Steve Kirsch says unpublished statistics from 182 patient records establish causation. The numbers may justify further study, but they cannot support the conclusion he draws.
Steve, the clustering is not in dispute. The question is whether it exceeds what we would expect without a causal effect. “You can’t explain it” does not provide that comparison.
I’m willing to examine parent reports and test your “Andrews effect,” but let’s define it and agree on the protocol before selecting cases. Contacting parents can verify dates and histories. It cannot create the missing denominator or fix selection bias.
If the goal is truth, I’m in. Let’s specify in advance what result would support or weaken each explanation.
Steve, I’d start by agreeing on exactly what we’re testing and what result would count against the hypothesis before looking at more data.
My suggested next steps:
Define the “Andrews effect” precisely and specify the predicted pattern.
Randomly select a prespecified sample of the clinic cases and independently verify vaccination dates, onset dates, and the nature of the regression with the parents and, where possible, records.
Establish the comparison we’re currently missing: how often children of the same ages and vaccination patterns would be expected to have a vaccination within the same two-day window absent a causal effect.
Preregister the analysis and the criteria for what would strengthen or weaken the vaccine hypothesis.
Publish the verified data, methods, and results so either of us can be wrong.
I’m happy to participate. The key for me is that we agree on the test before we see the answers, rather than interpreting the pattern afterward.
Steve, I think the best next step is to finish and lock the protocol, survey, and analysis plan before looking at the main-study responses. I’m willing to help with that.
My main concern is not whether the study produces a vaccine signal. It is whether we can agree beforehand on a design that gives that signal a fair chance to appear, but also gives the study a fair chance to show that vaccination is not the explanation.
Here is what I would suggest:
1. Lock the primary question and statistic.
Define exactly what we are testing. My understanding is that the primary hypothesis is excess child-specific synchronization between SORA onset and the child’s actual most recent pre-onset vaccination, with Days 0–2 as the primary window. We should also state in advance what result would weaken that hypothesis, not just what would support it.
2. Resolve the primary-window issue before proceeding.
The current overview says Days 0–2 will be the prespecified primary statistic, but the validation section also says the pilot will be used to choose the primary risk window empirically. I don’t think we should do both. If Days 0–2 is the hypothesis, keep it locked and use the pilot to determine whether parents can date onset accurately enough for a day-level analysis. If the pilot is going to choose the window, then the confirmatory main study needs to be genuinely separate and untouched when that choice is made.
3. Fully specify and simulation-test the permutation null before seeing the main data.
We should decide exactly what gets permuted, the mandatory strata, how delayed and catch-up schedules are handled, sparse and singleton strata, missing or month-only dates, tied dates, and how the “most recent vaccination before onset” rule works under permutation. Then simulate realistic null datasets to make sure the test actually maintains the intended false-positive rate under real vaccination schedules, age-dependent onset, date heaping, clinic differences, and delayed visits.
4. Make the SORA definition operational rather than primarily self-classified.
I would collect the underlying features and apply a locked classification algorithm afterward. Confirm that the lost skill or behavior was actually established beforehand, distinguish loss from plateau or fluctuation, record persistence and functional significance, and have a “possible/uncertain SORA” category. I would also document whether ASD was clinically diagnosed and the source of that diagnosis.
5. Capture competing events with enough timing to distinguish them.
Fever, illness, medication, vaccination, and medical visits can occur together. If we only know that each happened sometime within five days, we may not be able to separate them. For major exposures, collect timing, duration, indication, and whether they were part of the same episode or medical encounter.
6. Keep imprecisely dated cases rather than throwing them away.
Exact-date cases can enter the confirmatory day-level test. Month-only or uncertain cases should still remain available for descriptive and age-at-onset analyses. Otherwise we risk preferentially selecting cases whose onset is anchored to a memorable event.
7. Prespecify recruitment flow and record verification.
For each clinic, report invitations sent, successful deliveries, starts, completions, developmental categories, SORA classifications, follow-up consent, records obtained, and exclusions. For record verification, randomly select from consenting cases under a prespecified rule and compare verified with unverified respondents.
8. Freeze and preregister the protocol and statistical analysis plan before the first main-study analysis.
Primary hypothesis, primary window, SORA algorithm, permutation procedure, exclusions, missing-data rules, site effects, multiplicity, sensitivity analyses, and a commitment to report the primary result regardless of direction.
After that, run it and publish the data and code.
I’m not asking you to design a study that makes a vaccine association harder to find. I’m asking for one where, if a strong association survives, the obvious methodological objections have already been dealt with.
If we can agree on that principle, I’m interested in participating in the process. I’d rather help define the test before we know the result than argue about the interpretation afterward.
the tl;dr here is that you cannot explain the clustering.
https://alter.systems/p/a29f5a5a-9956-4660-b9bc-bc16ab2b06e2
Would you like to contact the parents directly to verify an Andrews effect because I'm sure you don't trust me to do that.
Or you can randomly pick a few of my followers and see if there is an Andrews effect.
Let me know if you are more interested in throwing darts or finding truth!
Steve, the clustering is not in dispute. The question is whether it exceeds what we would expect without a causal effect. “You can’t explain it” does not provide that comparison.
I’m willing to examine parent reports and test your “Andrews effect,” but let’s define it and agree on the protocol before selecting cases. Contacting parents can verify dates and histories. It cannot create the missing denominator or fix selection bias.
If the goal is truth, I’m in. Let’s specify in advance what result would support or weaken each explanation.
what do you suggest for next steps?
Steve, I’d start by agreeing on exactly what we’re testing and what result would count against the hypothesis before looking at more data.
My suggested next steps:
Define the “Andrews effect” precisely and specify the predicted pattern.
Randomly select a prespecified sample of the clinic cases and independently verify vaccination dates, onset dates, and the nature of the regression with the parents and, where possible, records.
Establish the comparison we’re currently missing: how often children of the same ages and vaccination patterns would be expected to have a vaccination within the same two-day window absent a causal effect.
Preregister the analysis and the criteria for what would strengthen or weaken the vaccine hypothesis.
Publish the verified data, methods, and results so either of us can be wrong.
I’m happy to participate. The key for me is that we agree on the test before we see the answers, rather than interpreting the pattern afterward.
see: https://github.com/skirsch/autism/tree/main/protocol/v2 for the protocol and the survey and the process. This is not finished yet.
Now have a 3 prong approach to test the hypothesis.
Steve, I think the best next step is to finish and lock the protocol, survey, and analysis plan before looking at the main-study responses. I’m willing to help with that.
My main concern is not whether the study produces a vaccine signal. It is whether we can agree beforehand on a design that gives that signal a fair chance to appear, but also gives the study a fair chance to show that vaccination is not the explanation.
Here is what I would suggest:
1. Lock the primary question and statistic.
Define exactly what we are testing. My understanding is that the primary hypothesis is excess child-specific synchronization between SORA onset and the child’s actual most recent pre-onset vaccination, with Days 0–2 as the primary window. We should also state in advance what result would weaken that hypothesis, not just what would support it.
2. Resolve the primary-window issue before proceeding.
The current overview says Days 0–2 will be the prespecified primary statistic, but the validation section also says the pilot will be used to choose the primary risk window empirically. I don’t think we should do both. If Days 0–2 is the hypothesis, keep it locked and use the pilot to determine whether parents can date onset accurately enough for a day-level analysis. If the pilot is going to choose the window, then the confirmatory main study needs to be genuinely separate and untouched when that choice is made.
3. Fully specify and simulation-test the permutation null before seeing the main data.
We should decide exactly what gets permuted, the mandatory strata, how delayed and catch-up schedules are handled, sparse and singleton strata, missing or month-only dates, tied dates, and how the “most recent vaccination before onset” rule works under permutation. Then simulate realistic null datasets to make sure the test actually maintains the intended false-positive rate under real vaccination schedules, age-dependent onset, date heaping, clinic differences, and delayed visits.
4. Make the SORA definition operational rather than primarily self-classified.
I would collect the underlying features and apply a locked classification algorithm afterward. Confirm that the lost skill or behavior was actually established beforehand, distinguish loss from plateau or fluctuation, record persistence and functional significance, and have a “possible/uncertain SORA” category. I would also document whether ASD was clinically diagnosed and the source of that diagnosis.
5. Capture competing events with enough timing to distinguish them.
Fever, illness, medication, vaccination, and medical visits can occur together. If we only know that each happened sometime within five days, we may not be able to separate them. For major exposures, collect timing, duration, indication, and whether they were part of the same episode or medical encounter.
6. Keep imprecisely dated cases rather than throwing them away.
Exact-date cases can enter the confirmatory day-level test. Month-only or uncertain cases should still remain available for descriptive and age-at-onset analyses. Otherwise we risk preferentially selecting cases whose onset is anchored to a memorable event.
7. Prespecify recruitment flow and record verification.
For each clinic, report invitations sent, successful deliveries, starts, completions, developmental categories, SORA classifications, follow-up consent, records obtained, and exclusions. For record verification, randomly select from consenting cases under a prespecified rule and compare verified with unverified respondents.
8. Freeze and preregister the protocol and statistical analysis plan before the first main-study analysis.
Primary hypothesis, primary window, SORA algorithm, permutation procedure, exclusions, missing-data rules, site effects, multiplicity, sensitivity analyses, and a commitment to report the primary result regardless of direction.
After that, run it and publish the data and code.
I’m not asking you to design a study that makes a vaccine association harder to find. I’m asking for one where, if a strong association survives, the obvious methodological objections have already been dealt with.
If we can agree on that principle, I’m interested in participating in the process. I’d rather help define the test before we know the result than argue about the interpretation afterward.