Another day, another brutally flawed study going around flu bro (bra?) twit.
Here's the study: the author's tested residents of Santa Clara Co, CA to see if they had antibodies to the virus. Exactly 50 out of 1,500 did. That's 1.5%. (thread 1/)
Here's the study: the author's tested residents of Santa Clara Co, CA to see if they had antibodies to the virus. Exactly 50 out of 1,500 did. That's 1.5%. (thread 1/)
Alas, the authors decided that 1.5% was not enough. They tested too many Karens and not enough Joses, so based on the presumption that the true prevalence of disease (something they guessed at) is much higher in males and minorities, they raised their estimate of prevalence. 2/
Based on test characteristics provided by the manufacturer, and 67 blood samples at Stanford, the authors believed that the test has almost no false positives, but many false negatives, so they raised their estimate of prevalence even higher. 3/
What the authors do not account for, the biggest confounder of all, is that they tested only 3,300 of 1,928,000 people (just 0.17%, compared to 66% in the Gangelt, Germany study) in the county, and therefore their study was subject to the possibility of massive selection bias. 4/
How did the authors find patients to test? They tested volunteers who showed up from a Facebook ad (!!!). The small # of people who showed up may very well have been disproportionately motivated to show up because they had recently had covid-like symptoms. 5/ 
The authors collected data about prior clinical symptoms, but they do not share them in the publication. Why withhold this? Presumably, this data did not help to support the conclusions of their study. 6/
Also, rather surprisingly, unlike the Gangelt group, these authors did not bother to do rt-PCR testing on their volunteers, to detect patients with an active covid infection that were asymptomatic or presymptomatic. 7/
True population-level estimates of prevalence can only be determined by selecting a significant portion of a population AT RANDOM. The likelihood that this study was ruined due to selection bias is way higher than the likelihood of covid prevalence being >50X positive cases. /fin
for those interested, the pre-print is here:
medrxiv.org/content/10.110…
medrxiv.org/content/10.110…
and apologies for the typo on my first tweet in the thread. it was 50/3,300 to get to 1.5%.
(applications for volunteer editor are now being accepted)
(applications for volunteer editor are now being accepted)
last one. the other nuts thing about "Hey, maybe 4% of the people in the county have had this!" is that in a group of very self-selected, symptomatic people getting diagnostic PCR tests BECAUSE THEIR DOCTOR THINKS THEY MIGHT HAVE COVID, the +ve rate there is barely over 10%
