1. A thread on selection bias, COVID testing, and a failure to detect community spread in the US that cost us a month or more in the pandemic response.

In my forthcoming book with @JevinWest, we devote an entire chapter to a class of statistical pitfalls known as selection bias.
2. Selection bias arises when you try to sample individuals to understand what is going on in a population, but the individuals that you sample differ systematically from the population of individuals eligible for your study.
3. One of the White House talking points during February was that there was no background community spread of #COVID19 in the United States.

How did we know this? One might hope that COVID testing would pick up any such cases — or at least detect clusters should they emerge.
4. But in February, the CDC's testing policy required that to even be tested, a suspected US case must have either traveled from Wuhan (later: other areas of China as well) or have been in contact with a positively diagnosed case.

Think about this for a moment.
5. Given these rules for testing, there was *no way* of detecting background community spread.

Only travelers from China (who do not constitute community spread) and contacts of cases (not background community spread) were eligible for testing in the first place.
6. This is an extreme case of selection bias. When your sampling criteria exclude any possible cases of background community spread, there is no chance of observing such spread, even if it is occurring broadly.

That's exactly what happened. It was occurring, and we missed it.
7. I won't venture to speculate about which way Hanlon's razor would slice in this particular case. But given the importance of detecting community spread and the impossibility of doing so under the CDC's testing rules, this was a huge and easily avoidable mistake.
8. Even worse were those who stood up and claimed on the basis of this testing policy that there was no evidence of community spread in the US.

It's true—but it's like saying, based looking at a light source, I have no evidence of UV emissions.

No shit. I can't see in the UV.

Create an account for weekly updates and features such as bookmarks & reading history.