A difference I see here from a lot of our work and the nowcasting literature is assuming final reported cases are known without reporting error. This approach allows the use of a binomial observation model. Instinctively I think this approach is less representative than the Poisson/NG sum approach but might be interesting to support as an option (I assume it has different failure modes which could be useful). This difference might be related to some of the work @johannes and co are doing looks at nowcasting obs models.
Some other apparent differences:
- only truncation handling and doesn’t also have support for latent reporting delays.
- uses a priori fixed piecewise constant model for all time evolving parameters vs allowing this to be fit by the model.
- it looks like the main approach is using a hand rolled gibbs approach
- its treating I as an integer (which is what motivates the gibbs approach) vs propagating expectations
- it looks like it is ignoring primary censoring of the delay distributions
So after having read I think the features that aren’t a subset of epinowcast are:
- treating final counts as known and so using a binomial of beta-binomial obs model
- treating I as a discrete parameter.
I think for the latter we want to avoid at least until the Julia framework is in place so we can do mixed NUTs and gibbs using robust tools but we could look at moving up work on approximating this with uncertainty on I that is approx Poisson (which @adrianlison looked at I know and came away not liking).
As I said above I am not sure that treating final counts as known totally makes sense but it is in the roadmap for non-joint delay estimation and so it could be quite easy to support these obs models (I would ideally see a comparison of the trade-offs before we commit to the dev work).
