John Ioannidis, of Most Published Research Findings Are False fame has again updated his preprint on infection-fatality rate for COVID-19 Kudos to him for updating, let's again look at what's happened 1/npic.twitter.com/nE4t1S6rQi
You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
8/n Let me illustrate with an example This study has been included to calculate IFRs for 4 regions of China - Hubei (not Wuhan), Chonqing, Sichuan, and Guangdongpic.twitter.com/fZFInUGuHU
9/n So the author has taken the percent positive of antibody tests for COVID-19 from each of the regions represented in the study, and used that as a population estimate for the entire region to calculate IFR But is this reasonable?
10/n The study only tested two groups: healthcare workers and people on dialysis Now, Ioannidis excludes any testing on healthcare workers, but dialysis patients are...fine?pic.twitter.com/4tYTh4dwNT
11/n And these high numbers of seropositive estimates led to inferred IFRs for these four places in China of 0.00%!pic.twitter.com/zkMNza7NNo
12/n If nothing else, the numbers here imply that 99.9996% of all infections in Chongqing were asymptomatic (500 official cases, widespread testing, but seropositivity of 3.8% in the study implying 12 million 'true' cases) Is this plausible???pic.twitter.com/qSm99jQ0Kw
13/n There are also some numbers in this revised paper that are wrong This figure should read 44%, not 47%pic.twitter.com/xccqCxh6Om
14/n Moreover, in the example highlighted above, the IFR calculated is for Brooklyn, but this was only true for a tiny subset of 240 patients in this 28,523 patient study. The IFR calculation should've been for the whole of NYC, not just Brooklyn!pic.twitter.com/178IRcCQ3x
15/n There are also some worrying inconsistencies in how Ioannidis has split up studies that sampled multiple places within countries
16/n For example, the ENE-COVID and Brazilian studies, which sampled entire countries by region, are only summed up as a single valuepic.twitter.com/6295tiG91O
17/n On the other hand, several studies that sampled multiple regions (but found MUCH lower IFRs) in other places are split up by area I cannot see any explanation for this in the paperpic.twitter.com/sdMyr7qGCo
18/n On top of this, we've got another problem - collinearity The basic issue is that you shouldn't lump multiple samples of the same group of people together into one study
19/n But now, in the study we have Wuhan (A), Wuhan (B), and Hubei (not Wuhan) It's very poor statistical practice to lump all these estimates together like this
20/n Similarly, we have two estimates from Spain. One is the ENE-COVID study, a rigorous randomized seroprevalence study that is the envy of the world The other is a non-random sample of pregnant women at one place in Barcelona These are given EQUAL WEIGHTS in the analysispic.twitter.com/j2r5CLKGlj
21/n The Spain example is even more of a problem because the ENE-COVID (the rigorous study) implies an IFR in Barcelona of ~1% The survey of pregnant women implies ~.5% Guess which one is used?pic.twitter.com/1JO5aNfrbV
22/n Now, all of this collinearity is particularly troubling for that 0.27% estimate that I mentioned way back at the start of the thread
23/n If we get average the collinear results - where we've included the same study or the same sample multiple times - the median jumps immediately to 0.35% That's quite a bit higher!
24/n But there are more corrections to be made. In several places, the IFR that is in this paper does not match the IFR calculated by the study authors
25/n For example, Geneva. The original authors calculated an IFR of 0.64%, but this is downgraded to 0.45% in the paperpic.twitter.com/lEEG7LPjez
26/n And this is not the only example. Another study tested over three weeks and found seroprevalence of 3.85%, then 8.36%, then 1.46%. Overall 3.53% The 8.36% figure is used, giving 5x more infections than the study itself found, and the lowest IFR possiblepic.twitter.com/ZeC8arsL7P
27/n Taking all this into account, let's look at the IFRs for only those studies using representative population samples that were correctly calculated
28/n Here's the revised table. The lowest IFR is, again, Ioannidis' own study, at 0.18%. Nearly half of the estimates are above 1%, and they range all the way up to 1.63% (!)pic.twitter.com/7xU7DGrq2Q
29/n Somehow, for the third time running, there are innumerable decisions made in the paper that seem to only ever push down the IFR, rather than produce the best estimate
30/n As I've outlined, there are also a number of simple errors that make this very problematic as an estimate of the IFR (or the IFR range) for COVID-19
31/n All that being said, the discussion is now MUCH better, and really engages with some of the things I (and others) discussed in previous threads. Too much to go over here, but well worth a read
32/n Ioannidis has also now included some of the government-conducted studies in the paper, which is good to seepic.twitter.com/VRLXEr8geQ
33/n All in all, some definite improvements, but a lot of things still in the paper that are really hard to reconcile with best practice
34/n The one thing I would point out - this from earlier in the thread is a classic example of moving the goalposts. The influenza comparison was clearly wrong, so now we have another comparison which is bad but slightly less wronghttps://twitter.com/GidMK/status/1283232032085032961?s=20 …
35/n imo much better practice would be to acknowledge that COVID-19 is probably substantially more lethal than influenza, but that quantifying this difference is somewhat challenging
36/n Also, another statement that is incorrect and has remained in each version - that disadvantaged populations/settings are uncommon exceptions in the global landscape This remains simply untruepic.twitter.com/8M8QjQ6ZWv
37/n Also, you can find my personal best estimate in the paper that @LeaMerone and I authored on IFR here. A reasonable guess for most areas seems to be 0.5-0.8%https://www.medrxiv.org/content/10.1101/2020.05.03.20089854v4 …
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.