Thaddeus Michaels: Fixating on bad evidence

This is an article by Thaddeus Michaels (not his real name), who is a data scientist.

As a data scientist, I have a huge amount of respect for anyone who takes a quantitative approach to tackling a problem. Statistics, data analysis – even forecasting and modelling – are there to bring a sense of scale and proportion to a problem, which verbal reasoning often struggles to provide.

But what is becoming apparent in the Covid-19 debate is that quantitative analysts are profoundly unaware of biases in their own reasoning, which clash with the basic anecdotal evidence people are seeing with their own eyes.

When quantitative analysis clashes with anecdotal feedback, this very often means there is something wrong with your quantitative problem formulation – perhaps a false assumption, or a missing variable, or an error in one of the inputs.

It is clear now that various things are happening which contradict the ‘mainstream’ position of epidemiologists. The most obvious is the total lack of evidence of a second wave in places that have largely pulled out of lockdown.

London has seen packed tube trains now for several weeks. Many people aren’t wearing masks. People flout lockdown restrictions for private gatherings, or huddle together in parks for picnics with other households. Anyone with a pair of eyes can see this.

And yet the antibody tests come back and say only 20% of Londoners have been infected. But herd immunity requires 60% to be infected!

The quantitatively minded researcher says: this antibody test is well-designed, well-sampled, and the statistic has been rigorously calculated.

What that same researcher should also be saying is: even if the test has been conducted correctly, is it possible that for whatever reason this is not capturing the thing I am trying to measure.

Is % of population with antibodies a good proxy for (a) % of population who have had the disease, and (b) % of population who are resistant to the disease?

Much of the pro-lockdown analysis proceeds on the basis that it is a very good proxy for both of these things, and that we are sleepwalking into a second wave.

Moreover, the magic 60% figure is calculated on the basis of the famous R0. Measles has an extremely high R0 of ~15, which means over 90% of the population must be immune for herd immunity to be attained.

But R0 is a badly specified measure – it is very widely distributed (many infected individuals infect zero other people, and some superspreaders infect 100+), it changes continually over time, and it is strongly influenced by contextual factors. Given that many of these contextual factors (social distancing, wearing of masks, public perception of danger, weather) are continually changing themselves and are locked into a feedback loop with the spread of the epidemic, it seems completely impossible to ever pin R0 down in the middle of an epidemic.

So we have a herd immunity target, 60%, based on a number we know to be extremely unreliable.

And we have a measure of progress towards that target, the antibody tests, which may not fully measure that progress.

And we have anecdotal evidence (and indeed statistical evidence – deaths, hospitalisations) staring us in the face that contradicts the mainstream analysis of these two data points.

Competent analysts should, instead of explaining this away as mischief-making by anti-lockdown trolls, ask themselves why this might be. We now have several plausible explanations:

1. Many more people have pre-existing immunity (or resistance to severe outcomes) afforded by mechanisms other than SARS-CoV-2 antibodies.

2. R0 is being kept lower than expected by mild non-lockdown measures (e.g. handwashing, self-isolation of symptomatic patients).

3. There could be something wrong with the antibody tests – sampling is extremely difficult to get right; calculating the sensitivity and specificity of a test when no reliable “ground truth” exists is also extremely hard.

Far too much of the debate now descends into people throwing numbers at each other, without considering how flawed those numbers might be.

Update: Hector says: Please keep the submissions coming in, I have a dozen articles I want to write but things are so hectic at the moment with everything that’s going on that there’s plenty of space here for your articles to go up. Also, please help out with donations or a Patreon sub if you can, or at least buy my book, my life has been taken over by Covid (my Twitter page is where most of the recent action has been), and I don’t have any money coming in other than from book sales or donations.

Share this article on social media:

21 thoughts on “Thaddeus Michaels: Fixating on bad evidence

  1. Hector, I bought your book and loved it. Can’t wait for Vol 2! Great blog, keep up the good work, don’t let the turkeys get you down.

    Regards, Nick.
    Victoria, BC, Canada

  2. Those doctors reporting success with Hydroxychloroquine (HCQ) make clear that it works when used soon after symptoms have developed and in conjunction with AZT and zinc. Evidence is widespread that those trials of that cocktail on early symptoms have proved successful. And yet larger, more publicised trials using HCQ on its own and with an antibiotic, do not use zinc and are almost always tested on those already hospitalised with different degrees of serious illness. Indeed, much of the negative ‘conclusion’ around HCQ has resulted from deaths of already seriously ill patients for whom not only will HCQ not work (without zinc) but worse, doses in some case have been so high as to kill patients. These of course have been recorded as deaths following HCQ, ergo HCQ, a drug that has been used for 60 years, is now labelled a killer by those with vested interests in its failure as a cheap, readily available cure and/or are openly hostile politically to those leaders who have promoted the use of the HCQ cocktail. The data reported in The Lancet certainly raises more questions than it answers, not least of all why test HCQ on dying patients without zinc, given HCQ is only a facilitator of the zinc which hopefully deals with the fatal effects of cytokine storm. There are now reports of judicial investigations into these trials as doses of over 800mg administered to dying patients, (many times recommended dosage) arguably constituted manslaughter or even murder.

  3. Good article. Just one minor comment on this:

    “London has seen packed tube trains now for several weeks. Many people aren’t wearing masks. People flout lockdown restrictions for private gatherings, or huddle together in parks for picnics with other households. Anyone with a pair of eyes can see this.”

    Is it possible that the people doing this are mainly the same people, and so they have either been infected early on or are somehow immune?

  4. “Is it possible that the people doing this are mainly the same people, and so they have either been infected early on or are somehow immune?”

    But how could they contrive that, Dene? I can see that the people using Underground trains might be more or less the same people from day to day and week to week – but how could anyone tell what the overlap might be between them and those with immunity?

    Unless of course a hell of a lot of people have immunity…

  5. @Thaddeus
    Good piece, I agree

    A nice corollary to “I Believe and Worship the Holy Prophet Model which tells me what will happen”

    Maybe they’re people who’ve looked at evidence, not msm, and concluded “I’m <30, very low risk of death, I'm not stopping being happy"

  6. @Jonathan Anthony
    Spot on. HCQ works best as a CV19 prophylaxis. Until Trump supported it, all was fine. Since then establishment have done all they can to discredit HCQ even if it results in people dying

    Safe: hundreds of millions, maybe billions have been taking HCQ/CQ daily as malaria prophylaxis for >50 years

    Chloroquine, an antimalarial drug discovered in 1934 and introduced generally in 1947, is probably the
    drug to which humans have been most exposed. With an adult treatment dose of 1.5g for malaria, an
    annual global consumption of hundreds of metric tonnes for over 50 years, and an elimination half-life of
    approximately one month, the average person in many tropical countries once had detectable chloroquine
    in their blood. Chloroquine has a very large apparent volume of distribution because of extensive tissue
    binding and slow elimination (6-8). Plasma concentration profiles with daily dosing are determined mainly
    by distribution rather than elimination. The main metabolite desethyl chloroquine also has significant
    biological activity. Chloroquine is inexpensive and simple to administer. It remains a first-line treatment
    for non-falciparum malaria and is on the World Health Organization’s List of Essential Medicines3.

    Chloroquine has been used extensively as continuous chemoprophylaxis against malaria for individual
    periods often exceeding five years and has been the prophylactic drug of choice in pregnancy (9). It is safe
    in all age groups. In addition to its antimalarial use both chloroquine, and the closely related and slightly
    more hydrophilic hydroxychloroquine
    , are used in continuous daily dosing for rheumatoid arthritis,
    systemic and discoid lupus erythematosus and psoriatic arthritis. Chloroquine at a dose of 2.4mg base/kg
    (155 mg)/day for years is used for rheumatoid arthritis. Chloroquine given at the correct dose has an
    excellent safety profile. It has even been added to salt to prevent malaria by mass exposure

    GPs etc are going to be inundated with arthritis, lupus, malaria patients screaming “You’re killing me”. Malaria areas too scared to take. More deaths OK if Trump gone. They’re cnuts

  7. I think this explanation by a renowned mathematician and scientist offers us a useful warning about the applicability of models in science.

    “As a scientist I make a sharp distinction between models and theories. A theory is a construction, built out of logic and mathematics, that is supposed to describe the actual universe that we live in. A model is a construction that describes a much simpler universe, including some features of the actual universe and neglecting others. Theories and models are both useful tools for understanding nature. They are useful in different ways, and it is important to keep the difference in mind. A theory is useful because it can be tested by comparing its predictions with observations of the real world. On the other hand, a theory may be useless because its consequences are too complicated to be predicted. A model is useful because its behaviour is simple enough to be predicted and understood. On the other hand, a model may be useless because it leaves out too much and loses any connection with reality. As we explore the universe, we move out from well-trodden ground into the unknown. On well-trodden ground we build theories. On the half-explored frontier we build models”.

    – Freeman J. Dyson, “The Sun, the Genome, and the Internet” p. xiv

  8. Pcar and Jonathon; agree 100%. This is scandalous. I die a little everytime I read the phrase ‘the controversial Raoult in France’. This is one of the top men in his profession in the world. In France Veran, who hates Raoult with a vengeance has used this lousy US report to try to stop all the good work Raoult has done to lower the French death rates ever since Macron put him in charge of using HCQ with zinc in French hospitals.
    These people are literally killing thousands of people with their bias. I think its a combination of enormous pressure from US sources because of Trump’s backing, but also the money is to be made from backing new untested drugs which sell for big money compared to HCQ.
    There are a lot of scandals emerging from this virus fiasco, the HCQ one is amongst the biggest.
    Re the article, the only stat that is more or less unsoiled is the excess death number. That is shouting in Europe, its over, its gone. So of course today we have a ‘Hancock half hour’ saying they will use the law and ppolice to enforce isolation if necessary with the new ‘track and trace’. We will never be rid of these new CV Laws, a huge chunk of personal freedom is gone for good.

  9. Good article.
    On the point about what herd immunity might or might not be, I completely agree with the uncertainty. Its worth remembering that children do not fall ill and may not even pass the virus onwards. It appears they have some innate immunity, which means they might not even develop antibodies if exposed to Covid19. What it also means is that they are effectively already part of the immune herd and will therefore lower the threshold of everyone else (to get to the magic 60%).
    Even further, what if innate immunity declines with age ? It stands to reason that a significant portion of teenagers and people in their 20’s may also have the same property. And maybe even some in their 30’s and 40’s. A declining % but still relevant. This may also explain why Africa sees no problem, With over 40% under the age of 15, maybe these countries are at herd immunity already. That’s why there is no epidemic.
    What do you think ?

  10. Words mean things. When human beings are happy to adopt the Political Class language of “herd immunity”, they are willingly getting down on all fours and going “Moo”.

    For Goodness sake, people! You are not cattle — even though your Betters think of you that way. Show some self-respect! Stand up! The expression the politicians & bureaucrats do not want you to use is “population immunity” — as in a population of free men & women. Their resistance & displeasure is the best reason for adopting the proper term “population immunity”.

    Ok., now that we have dumped the demeaning language, let’s talk about the numbers. One of the very few undeniable real-world data points out there is the Diamond Princess, where we know that 80% of those exposed to C-19 were NOT infected. There are similar reports from the US carriers — where the complete population was both exposed and tested. So if it takes 60% population immunity to stop a virus like C-19 from creating a pandemic, why are we not already at that point?

    The fact that modelers do not want to address the Diamond Princess real-world example suggests that their models are not able to match reality, i.e. that their models are missing important factors or are based on inaccurate assumptions.

  11. “Is % of population with antibodies a good proxy for (a) % of population who have had the disease, and (b) % of population who are resistant to the disease?”

    Absolutely; so well said.

    We are so ignorant of so many things; yet we fill in our knowledge holes with bluster, hubris and made-up stuff: trust me, I’m an expert.

    “The fact that modelers do not want to address the Diamond Princess real-world example suggests that their models are not able to match reality,”

    Indeed; so well said

    (=short-hand for: . … we are sick of “experts are warning ….” ….. totally over “experts”

  12. @Gavin L: “For Goodness sake, people! You are not cattle”

    ‘Herd’ is right because the virus concerns man purely as an organism / beast / natural kind not a person / rational being / moral agent categories which make no sense biologically. We don’t use different terms for our bodily parts than for other mammals. The categorical difference between humans and other mammals is moral / philosophical not biological. Zoologically man is a predatory pack animal.

  13. @Jonathan Anthony,

    Deaths in clinical trials are always recorded, and the primary analysis of those is always based solely on number of dead patients. The opinion of the treating physician or anyone else on what caused the death is far less important. There’s no point contaminating a very hard endpoint with very soft data, and death (from any cause) is not only the hardest endpoint you can have in a clinical trial, it’s also one we are always very interested in.

    That said, the recent Lancet paper is a poor excuse for stopping ongoing trials early. Those trials will (should) have their own unblinded safety committees and pre-defined stopping criteria, which won’t involve consideration of a hodgepodge result of retrospective analyses of nonrandomized patients across multiple countries, doubtless with tons of outcome data missing.

    I know it’s very tempting to stick together a bunch of semi-remembered academic facts about what does what in a cell culture, or even physiologically, and say that must be the treatment of choice. Clinical trials are there to find out if it does what it says on the tin. Unfortunately, most of the time, it does not. Also consider that very few viruses are susceptible to any form of treatment other than palliative, and there is no reason to think coronavirus would be different, so the chances are higher than average that all the various quacktails of HCT will fail on meaningful endpoints. But hey, that’s why the trials are being done, with oodles of government money being thrown at them, and the evil big pharma that is supposedly undermining them because they hate the Orange Man preparing to ramp up production.

    Perhaps it would just be better if the armchair medics and Sunday scientists who have overnight become clinical research experts would just shut up and let those of us who do know what we are doing get on with the job of finding out what, if anything, works. After all, how to treat Covid patients should be a data-driven decision, not one determined by political views of either the left or the right.

  14. Just a couple of responses to several people who’ve made broadly similar points.

    Although R0 in, say, Ferguson’s model is quoted as a single number, it’s not implemented as such. The models cover a range of R0 values which is actually quite wide. The quoted R0 is the mean value.

    Infections on the Diamond Princess are said to have reached 20%, which I believe is corret, It’s then claimed that this shows that the infected percentage is limited to 20%. When you look at the data, you’ll find that the infected percentage reached 20% on the day before disembarkation. Up to that day, R0 was ~2.4. Thereafter, since everyone disembarked, R0 among DP passengers dropped to zero (I don’t know whether subsequent infections of passengers were tracked). DP shows only that spread of the disease stops when people are separated, which is not a surprise.

  15. Is the model even important?

    Once we know the population IFR is close to that of influenza, perhaps as terrifying as 3 times as bad as influenza, we know it isn’t something worth destroying civil society for. R0 and other things are then totally irrelevant. All we need to care about is who is at elevated risk of death or complications (and we’ve known that for at least a month) and use the unprecedented productivity of our economy to reduce their risk of infection to ALARP. Isolate at home with the army bringing you food and bog roll once a week, for example, and use the time won to work out more humane ways of reducing infection risk for high-risk people.

  16. @JimW
    Re: Kindergarten Gruppenführer Hancock the Punisher and new ‘track and trace’

    Welcome to North Korea, UK Region

    Personal information collected by NHS Test and Trace to be kept for 20 years
    And there is ‘no absolute right’ for people to delete their personal data after the pandemic has passed

    `The Party seeks power entirely for its own sake. We are not interested in
    the good of others; we are interested solely in power – pure power.´

  17. @Gavin Longmuir
    Agree, well said

    Numbers? 7,800,000,000 in world – CV-19 deaths ~350,000 That’s 35 per 7,800,00 ie Nothing

    CQ/HCQ & DoxyC appear to be a CV19 prophylaxis, used as a treatment they can help recovery not cure – usual contra-indications & dose apply. Stop distorting prevent/help by using cure to discredit

  18. Pcar, the clinical trials will tell us if your claim is true. If so, then wonderful! Until then it is woo quackery “alternative” “medicine”.

  19. A lot of care home patients are likely to be suffering from arthritis or related diseases for which HCQ has been prescribed.

    I wonder if a study would show up a variation between their death rates and those of people who were not receiving the drug?

  20. “Stop distorting prevent/help by using cure to discredit”
    Hear hear!
    When they announced the trial, I’d assumed they’d use it to discredit HCQ. Didn’t take them long. Follow the money.

    A good article, raising questions that would ideally be reverberating around the MSM.

    As has been pointed out, we’ve known for weeks who’s vulnerable. With the highly notable exception of children, they are the same people who are vulnerable in any flu season. They can be shielded (if they so desire) while the rest of us should be allowed to make up our own minds re risk and get on with our lives before everyone is unemployed.

    I am sick of being treated like an idiot. You can now have your grandkids visit, provided you have a garden and they clean the bathroom after the’ve used it. WTF?!

    You’ll have to work out how far down the conspiracy rabbit hole you personally want to go because I’m afraid some of the more alarming theories are rather convincing in the light of the Monty Pythonesque rules, coupled with the obvious lies, deceit and propaganda from our glorious leaders and their MSM partners.

  21. Great post. Antibody tests are usually calibrated against blood samples from people who tested PCR positive, and probably had fairly “productive” infections.

    But we also know that blood donors in Milan ( who, according to the guidance in place there, should not have been donating blood if they had recently had symptoms of a flu-like illness, had a prevalence of a few % around the start of February. So at least some asymptomatic cases do result in antibodies.

    Other cases (whether symptomatic or not) may beat the infection with the “wrong” antibodies (that match different parts of the virus, learned from other coronaviruses). Perhaps 50% of people have this kind of cross-immunity, but antibody tests usually have a higher measured sensitivity than that (around 80% to 90%). You probably make some “spike” antibodies as well (the kind the tests look for) even if you had some cross-immunity. If exposure is dramatically higher than seroprevalence we have to explain why the measured sensitivity of the tests would be so much lower in the real world than in lab samples. Viral load is the most likely candidate– with a high viral load there is a higher chance that you will make “spike” antibodies whatever pre-existing cross-immunity you have and however effective your innate immunity. But this is speculation. What I would like to see is a seroprevalence study done of people in Iceland who tested PCR positive, including symptomatic and asymptomatic cases, but who acquired in the infection in the community with the sort of viral loads that are typical out there.

    Until there is better evidence for this the best assumption is that exposure may be 10% or 20% higher than seroprevalence.

    As for point 2 (handwashing and light measures make a big difference), I think this is likely to be significant. An infected person is infectious for about 4 days. If you managed to stay at home and not infect anyone for just 2 of those 4, you would cut R0 in half, which would take the herd immunity threshold from 60% to 20%. Never mind “self-isolating for 14 days”. Such measures would have made a difference. The only question is whether they did, and this depends on how late they are. The later in an epidemic you do these things the less you need them.

    The elephant in the room is the roughly 50% of asymptomatic cases we know about. Although we read scary headlines about asymptomatic “super-spreaders” like this was a bad thing (and it’s true these people won’t be “self-isolating”) they are also likely to be shedding much less virus and be much less infectious. Several studies have shown that children, usually extremely efficient spreaders of viruses, have trouble infecting anyone with SARS2, and they are generally asymptomatic. It’s reasonable to believe it’s the same for the asymptomatic adults (in general and on average). This would also cut R0 in half.

    The fact that seroprevalence seems to approach 10% (or 20% in more crowded areas) just about everywhere strongly implies that this is close to its final equilibrium level. In an epidemic the number who are immune follows a “reverse sigmoid”– it starts off slowly, then shoots up very rapidly, and then levels off. If you had arrested an epidemic whose true equilibrium was about 60%, with lockdowns, you would find a much bigger range in different places. Some hotspots would be close to 60%, others maybe up at a few %.

Comments are closed.