Sam Abbott (and one Adam Howes) issues and comments

[athowes]

Issue for Sam Abbott issues / comments. Remains to move these out to relevant repos / filter for any that remain useful:

1. Relevant to developers discussion: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011653
2. https://broom.tidymodels.org/: An example of a converter based approach is scoringutils: https://epiforecasts.io/scoringutils/. The idea there is that you would write a as_forecast_sample. method to convert from your model output to the format that scoringutils wants and then be able to do forecast scoring etc.
3. I think expressing delay distributions is probably a fairly common problem both in terms of the methods and in terms of tooling. Flagging some resources here: primarycensoreddist: Tools for expressing double censored delay distributions in R and stan. EpiNow2 has a new interface based on Distributions.jl for working with distributions. There are plans to pull it out as its own tool which needs external feedback and contributions: Possibility to generalise probability distribution interface? epiforecasts/EpiNow2#555
4. As per the developer discussion this is the repo we used to share methods for Rt estimation for the 2022 mpox outbreak: https://github.com/epiforecasts/nowcasting.example with the idea to compare across different methods and provide analysts some choices (though biased to a certain subset of methods). The epinowcast one evolved here: https://package.epinowcast.org/articles/single-timeseries-rt-estimation.html I think the epiverse has some thinking here: https://epiverse-trace.github.io/ (though not sure where). The EpiNow2 version was generalised here: https://github.com/seabbs/cste-forecasting-workshop. This idea reminds me of the epirecipes project https://github.com/epirecipes (lots of different implementations of SIR models but without a formal comparison)
5. For interest: https://cdcgov.github.io/Rt-without-renewal/dev/
6. Epinowcast: Needs guidance: @cmilando what guidance!! I have a feeling you have have run into needing a new feature here so that could be a cool thing to do as an outcome
7. https://link.springer.com/article/10.1007/s13253-018-00348-w I think this paper is a good comparison of methods. The point I think we can learn from here is that it involves getting the developers of each method to implement it for the same dataset. A standard problem that comes up in comparisons is that the person writing the comparison understands one method better than others so does a better job with that. I think we have the opportunity here for "the best version" of each Rt package to be used in some kind of comparison (with input, either writing the code or signing off on it) from the package developers / people best placed to do a good job.
8. Generating new synthetic data partially based on real data. vs. directly from purely synthetic discussions as flagged in the developer conversation: https://github.com/epiforecasts/EpiNow2/blob/main/inst/dev/recover-synthetic/rt.R. Is this like fit to real data and then simulate from posterior? As you mentioned during F2F conversation. Yes kind of but with an additional step where you change the posterior Rt estimate for another simulated Rt trajectory (but everything else is from the posterior)