Hacker News Reader

I think Ioannidis included that very low number as an example of how ridiculous the statistics can look when using the Diamond Princess as a sample and factoring in the relevant sources of uncertainty.

<<Projecting the Diamond Princess mortality rate onto the age structure of the U.S. population, the death rate among people infected with Covid-19 would be 0.125%. But since this estimate is based on extremely thin data — there were just seven deaths among the 700 infected passengers and crew — the real death rate could stretch from five times lower (0.025%) to five times higher (0.625%). It is also possible that some of the passengers who were infected might die later, and that tourists may have different frequencies of chronic diseases — a risk factor for worse outcomes with SARS-CoV-2 infection — than the general population. Adding these extra sources of uncertainty, reasonable estimates for the case fatality ratio in the general U.S. population vary from 0.05% to 1%.>>

Mainly I think he wants more testing, but people are cherry picking his argument to try to downplay the risk.

His key point from the overall piece is something like "we need more testing, including randomly sampled data. Let's not give up on testing while we ratchet up the countermeasures."

Evidence:

Second graph- <<At a time when everyone needs better information, from disease modelers and governments to people quarantined or just social distancing, we lack reliable evidence on how many people have been infected with SARS-CoV-2 or who continue to become infected. Better information is needed to guide decisions and actions of monumental significance and to monitor their impact.>>

in the middle- <<The most valuable piece of information for answering those questions would be to know the current prevalence of the infection in a random sample of a population and to repeat this exercise at regular time intervals to estimate the incidence of new infections. Sadly, that’s information we don’t have.>>

concluding graph- <<If we decide to jump off the cliff, we need some data to inform us about the rationale of such an action and the chances of landing somewhere safe.>>

---

Coming from a different angle, the latest estimate from Wuhan is 1.4% (.9-2.1)

https://www.nature.com/articles/s41591-020-0822-7?mod=articl...

Well given how widely people have carried his "0.05%" as lower bound, I'd imagine he'd have said something if he didn't really think it was a "reasonable estimate for the case fatality ratio in the general U.S. population", as he said.

I do agree though that more data is critical. Although we have far more than enough evidence now to know that we need to act strongly at least for now.

> we are missing the VAST majority of all cases (due to them being mild) and only seeing the severe ones / deaths

This is entirely possible, and is the conclusion of a different, unrelated paper published today. [0]

> But there's a simple check to this - how many deaths does Italy have per day in a bad flu season vs how many Covid-19 deaths per day are they having now?

This is a good idea, and a reasonable metric, but it's not a good way to compare the CFR, because there are too many confounding factors.

For example, it's possible that the early lockdown in Northern Italy actually increased the number of deaths by expanding the window of exposure to elderly populations (whose younger relatives normally would have been out of the house all day). It's possible that, for some reason, Italians are particularly susceptible to SARS-CoV-2. It's possible that the weather pattern or air quality had some strange impact.

I'm not saying any of these things are true; I'm saying that it's not a great way to compare CFRs or speculate about the number of mild or asymptomatic cases.

0: https://www.reddit.com/r/PrepareInsteadOfPanic/comments/fljz...

> it's possible that the early lockdown in Northern Italy actually increased the number of deaths

So in China the lockdowns stopped the Virus from spreading, but in Italy it increased the spread? If the young are all at home giving it to their grandparents, you think it already spread to almost all young people, at pretty much the same point in time, and then lockdowns started and they gave to elderly who somehow mostly didn't get it before that?

> It's possible that, for some reason, Italians are particularly susceptible to SARS-CoV-2. It's possible that the weather pattern or air quality had some strange impact.

If we are considering these as possibilities than we'd better prepare for the U.S. (or any other country) have genetic predispositions or weather patterns that make it even more deadly here than Italy!

I just think my method shows that barring highly unlikely scenarios, this thing is at least 5x as bad as flu, not a lower bound of 0.5 as bad. So we should act accordingly.

> I just think my method shows that barring highly unlikely scenarios, this thing is at least 5x as bad as flu, not a lower bound of 0.5 as bad. So we should act accordingly.

I don't think your method allows any conclusion of the comparison between SARS-CoV-2 and A(H1N1) or whichever flu is in Italy this year.

And it depends what "act accordingly" means. I think it's unwise to take extreme measures like school closings or shelter-in-place, which we know will have their own adverse impacts, without first retracting the flawed science that arrived at the 3.4% CFR and then seeking to replicate the study which I linked above.

I mean, do you think that study is equally flawed? Do you think that the researchers who submitted it for review today simply didn't think to compare the number of deaths yesterday with the mean deaths from flu?

All the people I've ever seen calling the 3.4% figure "flawed science" only seem to ever attack it for being too high. There's very rarely ever voices from the other direction of calling it too low.

People underestimate how many cases Hubei actually managed to treat. They built two purpose built hospitals where every bed was an ICU and converted stadiums and hotels to centralized quarantine facilities so the less acute cases could be near but not in hospitals but could be rapidly shuttled there if the symptoms got worse. They flew in over 40K+ medical workers from all over China to treat less than 80K cases, 80% of which were "mild" which means 2.5 medical workers per non-mild case.

What we really need is the cohort from China who had hospitalizatable symptoms and were denied any sort of treatment, what the range of outcomes in those cases? Even then, it's not a natural experiment because China is practicing triage, same as Italy, and so we need to compare like to like as much as possible.

If the models are correct at assuming that our hospitals are going to get more than 10X overwhelmed, then this is the real number that matters. What's the rate of death for people who have 0 medical intervention.

The reason why is, if the healthcare system isn't overwhelmed, then the CFR is not an important number because the total number of deaths is small enough that variances in that range are trivia. If the healthcare system is overwhelmed, then any data we get from non-overwhelmed systems, which is all people appear to be using, is utterly bollocks.

If people are suggesting that 70% of people in their country could be infected in a year, then we should be using a CFR closer to 10 - 15% as the true number and that means we could be imagining a world where 7 - 10% of all people are dead by this time next year.

Oh heck yeah, the CFR can be closer if there isn't capacity to help people who can be saved by intervention.

But the specific flaw in the 3.4% number is that the denominator was calculated without any real basis for believing that the number of confirmed cases was close to the number of cases (or that it was similarly close in various regions).

That's the part I want to see revisited, and which is of course the topic of a great deal of discourse over the past 48 hours or so.

It's already been linked elsewhere in the thread, but here's one of the widely-circulated commentaries calling the 3.4% CFR into question (in fact making the case that it's meaningless):

https://www.statnews.com/2020/03/17/a-fiasco-in-the-making-a...

I think you underestimate the length of the flu season and the numbers for COVID-19 are therefore much worse. At least 4 months, but really 5 with one of them more mild so 4.5 months for the flu season. Also we should use the average year since harsh flu years have been newsworthy tragedies in their own right. So 111 per day in all Italy. Lombardia is about 16.3% of the population (wikipedia.it), and has about 50% of the deaths (repubblica.it). So it is roughly 235/16.3 , or 13x worse (edit: number changed because Italian flu average is only 8000 per official source below), so roughly (235/9.65) 24x more than flu once healthcare is overrrun by my best estimate.

However flu does not overwhelm ICU and there are tangential deaths from lack of available healthcare which might double these numbers. Who knows, but it not going to be insignificant at all.

edit: In fact I see Italy counts flu season from October to February inclusive, so I think my estimate is correct.

https://www.epicentro.iss.it/influenza/sorveglianza-mortalit...

I agree, I used 3 months, 1/4, worst year, etc. specifically to show that COVID-19 has to be at least that bad. I agree that it's likely far far worse.

Another scary comparison is to overall mortality from all causes. 475/1734 (based on 633,000 deaths in 2018) is equal to 27% of all other deaths per day in Italy, period. If/when it takes off outside of Lombardia, it could well be 100%+.

Your analysis could go wrong if the two diseases have different rates of spread. Eg to continue your most favorable example that everyone in Italy has covid now, the death rate could be order the deaths so far over total population. The flu takes three months to burn out, had a death rate of 25k / population, assuming everyone got that.

That is true. But I think it shows the impossibility of a fatality rate half that of the flu. At least one small Italian town tested every resident, and had 3% positive rate. And more than 80% of cases are unresolved, so many more deaths will come, even from just the cases they've already found.

I mean honestly because of the geographical density of acute cases (this is relative to total cases, not a population normalized value) I can't help but believe that there is in Italy a regional concentration of some sort genetic susceptibility, especially among the elderly, who are less cosmopolitan.

Spread those over 3 months for flu season, and that's 277 deaths per day.

Fly deaths aren’t evenly spread out like that. There is a definite flu “peak” which then tails off.

It would be safer to assume that your peak is at a minimum of 2-3x the average.

> 0.05% is just ridiculous.

It is absolutely not ridiculous as a lower bound!

Doesn't my analysis of Italy prove lower bound can't possibly be lower than 5x flu death rate, around 0.5%?

No, it does not. Your argument is plausible as an estimator of the CFR, but cannot be used to establish a lower bound.

One reason for that is that you cannot examine just one locality. Some will be high, some will be low -- fluctuations are just the nature of statistics. It is necessary to average the hot spots and the cold spots to estimate accurately.

Basically, this is a really hard problem and the error bars are far larger than most people realize, especially in the media.

Can't you say that at least for Italy as a whole, this is killing 2x as many as the worst flu seasons? Italy has a population around ~60 million, law of large numbers makes it impossible to have error bars THAT big. You'd have to assume that Italians have genetic predispositions to die from COVID-19 far higher than the flu.

> Can't you say that at least for Italy as a whole, this is killing 2x as many as the worst flu seasons?

As I said above, I agree that this argument is plausible. I think your argument is at least as likely to be correct as any other, and probably not too far off the mark for the "true" CFR.

> Italy has a population around ~60 million, law of large numbers makes it impossible to have error bars THAT big

However this is not true. The law of large numbers helps with statistical errors. It does zero, absolutely zero, to control for systematic errors. There are so many factors in play that it is extremely hard to declare anything "impossible" with this pandemic.

Can you drop sources for your numbers? Thanks!