I think you were labouring over the wrong denominator*.
As I understand Steve Kirsch:
Let’s have the observation window end at t = 0.
Then, let Bill receive an intervention at t = -v (the event).
Bill’s death occurs at t = -d (with d necessarily <= v).
Then Bill’s daisy-pushing ratio is d/v.
We could base our analysis upon the daisy-pushing ratio, but Steve, and everyone else, has used the breathing ratio, r = 1 – d/v.
For a cohort of Bills, in the absence of a correlation between the intervention and death (hence between v and d), the expectation of r is 0.5, providing deaths are not otherwise skewed.
To remove any skew, how about we find the cohort of people who have never receive the intervention, and remained alive at the time Bill received his? We can calculate the average breathing time, for them (being the average of (v – d’), where -d’ denotes each such individual’s time of death). Then, instead of analysing Bill’s breathing ratio, r, we can look at his effect ratio, f = (v – d)/Avg(v – d’). Proceeding on this basis, the average of f, for all Bills, will be 1, when the intervention has no effect, less than 1, where the intervention appears to truncate life, and greater, where it appears to extend life.
But, back to r. r < 0.5 suggests that the intervention is life-shortening.
The “boosting” problem is genuine. If Bill is absolutely determined to receive a booster every ten days, say, then there was never any possibility that his death would be more than ten days after his most recent booster. This makes it all but inevitable that the average of r, for the intervention group, will be less than 0.5 We could try to move forward by analysing an effect ratio, as above, but for which the non-intervention cohort consisted of people who died within that ten-day period. However, we simply don’t know the scheduled dates of each Bill’s first, postmortem, booster. Possibly, we can assume a minimum boosting interval (which may well be the time imposed between dose #1 and dose #2, of the original “fully vaccinated” programme), and construct each Bill’s comparison cohort based upon it.
After Steve switched to taking dose #1 as the intervention, and accepting his dismissal of skew, there is no flaw in his logic exposed by the discussion, thus far.
However, suppose the intervention has material effect in blocking infection. This will manifest itself in a reduction of infection-ascribable deaths. A consequence of this is that deaths in the intervention group will comprise disproportionately more, with infection as cause, shortly after the intervention (where the infection occurred just prior to the intervention), than longer after the intervention. This, of course, would move r below 0.5, just as Steve observed it to be, and, indeed, would move f below 1.
Thus, Steve’s finding is consistent with a 100% safe vaccine which has material efficacy against infection. Which is, undoubtedly, exactly what is being sold to us!!!
* Your denominator was the full, two-year, observation window.
There are two denominators. In the terminology of the spreadsheet at the time of the original posting and my "solution" post, there was vda/vdo (vaxxed days alive/observed) and uda/udo (unvaxxed days alive/observed) for the vaccinated. "Unvaxxed" just meant from January 1, 2021 to the day the report was submitted; it was how the original post calculated the unvaxxed alive/observed ratio but the same number was still valid for the vaccinated (ie time alive/observed from date, same metric). But the function was messed up and so I had to virtualize that value using blocked mid-month days.
"To remove any skew, how about we find the cohort of people who have never receive the intervention, and remained alive at the time Bill received his?" That's already mentioned in my post - "Lots of studies work this way, finding a demographically matched control and watching them from the same day as the vaccinated subject."
But is not "vaxxed days observed" (vdo) purely the period commencing at the intervention time and ending with the end of the observation window, rather than the full width of the window? (This is "v", in my terminology.)
"vaxxed days alive" (vda) is unequivocally the period between intervention and death.
Both being so, why would it be impossible for the average of the ratio, AVG(vda/vdo) (or, equivalently, the ratio SUM(vda)/SUM(vdo)) to come in as > 0.5 as you seemed to be suggesting? For an irrelevant intervention, the ratio (in the absence of other skew) would be 0.5, with a ratio greater than this suggesting the intervention is life-extending.
"'Unvaxxed' just meant from January 1, 2021 to the day the report was submitted" - Maybe I should have tried digging into the spreadsheet! Did you mean "day of death"? The submission date doesn't seem to be material. For the unvaxxed, udo is just the length of the (two-year) observation period, isn't in?
The unvaxxed are "observed" in their category (whether dead or live) for the whole two years. Whereas, the vaxxed are "observed" in their category (again, dead or alive) only from their injection time to the end of the two year study period.
Isn't the solution to simply compute total alive time from each (vaccinated at all / unvaccinated) group in the 2 year period, rather than time after any event? Deaths in any age group ought to be distributed the same in the V and U/V groups.
In other words, make one of your graphs for V and U/V. Both cover the whole time span. U/V ought to have death line (green) slope approximately equal to 1, V ought to have slope greater than one. This aggregate shift in slope, i believe, approximates the statement "V, all else being equal, increases likelihood of death."
There's a lot of complex reasons why that would not necessarily be revealing of anything. For example population of vaxxed grows bigger through time and unvaxxed smaller, does that mean theres a late and early deaths bias for both, respectively? OTOH this isn't a scenario where you are tracking all deaths in the population, it's just reported deaths, so none of those biases may actually be that strong. All you really know about these deaths is that they happened. Not actually "how often" they happened in a tracked group.
Even if you normalize to total deaths you don't know "how often" because you don't know what the number of survivors is. So let's say you normalized to total deaths, and it looked like the unvaxxed were dying faster. But what if IRL they only died at 1/10 the rate? Then the correct graph would only be 1/10 as high vs the unvaxxed. You have no way to know.
If we count in the early death bias for U/V and late bias for V, that would counter the trend we would be looking to find. It would flatten the slope of the V group. If the signal shows up despite that flattening, then we have found a strong signal that V experienced a death-causing event. That's all Steve wants to find anyway -- not a precision instrument, but only a rough signal.
No normalization necessary. Time vs. Lives. The boundary between life & death should have slope about 1. If an event causes death, that slope increases for the cohort that experienced that event. Serious war would make it nearly vertical for a segment of the population on the losing side. A worldwide deluge would make it completely vertical for those not in boats.
A shot causing death should increase its slope, perhaps only slightly. All you need to know is whether a person received any number of shots. Perhaps you could also make separate graphs for number of shots, to see whether n shots increases slope to, say, cn from 1.
I look forward to the impending update. However, I don't think "trivial fix which is simply to use a specific shot #" answers the model. This is why I specifically do not limit the model / argument to the booster "re-up" problem, it applies to any event. I don't mind having my model refuted at all, it would be a pleasure - maybe your update will have something to change my mind about the vaccine number question.
*edit: Maybe I see where you are going with "can’t pick a start time that has implicit knowledge of the date that the person died" but I will wait to reply further.
Ah, I see the airtable uv days alive has been fixed, that makes using the data a lot easier
Thank you for this challenge - I found it compelling.
Having thought about it all day I think you may have goofed after all.
To make the problem simpler we can assume that everyone was vaccinated after one year. If deaths occur evenly as in your hypothesis, at the point of vaccination, half of the deaths will already have occurred. Being vaccinated is therefore a marker of being alive at a later date.
If we took your square diagram and instead of having random events made all the events line up at the half way point then you can see that the denominator for the vaccinated should not include those deaths that have already occurred. Instead a new square can be drawn that includes only the top left hand corner of your original square. From this mini square the ratio of life lived to potential life lived would still be 0.5 for that subgroup.
There is a bias created, however. It is a bias in the unvaccinated. For them the first year passes with half of the cohort dying. At the point of vaccination a large proportion of the cohort moves into the vaccinated category. The remaining unvaccinated cohort go on to have deaths in the second year but these will be far fewer than those that occurred in the first year because there are fewer people who remain unvaccinated for the second year. Therefore, the deaths in the unvaccinated cohort will be disproportionately in the first year creating an artificially low ratio.
I think the methodology is sound but that does not mean there is no room for there being a bias in the data. The bias that concerns me is that it is possible that a proportion of Steve's followers had relatives who did die coinicidentally after vaccination. Because that subgroup might be over-represented there may be a bias that reduces the ratio for the vaccinated. Indeed, I do not think the ratio could fall as low as 0.3 without there being some such bias present.
P.S. I'm looking forward to your part 3 on the worry window. It has been over 18 months of not having had that hypothesis challenged and it needs to be.
True, but that bias would mainly be necessarily applicable if following both survivors and die-ers. So your survivors are banking all this immortal time before vaccination, it totally distorts their death rate. But these people have already died. So it's not as obvious that vaccination creates an expectation of life, since we aren't seeing any alive people to begin with (the big exception is that the winter 20/21 deaths are unvaxxed, because no one is vaxxed yet, as shown in the breakdown below).
For this reason, I started with looking at deaths per month to get a feel for whether comparison to a even distribution proof was adequate or if I would have to model something more difficult like "can expected .75 be made to go to .31" given assumed or real vax dates."
In particular, vaxxed alive/observed climbs throughout the calendar, and lands in .96 for December 2022. That matches an expectation of little survivorship bias.
Mortality rates at the time as well as unvaxxed per month survey deaths can both predict the immortal time credit that is given to the vaxxed. The per month unvaxxed is arguably best because "recall bias" should apply to both groups (thereby softening immortal time bias).
Most of the savings would be in January to April:
2021 Unvaxxed deaths:
J 20
F 13
M 11
A 9 -> =38%
M 5
J 9
J 3
A 20
S 15
O 11
N 10
D 14 -> = 100%
2021 Vaxxed:
J 1
F 5
M 9
A 12 -> =13%
M 15
J 24
J 18
A 28
S 23
O 22
N 23
D 34 -> = 100%
There aren't many people dying between May and July since that's between Alpha and Delta waves. And by the end of this period, the adult vax rate is basically set. Presumably, you could start the clock from there, subtracting the extra observed days, and reproduce Kirsch's original result.
Having said all that I think there is room for something resembling the bias you describe. The vaccinated had more than one dose and the measure was taken from the last dose not the first. If the clock had started with the first dose we'd expect a ratio of 0.5. Because some clocks started with later doses they were measured over a shorter time span and would have an artificially lower ratio.
Yes, that would be the expected effect of boosters - however, it seems like this problem was minimal give that vaxxed alive/observed is .96 in December 2022.
*edit: Wait, that's just an artifact of the cutoff being immediately after va begins. So if boosters depress va/vo it will be hidden immediately before cutoff. But definitely could be depressing va/vo in September and October.
Have you thought about having a debate about this with Steve Kirsch? It will be interesting to see Norman Fentons take on this? If there is anything wrong with this survey HE will definitely find it. However it turns out, I think we should all be thankful that at least we can talk about it? We need to find the truth for the future, or as near as we can get anyway! You ALL need to pat yourselves on the back for trying, under impossible circumstances, to make sense of all this!
I saw Steve's post this morning(and the original for the survey back whenever) and wasn't sure what I didn't like about it. I read your post and still didn't get it. I write this here because it is quieter and more friendly to skeptical opinion. BTW I like Kirsch and the way he is trying to fight this is not meant as a shot at him.
One of the best tricks to understanding a math problem is to game it out for a specific case. I will draw out Steve's survey for one data point.
So, let's imagine that the total period of the study is 100 days, I think it is actually from Jan 1 2021 to the date he started collecting data but replacing things with powers of ten let's us look at statistics without getting bogged down in arithmetic. So, we take our case and have him die on day 10. He lived 10 out of a possible 100 days for a value of .1
Now, let's take the same guy and vaccinate him on day 5. He still dies on the same day, day 10 of the study which is 5 days after vax. From Vday to end of study is 95 days. He lived 5 days after vaccination out of a possible 95 days for a value of .052. Died on the same day but because the numerator is smaller, shrinking it and the denominator both by 5 days has a bigger effect on the numerator.
The vaccine would actually have to extend his life by 4.5 days to 9.5 days post Vax to get the same .1 result.
If I understand the study right, then that is my analysis. I'm not sure I understand Steve's method though. If this helps anyone collect the prize please contribute to Feeding Jon's Kids a wonderful charity that I will be starting soon to help feed impoverished children who live in my house so that I can use the money that is currently feeding them to buy better booze, I mean for some noblish purpose.
That's also 100% accurate. I use the perfectly smooth death distribution model for my proof but you can also demonstrate the effect on any discrete death. All that matters is you ask "what happens under the null hypothesis that this death was fated?" And you find that starting the clock from injection vs. the beginning of the observation window *must* lower the alive/observed ratio. It is impossible to not reduce it.
I like that he brings these things up and he provides some interesting points of view but many on closer inspection turn out be 🤷
I couldn’t really follow his methodology at a glance, sounded good but when you start taking ratios at different points in time and taking averages etc. it’s a good way to mess things up if your not careful. Not saying he has, but if what you've written is a fair assessment of his method, I can totally see the issue. I can clearly follow your example but find it hard to relate it directly to Steve’s analysis mainly because I can’t follow Steve’s logic fully (could be lack of trying on my part).
I don’t really want to spend the time unraveling his post. I’m sure others like yourself did/will do a better job at it. I hope Norman Fenton or Clare Craig will do a deep dive after being name dropped. Even if they agree with his methodology they would be able to show why his critics are wrong.
I look forward to more skeptical discussion of his findings, not because I disagree with him, I hope he has found a silver bullet but it has to stand up better scrutiny than just “put up or shut up.”
As for relating my model to Kirsch's proof, when you go replicate his filters in the airtable it's clear that the denominator for the unvaccinated is 1/1/21 to the day the reports are filed. So it's a perfect replication of my "no event" model, although the distribution isn't perfectly even, but that's life.
Conversely, there's some weirdness with the values for the vaccinated which might be why he didn't notice this obvious flaw. If you want to find out what the vaccinated "time alive" would be from 1/1/21, you have to independently hash it from 1/1/21 to time of death. But I just lumped all monthly deaths together and used the days from 1/1/21 to middle of the month to virtualize this value. It results in a clear signal that the vaccinated are not dying more "quickly," which would have been a red flag if noticed.
Now I’m verging on being obsessive and I appreciate the tone of your comment, so I don’t want this to sound like I’m piling on. But this entire “study” (and I’m being generous even putting it in quotes) is complete garbage, before you even dissect the finer points of the methodology. I always enjoy Brian’s “begging the question” approach to debunking (which Kirsch is basically demanding as a response), but I feel like it might bury the lede here a bit. Which is this is a self-reported survey, with leading questions, from a biased audience that contains a grand total of 1,634 responses. I’d say it’s statistically insignificant, but that’s still too flattering as it’s an insult to statistics.
It’d be one thing if it was presented as a “well, look at this, kind of interesting, right?” But as some sort of unassailable proof that is so damning that the mainstream media has to ignore it? It’s either extreme hubris or a deliberate con.
And again, respectfully, I don’t think you need to dig into this very deeply to see the inherent problems. It’s why Kirsch poisons the well to begin with by claiming that anyone who raises the obvious objections is somehow being disingenuous. Dude’s a smooth operator. How much of that is his inherent personality or whether he knows what he’s shoveling is anyone’s guess.
I think we may be more in agreement than you think. I was just being more generous to Steve in my assessment. I find his approach interesting but didn’t want to criticize it too much without giving it due consideration.
No doubt and I was being genuine in saying I appreciated the tone of your reply. It’s frankly one that I should adopt more often. I seem to be especially uncharitable this evening and I’m probably being a bit harsh on Mr. Kirsch. Using small sample size like this is just nails on a chalkboard to me though and I’m getting old and cranky when it comes to rhetorical games.
You can also prove this wrong a different way: if you assume a constant probability p of death each day, the resulting distribution of deaths is not uniform, because, with the exception of Damar Hamlin, people generally only die once and then stop. So, deaths will be weighted more strongly towards the start. The odds of dying on the first day is p, but of dying the nth day would be ((1-p)^(n-1))*p. The other problem as you also point out is that death rates are not uniform over the year (and there could be confounding correlation between time of vax and overall death rates)
Is this Groundhog’s Day? Didn’t we go through this circus act like a year ago? With the same incredibly small sample size with self-reported results submitted by an audience that is definitely not a random distribution? Is this what handwaving looks like?
I’d probably reflexively puke in my mouth less if it wasn’t presented like a “get rich quick” or “miracle cure” scam website. COMPLETE with BOLDED TEXT and APPEALS TO AUTHORITY to demonstrate that IT IS IMPOSSIBLE I am wrong! And if I am, you CAN WIN BEN STEIN’S MONEY!!!
Maybe I’m just cranky, but I’m past the point where these things are just a cute character quirk, and their heart is in the right place, yada yada. If this is what the being anti-vax means, I don’t even care which side wins. You’re just exchanging one group of egotistical sociopaths for another (which is on brand for humanity, give a cursory review of history).
I haven't ever minded Kirsch's stylistic red flags. Mostly I find him to err in the opposite direction, utterly besotted of Fenton's credentials regardless of how many times it lands him a one way ticket to Bias Illusion Burn Town.
So if I were to get the analogy and the setup correct, it appears that the proof was already constructed in such a way that the inclusion of any data point would inherently bias the results in favor of shortened life?
I tried reading through the first portion (Edit: of Steve's post) and to be honest I'm caught off-guard a bit by the early comments which appear more to prop up the supposed data.
The use of his own survey should raise serious criticisms as well. I think I've heard this comment before, but I think there's a saying that "the people who answer surveys usually aren't the people you want answering surveys", so I'd be curious how you control for such biases. The causes of death (and key symptoms) included in the survey, I feel, are also inherently biased in favor of arguing a relationship between a relative's death and the vaccines. I certainly wouldn't list vaccines with (the "cure"). Again, that's already biasing results. I haven't produced my own surveys, but I find that the survey is sprinkled with far too many *wink, wink; nudge, nudge* towards the vaccines.
Right, there's obvious bias problems in the survey. Repliers were supposed to report the "closest" person to them that had died. Obviously that leaves wiggle-room for someone to highlight a case with more recent injection vs. another case with less recent. But this wouldn't get to the heart of the problem of why it's a no-win design.
Boosters exaggerate the problem as well, like I say, but don't get to the heart of it. The way the denominator is designed doesn't overtly penalize boosting. So it would be harder to model the detriment there.
Which is why the proof is based on a simple any event model.
This is why we need more people raising some skepticism for posts. Having that discourse may encourage more stringent surveys and cleaner results. The problem is that most posts may just be met with people already starting with the assumption that what is being presented is inherently infallible, or people already just agreeing with the premise right off the start.
The only you will get Kirsch to admit he’s wrong is to apply Captain Kirk wits to the Kobayashi Maru dilemma- you would need someone to trick him into admitting he’s made a mistake.
On that discussion we had about the message that all vaccines warrant further investigation- Andrew Bostom I think did a good job of cracking open the door with this post about the ineffective flu vaccines.
We'll see. I mean in this case, he's engineered some type of automatic bet deal so anyone could supposedly use this post to 10X their money if they wanted. I don't have the minimum bid on hand, and I found the nature of what is actually supposed to be betted on hard to parse.
Thanks for highlighting that! My impression of the flu vaccines is that they just put off infection for a year. After that, the vaxxed are in natural immunity debt and start to catch up. In the end it's a wash. It doesn't have anything to do with OAS, or strain mismatches (vaccines still tend to work on not-previously-vaccinated regardless of being a perfect match to seasonal drift, only antigenic shift totally breaks vaccines). It's not some big paradox, it's simple futility.
I went and re-read. It's not sufficient to simply show that the survey proof is wrong. Looks like the whole exercise is a Russian Nested Kobayashi Maru Doll
Despite the caps locks, the conditions of the bet seem unwinnable.
First, Fenton is the judge, despite the fact that he has apparently already blessed the survey proof and so what you are debating is whether Fenton is wrong. That's crazy.
Second, there is a further requirement of proving the vaccines save lives either with the survey set or a different resource. Actually, I think someone could pull that off (obviously it would only be showing healthy user bias), but again, the unfair judge problem.
Was your main criticism similar to this comment by Prof Norman Fenton?
https://stevekirsch.substack.com/p/game-over-medicare-data-shows-the/comment/13112807
Alternate link in case the above doesn't load: https://archive.is/smwPc
I think you were labouring over the wrong denominator*.
As I understand Steve Kirsch:
Let’s have the observation window end at t = 0.
Then, let Bill receive an intervention at t = -v (the event).
Bill’s death occurs at t = -d (with d necessarily <= v).
Then Bill’s daisy-pushing ratio is d/v.
We could base our analysis upon the daisy-pushing ratio, but Steve, and everyone else, has used the breathing ratio, r = 1 – d/v.
For a cohort of Bills, in the absence of a correlation between the intervention and death (hence between v and d), the expectation of r is 0.5, providing deaths are not otherwise skewed.
To remove any skew, how about we find the cohort of people who have never receive the intervention, and remained alive at the time Bill received his? We can calculate the average breathing time, for them (being the average of (v – d’), where -d’ denotes each such individual’s time of death). Then, instead of analysing Bill’s breathing ratio, r, we can look at his effect ratio, f = (v – d)/Avg(v – d’). Proceeding on this basis, the average of f, for all Bills, will be 1, when the intervention has no effect, less than 1, where the intervention appears to truncate life, and greater, where it appears to extend life.
But, back to r. r < 0.5 suggests that the intervention is life-shortening.
The “boosting” problem is genuine. If Bill is absolutely determined to receive a booster every ten days, say, then there was never any possibility that his death would be more than ten days after his most recent booster. This makes it all but inevitable that the average of r, for the intervention group, will be less than 0.5 We could try to move forward by analysing an effect ratio, as above, but for which the non-intervention cohort consisted of people who died within that ten-day period. However, we simply don’t know the scheduled dates of each Bill’s first, postmortem, booster. Possibly, we can assume a minimum boosting interval (which may well be the time imposed between dose #1 and dose #2, of the original “fully vaccinated” programme), and construct each Bill’s comparison cohort based upon it.
After Steve switched to taking dose #1 as the intervention, and accepting his dismissal of skew, there is no flaw in his logic exposed by the discussion, thus far.
However, suppose the intervention has material effect in blocking infection. This will manifest itself in a reduction of infection-ascribable deaths. A consequence of this is that deaths in the intervention group will comprise disproportionately more, with infection as cause, shortly after the intervention (where the infection occurred just prior to the intervention), than longer after the intervention. This, of course, would move r below 0.5, just as Steve observed it to be, and, indeed, would move f below 1.
Thus, Steve’s finding is consistent with a 100% safe vaccine which has material efficacy against infection. Which is, undoubtedly, exactly what is being sold to us!!!
* Your denominator was the full, two-year, observation window.
There are two denominators. In the terminology of the spreadsheet at the time of the original posting and my "solution" post, there was vda/vdo (vaxxed days alive/observed) and uda/udo (unvaxxed days alive/observed) for the vaccinated. "Unvaxxed" just meant from January 1, 2021 to the day the report was submitted; it was how the original post calculated the unvaxxed alive/observed ratio but the same number was still valid for the vaccinated (ie time alive/observed from date, same metric). But the function was messed up and so I had to virtualize that value using blocked mid-month days.
"To remove any skew, how about we find the cohort of people who have never receive the intervention, and remained alive at the time Bill received his?" That's already mentioned in my post - "Lots of studies work this way, finding a demographically matched control and watching them from the same day as the vaccinated subject."
OK.
But is not "vaxxed days observed" (vdo) purely the period commencing at the intervention time and ending with the end of the observation window, rather than the full width of the window? (This is "v", in my terminology.)
"vaxxed days alive" (vda) is unequivocally the period between intervention and death.
Both being so, why would it be impossible for the average of the ratio, AVG(vda/vdo) (or, equivalently, the ratio SUM(vda)/SUM(vdo)) to come in as > 0.5 as you seemed to be suggesting? For an irrelevant intervention, the ratio (in the absence of other skew) would be 0.5, with a ratio greater than this suggesting the intervention is life-extending.
"'Unvaxxed' just meant from January 1, 2021 to the day the report was submitted" - Maybe I should have tried digging into the spreadsheet! Did you mean "day of death"? The submission date doesn't seem to be material. For the unvaxxed, udo is just the length of the (two-year) observation period, isn't in?
The unvaxxed are "observed" in their category (whether dead or live) for the whole two years. Whereas, the vaxxed are "observed" in their category (again, dead or alive) only from their injection time to the end of the two year study period.
Brian, there were a number of errors in the article which are now being fixed. Check back in a few days.
I don't see anything in today's update to suggest that my model is limited to boosters.
I have updated my post to show that the aggregated vaxxed alive/observed ratio replicates a weighted per-month average of vaxxed alive/observe ratios. Therefore, it does not show an association with the first dose but only the impact of shortening "alive" time. https://unglossed.substack.com/i/99917859/and-so-my-response-to-the-february-semi-rebuttal
Isn't the solution to simply compute total alive time from each (vaccinated at all / unvaccinated) group in the 2 year period, rather than time after any event? Deaths in any age group ought to be distributed the same in the V and U/V groups.
In other words, make one of your graphs for V and U/V. Both cover the whole time span. U/V ought to have death line (green) slope approximately equal to 1, V ought to have slope greater than one. This aggregate shift in slope, i believe, approximates the statement "V, all else being equal, increases likelihood of death."
There's a lot of complex reasons why that would not necessarily be revealing of anything. For example population of vaxxed grows bigger through time and unvaxxed smaller, does that mean theres a late and early deaths bias for both, respectively? OTOH this isn't a scenario where you are tracking all deaths in the population, it's just reported deaths, so none of those biases may actually be that strong. All you really know about these deaths is that they happened. Not actually "how often" they happened in a tracked group.
Even if you normalize to total deaths you don't know "how often" because you don't know what the number of survivors is. So let's say you normalized to total deaths, and it looked like the unvaxxed were dying faster. But what if IRL they only died at 1/10 the rate? Then the correct graph would only be 1/10 as high vs the unvaxxed. You have no way to know.
If we count in the early death bias for U/V and late bias for V, that would counter the trend we would be looking to find. It would flatten the slope of the V group. If the signal shows up despite that flattening, then we have found a strong signal that V experienced a death-causing event. That's all Steve wants to find anyway -- not a precision instrument, but only a rough signal.
No normalization necessary. Time vs. Lives. The boundary between life & death should have slope about 1. If an event causes death, that slope increases for the cohort that experienced that event. Serious war would make it nearly vertical for a segment of the population on the losing side. A worldwide deluge would make it completely vertical for those not in boats.
A shot causing death should increase its slope, perhaps only slightly. All you need to know is whether a person received any number of shots. Perhaps you could also make separate graphs for number of shots, to see whether n shots increases slope to, say, cn from 1.
Right - I misread your previous comment and thought you were proposing something more like a histogram.
I look forward to the impending update. However, I don't think "trivial fix which is simply to use a specific shot #" answers the model. This is why I specifically do not limit the model / argument to the booster "re-up" problem, it applies to any event. I don't mind having my model refuted at all, it would be a pleasure - maybe your update will have something to change my mind about the vaccine number question.
*edit: Maybe I see where you are going with "can’t pick a start time that has implicit knowledge of the date that the person died" but I will wait to reply further.
Ah, I see the airtable uv days alive has been fixed, that makes using the data a lot easier
Oh Brian, Brian, Brian...you're such a party pooper. 😄
Thank you for this challenge - I found it compelling.
Having thought about it all day I think you may have goofed after all.
To make the problem simpler we can assume that everyone was vaccinated after one year. If deaths occur evenly as in your hypothesis, at the point of vaccination, half of the deaths will already have occurred. Being vaccinated is therefore a marker of being alive at a later date.
If we took your square diagram and instead of having random events made all the events line up at the half way point then you can see that the denominator for the vaccinated should not include those deaths that have already occurred. Instead a new square can be drawn that includes only the top left hand corner of your original square. From this mini square the ratio of life lived to potential life lived would still be 0.5 for that subgroup.
There is a bias created, however. It is a bias in the unvaccinated. For them the first year passes with half of the cohort dying. At the point of vaccination a large proportion of the cohort moves into the vaccinated category. The remaining unvaccinated cohort go on to have deaths in the second year but these will be far fewer than those that occurred in the first year because there are fewer people who remain unvaccinated for the second year. Therefore, the deaths in the unvaccinated cohort will be disproportionately in the first year creating an artificially low ratio.
I think the methodology is sound but that does not mean there is no room for there being a bias in the data. The bias that concerns me is that it is possible that a proportion of Steve's followers had relatives who did die coinicidentally after vaccination. Because that subgroup might be over-represented there may be a bias that reduces the ratio for the vaccinated. Indeed, I do not think the ratio could fall as low as 0.3 without there being some such bias present.
P.S. I'm looking forward to your part 3 on the worry window. It has been over 18 months of not having had that hypothesis challenged and it needs to be.
True, but that bias would mainly be necessarily applicable if following both survivors and die-ers. So your survivors are banking all this immortal time before vaccination, it totally distorts their death rate. But these people have already died. So it's not as obvious that vaccination creates an expectation of life, since we aren't seeing any alive people to begin with (the big exception is that the winter 20/21 deaths are unvaxxed, because no one is vaxxed yet, as shown in the breakdown below).
For this reason, I started with looking at deaths per month to get a feel for whether comparison to a even distribution proof was adequate or if I would have to model something more difficult like "can expected .75 be made to go to .31" given assumed or real vax dates."
In particular, vaxxed alive/observed climbs throughout the calendar, and lands in .96 for December 2022. That matches an expectation of little survivorship bias.
Mortality rates at the time as well as unvaxxed per month survey deaths can both predict the immortal time credit that is given to the vaxxed. The per month unvaxxed is arguably best because "recall bias" should apply to both groups (thereby softening immortal time bias).
Most of the savings would be in January to April:
2021 Unvaxxed deaths:
J 20
F 13
M 11
A 9 -> =38%
M 5
J 9
J 3
A 20
S 15
O 11
N 10
D 14 -> = 100%
2021 Vaxxed:
J 1
F 5
M 9
A 12 -> =13%
M 15
J 24
J 18
A 28
S 23
O 22
N 23
D 34 -> = 100%
There aren't many people dying between May and July since that's between Alpha and Delta waves. And by the end of this period, the adult vax rate is basically set. Presumably, you could start the clock from there, subtracting the extra observed days, and reproduce Kirsch's original result.
To put it more simply, if you had a dataset of 100 dead unvaccinated people, what proportion would you think died in 2021 compared to 2022?
Likewise of 100 dead vaccinated people what proportion would have died in Jan/Feb 2021?
Having said all that I think there is room for something resembling the bias you describe. The vaccinated had more than one dose and the measure was taken from the last dose not the first. If the clock had started with the first dose we'd expect a ratio of 0.5. Because some clocks started with later doses they were measured over a shorter time span and would have an artificially lower ratio.
Yes, that would be the expected effect of boosters - however, it seems like this problem was minimal give that vaxxed alive/observed is .96 in December 2022.
*edit: Wait, that's just an artifact of the cutoff being immediately after va begins. So if boosters depress va/vo it will be hidden immediately before cutoff. But definitely could be depressing va/vo in September and October.
I don't think you goofed here. You should be able to write a simple python script and illustrate this.
I really hope Steve gets to hear about your criticism in time.
Have you thought about having a debate about this with Steve Kirsch? It will be interesting to see Norman Fentons take on this? If there is anything wrong with this survey HE will definitely find it. However it turns out, I think we should all be thankful that at least we can talk about it? We need to find the truth for the future, or as near as we can get anyway! You ALL need to pat yourselves on the back for trying, under impossible circumstances, to make sense of all this!
I had posted a link to the google sheet in his post's comments, he skipped over it, replying to the comment before and liking the one above.
😉 I track my unwanted comments that way, too.
By reading this it seems you do like to play with fire.
I saw Steve's post this morning(and the original for the survey back whenever) and wasn't sure what I didn't like about it. I read your post and still didn't get it. I write this here because it is quieter and more friendly to skeptical opinion. BTW I like Kirsch and the way he is trying to fight this is not meant as a shot at him.
One of the best tricks to understanding a math problem is to game it out for a specific case. I will draw out Steve's survey for one data point.
So, let's imagine that the total period of the study is 100 days, I think it is actually from Jan 1 2021 to the date he started collecting data but replacing things with powers of ten let's us look at statistics without getting bogged down in arithmetic. So, we take our case and have him die on day 10. He lived 10 out of a possible 100 days for a value of .1
Now, let's take the same guy and vaccinate him on day 5. He still dies on the same day, day 10 of the study which is 5 days after vax. From Vday to end of study is 95 days. He lived 5 days after vaccination out of a possible 95 days for a value of .052. Died on the same day but because the numerator is smaller, shrinking it and the denominator both by 5 days has a bigger effect on the numerator.
The vaccine would actually have to extend his life by 4.5 days to 9.5 days post Vax to get the same .1 result.
If I understand the study right, then that is my analysis. I'm not sure I understand Steve's method though. If this helps anyone collect the prize please contribute to Feeding Jon's Kids a wonderful charity that I will be starting soon to help feed impoverished children who live in my house so that I can use the money that is currently feeding them to buy better booze, I mean for some noblish purpose.
That's also 100% accurate. I use the perfectly smooth death distribution model for my proof but you can also demonstrate the effect on any discrete death. All that matters is you ask "what happens under the null hypothesis that this death was fated?" And you find that starting the clock from injection vs. the beginning of the observation window *must* lower the alive/observed ratio. It is impossible to not reduce it.
Is this just Steve “jumping the gun” again?
I like that he brings these things up and he provides some interesting points of view but many on closer inspection turn out be 🤷
I couldn’t really follow his methodology at a glance, sounded good but when you start taking ratios at different points in time and taking averages etc. it’s a good way to mess things up if your not careful. Not saying he has, but if what you've written is a fair assessment of his method, I can totally see the issue. I can clearly follow your example but find it hard to relate it directly to Steve’s analysis mainly because I can’t follow Steve’s logic fully (could be lack of trying on my part).
I don’t really want to spend the time unraveling his post. I’m sure others like yourself did/will do a better job at it. I hope Norman Fenton or Clare Craig will do a deep dive after being name dropped. Even if they agree with his methodology they would be able to show why his critics are wrong.
I look forward to more skeptical discussion of his findings, not because I disagree with him, I hope he has found a silver bullet but it has to stand up better scrutiny than just “put up or shut up.”
As for relating my model to Kirsch's proof, when you go replicate his filters in the airtable it's clear that the denominator for the unvaccinated is 1/1/21 to the day the reports are filed. So it's a perfect replication of my "no event" model, although the distribution isn't perfectly even, but that's life.
Conversely, there's some weirdness with the values for the vaccinated which might be why he didn't notice this obvious flaw. If you want to find out what the vaccinated "time alive" would be from 1/1/21, you have to independently hash it from 1/1/21 to time of death. But I just lumped all monthly deaths together and used the days from 1/1/21 to middle of the month to virtualize this value. It results in a clear signal that the vaccinated are not dying more "quickly," which would have been a red flag if noticed.
That makes sense. I can see the relationship better now. I’d assumed he accounted for this somehow.
Now I’m verging on being obsessive and I appreciate the tone of your comment, so I don’t want this to sound like I’m piling on. But this entire “study” (and I’m being generous even putting it in quotes) is complete garbage, before you even dissect the finer points of the methodology. I always enjoy Brian’s “begging the question” approach to debunking (which Kirsch is basically demanding as a response), but I feel like it might bury the lede here a bit. Which is this is a self-reported survey, with leading questions, from a biased audience that contains a grand total of 1,634 responses. I’d say it’s statistically insignificant, but that’s still too flattering as it’s an insult to statistics.
It’d be one thing if it was presented as a “well, look at this, kind of interesting, right?” But as some sort of unassailable proof that is so damning that the mainstream media has to ignore it? It’s either extreme hubris or a deliberate con.
And again, respectfully, I don’t think you need to dig into this very deeply to see the inherent problems. It’s why Kirsch poisons the well to begin with by claiming that anyone who raises the obvious objections is somehow being disingenuous. Dude’s a smooth operator. How much of that is his inherent personality or whether he knows what he’s shoveling is anyone’s guess.
I think we may be more in agreement than you think. I was just being more generous to Steve in my assessment. I find his approach interesting but didn’t want to criticize it too much without giving it due consideration.
No doubt and I was being genuine in saying I appreciated the tone of your reply. It’s frankly one that I should adopt more often. I seem to be especially uncharitable this evening and I’m probably being a bit harsh on Mr. Kirsch. Using small sample size like this is just nails on a chalkboard to me though and I’m getting old and cranky when it comes to rhetorical games.
You can also prove this wrong a different way: if you assume a constant probability p of death each day, the resulting distribution of deaths is not uniform, because, with the exception of Damar Hamlin, people generally only die once and then stop. So, deaths will be weighted more strongly towards the start. The odds of dying on the first day is p, but of dying the nth day would be ((1-p)^(n-1))*p. The other problem as you also point out is that death rates are not uniform over the year (and there could be confounding correlation between time of vax and overall death rates)
Welcome Brian,
ICANDECIDE.ORG, TRIALSITENEWS.COM, FIGHTPHARMACORRUPTION.COM, WWW.NCBI.NLM.NIH.GOV, HEALTHYPROTOCOLS.COM, CHILDRENSHEALTHDEFENSE.ORG, WWW.NVIC.ORG, SANEVAX.ORG, MERYL DOREY @ AVN.ORG.AU
Is this Groundhog’s Day? Didn’t we go through this circus act like a year ago? With the same incredibly small sample size with self-reported results submitted by an audience that is definitely not a random distribution? Is this what handwaving looks like?
I’d probably reflexively puke in my mouth less if it wasn’t presented like a “get rich quick” or “miracle cure” scam website. COMPLETE with BOLDED TEXT and APPEALS TO AUTHORITY to demonstrate that IT IS IMPOSSIBLE I am wrong! And if I am, you CAN WIN BEN STEIN’S MONEY!!!
Maybe I’m just cranky, but I’m past the point where these things are just a cute character quirk, and their heart is in the right place, yada yada. If this is what the being anti-vax means, I don’t even care which side wins. You’re just exchanging one group of egotistical sociopaths for another (which is on brand for humanity, give a cursory review of history).
💯
I haven't ever minded Kirsch's stylistic red flags. Mostly I find him to err in the opposite direction, utterly besotted of Fenton's credentials regardless of how many times it lands him a one way ticket to Bias Illusion Burn Town.
So if I were to get the analogy and the setup correct, it appears that the proof was already constructed in such a way that the inclusion of any data point would inherently bias the results in favor of shortened life?
I tried reading through the first portion (Edit: of Steve's post) and to be honest I'm caught off-guard a bit by the early comments which appear more to prop up the supposed data.
The use of his own survey should raise serious criticisms as well. I think I've heard this comment before, but I think there's a saying that "the people who answer surveys usually aren't the people you want answering surveys", so I'd be curious how you control for such biases. The causes of death (and key symptoms) included in the survey, I feel, are also inherently biased in favor of arguing a relationship between a relative's death and the vaccines. I certainly wouldn't list vaccines with (the "cure"). Again, that's already biasing results. I haven't produced my own surveys, but I find that the survey is sprinkled with far too many *wink, wink; nudge, nudge* towards the vaccines.
Right, there's obvious bias problems in the survey. Repliers were supposed to report the "closest" person to them that had died. Obviously that leaves wiggle-room for someone to highlight a case with more recent injection vs. another case with less recent. But this wouldn't get to the heart of the problem of why it's a no-win design.
Boosters exaggerate the problem as well, like I say, but don't get to the heart of it. The way the denominator is designed doesn't overtly penalize boosting. So it would be harder to model the detriment there.
Which is why the proof is based on a simple any event model.
This is why we need more people raising some skepticism for posts. Having that discourse may encourage more stringent surveys and cleaner results. The problem is that most posts may just be met with people already starting with the assumption that what is being presented is inherently infallible, or people already just agreeing with the premise right off the start.
The only you will get Kirsch to admit he’s wrong is to apply Captain Kirk wits to the Kobayashi Maru dilemma- you would need someone to trick him into admitting he’s made a mistake.
On that discussion we had about the message that all vaccines warrant further investigation- Andrew Bostom I think did a good job of cracking open the door with this post about the ineffective flu vaccines.
https://twitter.com/andrewbostom/status/1619038258209619979
We'll see. I mean in this case, he's engineered some type of automatic bet deal so anyone could supposedly use this post to 10X their money if they wanted. I don't have the minimum bid on hand, and I found the nature of what is actually supposed to be betted on hard to parse.
Thanks for highlighting that! My impression of the flu vaccines is that they just put off infection for a year. After that, the vaxxed are in natural immunity debt and start to catch up. In the end it's a wash. It doesn't have anything to do with OAS, or strain mismatches (vaccines still tend to work on not-previously-vaccinated regardless of being a perfect match to seasonal drift, only antigenic shift totally breaks vaccines). It's not some big paradox, it's simple futility.
I went and re-read. It's not sufficient to simply show that the survey proof is wrong. Looks like the whole exercise is a Russian Nested Kobayashi Maru Doll
Hahaha, yeah, well, I’m just Saavik in this storyline, so I wont be much help.
Despite the caps locks, the conditions of the bet seem unwinnable.
First, Fenton is the judge, despite the fact that he has apparently already blessed the survey proof and so what you are debating is whether Fenton is wrong. That's crazy.
Second, there is a further requirement of proving the vaccines save lives either with the survey set or a different resource. Actually, I think someone could pull that off (obviously it would only be showing healthy user bias), but again, the unfair judge problem.
Shared.
Thanks!