If the questions are about current romantic relations, and some respondents are not, why would you want to impute values? Also, why someone is or is not in a romantic relationship doesn't seem random, unless your sample has an odd way of choosing how people are matched up. So, I think some more info about your project is necessary to get more thoughtful responses (and more about the covariates you have if you are trying to use some imputation scheme).

Hello Karolina,

In your scenario, the data are missing for a specific reason, though this may well be ignorable. However, since there may be qualitative differences between persons in and persons not in a romantic relationship, it would be wise to include that as a grouping variable in your data set. This could later be used to compare relationships/differences/structures between the two batches.

If your goal was to evaluate relationships among variables only for those who were in a romantic relationship at the time they responded, then you should omit those cases/participants were not in a romantic relationship.

Good luck with your work.

@Daniel Wright, @David Morse, thank you for your responses.

I will try to explain some details - the project is aimed at verifying the structure of a questionnaire that is supposed to consist of three correlated factors - two of them are about peer relationships and one is about romantic relationships. I want to perform a CFA on an adult sample.

The most logical thing would be to include only the participants who are in a romantic relationship and to exclude those who are not - that way I don't have to deal with missing data and think of ways of handling those. I was curious, however, whether it would be a good idea to use the whole sample without deleting anyone and rely on methods like Full information maximum likelihood estimation. Although, as far as I know, FIML requires the data to be MAR or MCAR and I wasn't sure if that's the case here. I hope this clears things up.