I stand accused of excess negativity on studies showing that covid19 is already extremely widespread, and therefore, that the infection fatality rate is very low.

The accusation is that I am quick to retweet pessimistic studies, slow to accept optimistic ones.

This, it is being suggested, is because my Dad is sick with it.

I don't think that's right, exactly, but I can see why people think it.

So let me explain where my thinking comes from, and why I do tend to find the super high estimates of infection prevalence (>40% in NYC) and correspondingly low estimates of infection fatality rates (<0.3%) really unconvincing.

It's not because of my Dad who is--great news!--doing splendidly so far, and whose condition might just as easily bias me towards wishful thinking and eagerly grasping every optimistic data point.

It's just I don't really think of this as a "dueling mathematical models" contest

I've been in plenty of those fights. I enjoy them. But in this case, the optimistic model doesn't have to best the pessimistic model in a regression battle. It has to match reality.

Very easily observable reality: the number of dead people in hard-hit regions. Like New York City.

If your estimate of the infection fatality rate requires everyone in New York City to have already been infected, I do not believe your estimate.

If it requires 50% to already have been infected, I'm deeply skeptical but open to the possibility that I'm wrong. But I'm gonna want some proof.

Assuming that there will be no furhter deaths in NYC, and that 400,000 people have left NYC one way or another--students, kids who lost jobs and went home, rich people bugging out to the second home--then a 100% infection rate would put an absolute lower bound on the IFR of 0.13%

These assumptions are patently ludicrous.

Nor do I think that NYC--which is, contra stereotypes, not a pestilent hellhole full of the almost-dead, but a city that is younger than the US average, with lower prevalence of comorbidities like hypertension than the US average--has an IFR that is 2-3x higher than, say, LA.

(Which is also younger than the US average, with a below-average prevalence of hypertension, but not really all that different in terms of important demographic characteristics like age, poverty rate, obesity, hypertension, etc)

Basically, I have a very strong prior, which is simply this: a disease that is not much worse than seasonal flu does not do to hospitals what covid19 did to NYC, and Italy, and Spain, and France, and China, and the UK.

When a study comes in suggesting that this might have a fatality rate that is not much worse than seasonal flu, my prior causes me to look for the reasons that this is incorrect.

Maybe this is confirmation bias. but look: I have a very strong prior that unsupported objects fall towards the ground. If you say your study shows that I can jump from 80 story windows, sans parachute, and not die, I will look for the reasons your study is wrong. i will not jump

So when people freak out about the models, and whether ICL was nonsense and IHME is bad--I was never looking at models. I was looking at hospitals.

Your study may well beat the models. But if it doesn't explain the hospitals, I'm probably not going to buy in.

When I used to give talks to libertarian students I used to remind them not to get so excited about their efficient market models that they end up disproving the existence of extended warranties.

Similarly, if you've got a nice, scary model: don't accidentally prove that Iceland doesn't exist.

But if you've got a cheery serological study: the City of New York with its overflowing hospitals and five-digit death rate have to be within your error bands