Discretisation

[sbfnk]

The current discretisation of delay distributions integrates between integer boundaries [0, 1], [1, 2], ...

Recent work has shown that this leads to bias. A better option would be to discretise as [0, 1], [0, 2], [1, 3], ...

[seabbs]

• Relative to the underlying continuous distribution that may be used as a prior.

[jamesmbaazam]

Can you link some of the recent work here?

[seabbs]

https://github.com/epinowcast/epinowcast/blob/05b4f64be8ecb5f319c4cffb970698e91434c3a5/inst/stan/functions/discretised_logit_hazard.stan#L35 and see the package docs for the maths behind it.

The parts are all put together here: https://github.com/epinowcast/epinowcast/blob/05b4f64be8ecb5f319c4cffb970698e91434c3a5/inst/stan/functions/discretised_logit_hazard.stan#L127

The EpiNow2 version of this that has issues is:

EpiNow2/inst/stan/functions/pmfs.stan

Line 7 in af8c1f3

vector discretised_pmf(real mu, real sigma, int n, int dist) {

[jamesmbaazam]

Thanks, Sam. I'll take a look.