Page 3 of 3

Re: 2LevelImpute

Posted: Mon Oct 06, 2014 10:19 am
by richardparker
Hi - I ran your query past one of my colleagues, Harvey Goldstein, who provided the following comments:

The real issue is whether the MOI has random coefficients. If it does and these have no missing data they can be treated in one of two ways. They can be included as responses in the IM so that the issue of random coefficients in the IM does not arise since any other covariates in the IM do not have random coefficients in the MOI and hence you should not use them in the IM. They can also be included as covariates in the IM and here you should have random coefficients to ensure consistency with the MOI. In your longitudinal data the random coefficients in the MOI are with respect to time - there is no missing data with respect to time, but we could have another variable that predicts the level 1 outcome that is missing for some time points. Then the IM could be a bivariate model including random coefficients for both the original outcome (and possibly the covariate). But the original outcome random coefficients are there because that is the model for the outcome. Alternatively we could have just the covariate as response in the IM where now the predictors are the original outcome (no missing values) and time and whether or not there are random slopes for time depends on whether a model for the covariate as a function of time requires these. In fact in this situation my view is that the bivariate response IM would work better and you could explore whether slopes for the covariate response were needed.

He hopes that helps, rather than adds to the confusion!

Best wishes,