trends from count regression models

[vpnagraj]

just read about a couple of very relevant, promising new R pkgs with methods we could explore:

• https://github.com/reconhub/trending
• https://github.com/reconhub/trendeval

the interesting thing here is that these tools provide automated model selection and prediction intervals for regression on count outcomes (e.g. Poission, negative binomial).

how does this differ from the time series approach? well, strictly speaking all of the C19FH targets are counts but the methods we're using in SigSci-TS weren't developed for count data:

https://otexts.com/fpp3/counts.html

food for thought as we slide into task 3 ...

[stephenturner]

Neat. I haven't followed much of what reconhub folks have been doing. The count thing - I thought of that too. Used to have issues with things that were zero inflated or acted weird around zero, but with numbers so high, I'm not too worried about any kind of transformation or whatever until we start getting numbers near zero (at that point we'll lay this project down).

Anyway, keep the issue open, I'll look at these sometime.

[vpnagraj]

underway at https://github.com/signaturescience/focustools/tree/count-reg

[vpnagraj]

closing this issue for now. the count-reg branch is still open but we may not drop it since it is looking unlikely that we will include a function in the package to do count regression modeling. if we decide otherwise, will reopen this issue