3 ms·
There are well-established frameworks in statistics and decision theory for handling this kind of situation, some of them in continuous successful use for over
by nerdponx 2y ago
There are well-established frameworks in statistics and decision theory for handling this kind of situation, some of them in continuous successful use for over a century. It's not exactly a solved problem, but we aren't fumbling around in the dark either.
- nradov 2y agoThose well-established frameworks in statistics and decision theory are unfortunately mostly irrelevant in complex medical cases like this. There is no reliable source of truth to index against. We often don't have a reliable scientific basis for taking the result of a particular quantitative test and calculating the probability that a patient has (or doesn't have) a particular condition. Estimates of selectivity and sensitivity for particular tests are often wildly off. And when there are multiple tests involved we often don't know the correlation coefficients. Even the definitions of particular conditions are often somewhat arbitrary and not particularly biologically relevant. Software developers often have this unrealistic fantasy about how things ought to work in medicine and biology. They reality is far messier.
- nerdponx 2y agoFortunately I'm not a software developer! Also yes, you're right about having good estimates for medical tests. Mainly I take issue with your claim that "Running most tests twice won't necessarily give you any useful additional information." As long as it is directionally associated with the thing you're trying to test for (i.e. it has any medical value at all), then running it twice absolutely should give you additional information. "Not necessarily" I guess is true, but again, probability gives us the tools to reason about situations like this even if we can't estimate population parameters reliably.