This is a follow up to my previous post about a paper that had some very surprising result about non-pharmaceutical interventions (NPI) and how strict lockdowns seemed to have less impact than light lockdowns.
The crux of this entire thing is that they feed delta(log(cumulative)) into a linear regression model for the log of the growth rate (g) but there is no linear relationship between delta(log(cumulative)) and g when g is moving around. There is if you use delta(log(daily)) but they didn’t. The result is that it is heavily biased towards NPIs which happen earlier and of course everywhere starts off by trying light NPIs and switches to heavier later. Hence the paper’s surprising results.
What’s new in this post is that I’ve fired up R and recreated the linear regression from the paper. The results are spectacular. I had no idea how biased this was.
The R notebook with several simulated epidemics is over here. The highlights are:
- 2 equally effective NPIs come out with an estimated impact on
gof -0.2063621 and -0.0646044 respectively (with the later one losing out) - An NPI that increases growth by 5% looks better than an NPI that immediately stops the epidemic!
That’s right, with this broken methodology, an NPI that makes things worse beats an NPI that ends the epidemic, simply because it happens earlier.
