Case study on reconstructing infections for COVID and Ebola #545
Conversation
adamkucharski
changed the title
Use rt implementation and add explanation about low counts.
sbfnk
force-pushed
the
main
branch
2 times, most recently
from
7568882 to
23229a8
Compare
|
Thanks for this PR, @adamkucharski! Really nice to see this application, and great to surface this use case of the package model. In terms of where this sits it might be best to host it outside the package and add it to the repository of case studies so as not to keep the package documentation maintainable and not to add unneeded dependencies here: in this case it would we'd have to add With this in mind I'll close the PR and invite you to either put it in its own repository (which would enable full reproducibility with e.g. |
|
That sounds sensible (hadn't appreciated the dependency headaches that vignette-only packages would cause). Will look at moving to separate repo, and updating with the new interoperability functionality coming down the line for epiparameter/EpiNow2. In meantime, would be interested to hear thoughts on the |
Opened an issue in #611 and will propose a fix. For what it's worth in this case the problem can also be fixed by using a zero-centered prior ( |
|
For reference, have moved draft to this branch: https://github.com/epiverse-trace/howto/tree/additional-inference |
This is a draft PR with a vignette illustrating how to reconstruct infections from delayed outcomes for COVID and Ebola. This code is likely to be streamlined further with incoming functionality to
epiparameterandEpiNow2– notes are given in bold in the.Rmd. There is also an issue with current estimation of infections for Ebola, which are estimated to exponential but should be a small outbreak – would value comments/suggestions!Subsequent note: I explored the issue of mis-estimation of infections in more detail. Initially wondered if it might be a period with zero initial cases in the dataset, but omitting these did not fix the problem. So instead compared the $R_t$ model with the
rt=NULLnon-mechanistic version.This is the (erroneous) estimate I got with
rt=NULL:And the plausible version with $R_t$ implementation:
I have updated the vignette to include my understanding of why it might be preferable (estimation using low counts works better with a more constrained model process), but there may well be a better explanation. In any case, probably useful to show users a case study both with large counts (COVID) and small counts (Ebola) to demonstrate functionality – we use similar examples in the {cfr} package to show advantages/limitations of different approaches.