New preprint: Estimating epidemiological delay distributions for infectious diseases

New preprint with the alternative title “What I wish I had known about delay distribution estimation in January 2020”

Featuring Poppy the puppy as a guest contributor!

Great collab w/ @sangwoopark, Andrei Akhmetzhanov, @kcharniga2, Anne Cori, Nick Davies, @dushoff, @sbfnk, @kgostic, Brian Grenfell, Natalie Linton, Marc Lipsitch, @adrianlison, @OvertonC, and Thomas Ward.

Not interested in my late-night rambling?

Go straight to the paper →

(if anyone has thoughts on a nice non-profit open-access journal that would take something like this please reach out!)

Note: This work is the basis of the new {epidist} package which we are spinning up as another part of the epinowcast tooling. Any help very much wanted! We are currently still in the package setup stage after moving the core code out of the paper analysis repository.

1 Like

This is really good stuff!

In the integrals over unobserved exact event times, how do you implement that the secondary must occur after the primary?

Are the distribution and density functions implicitly defined to be zero in that case? In which case how does this work programatically.

It looks like we ended up dropping zero, though I can’t remember how/when we decided on that. I remember having this discussion long time ago…

Programatically, I don’t think there’s a good way to deal with this actually…

The simplest way to deal with this would be to write something like:

p_time ~ uniform(p_min, p_max)
s_time ~ uniform(s_max, p_time)

but if we had p_min < p_time < s_min < p_max, then we would still be sampling from zero densities (though I think it makes sense to assume s_min = p_min and this probably won’t make a huge difference practically?). I feel like we had this implemented at some point but not sure why/when we took it out, unfortunately.