analyse_mi_data() when the exact imputation method is unknown.

[wwojciech]

Hi,

This issue concerns the analysis step:

# Step 4: Analyse each imputed dataset
ana_obj <- analyse_mi_data(data = ADMI, vars = vars, method = method, fun = ancova)

The analyse_mi_data() function requires the method argument, which assumes that the user knows the imputation method that was used to create ADMI. However, this is not always the case.

Issues:

1. How should this parameter be initialized when the imputation method is unknown?
2. My understanding is that method is used to select the appropriate pooling method;
   2.1. Is it used for any other computations as well?
   2.2. If its sole purpose is to determine the pooling method, would it make sense to allow users to specify the pooling method directly? For example:

   pooling_method = get_pooling(method)
   

   where get_pooling() would be a helper function that returns a pooling method corresponding to a given imputation method, while still allowing users to specify the pooling method explicitly when the imputation method is unknown.

Thanks for any clarification or comments.

[bailliem]

Thanks @wwojciech ; will review and get back asap.

[bailliem]

Thanks again for your patience @wwojciech and for the clear analysis. Sorry on my part, work got in the way of responding quickly. I agree with you on both points. method is used for exactly two things in analyse_mi_data(): extracting n_samples to validate the number of imputations in the data, and selecting the pooling approach attached to the analysis object.

My proposed resolution, along the lines you suggested:

Add a pooling argument ("rubin", "bootstrap", "jackknife", "bmlmi") so you can supply either method (unchanged behaviour) or pooling directly when the imputation method is unknown. When only pooling is given, the sample-count check is skipped anber of imputations.
Export a get_pooling(method) helper — essentially your proposal — as the single source of truth for the method → pooling mapping.
For an ADMI dataset of unknown provenance, pooling = "rubin" will be the documented recommendation, since the resampling-based options assume a sample ordering that can't be verified without the method object.
This is also being designed so it can be lifted into {rbmi} core largely as-is, per openpharma/rbmi#589. I'll implement a sketch of this shortly. @tobiasmuetze and @luwidmer FYI