Read on Twitter
I’ve seen lots of folks confused about the IHME uncertainty bands: why does it look like the estimate for tomorrow is more uncertain than the estimate for months from now? Two points:
this is an artifact of the model choice
the uncertainty is not quantifying what you think https://twitter.com/LucyStats/status/1247999935531888641
this is an artifact of the model choice the uncertainty is not quantifying what you think https://twitter.com/LucyStats/status/1247999935531888641
the IHME model is a *curve fitting* exercise - the curve they’ve chosen is Gaussian (normal), this comes with many assumptions! E.g.:
it’s symmetric, meaning the rate of increase is assumed to be the same as the rate of decrease
the biggest uncertainty is at the peak
They are using the observed data (solid line) and not reporting uncertainty since it is known, and predicting the future data (dashed line) with uncertainty. Since they predict we are riiiight at the peak, it looks like we are the most uncertain about tomorrow
If you look at this screenshot I took a few days ago before their predicted peak, you see the maximum uncertainty was a few days into the (then) future because, again, with this particular curve the maximum uncertainty is at the peak
the confusion this generates highlights why the model assumptions play such an important role in how we evaluate the model itself - it is really not enough to look at the bands and try to understand the implied uncertainty, there is a lot more going on under the hood!
