Everyone on even-remotely-stats-adjacent Twitter is appalled at this. I'm going to try to explain why in a thread.
1) Fitting models here is hard because the data is super noisy & changes direction every few days.
1) Fitting models here is hard because the data is super noisy & changes direction every few days.
2) If you have really noisy data, you could estimate a model to fit it closely, but that's likely to be overfitting (e.g. here, we don't *actually* think deaths are fluctuating so much). So instead, you have to make assumptions about what the data looks like absent the noise.
But those assumptions are important to get a model that fits well out of sample. So this should be guided based on your epidemiological knowledge, your understanding of how the data was collected, and analyses of model fit.
3) Now, the "cubic fit" part: a) you're assuming a very particular functional form. And b) using polynomials to extrapolate is particularly iffy because the edge behavior of polynomials is notoriously finicky (en.wikipedia.org/wiki/Runge%27s…).
This is a bad way to fit a flexible model.
This is a bad way to fit a flexible model.
4) This *could* be something more complicated like a gam with a cubic spline, but all we have to go on is the phrase "cubic fit", so... who knows? And it may well have similar issues.
5) I think a lot of the frustration here is that this hugely policy-relevant model seems like it might be based on something my third stats class told me not to do, and we don't have the information necessary to evaluate it.
