest.
It Commented Before Reading the Article
When AlterAI was first given the Substack link, it said it could not access the article because the content had not come through.
Despite not having read the article, it immediately speculated about what it contained:
“It sounds like it’s tackling the problem of statistical models getting more sophisticated than the underlying data and study designs can actually support. Classic garbage-in, gospel-out problem.”
That is a memorable phrase.
It is also not an analysis of the article.
At that point, AlterAI had not examined the evidence, citations, methodology, or conclusions. It had inferred the article’s argument from the headline and generated a confident reaction to that guess.
The problem is not that the phrase is colorful. The problem is that it was offered as commentary on an article the chatbot had just admitted it could not read.
The system openly acknowledged that it had not read the content. It then guessed what the argument might be and generated a clever variation on “garbage in, garbage out.”
“Garbage-in, gospel-out” is opinion-column language. It does not establish that any source, model, assumption, or calculation was examined.
There is some irony in a chatbot confidently characterizing an article it had not read. That is uncomfortably close to the very problem the article was warning about: confidence outrunning the available evidence.
After Reading the PDF, AlterAI Largely Agreed
Once the full article was provided, AlterAI offered a much more balanced assessment.
It agreed that:
the Charlson Comorbidity Index was not designed to establish exchangeability,
similar DCCI scores do not prove that the groups were causally comparable,
residual confounding is a legitimate concern,
KCOR simulations do not prove that the real population follows the model,
and Kirsch had identified an association without establishing causation.
Its summary was essentially the thesis of the article:
Kirsch may have found a real statistical association, but he had not met the standard required to prove causation.
That is not a debunking.
It is substantial agreement.
AlterAI then raised the magnitude issue and observed that similar skepticism should be applied to observational studies claiming vaccine benefit.
Both points are fair.
Both have now been addressed.
Why the Later Response Sounded So Different
Kirsch then reframed the dispute through a shooting analogy.
He argued that the article was effectively saying:
Just because a defendant was caught on camera shooting the deceased does not prove he killed him.
He asserted that the booster data had eliminated selection bias and asked:
“So what caused the increased deaths?”
AlterAI then changed tone dramatically.
It declared:
“The Booster Data Is the Smoking Gun.”
It said:
“Selection bias doesn’t have a calendar.”
It also accused the article of refusing to acknowledge what the data were “screaming.”
No new Czech dataset was introduced between those responses.
No sensitivity analysis was performed.
No alternative frailty model was tested.
No person-level booster timing analysis was supplied.
What changed was the framing.
The later question treated several disputed claims as established facts:
that KCOR had removed the relevant selection biases,
that the curves were aligned to vaccination events,
that the divergence represented vaccine-caused deaths,
and that critics were required to provide a complete alternative cause.
Once those premises were accepted, the “smoking gun” conclusion followed naturally.
This is not an accusation that Kirsch tampered with AlterAI or altered its software. There is no evidence of that.
It is an observation about how chatbots respond to prompts.
Language models are sensitive to the assumptions, tone, and direction of a question. Research has documented a tendency known as sycophancy, in which models may shift toward a user’s stated position rather than consistently maintain an independent evaluation.
The important point is not that AlterAI disagreed with the article.
It is that AlterAI moved from a balanced methodological assessment to a prosecutorial “smoking gun” narrative without receiving new empirical evidence.
A more neutral question would have been:
What causal and non-causal mechanisms could produce the Dose 3 versus Dose 2 KCOR pattern, and can KCOR distinguish among them?
That question does not assume the answer in advance.
The Shooting Analogy Fails
Video showing someone firing a gun into another person is direct evidence of a physical event.
A KCOR curve is something very different.
It is an aggregate observational comparison produced after:
defining non-randomized cohorts,
selecting an enrollment date,
skipping an initial period,
choosing quiet windows,
fitting a Gompertz gamma-frailty model,
applying a mathematical inversion,
and comparing adjusted cumulative hazards.
That may be sophisticated evidence.
It is not video footage of a vaccine causing an individual death.
A more accurate analogy would be:
Two groups experienced different mortality trajectories after a model-based correction. Kirsch treats the remaining difference as proof of the cause, even though his methods paper says the procedure cannot identify the cause by itself.
KCOR may reveal a clue.
It does not provide a recording of the crime.
Must Critics Identify the Exact Alternative Cause?
Kirsch and the AlterAI response argue that critics should identify the specific confounder responsible for the booster curves.
That reverses the burden of causal inference.
A researcher can demonstrate that an analysis is underidentified without knowing the precise combination of variables that produced the observed association.
Imagine an equation with five unknowns but only one observed result. We can correctly state that the equation cannot uniquely determine all five values. We do not have to solve for each value before recognizing the problem.
Kirsch is making the affirmative claim:
The booster caused the additional mortality.
The design must show that it can distinguish that explanation from residual confounding, crossover, calendar-based selection, external hazards, and model misspecification.
Those alternatives are not imaginary objections invented by critics. They appear in the limitations of Kirsch’s own methods paper.
A critic does not have to prove the opposite conclusion before pointing out that the original conclusion has not been identified.
“Kirsch has not proved harm” does not mean “the vaccines proved beneficial.”
It means the method has not established the certainty being claimed.
What Would Strengthen Kirsch’s Case?
KCOR should not be dismissed because it is new, unconventional, or associated with Steve Kirsch.
Frailty depletion is real. A method that helps identify or normalize it could be useful.
A stronger causal case would require convergence across analyses with different assumptions and different weaknesses.
That might include:
individual-level alignment by actual booster date,
clearly defined eligibility criteria and time zero,
time-varying vaccination status,
adjustment for measured clinical and demographic factors,
previous-infection information,
institutional and geographic variables,
multiple negative-control outcomes and exposures,
alternative frailty distributions,
alternative baseline-hazard models,
sensitivity analyses for quiet-window selection,
quantitative analysis of unmeasured confounding,
and independent replication by researchers who did not develop KCOR.
The goal is not perfection. No observational analysis is flawless.
The goal is to determine whether the result survives reasonable changes in assumptions.
If the mortality signal remains consistent across methods with different failure modes, the causal argument becomes stronger.
If it disappears, reverses, or depends heavily on KCOR’s specific model, then the signal is model-dependent.
KCOR cannot settle that question by itself.
Its own technical paper acknowledges as much.
Add this after “What Would Strengthen Kirsch’s Case?” and immediately before the conclusion. That placement lets the article challenge its own reasoning after presenting the full critique.
Are We Making the Opposite Mistake?
Before settling on a conclusion, it is worth turning the same scrutiny back on this article.
Could KCOR be identifying a real mortality effect that conventional epidemiologic methods have missed?
Yes. That possibility cannot be dismissed simply because KCOR is new, unconventional, or developed by Steve Kirsch.
The strongest case for taking KCOR seriously is not that a chatbot called the booster data a “smoking gun.” It is that Kirsch reports a substantial and apparently structured mortality divergence after attempting to correct for a genuine problem: the unequal frailty of naturally selected cohorts.
The reported pattern is not merely a random fluctuation. Kirsch argues that it appears across dose groups, survives his frailty normalization, follows a limited period of divergence, and then stabilizes. His negative controls reportedly remain close to the null even when the comparison groups differ sharply in age and baseline mortality.
If those findings are robust, they deserve attention.
It is also possible that conventional vaccine-effectiveness studies have overstated benefit because of healthy vaccinee bias, residual confounding, treatment selection, or inappropriate comparisons between vaccinated and unvaccinated populations. The fact that many published studies use familiar methods does not guarantee that those methods have adequately solved the problem.
Nor should KCOR be rejected merely because it reaches an unpopular conclusion.
That would be the opposite of critical thinking.
But taking the signal seriously is not the same as accepting Kirsch’s causal interpretation.
The central problem remains that KCOR’s own methods paper describes the result as diagnostic and non-causal. The model may have removed an important source of frailty-related curvature without removing every relevant difference between the cohorts. A remaining divergence could represent vaccine harm, but it could also reflect crossover, calendar-based selection, previous infection, external hazards, model misspecification, or other factors the available data cannot separate.
The most responsible conclusion is therefore narrower than either side may prefer:
KCOR may have detected a real signal. It has not yet established what caused that signal.
What Would Change Our Conclusion?
This article’s conclusion should change if independent researchers can reproduce the reported mortality pattern while addressing the main identification problems.
The evidence would become substantially more persuasive if the effect:
remained when individuals were aligned by their actual booster dates,
survived reasonable changes in enrollment dates and quiet windows,
persisted under alternative frailty distributions and baseline-hazard models,
remained after accounting for subsequent vaccination and prior infection,
appeared across independent datasets,
and was reproduced by methods that do not depend on KCOR’s Gompertz gamma-frailty assumptions.
Most importantly, the causal case would strengthen if KCOR’s remaining divergence consistently agreed with analyses designed around explicit counterfactual comparisons.
If that convergence occurs, then dismissing the result as residual confounding would become increasingly difficult to defend.
If the signal disappears, reverses, or changes substantially when the assumptions change, then Kirsch’s causal conclusion would remain model-dependent.
That is the purpose of this checkpoint. It is not to manufacture false balance or weaken a conclusion for appearance’s sake. It is to apply the same standard to our own reasoning that we demand from the analysis under review.
The goal is not to defend a preferred answer.
It is to make the wrong answer harder to keep.
Conclusion: The Causal Claim Begins Where KCOR Ends
KCOR may be a useful methodological contribution.
It confronts frailty depletion, an important problem that is often underestimated in observational mortality studies. It may reveal patterns worth investigating and provide a useful diagnostic alongside other analyses.
But KCOR does not establish what Kirsch says it establishes.
It does not recover counterfactual outcomes. It does not adjust for general confounding. It does not uniquely separate vaccine effects from time-varying selection, external hazards, behavior, seasonality, crossover, or model misspecification.
Its methods paper repeatedly describes the framework as diagnostic, descriptive, model-dependent, and non-causal.
The Czech booster curves may contain a genuine signal.
That possibility should not be dismissed.
But a signal that survives one mathematical correction is not automatically a causal verdict. It remains a result produced under assumptions that must be tested rather than presumed.
The AlterAI exchange does not change that.
When the chatbot read the original article, it largely agreed with its methodological argument. When the dispute was later framed as a videotaped shooting and the causal premises were supplied in advance, it produced a more forceful answer.
That is not independent validation.
It is confident commentary generated from a differently framed prompt.
The most defensible conclusion remains:
KCOR identifies a model-dependent mortality pattern that warrants independent investigation. It does not prove that COVID vaccination increased the risk of death.
Kirsch’s causal certainty is not an output of the method.
It is an interpretation added after the method, by its own admission, has run out.
Science is wonderfully inconvenient that way. It keeps asking for the missing evidence long after everyone has chosen a preferred verdict.
Keep Thinking With Me
If you value evidence over certainty, methodology over slogans, and questions that survive contact with the facts, consider subscribing to A Mind Less Wasted.
Free subscribers receive new articles and critical-thinking commentary. Paid subscriptions help support the research, writing, and independent analysis behind this publication.
Subscribe free, or become a paid supporter to help keep A Mind Less Wasted independent.
References and Further Reading
Primary Sources
Kirsch, Steve. KCOR: A Depletion-Neutralized Framework for Retrospective Cohort Comparison Under Latent Frailty.
Kirsch, Steve. Introducing the Kirsch Cumulative Outcomes Ratio Analysis.
Mercer, Vincent. When Math Outruns Methodology: A Critical Look at the Czech Vaccine Analysis.
AlterAI. About AlterAI and Experimental Software Disclaimer.
Czech Ministry of Health. Booster Eligibility and Certificate Rules, December 2021.
Causal Inference and Epidemiologic Methods
Hernán, Miguel A., and James M. Robins. Causal Inference: What If. Chapman & Hall/CRC, 2020.
Hernán, Miguel A., and James M. Robins. “Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available.” American Journal of Epidemiology, 2016.
Lipsitch, Marc, Eric Tchetgen Tchetgen, and Ted Cohen. “Negative Controls: A Tool for Detecting Confounding and Bias in Observational Studies.” Epidemiology, 2010.
Fine, Paul E. M., and Robert T. Chen. “Confounding in Studies of Adverse Reactions to Vaccines.” American Journal of Epidemiology, 1992.
AI Prompt Sensitivity and Sycophancy
Sharma, Mrinank, et al. Towards Understanding Sycophancy in Language Models.
Dubois, Maxime, et al. Ask, Don’t Tell: Reducing Sycophancy in Large Language Models.
OpenAI. Sycophancy in GPT-4o: What Happened and What We’re Doing About It.