Thread by @GidMK

Fascinating study demonstrating the issues with selection bias in seroprevalence estimates

Using a selected sample of participants, the estimated prevalence of past COVID-19 infection doubled (!) nature.com/articles/s4146…

The study is really interesting. They used an existing representative sample of people aged >30 to estimate the population prevalence of antibodies to SARS-CoV-2

They then added a second group. These were people who had not previously signed up to the existing cohort, but were eligible

In Group 1, 0.97% of people had antibodies to the virus

In Group 2, this doubled to 1.94%

Even more interesting, this difference did not disappear even when adjusting for age, sex, or reported past symptoms of COVID-19

The only major difference between the groups? Thinking you'd been exposed to COVID-19 in the past

Two take-homes:

1. Selection bias is a big problem
2. Adjusting for demographics and symptoms may not be adequate to correct for this bias

What this means is that if you recruit people to a seroprevalence study in a biased way (say, by telling them that they can go back to normal life if they get a positive result), you might end up with a massively inflated estimate

This is important in IFR calculations

If we used the representative sample, we get an IFR of ~0.8%

Using the biased sample, it's halved to ~0.4%

Big difference!

Thanks @MikeDeeeeeee for pointing out the research