How to handle data with missing continuous outcome?

[mpetukh]

Hi, we have a three-arm experiment with a loss to follow up. Chapter 7 from tlverse book https://tlverse.org/tlverse-handbook/tmle3.html says the following:

 Missing outcomes are efficiently handled by the automatic calculation (and incorporation into estimators) of inverse probability of censoring weights (IPCW); this is also known as IPCW-TMLE and may be thought of as a joint intervention to remove missingness and is analogous to the procedure used with classical inverse probability weighted estimators. These steps are implemented in the process_missing function in tmle3: processed <- process_missing(washb_data, node_list) washb_data <- processed$data node_list <- processed$node_list

I have been using the above until I realized that process_missing just drops the records with missing outcome and does not seem to produce IPCW.

So instead of using process_missing I defined learners for delta_Y in learner_list. This worked for a dichotomous outcome but did not work for a continuous outcome.

Is using delta_Y the correct way to deal with missing dichotomous outcomes? And if yes, what should I do with missing continuous outcomes?

Thank you.

[rachaelvp]

Thanks for filing this issue! We’ll figure this out. Could you please send your R sessionInfo as well as a reproducible example, so we can more easily replicate the error your seeing and fix this?

[imalenica]

Resolved in https://github.com/tlverse/tmle3/tree/fix-ltfu