Return frequentist MMRM model object

[gowerc]

Hi @wolbersm ,

We have been asked a few times by Paul to add return the model object of the MMRM fit for the bootstrap methods for diagnostic / residual purposes.

Talking to @nociale our current proposal would be that we return the MMRM fit to the original dataset (only) for the condmean jackknife + bootstrap methods, for all other methods (bayesian, approx bayes, blmi) we would carry on returning nothing.

What do you think ?

[wolbersm]

We need to fit an MMRM model to the original data for all methods, right?
Therefore, I'd suggest to return the original MMRM fit for all methods.
Does this work or would it cause extra complications.

One totally separate issue is that the MMRM to the original data does exclude "non-MAR" post-ICE data but I guess this is unavoidable. Also, it's not 100% clear to me how to do good residual diagnostics for an MMRM model but I leave that to Paul for now.

[gowerc]

@wolbersm ,

Currently we do not fit the MMRM to the original dataset for the bmlmi or approx_bayes methods. Additionally I'm not sure it makes any sense to return it for the bayes method as its only used to initalise the stan sampler.

[wolbersm]

Thanks for clarifying @gowerc
I am not 100% sure what Paul wants to do with the MMRM fit in any case but after some consideration, your original proposal sounds reasonable.

[nociale]

@gowerc @wolbersm Just to be clear, currently drawsObj$fit is NULL for all the methods except bayes for which the stanfit object is returned. We could return the fit object from the glmmTMB() function for the MMRM fit on the original dataset for the condmean method (either bootstrap or jaccknife). drawsObj$fit would still be NULL for bmlmi and approxbayes. Even though we do fit MMRM on the original dataset also when these methods are called, in these cases (as for bayes) this fit serves only for initialization / setting starting values, so I would personally not output this fit object.

[wolbersm]

Thanks, as long as the content of drawsObj$fit is properly documented, this sounds ok for me.

[gowerc]

I'm not sure we would be able to document it other than saying its a "glmfit" object. Otherwise we have a moving target that our documentation is dependent on another teams implementation details.

[wolbersm]

Yes, I think it's fine to just say that this contains the result of the imputation model fit returned by glmmTMB() or stan().

[delmarp]

sounds good. I would also return the full object from the ancova function for the same reason to check residuals

[pengguanya]

Just add here the relevant functions in case I forgot. There is a kind of chain relation among them

1. draws.condmean
2. get_draws_mle
3. get_mmrm_sample
4. fit_mmrm_multiopt
5. fit_mmrm
6. glmmTMB