Support for clustered bootstrap and jackknife - Currently preparing a PR

[mgondan]

Is your feature request related to a problem? Please describe.
I have a psychotherapy study (very similar to antidepressant_data) in which multiple patients are treated by the same therapist.

Describe the solution you'd like
I would like to add clustered bootstrap and clustered jackknife to obtain better confidence interval coverage. In fact, I did so already, see https://github.com/mgondan/rbmi

Describe alternatives you've considered
none

Additional context
I a currently cleaning up my fork of rbmi. When I am done, would you consider a PR (and help me a bit making it ready for your package, tests, vignette, examples)?

[gowerc]

Hi @mgondan,

Apologies for the delay in replying.

Do you have any accompanying literature that can help justify the methodology ? At least our internal statisticians raised a query over it :

the hierarchical structure in the data is assumed when doing the bootstrap, but this is either ignored by the imputation model, or even if taken into account (e.g. as a covariate) it may not be fully compatible with the bootstrap procedure. I am not quite clear whether this is fully justifiable from a theoretical perspective.

If there isn't any available literature our position would be that we are not prepared to adopt this directly into rbmi; However, in this case we would be willing to help work with you in order to support an extension package. That is we would be willing to help develop / provide the relevant hooks that you might need in order for you to create a package that combines with rbmi to provide the additional method.

Please let me know if this would be of interest.

[mgondan]

Great, thanks! I happily accept your second offer. I guess it is easiest if I provide a minimalistic PR, so that you can see if it interferes with the main functions of your package.

I will also do a bit of research to verify if a cluster bootstrap/jackknife alone is sufficient for good CI coverage with clustered data. But that is a second step that can be dealt with separately.