Discretise with two-day censoring windows

[sbfnk]

This has been shown to reduce bias in the discretisation. More detail in https://package.epinowcast.org/dev/articles/distributions.html

Closes #411

[sbfnk]

Some delay-related tests failing - will revisit after merging #504

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 19b5707 is merged into main:

• ✔️default: 33.8s -> 28.6s [-31.69%, +1.19%]
• ✔️no_delays: 38s -> 37.6s [-19.14%, +17.2%]
• ✔️random_walk: 9.78s -> 9.13s [-17.13%, +3.75%]
• 🚀stationary: 21s -> 17.2s [-32.28%, -4.74%]
• ✔️uncertain: 55.4s -> 55.9s [-24.82%, +26.41%]
  Further explanation regarding interpretation and methodology can be found in the documentation.

[jamesmbaazam]

I've crosschecked this against the epinowcast vignette on discretised delays and it looks good to me.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 3b3b2a7 is merged into main:

• 🚀default: 38.3s -> 31.4s [-30.03%, -6.3%]
• ✔️no_delays: 40.8s -> 41.3s [-28.71%, +30.99%]
• 🚀random_walk: 10.1s -> 9.07s [-19.9%, -0.5%]
• 🚀stationary: 21.7s -> 18.5s [-25.33%, -3.98%]
• ✔️uncertain: 1.06m -> 50.4s [-47.55%, +6.52%]
  Further explanation regarding interpretation and methodology can be found in the documentation.

[sbfnk]

I've crosschecked this against the epinowcast vignette on discretised delays and it looks good to me.

Thanks! Added you as a reviewer - if happy please re-approve.

[seabbs]

Sorry i haven't had a chance to fully review but the stan code looks in line with what I expect. Happy to see this merged once @jamesmbaazam is happy (I think we want a big news note about this at the least though as it will change what people are getting).

I think this is true but also we should double check that the R side discretisation implementation is either the same approximation or the more accurate simulation base approach epinowcast uses.

Note both of these should be replaced soon with and at least R side standalone package that handles this all numerically.

[sbfnk]

I think this is true but also we should double check that the R side discretisation implementation is either the same approximation or the more accurate simulation base approach epinowcast uses.

Yes it should be - we're already testing that in

EpiNow2/tests/testthat/test-dist.R

Line 4 in 93c072d

test_that("distributions are the same in R and stan", {

[jamesmbaazam]

I'm happy for this to be merged.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 93c072d is merged into main:

• ✔️default: 38.4s -> 36.3s [-36.3%, +25%]
• ✔️no_delays: 48.7s -> 39.7s [-65.74%, +28.97%]
• ✔️random_walk: 10.3s -> 10.6s [-21.68%, +26.07%]
• ✔️stationary: 20.6s -> 18.2s [-25.65%, +2.42%]
• ✔️uncertain: 58.9s -> 53.7s [-28.07%, +10.44%]
  Further explanation regarding interpretation and methodology can be found in the documentation.