6/n The author then concludes that the IFR "tended to be much lower than estimates made earlier in the pandemic", which is odd because his own estimates made earlier in the pandemic (in May) were...lowerpic.twitter.com/7awIA4gEoW
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
18/n Then we have blood donors, who again may give an erroneous result. These are people who, DURING A PANDEMIC are happy to go out and about and give blood. It is quite possible that they are MORE likely to have been infected than the general population!pic.twitter.com/0ME5x1br1B
18.5/n There are also a lot of included studies from places in which there is almost certainly an enormous undercount of deaths For example, India, where the official death counts may represent a substantial underestimatehttps://www.bmj.com/content/370/bmj.m2859 …
19/n A very basic, reasonable thing to do would be to conduct a sensitivity analysis excluding these biased estimates, to see what happens when you only use representative population estimates Which we can do
20/n If we take the median of only these somewhat good-quality studies (some of them still aren't great, but at least they're not clearly inappropriate), we get a value of 0.5% Double the estimate of 0.27%pic.twitter.com/tQSICfngT9
21/n I thought at this point I'd briefly look at blood donor studies, because they are an interesting case study The author argues that these should be included because, due to "healthy volunteer bias", at worse any estimate should bias the IFR results upwardspic.twitter.com/zfMeDdKKgn
22/n Well, we can now actually test this theory and see if it is true. Enough studies have been done that we have COVID-19 seroprevalence estimates from BOTH blood donor studies AND representative samples and compare them
23/n For example, in England an ongoing study on blood donors by PHE estimates that 8.5% of the population has developed antibodies to COVID-19 However, the ONS with their massive randomized study puts the figure at 6% insteadpic.twitter.com/ocnYbQoazk
24/n In Denmark, a robust population estimate put the figure at 1.1%, while their blood donor study estimates 1.9% have been infected previouslypic.twitter.com/BT1eqgjtuT
25/n Indeed, in every location where both a non-probabilistic, convenience sample has been taken (not just blood donors) AS WELL AS a well-done population estimate, the convenience sample overestimates the seroprevalence
26/n We have a new paper that we're working on that suggests that using such estimates will usually overstate the true seroprevalence by a factor of about 2x Which means the true IFR would be double the number computed from such studies
27/n There are also still, after many revisions, studies that have been excluded inappropriately from the estimates This study from Italy, for example, which produces an estimate of 7% (!) for IFR in the regionpic.twitter.com/41OGhRzSoN
29/n Similarly, there are numerous country-wide efforts not looked at in any way, such as the large population studies conducted in Italy (150,000 participants) and Portugal (2,300 participants)pic.twitter.com/Ur07B0QRpM
30/n And while there is a very brief discussion of the variation in IFR by region, the main component (age) - as we have demonstrated - was barely addressed, with the author instead focusing on vague speculation about healthcare systemshttps://www.medrxiv.org/content/10.1101/2020.07.23.20160895v6 …
31/n We can actually see how age of those infected impacts IFR quite neatly from some of the studies in this review Qatar (0.01%) and Spain (1.15%) look very different, right?pic.twitter.com/kwAOJ4QzCX
32/n Wrong! In fact, the difference here is entirely explained by age! In Qatar, infections have mostly been limited to the immigrant worker population (<40 years), with this group representing more than 50% of infectionshttps://twitter.com/GidMK/status/1300938689535565824?s=20 …
33/n Since this group is at a very low risk of death from COVID-19, the population IFR is MUCH lower than in Spain, where infections among the elderly have been much more common
34/n All of these errors are a shame, because to a certain extent I agree with the author IFR is NOT a fixed category. In the metaregression linked above in the thread, we demonstrated that ~90% of variation in IFR between regions was probably due to the age of those infected!pic.twitter.com/fj0k5oyW2B
35/n Unfortunately, Prof Ioannidis appears not to have read this study, but if you are interested here is the preprint version to perusehttps://www.medrxiv.org/content/10.1101/2020.07.23.20160895v6 …
36/n Anyway, there are numerous errors remaining in the text that I haven't pointed out, but if you've reached this far in the thread I'm sure you're tired of me telling them to you straight up. Have a really careful look and see if you can find them!
37/n (As a start, there is now a representative population estimate from Wuhan out that implies an IFR SUBSTANTIALLY lower than the ones inferred in this paper from samples including hospitalized patients)
38/n Regardless, the main take-home remains, unfortunately, that this paper is overtly wrong in a number of ways, it does not adhere to even the most basic guidelines for this type of research, and thus the point estimate is probably wrong
39/n Sorry, typo in tweet 37 - should read an IFR SUBSTANTIALLY *higher*, not lower. The SEROPREVALENCE is lower (at ~2%) which implies an IFR of ~1.2%
40/n Oh, on an unrelated sidenote, it's quite funny that the author spends some time arguing that using a median is more appropriate than doing a R-E meta-analysis (as @LeaMerone and I did), so I quickly calculated the median for our study and it is higher at 0.79% for IFR
pic.twitter.com/QTkJKNzMnb
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.