I can't see that domain, it's blocked on our corporate network, but those seem like odd numbers to have picked. Lateral flow tests have a 0.3% false positive rate, and a 23% false negative rate. With those numbers, at a 5% infected rate, 7% of all positive test results will be false, rather than 34%.
But that ignores a major factor, which is that testing isn't uniformly distributed. It's inevitably going to be skewed towards people who are more likely to be infected. That means the true number will be less than 7%.
The danger with testing really isn't the false positive rate. It's the false negatives.
> In the lab ... which is not the same as the real world
No, I quoted the real-world value intentionally. The lab value is 0.06%. Depending on when you wrote the page it's understandable that you might not have had access to that yet, but it turns out that LFT's are really, really good at not throwing a false positive.
> Only if you assume there is no cost to false positives.
No, you don't need to assume there is no cost to false positives, only that it is less than the cost of false negatives in aggregate. And that's not hard: false negatives cause infections which compound exponentially, whereas false positives have a linear cost. That alone would make false negatives more costly, but when you account for the probability of each it's really not close.
Stuff like this[1] ought to have some impact on the "real world" results, right?
> false positives have a linear cost.
That's a strong assumption. Assuming constant testing happens in an environment of quarantining of the person who tested positive and their contacts and possibly their contacts (which is the goal of the whole trace & quarantine fever of 2020), it is hard to argue that the costs of false positives do is linear in the number of false positives. They are hard to observe and record because they are mostly in foregone productive activities and negative impacts on mental health, but that does not make them go away.
> Stuff like this[1] ought to have some impact on the "real world" results, right?
The real world results, as measured, showed a 0.3% false positive rate. If your [1] has a measurable effect, it's included in that 0.3%.
> > false positives have a linear cost.
> That's a strong assumption. Assuming constant testing happens in an environment of quarantining of the person who tested positive and their contacts and possibly their contacts (which is the goal of the whole trace & quarantine fever of 2020), it is hard to argue that the costs of false positives do is linear in the number of false positives.
It's not hard at all. It literally follows from the problem setup. If one person has 4 contacts, and each of those contacts has 4 more, and you quarantine all of them, then that's 17 people quarantined from a single positive test. That "17" is a linear constant. There's no feedback mechanism: one positive test leads to 17 quarantines only, and you're done. It's linear. It might have a higher constant factor than you personally think is justified, but that's a very different question.
> The real world results, as measured, showed a 0.3% false positive rate.
Try as I did, I can't find a source for this claim. The best I could find was[1] which seems to indicate a 1.1% false positive rate for people without symptoms (keep in mind my argument is against widespread, unconditional testing). PCR results in the real world seem to vary a lot, but a 3% false positive rate seems appropriate.
> 17 people quarantined from a single positive test. That "17" is a linear constant.
That is a faulty argument. You are assuming total costs of a trace/quarantine regime are linear in the number of people prevented from participating in society and productive activities.
The specificity of the test was recorded as 99.68% - the
overall false positive rate was 0.32%, although this was
lowered to 0.06% in a lab setting.
> PCR results in the real world seem to vary a lot, but a 3% false positive rate seems appropriate.
Not to me. Again, first page of google: https://www.ons.gov.uk/peoplepopulationandcommunity/healthan... puts an _upper_ bound on the PCR false positive rate of 0.08%, and that would require all the positives in that survey to have been false. And that's a real-world result across 208,730 samples.
But given the existence of LFTs and how cheap and immediate they are, I don't think anyone's talking about population-wide PCR testing anyway.
> You are assuming total costs of a trace/quarantine regime are linear in the number of people prevented from participating in society and productive activities.
For the proportion of the population removed at any given time? As close to linear as makes no difference, yes. Is your argument that there's a quadratic factor in network effects? You haven't explained what your cost model actually is. For your argument to work you need those costs to follow an exponential curve, and there's just no mechanism for that to happen.
But that ignores a major factor, which is that testing isn't uniformly distributed. It's inevitably going to be skewed towards people who are more likely to be infected. That means the true number will be less than 7%.
The danger with testing really isn't the false positive rate. It's the false negatives.