I have a question about using _mi_ for multiple imputation for panel data. All the missing values are baseline covariates (time-invariant), so in long form, they would be repeated for each observation in the panel. Let’s say we have times t=0,1,2. For imputation, we only want to use other baseline covariates plus the first (t=0) response, and possibly t=0 values of time-varying covariates.
Using Stata’s _mi_ system, I can imagine two ways to do this:
Using Stata’s _mi_ system, I can imagine two ways to do this:
- Put the data in wide form, and then set up _mi_ to only use other baseline values plus t=0 values in the imputation. The question is then, if, post-imputation, I transform to long form, does _mi_ handle that?
- Alternatively, I could keep the data in long form [which will ultimately provide more flexibility], and then only do imputation for t=0, with an _if_ statement for mi impute. After that, I could use _egen_ to propagate the imputed values across all observations in the panel. This also is of concern because it’s not clear if those other observations represent “passive” imputations to _mi_ or something else.
