Scale-wise testing for "missing completely at random"?

Hi there! :)

I want to impute my missing data by using maximal likelihood (ML) estimation in SPSS. I've read that concerning questionnaire data it's recommended to impute the missing data scale-wise, since the intercorrelations of the items of each scale lead to more realistic estimations of the missing values within that scale.

Here is my question: Do I need to test for "missing completely at random" (Little's MCAR test) for each scale separately or incorporate the whole questionnaire at once?

I'm happy for any helpful suggestions!


Phineas Packard
Don't bother. Almost certainly (unless the data is missing by design or you randomly deleted values) it will not be MCAR. IMHO It is a silly test and a waste of your time.

I also disagree that imputing scale-wise is a good idea. For me the real choice should be do you want to impute via restricted (Model based imputation) or unrestricted (shove it all in there without consideration of any model). I see benefits and negatives in both though for the most part people using SPSS will use unrestricted. As such I think it makes most sense to feed it as much information as possible.