Functionality for imputation of single post-baseline

[tobiasmuetze]

Is your feature request related to a problem? Please describe.
The {rbmi} package requires two post-baseline timepoints for imputing missing value. I've seen study designs where only a single post-baseline visit, i.e., the primary timepoint of interest, is available. My understanding is that reference-based imputation can also be conducted in these settings.

Describe the solution you'd like
Allow reference-based imputation in studies with only a single post-baseline visit.

Describe alternatives you've considered
Technically, one can add the baseline value as a dependent variable and this would solve the problem.

[tobiasmuetze]

@wolbersm I'd be interested in your thoughts on whether this is a relevant setting that should be handled by {rbmi}.

@gowerc Any thoughts on whether it would be technically difficult to implement this?

[wolbersm]

Hi @tobiasmuetze
Thanks! Agree that the problem is relevant but I am not sure whether rbmi is the right solution for the following reasons:

• I think imputation options for this simple setting are quite limited. E.g. unless I am mistaken, all reference-based imputation methods coincide and I guess one could implement them by simply setting the treatment group for subjects with missing outcomes in the active treatment group to control for the sake of imputation and then do a standard MAR imputation which many other packages can do.
• My fear is that this would complicate implementation quite a bit but @gowerc is more qualified for this part.

PS: To make your alternative work "cleanly", I think one would have to constrain baseline means to be the same across treatment group but, unfortunately, I don't think rbmi supports this currently.

[gowerc]

Sorry if a stupid question but are you meaning that there is only 1 observation per subject being used for the outcome of the model (assuming baseline is being used as covariate) ? Assuming this is right I'm honestly now sure off the top of my head how hard this would be for the implementation, I'd need to dig into the code again.

[tobiasmuetze]

are you meaning that there is only 1 observation per subject being used for the outcome of the model (assuming baseline is being used as covariate) ?

Yes, that's the setting of interest.

Assuming this is right I'm honestly now sure off the top of my head how hard this would be for the implementation, I'd need to dig into the code again.

Thanks - that would be interesting to know. I appreciate @wolbersm point that other software could be used, but for some companies' workflows might be good to have this covered by {rbmi}.