# Multiple imputation and factor analyses

I have conducted a survey to collect my thesis data. The data naturally contains some missing values. I want to use multiple imputation but as I want to do a factor analysis this seems a little problematic afterwards because of my analysis plan (depicted below):

• Impute missing values
• Check for univariate and multi variate outliers and if I can justify to remove them, I would remove them.
• Check factor structures of latent variables with factor analyses and decide whether to keep all questions or to throw out questions that don't measure the factor.
• if all latent variables have a satisfying factor structure, I would calculate sums scores
• if I have calculated sum scores, I would construct interactions terms

Usually I think you would already include interaction terms in your model before you do multiple imputation, but in my case I haven't checked the factor structures yet. I also don't want to do factor analyses on 50 different imputed datasets. Would you recommend me in this case to average the values imputed on the 50 different datasets and as such use a single imputation technique? Or would you have other recommendations?

• I also don't want to do factor analyses on 50 different imputed datasets. Sensible. I don't think multiple imputation (MI) is a wise way to go with factor analysis. FA is mostly exploratory, not inferential; but MI is done by fastidious people to assure realistic confidence intervals for population estimates. A FA study can be content with a good (noise adding) single imputation. – ttnphns Mar 16 '16 at 19:23
• If imputation is at all indicated. A respondent with large percent (say, 25%+) of missing values is dubious, - is he a good candidate for the survey, wouldn't it be better to delete such individuals completely? Or, maybe, it is a sign that the questionnaire is bad, and so no FA can yield sensible results? – ttnphns Mar 16 '16 at 19:23