Factor command exploratory factor analysis: missing values

[Cecily Josten]

Hi! I am using the command 'factor' to perform an exploratory factor analysis to develop a scale. Of my 400 total observations, only 310 remain in the factor analysis because Stata drops observations with at least 1 missing value. Instead I would like to use those observations even if the respondents did not answer all questions. Can I impute using a mean value or would that be too imprecise? What can I do instead or what is common practice in this case?

[Maarten Buis]

What factor analysis uses is the correlation between two variables. If you start plugging in means, then you will change those correlations, and thus the result of the factor analysis, in a way that has nothing to do with your observed patterns of association. So that is a very very very bad idea. You could look at multiple imputation, see help mi

[Bruce Weaver]

Hello Cecily Josten. You may find this UCLA page helpful.

I was surprised (and pleased) to see my 2014 article with Hillary Maxwell cited. The 2009 Annual Review chapter by John Graham can be viewed here, by the way.

[Cecily Josten]

Thank you both! I am not so familiar with the multiple imputation command. Does it just impute the missing values? How can I then use the command factor (rather than factorformat)?

[Bruce Weaver]

From the UCLA web page I pointed to:

After running mi impute, the EM covariance matrix can be found in the saved results in r(Sigma_em) which we will then save to the matrix cov_em for use in factormat.

And here are some of the key bits of the code that is shown:

mi set mlong
mi register imputed item13-item28
mi impute mvn item13-item28, emonly
matrix cov_em = r(Sigma_em)
factormat cov_em, n(1396) fact(4) ml
rotate, varimax normalize blanks(.3)

[Cecily Josten]

Hi! When I run those commands I get very different results from running factor with missing variables. Could you maybe explain the difference between the commands? I am not sure it is doing the right thing.

[Bruce Weaver]

Hello Cecily. Are you based at a university or college? If so, I wonder if there is a knowledgeable local expert you could consult. I say that, because it seems there are more issues to sort out than can be done quickly and easily through a few exchanges in a forum like this one. (Other forum members may see things differently, of course!)

[Lydia Palumbo]

Hello,

I am interested in this thread. I followed the approach described earlier but I am not sure how to predict the factor scores for imputed values. Is there any suggestion? Thank you. Lydia