Update (Patreon)
Content
Hey everyone,
I wanted to share a bit about what videos have been brewing and what to expect in the hopefully-not-too-distant future.
The next video out will discuss how to reason about the probabilities surrounding medical tests. At the base, I think it would be helpful to simply visualize what all the relevant jargon means, terms like sensitivity, specificity, PPV, etc. But beyond that, it seems worth clarifying what exactly these tests tell us, what they don't, and to address some of the misconceptions and fallacies around them.
At the heart, it's another video about Bayes' rule, which has been covered on the channel before. But with such a concrete and topical example this gives a good opportunity to talk about this rule not just as a formula, but as a way to think, and to hopefully have it sink in more viscerally than it might for more abstract examples. Also, even for those who are very comfortable with Bayes' rule, there remain a number of fallacies and misconceptions around applying it to medical tests.
Of course, the risk with such a topical example that matters a lot in people's lives is that it's terrifying to say anything concrete or put up any hard numbers. There's so much uncertainty around these numbers, such a wide variety of constantly changing tests looking for different things, and so much nuance in the underlying reasons for false negatives and false positives that plugging in hard numbers runs the risk of being downright irresponsible without appropriate caveats. On the other hand, it would be deeply unsatisfying to completely throw up our arms and say nothing about how the central lessons of the video apply to current tests.
I've talked with a few doctors and an epidemiologist about this topic, and what's felt substantive about these conversations is not so much a reduction in the error bars around, say, the sensitivity of a given RT-PCR test, but instead some added clarity on why things are so inherently uncertain. What I feel most comfortable talking about is how all the relevant uncertainties related to each other, and how certain assumptions get us to certain conclusions, so long as it's crystal clear what's being assumed. Anyway, I'll save that all for the video itself.
--
Before I turned to this I had been working for longer than is reasonable on the third chapter of the "probabilities of probabilities" sequence. As it happens, this also involves Bayes' rule, but in the continuous setting of updating probability density function rather simple yes/no hypotheses (e.g. whether or not you have a given disease). More specifically it aims to explain the beta distribution. I can't put my finger on why exactly that one has proven so difficult to write, especially given that in the end, I hope for it to come off as a relatively simple concept.
--
One other probability-related topic on the burner is to explain why pi shows up in the formula for a Gaussian distribution. There is a standard integral trick for showing why the area under the curve exp(-x^2) is the square root of pi, which is typically what's given as an explanation of pi's presence in a Gaussian. But the aim of this video would be to get at something deeper. The relevant integral trick only works because of certain properties that this function exp(-x^2) has with respect to a kind of rotational symmetry, so a satisfying answer here would involve explaining why we should expect that the function which comes up so much in statistics, i.e. the one which will result from the central limit theorem, happens to have this relevant property related to rotational symmetry.
Feel free to share thoughts you have about any of these topics, I'm all ears!
-Grant