MI generates negative values

[Vilma Antonov]

Hi!

I'm trying to impute missing creatinine values in a dataset using

mi set mlong
mi register imputed y1_creat_followup
mi impute regress y1_creat_followup gender crenfail age creat_atsurg y6_creat_followup , add(20) rseed(1234) force

I have missing values in variables y6_creat_followup and creatatsurg, but feel like these need to be in the model since they represent the creatinine values at surgery and creatinine values 6 year post surgery, and I want the creatinine values for 1 year post surgery.

My issue is now that multiple values become negative, as well as a lot of missing values (which I think is due to using independent variables with missing numbers, but still an issue). Any ideas of how to deal with this?

[Nick Cox]

Negative predictions that don't make clinical sense are a consequence of an inappropriate model. The historic solution to maintain positivity is to work on logarithmic scale in fitting and then go back to the original scale. The modern way to do that is to use a generalized linear model with logarithmic link. I can't judge whether that will solve one problem and create others at the same time.

[daniel klein]

You should never use option force with mi. Never.

Why do you not impute all missing values (either with a multivariate normal approach or via chained equations)? If for some reason, you do not want to use imputed values of all variables in the analyses, you can discard the respective observations later.

Negative predictions, i.e., imputed values, are not per se a problem. If you are worried about negative imputed values, you might want to use predictive mean matching.

[Vilma Antonov]

Thanks for the input, I changed strategy as proposed and went with pmm.