Distribution discretisation

[seabbs]

This looks like a really nice package. I like the use of function factories to keep things modular (and more generally the passing of submodels is really nice here). Looking forward to doing some more with it.

I was just reading through and noticed the use of this approximation for discretising double censored data.

evofr/evofr/models/renewal_model/model_helpers.py

Line 7 in e883784

def continuous_dist_to_pmf(dist):

Just wanted to flag that I think this can lead to some small approximation error. It is the plus in the figure below I believe where the target is the circle (this is from https://www.medrxiv.org/content/10.1101/2024.01.12.24301247v1.full-text). I imagine this doesn't impact work done with this toolkit but if interested in switching out to a more exact approach there is an R and stan implementation here https://primarycensored.epinowcast.org/articles/why-it-works.html which has both an exact numeric solution and some analytical solutions for common dists. We are in the middle of porting this to Julia and it seems like it would be straightforward to do so for numpyro as well (julia port for interest https://github.com/epiaware/primarycensored.jl (yes it has been slow going)). If interested in the porting very happy to have a chat about it at some point

Slight tangent but I see some cool nowcast stuff happening with hazards etc for the MLR nowcast. Is that written up/evaluated anywhere?