Thread by @gdemaneuf

Here is a thread that looks at the zoonosis evangelists main argument that:

** since zoonosis happens all the time we should just use that hypothesis as the default one - the burden of proof must be on the research-related side **

First let me state that this argument is a fallacy that makes the most of the fact that people don't intuitively have a good grasp of probabilities.

One can explain this this way: Suppose that there are two lotteries in China: a zoonosis lottery and a research-accident lottery.

Let's say that the zoonosis lottery sells 20 times more tickets over China, and also that each ticket has the same chance of winning a top prize (whatever the lottery).

So on average you get 20 zoonosis top prizes for one accident top prize across China.

But Wuhan is also where most of the research accidents happen because that's where most of the research in China is done.

So Wuhan effectively buys most of the accident tickets in China.

At the same time it buys ~ 1% of the zoonosis tickets in China (just one of many cities).

So if you are told that Wuhan won a top prize, which lottery do you think it won?

Simple: roughly 5 times more chance of winning the research-related accident lottery than the zoonosis lottery, since it purchased 5 times more accident-lottery tickets than zoonosis ones.

Main conclusion:

Once you know that Wuhan won the lottery, the probability of this being a research-related accident suddenly goes dramatically up compared to China as a whole: from 1 in 21 (China) to 5 in 6 (Wuhan).

It's a complete switch.

And the morality could be:

"Zoonosis happens all the time in China but research-related accidents happen in Wuhan."

Also, those with Machine Learning experience will recognise the 'happens all the time' argument as a naive majority-class classifier.

That's the kind of binary dummy classifier that does nothing else that saying that everything belongs to the majority class.

It's often right (by the very definition of the majority class) but it is completely useless as it does not even try to predict anything.

It's like a spam filter that would mark ALL your emails as spam because your receive on average more truly spammy emails than valid ones.

For some reason it's alright though when some virologists, epidemiologists and science writers turn into majority-class classifiers.
I am not sure why.

In Machine Learning it's fatally flawed because it does not consider the cost of getting it wrong.

Which by the way is exactly what these scientists and science writers are doing with SARS-CoV-2 origins, as they typically also tell us that the benefits of the research are well worth the risk.

If the cost of getting it wrong is small, then indeed everything can be a zoonosis.

@threadreaderapp compile