Imputing missing data for Likert scale variables

Question

I am in the process of imputing data (two variables that are 3 point Likert type responses and items of both variables yield multiple subscales). While these variables are ordinal variables, given that they are summed and/or averaged to create subscales, is it appropriate to simply use the SPSS MVA to impute item scores then round up the values to create original response options (e.g. 0=never 1=sometimes 2=often)? What are the recommended ways of dealing with Likert type missing data?

Answer 1

SPSS MVA procedure performs single imputation (either by regression or EM approach) only in variables treated by the procedure as scale (interval). So, if you consider your Likert scale as ordinal you could employ Ordinal regression to make predictions that you can use as imputation, instead of using MVA.

However, since SPSS 17, Multiple imputation procedure is available. It treats scale, ordinal and nominal variables accordingly, so that you don't have to bother about.

Answer 2

You could have a look at hotdeck imputation. The idea is simple: a missing value is replaced with a random draw from the non-missing values. There seems to be a macro for hotdeck imputation in SPSS. Otherwise, you can have a look at R. The method is e.g. implemented in the VIM package.