$2,000 challenge: find a mistake in the KCOR algorithm

For the Czech Republic record-level mortality data (DOB, DOD, vaccination dates), I claim KCOR is the best generic method for determining what the data actually implies about net mortality effects.

Executive summary

I’m offering $2,000 to the first person (by chronological order of replies to the pinned comment) who both:

1. identifies a material methodological flaw in KCOR v5.4 (Kirsch Cumulative Outcomes Ratio) when applied to retrospective observational record-level mortality data, specifically the Czech Republic dataset, and

2. identifies a more reliable approach to analyzing the same Czech dataset which produces as its output a net harm/benefit as a function of time since the intervention.

A material flaw is not a limitation or modeling preference. It is an actual mistake that, under realistic conditions, can be shown to produce a wrong inference on a control test (see To win).

Why this matters

Most COVID observational analyses fail because they:

• assume proportional hazards when they are not,

• falsely believe that they can use 1:1 matching to overcome the static HVE effect (this was demonstrated to be false in these papers from Denmark and Qatar)

• censor or reassign people in ways that change cohort mortality,

• confuse selection effects with treatment effects.

KCOR was designed specifically to address those failure modes using fixed cohorts, explicit hazard modeling, and cumulative outcomes.

KCOR description

The KCOR 5.4 algorithm is described in this Word document. The KCOR v5.4 code is here.

In brief (COVID / Czech application):

1. Fixed cohorts at enrollment (t = 0)
   Cohorts are defined by number of vaccine doses received as of a single enrollment date.

2. Skip the first 2 weeks
   Follow-up weeks 1–2 are excluded to reduce dynamic Healthy Vaccinee Effects (HVE) immediately surrounding enrollment.

3. Cohorts remain fixed (no censoring, no switching)
   Individuals remain in their original cohort even if they later vaccinate.
   
   No censoring is performed, because censoring changes cohort composition and therefore changes the mortality being measured.

4. Discrete-time hazard transform
   Weekly mortality is converted into an equivalent continuous-time hazard using:

\(h(t) = -\ln\left(1 - \frac{D_t}{N_t}\right)\)

where Dt​ is deaths during week t and Nt is the number alive at the start of week t.
(The implementation uses the exact version appropriate to the data structure; the above is the conceptual form.)

5. Slope normalization (shape adjustment)
   Cohort-specific hazard-shape differences are estimated and removed by fitting log h(t):

   • Younger cohorts: linear fit over the analysis window

   • Older cohorts: flexible 4-parameter model to capture curvature from frailty depletion without assuming gamma frailty

   Robust fitting uses quantile regression (typically τ=0.5; other values may be used when appropriate).

6. KCOR estimator
   KCOR is computed as the ratio of cumulative hazards over follow-up and compared to a baseline ratio measured during a non-COVID period after enrollment.

To win

To claim the prize, you must identify an actual methodological mistake that meets both criteria:

1. It is demonstrably wrong, and

2. It causes KCOR to fail a realistic control test, producing a materially incorrect inference on either:

   • a negative control where the true effect should be ~null, or

   • a positive control where the true effect should be detectable.

Secondly, you must also identify a superior method for analyzing this data that is more likely to yield a correct result.

Your argument must include a reproducible test (realistic simulation or real-data control) demonstrating the failure. Failure under absurd parameter choices (inconsistent with the KCOR doc/code defaults) doesn’t count. Creating unrealistic pathological cohorts don’t count. KCOR is designed for use with actual real-world data, so you must show the failure mode with realistic assumptions.

Also no dataset is perfect so asserting that if the underlying data is unreliable, the KCOR results will be unreliable will not be a successful argument.

The neutral arbiter of whether you won

You must be able to convince ChatGPT you won.

Feed in this article, your proof of the mistake, with the prompt: “does my explanation meet the criteria to win?” and post that ChatGPT response acknowledging it believes that you won.

I will have one opportunity to point out the flaw(s) in your argument and if you survive that, you get the prize. I’m actually happy to be corrected if I got it wrong.

What does not qualify

• “In some pathological constructed cohort the fit isn’t perfect.”

• “I don’t like quantile regression / 4 parameters / no censoring.”

• “Observational data can’t prove causality.”

• Generic claims about “unknown confounders” that can apply to any method.

• Any objection that does not produce a concrete failing control test.

Identifying known limitations also does not qualify. Examples of known limitations include:

• A mortality step-function exactly at enrollment that remains constant thereafter (this would be absorbed into baseline hazard).

• Very short-lived effects confined entirely to the excluded weeks.

• Effects that require individual-level counterfactuals not identifiable in observational data.

These are limitations of any fixed-cohort observational estimator and are explicitly acknowledged.

What does qualify

• A coding-independent conceptual error (e.g., violation of an invariance property, incorrect estimator logic, unavoidable signal deletion under realistic conditions),
  plus

• A control demonstration showing that KCOR produces a wrong conclusion.

You must find an actual mistake in the method, not a philosophical objection or alternative modeling preference or pointing out a design limitation. The method has known limitations set by parameters (such as the 4 week baseline period). A vaccine which doubles mortality on the day of the shot would not be detected, for example.

You must also identify a method superior to KCOR which can determine the COVID vaccine net harm/benefit as a function of time using the Czech retrospective observational data.

Notes

• “Materially incorrect inference” means: wrong sign, large magnitude error (e.g., >20%), or confidence interval excluding the truth.

• Acceptable control tests include (but are not limited to) cohorts generated from a Gompertz hazard with a gamma-distributed frailty mix of mean 1.

• “First person” is determined by chronological order of replies to the pinned comment.

• Payment via PayPal, Venmo, or Apple Pay; choice of the winner.

• Disclaimer: This is an informal public challenge, not a contract.

• This contest is open to everyone, including Paul Offit, Peter Hotez, UPenn Professor Jeffrey Morris, Susan Oliver, Dorit Reiss, David Gorski, Skeptical Raptor, Uncle John Returns, Dr. Neil Stone, etc.

Summary

If KCOR’s errors are as obvious as some have claimed, demonstrating it should be easy, shouldn’t it?