Suppose there's 0 Covid cases over the study period. Then this is exactly the result we expect to see. Approximately half the deaths being among the vaccinated, and half the deaths being among the unvaccinated. Because there's going to be heart attacks and so on in the two cohorts over the study window.
Suppose someone offered you a pill. They say with certainty that you can trust that it will prevent you from dying of Covid but increase your likelihood of dying of something else by so much that your overall likelihood of dying actually increases. Is that some pill you'd want to take?
edit: I just finally got your point about a period with zero covid cases. Kind of makes sense, but how could you ever use that to make a claim about a vaccines effectiveness if there is known zero challenge?
~50% of the time, the vaccinated group will have more deaths by chance, even if the vaccine is safe - assuming either little to no community spread during the study, or a small study of young people.
All this result shows is that the vaccine isn't effective at preventing car crashes, which is something we already knew.
There are 4 deaths from cardiac arrest in the vaccine arm, vs. 1 in the placebo; and 1 death from covid vs. 2 in the placebo.
Taken altogether, especially with what we now know about myocarditis/pericarditis risk from the vaccine, non trivial input. And I don't know if you're aware, but the original (dec 2020) approval basically rested on 4:1 placebo:vaccine severe cases. It was approved based on statistically insignificant data, but somehow the same counts of death is not even interesting.
Depending on your model, 21:17 may or may not be statistically significant. I didn't see any "official" modeling of that, especially given that "no deaths were attributed to the vaccine" - if your model says something can't happen, it obviously can't statistically happen .....
There is one "death" (which might as well be a car crash) out of the 29=(15+14) causes listed. All others are medical.
##
>>> import numpy
>>> import random
>>> N_vaccine_deaths = []
>>> for j in range(500):
... x = []
... x = [0 for i in range(46000)] + [1 for i in range(38)] # 38 deaths, ~46000 non-deaths
... random.shuffle(x)
... N_vaccine_deaths += [np.sum(x[0:int(len(x)/2)])]
>>> N_vaccine_deaths = np.array(N_vaccine_deaths)
>>> p = np.sum(N_vaccine_deaths >= 21)/len(N_vaccine_deaths )
>>> p
0.296
>>> x = []
>>> for i in range(5000):
... x += [sum([random.random() < (38/46000) for i in range(int(46000/2))])]
>>> p = np.sum(np.array(x) >= 21)/len(x)
>>> p
0.3446
> And I don't know if you're aware, but the original (dec 2020) approval basically rested on 4:1 placebo:vaccine severe cases. It was approved based on statistically insignificant data, but somehow the same counts of death is not even interesting.
I don't see how this is relevant. That was a long time ago and we didn't have much data, hence low statistical power, and the need to make a decision with an incomplete picture. We now have much more data from over 20 countries, and are able to draw more conclusive and confident conclusions about efficacy.
We do not have efficacy data from 20 countries. E.g. israeli data which I am intimately familiar with is incredibly biased to show vaccine efficiency. Official data touted based on observational data showed 95% efficiency, while at the same time airport data which is the only place with unbiased testing showed about 60%.
(That data is biased by demographics, but not by sampling).
As you said, there’s no statistically significant RCT data showing vaccines are dangerous, but also no RCT data showing they are effective in preventing severe disease. Statistical significance cuts both ways.
Can you expand on this? Why is the data biased by demographics? Is this data biased in the same way as the other countries that show a 90% effectiveness?
As I argued above, there is a bias that's working against vaccines, given that a larger proportion of old people are vaccinated which leads to an understatement of their effectiveness.
> while at the same time airport data which is the only place with unbiased testing showed about 60%
Link?
> [there's] no RCT data showing they are effective in preventing severe disease
Only because they haven't done an RCT that could possibly pick that up, probably because of some ethical argument that giving people a placebo may cause them to behave more recklessly (which I would disagree with).
Well, Israeli MOH rules are basically that unvaccinated people must test at least twice a week to continue normal life, whereas vaccinated people need not test even while showing symptoms, even when being admitted to a hospital ; This had variation over time (e.g., since Jan 2021, approx 8 months only unvaccinated had to test on hospital admission, and 4 months everyone), but that's the gist.
Furthermore, upon exposure to a verified case, the vaccinated can do a home test, and self report that they are negative, and not quarantine - whereas the unvaccinated must quarantine, and test in an offical supervised test, twice in a 7-day period in order to finish quarantine after 7 days, or spend 14 days in quarantine if they do not want to test. (This was shortened to 5 days last week, and also had some variations over time, but same gist).
There are many other biases. It's kind of ridiculous.
I don't know exactly how other countries run these things, but I suspect it's similar in other places, though the details differ.
>> while at the same time airport data which is the only place with unbiased testing showed about 60% > Link?
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3963606
It's biased by economical demographic traits, because the people who actually fly are not a representative sample; But since everyone who lands gets tested, there is no bias such as the ones I described above. (Well ... there is; a vaccinated person tests once on arrival and that's it; an unvaccinated person must test again in 10 days, and if positive in the 10-days after arrival is considered to have been postiive on arrival -- but compared to the above, this is a "small" bias).
> Only because they haven't done an RCT that could possibly pick that up,
Well, yes, but there isn't one. As I said, it cuts both ways - they haven't done a test that's powered enough to show vaccine ACM is worse than placebo ACM. Is that the only reason it doesn't exist? Maybe yes, maybe not. Cuts both ways.
Are you familiar with Fenton & Neil's work?
https://probabilityandlaw.blogspot.com/2021/11/is-vaccine-ef...
https://probabilityandlaw.blogspot.com/2021/12/the-impact-of...
They also have academic level papers, but the blog posts are short and concise.
Well there's no such bias in Australia, AFAIK. Everyone has the same testing requirements.
Let's forget vaccine effectiveness (VE) against hospitalization, because as far as Israel is concerned you may be right.
What about VE against mortality? How could this data be biased? I always focus on VE against mortality, since it's the least prone to bias. (EDIT: I see your link to Fenton & Neil, I will discuss below)
> [Airport testing]
That's interesting, but there are other explanations for this.
For example, natural immunity in a large % of the 'unvaccinated' population by the time that this study was done.
Another explanation could be old people self-selecting into the vaccinated population. That will again make vaccines seem weaker than they are in reality, especially in their ability to prevent infection.
I also note that the RCT which you like showed strong VE against infection (against an old variant).
> Is that the only reason it doesn't exist?
Probably yes? And if not, it's probably some other basic explanation, such as Pfizer not wanting to waste their own money on a trial if the FDA didn't require it. They are for profit, after all. Is there a need for some deeper reason? Occam's Razor.
> Fenton & Neil's work
Interesting.
Maybe this bias existed in that ONS report at a particular time period, I didn't read the full report, it's quite long. Maybe this bias exists in all the data sets being reported by different countries. It's possible.
But what about the VE against mortality in countries that have plateued vaccination rates, where such a bias is going to be minimal? Wouldn't we find evidence of a sharp drop in VE over a rolling 4-8 week window? Can anyone show me this happening to any of the big countries?
If someone can show me that - and the data public in a number of countries, so it should be possible to show - I will probably change my mind.
At least two ways:
(a) Fenton & Neil trick in the blog posts I linked - it applies to any measure. Note that people who died in the first 28 days post original shot are considered unvaccinated, despite all tests showing very significant antibody response in 12 days.
(b) Ignoring all cause mortality. Taken ad absurdum: Inject cyanide. 100% efficiency against corona death. Step back from absurdum: it's possible the vaccine induces "silent covid" by generating the spike but failing the PCR tests -- so that, in fact, as far as mortality goes, it is equivalent to getting sick, but as far as testing goes, it's never attributed to covid.
> That's interesting, but there are other explanations for this.
Yes, there's always other explanations for everything. You reach your conclusions because of your prior, which is "the vaccine is effective, let's see how and why". Which is fine, but you should acknowledge your prior rather then assume it is universal. I have a less informative prior, and the data is much less convincing under that prior.
> They are for profit
And have already paid billions of dollars in fines for past "for profit against health" actions. BMJ editor Peter Doshi has shown -- as soon as they've published -- that equally valid interpretations of Pfizer's own data can show 29% effectiveness. Now, there are two ways to interpret the data - one shows 95% effectiveness against mild disease (and no information about severe disease), and one shows 29% against mild disease (and no information about severe disease). Which one do you expect a for-profit company to champion?
> But what about the VE against mortality in countries that have plateued vaccination rates, where such a bias is going to be minimal? Wouldn't we find evidence of a sharp drop in VE over a rolling 4-8 week window? Can anyone show me this happening to any of the big countries?
It is, in fact, the current narrative, that vaccine efficiency drops significantly after 4 months or so, and that's why a booster shot is needed (Israel has already been administering the 4th shot for a month, Fauci is talking about doing same in the US). It is strictly in the negative territory for mild disease now, still positive (but dropping) for severe disease and mortalitly. Some of it is blamed on immune escaping variants, but the dropping VE is seen everywhere. Don't have the links handy, but that's why they are boosting, officially.
But the RCT showed that vaccine safety isn't a big part of the picture, so all-cause mortality won't be relevant. The difference in non-Covid deaths was statistically insignificant. Yes the mRNA vaccines are more likely to cause myocarditis than Covid itself, but so far I haven't seen evidence that it has translated into a big source of mortality. (It may have, I guess?)
> [rest]
Let's ignore VE against mild disease for now. I just want to nail down VE against death, which is less susceptible to sampling bias, manipulation, and so on.
We should be able to easily answer this question conclusively by finding 1 or 2 countries where vaccine administration has plateued (which is actually most countries now), and simply measuring the VE against death over the last 8 weeks. Has anyone something like that?
My guess is that the VE is still very high, which is fairly conclusive evidence that the vaccines work at preventing death from Covid and that VE against death can't all be explained by the Fenton & Neil trick. And since the vaccines were shown to be safe in the RCT, that is proof that they work at preventing death itself - especially in older populations where the risk of Covid is higher and the risk of myocarditis is lower.
> [vaccine immunity is in the] negative territory for mild disease now
IIRC, this was just a 5000 sample study in one European country, and the result was inconsistent between Delta and Omicron. Or has this finding been replicated? (I don't want to get into the weeds on this question, though, since I want to talk about VE against death).
> vaccine efficiency drops significantly after 4 months or so
Sure, I don't dispute this.
I disagree we should look at Covid death exclusively - that IS subject to attribution bias. I look (and have, in the last two years) only at all cause mortality.
Israel and the US have both experienced increase in all cause death in 2021 that is not attributed to either Covid or vaccines. What is it then? No one seems to know (cdc classifies it under “not sure”). Very significant excess (40% or so) in young ages in Israel, only in 2021, at times that correlate to big vaccine drives with a 1-2 months shift. No cause data yet.
The RCT was not powered enough to show safety or extreme lack of it. The celebrated safety paper (based on biased Israeli data), Balicer, Lifsitz et al is statistical garbage with a predetermined result (lumps 16-60 age group when it was already known 16-24 myocarditis risk is 1:4000 and at 50 assumed to be 1:100000, for example)
My interpretation is that the rushed approval missed some safety issues. The vaccine is likely worth it for >60 or people with comorbidities. For young healthy people …. The data doesn’t say that, and does say “something is weird”
We're now discussing the extent to which vaccines are killing lots of people. Your data in support of this is unexplained excess deaths in the US and Israel in 2021.
I remember seeing something here on HN about Israel's excess deaths implying vaccines caused deaths and thinking it was bogus. Please send me the link again.
As for US excess deaths, why can't this just be Covid? Does every state test every death?
Also, the RCT wasn't underpowered. On what basis do you make that assertion? If the vaccines were killing a large number of people, it should've shown up in the RCT. The only way the RCT could have been underpowered is if the vaccines only kill so few people, that you need a gigantic sample size in order to tease out that small effect. That's what underpowered means. Unless you think the vaccine is only killing people after a long period of time (but that would be a separate issue to being underpowered)?
Not exactly. What I'm saying is that the vaccine surely reduces the number of people whose death is attributed to covid, but I see no reason to believe it reduces any death _at all_, and it might even increase death. E.g., you seem to care if covid death is replaced by a flu death (reducing covid death, but not changing all cause), whereas I don't, regardless of vaccine safety.
> I remember seeing something here on HN about Israel's excess deaths implying vaccines caused deaths and thinking it was bogus. Please send me the link again.
I have no idea which link you are referring to. But data here is well summarised here https://twitter.com/OS51388957/status/1485609513428398081 - note the rise in 2021 in ages 20-39, which is not attributed to covid (and Israel has a crazy amount of testing, about 4% of the population each day at this point, averaging 1.5% all year long in 2021). The rise in 0-19 is only compared to 2020 (which is unusually low - quarantines&lockdowns did not save anyone in that age group from covid death, but they did significantly reduce traffic deaths, which a significant percentage of deaths in that age group). I don't have a handy reference for the per-month breakdown, but there is 30-40% excess in months following vaccine drives, and 0-10% in others.
> As for US excess deaths, why can't this just be Covid? Does every state test every death?
There are huge financial incentives to report covid deaths and hospitalizations, the point many people suspect _over_ reporting of covid deaths. States are not even required to test for covid to consider something covid, it's enough for the case to be "likely covid". What this section mostly includes is sudden deaths, where you didn't have any warning -- if you have warning, you usually have an idea for the cause. So you are claiming it's "sudden covid death", where people die from covid without any signs ahead of time, and which only happens starting mid 2021 or so -- does that make any sense to you?
> Also, the RCT wasn't underpowered. On what basis do you make that assertion?
On the basis of the proof you gave above yourself. 24% more deaths within 6 months of the vaccine, for a group of (essentially) healthy 18-50 year olds. Your calculation showed that if it increases death by 24% in that group, you wouldn't see it.
Let's say the vaccine saves 1% of the people over 60 that get it from covid death, but kills 0.01% of all people that get it at any age with "random" cause. What you would see in the general population is ... exactly what you see. Population all cause mortality within statistical variation, mostly; But significant acm increase in groups that tend not to die, such as 20-39.
I'm intimately familiar with Israeli data, residing here now (and .. Israel has some of the best population level tracking, so it's useful to look at). 2020 had lower all cause mortality per 100K than 2015, despite now vaccines. 2021 has higher all cause mortality per 100K than any year since 2010, with vaccines (pending final revisions, but they almost always go up' not down). Corona attributed death do not explain the difference.
Something is weird, and the main thing that changed is ... the vaccine drive.
IIRC, you have provided no explanation for a high VE against Covid-death post vaccination plateau. The only remaining explanation is that the vaccines are effective at preventing death from Covid.
The point you are making here is a separate one which is that the vaccine is actively killing people, unrelated to Covid, and that this downside probably offsets the upside of VE against Covid-death for people below 60.
> On the basis of the proof you gave above yourself. 24% more deaths within 6 months of the vaccine, for a group of (essentially) healthy 18-50 year olds. Your calculation showed that if it increases death by 24% in that group, you wouldn't see it.
No, if there were more people that were dying in both groups, a 24% difference would be highly statistically significant. Given that so few people were dying in both groups, a 24% difference is not statistically significant.
Imagine if there were 2 car crashes in a group of 30,000 and 1 car crash in another group of 30,000. That's a 100% difference. Massive, right? It should be statistically significant because it's 100%? No, it is indistinguishable from noise. An attempt to read signal into the 2 vs 1 car crashes here is not good statistics, despite a 100% difference in deaths.
The only valid conclusion from the RCT is that the vaccines seem safe over the 6 month window. (Technically speaking, we have failed to reject the null of vaccine safety).
> [US data]
I guess, for this, I will need to really dig into the data in order to give your case a fair go. Let's focus on the Israel data for now.
> [Israel data]
Why cant [1] for 20-29 year old males just be noise?
That bump is similar in magnitude to the 2017-2018 dip. What caused the 2017-2018 dip?
[1] https://twitter.com/OS51388957/status/1485609530612400136/ph...
Israeli data does not show a high VE as such, and in fact there is no vaccination plateau in many places thanks to boosting and stuff. I'm not familiar with the data in every country, if you want to bring a specific case, we can discuss. And no, that is not the only explanation - while death is never disputed, covid attribution often is.
> No, if there were more people that were dying in both groups, a 24% difference would be highly statistically significant. Given that so few people were dying in both groups, a 24% difference is not statistically significant.
I guess we have a misunderstanding. You have shown that if there was a 24% increase in death, it would not appear to be statistically significant. Ergo, this test (at n=44,000, with p_death=~1/1000) is not powerful enough to detect a 24% increase in death. In order to detect 24% increase in death, you would need a much larger n. Therefore, you cannot rule out a 24% increase in death based on the only RCT ever done. Your prior says it is therefore rejected. But again, that's your prior taking precedence, it is NOT the test powerful enough to show safety.
> No, it is indistinguishable from noise. An attempt to read signal into the 2 vs 1 car crashes here is not good statistics, despite a 100% difference in deaths.
Exactly. The RCT shows neither that the intervention is effective or safe. It doesn't show it is ineffective or unsafe. At the rate of 1/30K, it is underpowered, which is what I said.
> The only valid conclusion from the RCT is that the vaccines seem safe over the 6 month window. (Technically speaking, we have failed to reject the null of vaccine safety).
No, that is your prior. The only valid conclusion is that the test is not powerful enough to show 24% increase (or 40% increase) in death, even if it is there. You have a test, that, given n=44,000 and p_death=1/1000, will only show an increase in death if it tops 100% or so.
> I guess, for this, I will need to really dig into the data in order to give your case a fair go. Let's focus on the Israel data for now.
https://data.cdc.gov/d/9dzk-mvmi/visualization ; choose "start date" as dimension and "symptoms, signs .. not otherwise" (one before last) as measure.
> Why cant [1] for 20-29 year old males just be noise? That bump is similar in magnitude to the 2017-2018 dip. What caused the 2017-2018 dip?
Note these are cumulative graphs. the 2017-2018 dip was "repaid" later (some people call it "the dry tinder" hypotheses, a version of "reversion to the mean" - won't go into this), whereas 2021 isn't. It is different. The reason I do not believe it is noise is that I looked at the monthly data on the Israeli central bureau of statistics, and it concentrates in the months following the vaccine drive for that age group. I don't have the latest data in a handy form, but look at these graphs mid year:
https://twitter.com/RanIsraeli/status/1403798079699144706/ph... - title says "Mortality in Israel, ages 20-29, feb-01 to april-29, 2016-2021" (The reason apr-30 was excluded is because of an unrelated disaster that would have bumped 2021 higher by ~40 deaths[1])
https://twitter.com/RanIsraeli/status/1422674509535064070/ph... - this also reflects june data. Israel did vaccination by age bands, with 16-29 mostly feb-apr. (But not in the order you'd expect - a big chunk of 18-22 went first through the army, then rest of 20-29, then rest of 16-19)
> Why cant [1] for 20-29 year old males just be noise?
It could. But ... that's your prior speaking, not the data. If your prior is not "vaccines are safe and effective", but rather "I don't know if they are safe or effective", the data can be read either way, and if your prior is "vaccines are unsafe and ineffective", you can also support it with the data. I try to always use the most uninformative prior.
Moreover, the way he's set up the example exaggerates the effect. He's assuming that 78% of the population gets vaccinated between weeks 6 and 11! Not even Israel is that quick. This is a moot point, anyway, given the above.
Given Fenton & Neil can't explain it, what other explanation is there for a high VE against death, when that VE is calculated using 6-12 months of data?
> Exactly. The RCT shows neither that the intervention is effective or safe.
- The RCT shows that the vaccine is highly effective at preventing infection itself in a previous strain.
- The RCT can't possibly show that the vaccine is effective at preventing Covid mortality (due to lack of Covid circulation over study window).
- The RCT can't possibly show that the vaccine is safe within the bounds of a 20-50% death rate increase.
- The RCT shows that the vaccine isn't so dangerous as to be massively increasing the death rate (within the limitations of N=44,000 power).
> Israel's 20-29 mortality
It could still be noise. You dismiss the 2017-2018 dip because it was "repaid". However there's no mechanism for repayment. The 2017-2018 data is i.i.d. barring some understanding of what's driving it (e.g. knowledge of a war or something). The existence of the repayment is really further evidence that the bump in 2021 could be noise. That dip & repayment both show that this level of random variance is to be expected in this data - we've seen it twice.
BUT! Given that these are largely cardiac deaths, and we see it in 3 different age bands at the same time, and we know the mRNA vaccines cause myocarditis, I think you could be right about the interpretation.
I just want to see it in other countries. If it's really driving mortality in this way, we should be able to find it elsewhere.
Did you check the cdc link? Do you have a theory for the rising curve?
Re:the “repayment” - every single person can be modeled I.I.d, but the event of death of each person is an increasing failure rate function. Someone who is in bad condition will (a) die at a significant challenge, be it flu, Covid, or other reasons; and (b) will only die once. In that sense there’s repayment - a weak flu season gives some people 6-12 until the next bad season, but usually not much more if a bad flu year would have killed them. This is known as “dry tinder” hypothesis in epidemiology. It’s not a law of nature, or a certainty, but has been observed in the past in some cases.
Thank you very much for an enriching, respectful conversation. Really uncommon on this subject these days.
- vaccines are killing people
- undiagnosed covid / long covid is killing people
- something innocuous like the reporting methodology has changed and this category is acting like a residual
What makes me think it probably isn't vaccines is that the "Diseases of Heart" category didn't spike. The working hypothesis is that the mRNA vaccines are killing people primarily via myocarditis and heart attacks. I would've thought that this category would therefore have spiked if that was the case?
I'd like to see how many of these are young men, I think that that would be revealing.
If the car crashes in the treated group are high enough, you have to start asking if there is a mechanism where vaccination side effects are the cause.
This is why you don't drive or operate heavily machinery with certain drugs.