Lots of serology happening around the world, so I thought I'd do a thread collating some of the infection-fatality estimates for #COVID19 1/n

2/n INFECTION-fatality rate (IFR) is the rate at which people die when they get the disease

Unlike CASE-fatality rate (CFR), this includes people who are not tested (i.e. asymptomatic cases) and will be lower, generally

3/n Early infection-fatality estimates from China using a sample including 80% mild/asymptomatic cases indicated an IFR of 0.9%

4/n Some early modelling work from late January had a similar finding - given an identification rate of 9%, the IFR was estimate to be 0.5-0.8%

5/n Later estimates that also used data from China came to similar conclusions - an IFR of 0.66% in this paper

6/n Next up, we have evidence from the Diamond Princess (remember that? It was either one month or 1,000,000 years ago)

Based on infections in this closed environment, the estimate came back at 1.3%, but with a WIDE confidence interval

7/n In early March, we get the lowest estimate so far, from Oxford University, using a number of datasets to predict an IFR of 0.1-0.36%

8/n There's also a modelling paper from Imperial College that has a much higher estimate, predicting 1% IFR

9/n A preprint from Italy in early April looked at these previous estimates and argued that neither the very low estimate (0.1%) nor the very high (1.3%) seemed consistent with the data there

10/n And now, serology from New York appears to imply an IFR of 0.9% in the city and closer to 0.5% across the state https://twitter.com/GidMK/status/1253441361547689984?s=20

11/n This is in contrast with two serology studies in California, that implied a much lower IFR (but had significant flaws)

12/n I'll keep adding to this thread, but up until now the estimates range from 0.1% at the lowest to 1.3% at the highest

Almost all estimates are between 0.3-1%, very few argue for lower or higher than this

Virtually all include 0.3-0.5% in their confidence interval

13/n Something worth noting here is that IFR is likely to vary by place, due to demographics and healthcare systems

A range of 0.3-1% depending on place would actually be entirely unsurprising

Since I had a minute, I plugged this all into a random-effects model in Stata

Looks like the point estimate from these studies is a 0.68% fatality rate, 95% CI 0.33-1.03%

15/n If you know of any IFR estimates I missed, let me know!