MI in RCTs with Survival and multilevel outcomes

1

Hi everyone,

I want to use MI to analyze an RCT with a primary survival outcome and several multilevel (and survival) secondary outcomes. However, I’m stuck with the question whether I need one or two imputation models because of the specifics of survival / multilevel MI (e.g., considering the multilevel structure which is unnecessary for the survival outcomes). Hu et al. (2016) write a bit about a joint modelling approach when the outcome of the survival analysis really is termination of follow-up, e.g., because of death. However, my outcome is not correlated with the end of follow-up and a clinical parameter (relapse).

Would a joint MI approach, imputing via one model for both survival and multilevel outcomes, be correct or should I favor performing two separate MIs?

2

Great question and I hope someone can answer. For future studies, discrete time multistate models are more likely to handle the raw data for elegantly without the need for imputation.

3

For anyone interested in the solution: I discussed this with an expert telling me to fit separate models for survival and multilevel outcomes. It should also be considered to fit separate imputations models for each hypothesis (or whether a joint model for all multilevel outcomes would be appropriate (for all survival outcomes separately, too)).