Just looking at your picture "Daily New Cases in New York". Hmmm.... What the heck is that 7-day period doing in there? Clearly there is a significant amount of measurement bias, unless there is some "forcing function" having to do with peoples' weekly schedules -- e.g., religious services, weekend activities, etc. There have been suggestions that religious services have contributed to superspreading events. ---- I have a far bigger problem with this type of chart: my mental model has 2 dimensions (probability density as a function of time), rather than simply value as a function of time. My mental model has an entire probability distribution (PDF) for each moment (day) in time, so any particular chart such as the one you linked is merely a random walk through this 2-dimensional chart. This process is analogous to computing the quantum wave function and then "collapsing the wave function" to a particular set of observations. With these fat-tailed distributions, however, the distributions are so wide that they are extremely *shallow* in depth so as to sum to 1. As a result, any particular chart labeled "Daily New Cases in New York" sampled from this 2-D density v. time chart is going to have a vanishingly small probability by itself, so there are a gigantic number of charts that are essentially equally likely. How does nature choose one of these charts? How does anyone choose which one of these possible charts to believe? How do you attempt to fit such a 1-D chart onto this 2-D model, and how could you estimate the parameters, and how would you define "fit" ? At 11:01 AM 5/22/2020, Brad Klee wrote:
Ha ha, tired of "proof by # of Twitter followers", me too! Glenn Greenwald published via The Intercept a nice essay about this "pathology" on May 18.
As for Kermack-McKendrick, I am personally interested in using differential equations to create functions that fit data. So, naturally I would ask, how closely can SIR describe, for example, NYC cases:
This data set has exponential rise, exponential tail, and probably sharp enough peak, so it looks like a SIR fit might work. That could be a challenge for you. If you want to KILL SIR, why don't you just show that it doesn't do a good job of fitting this data set?
I will look at your analysis if you come up with something by the numbers.
Cheers,
Brad
On Fri, May 22, 2020 at 12:07 PM Henry Baker <hbaker1@pipeline.com> wrote:
I was obviously trying to be too cute; I'm tired of being told "you have to believe me because 400 of my closest friends have referenced my papers" aka "proof by # of Twitter followers" when my own lying eyes (or my own algebraic calculations) tell me something quite different. [...] At the risk of mixing metaphors, I'd like to drive Dr. John Snow's wooden pump handle through the heart of the Kermack-Mckendrick 'R0' differential equation models.