Vignette about implementing retrieved-dropout methods using rbmi

[nociale]

rbmi can support the implementation of retrieved-dropout methods. However, a vignette describing how these can be implemented is still missing.

The vignette should include:

• A very short introduction about the methodology (refer to stats_specs vignette or main references such as Guizzaro et al, and/or James Bell et al, and our publication about estimands in trial for Parkinson's disease (PD)).
• Short description of the estimand of interest (optional).
• Short description about the data used (simulate_data could be used as in antidepressant_data there is no post-ICE data). E.g. show rate of discontinuation and rate of post-ICE missing data for the data.
• The implementation of few different retrieved-dropout methods: could start from adding to the imputation model a "ICE_indicator*treatment_group" term (as in TV1-MAR in PD paper) to "time_since_ICE*treatment_group" (as in TV2-MAR in PD paper) to "ICE_indicator*visit*treatment_group" (as in MMRM2 in James Bell et al paper).
• Few comments about variance estimation (optional).

Out of scope: full evaluation of the different approaches, the vignette has as only purpose to show how to implement these methods using rbmi.

To be evaluated: whether to add something to stats_specs vignette, as retrieved dropout methods are mentioned only in section 2.2.3.

[nociale]

@wolbersm please find above a proposal for the additional vignette on the implementation of retrieved dropout methods using rbmi. Could you please review the proposal and suggest as needed? Thank you!

[wolbersm]

Hi @nociale

Thanks a lot! This is very much in line with what we discussed previously and looks very good.

Regarding examples:

• Totally agree to use simulate_data for data generation.
• Approaches: I think we should probably start with a basic MAR model (MMRM1/MAR1 in James Bell et al paper) and then extend to "time_since_ICEtreatment_group" (as in TV2-MAR in PD paper) and "ICE_indicatorvisit*treatment_group" (as in MMRM2 in James Bell et al paper). Personally, I'd skip the TV1-MAR model.

It would be good to try the examples out for both Bayesian MI and conditional mean imputation and I hope estimates & SE will indeed be similar. For the actual vignette, we can stick to one method and I'd opt for conditional mean imputation (but mention that other methods would also be valid).

[nociale]

@wolbersm thanks! I agree with you. I will use conditional mean imputation for the vignette but I will try to compare with Bayesian MI "outside" the vignette.

[gowerc]

I am assuming this has been resolved by the linked PR that we merged. Thus I will close. Please re-open @nociale / @wolbersm if there is more that needs to be done