Stata's multiple imputation by chained equations vs. Royston's -ice- command: Why are there differences?

[Klaus Steitzel]

Hi

This is a broad question unfortunately without a working example. The answer could simply be pointing me to the right point in the manual or somewhere else because it is unlikely that I am the first to ask this question.

My point of departure is described in these Stata list posts: Here or here. Stata users try to impute missing values in their data using -mi impute chained- and see that Stata sometimes does not impute all missing values because you can't impute missing values with predictors that have missing values themselves. I have an intuition that this is true and I can accept it for my inner peace (but can't claim that I fully get my head around it).

My question is then, why does Royston's -ice- command (e.g. https://doi.org/10.1177/1536867X0900900308) (which presumably also does MI by chained equations) not have the same issue? Why does -ice- always (at least I have never encountered it, again apologies for not providing an example) return completely imputed data sets?

I often perceive the Stata defaults as being well thought out and would like to understand why an older, user-written command seemingly provides a solution to my problem that a newer, native Stata command can't tackle?

Thanks so much for explaining this to me
KS

[Richard Williams]

Some say (or at least said) that ice is sometimes better than Stata's routines. See

https://www.statalist.org/forums/for...ple-imputation

[daniel klein]

Missing imputed values are usually due to incorrect syntax. Variables with missing values go on the left-hand side of the equals sign (if any); only fully observed variables with no missing values go on the right-hand side of the equals sign. That is, in

mi impute chained (uvarmeth) ivars = indepvars

distinguish carefully between ivars and indepvars. The latter must not contain missing values; if they do contain missing values, you will end up with missing imputed values.