Problems imputing only item non-response with multiple imputation

[Jaimie Monk]

Problem:
I am having problems imputing only item non-response (but leave unit response as missing) when using multiple imputation with MVN.

I have coded my data such that items I want imputed are coded with soft missing (.) and those I don't want imputed are coded with hard missing (.a). When I run my MI code I end up with a truncated data set. It appears Stata removes all values for those who have any data still coded as missing (.a).

Background:

• The dataset I am working with has several waves so unit non-response is different for different variables. I also have auxiliary variables that have unit-non response.
• I am using Stata 15 and am imputing approx 150 variables with 20,000 observations.
• The analysis I'm doing now is cross-sectional but will be changing to panel later.
• The code works fine when I run it to impute all missing data at once (both item and unit response together).
• I will be doing regression analysis with my dataset(s), which drop the missing data anyway. However, much of the missing data is caused by unit non-response in the auxiliary variables which means a much smaller (and biased) sample.

Code used:

mi set flong
mi register imputed [variable list]
mi impute mvn [variable list], add(20) rseed(4312)

Any help in understanding this would be very appreciated.