1
$\begingroup$

I have a non-ordinal categorical dependent variable with 3 choice outcomes and 20 ordinal categorical predictors and want to do a multinomial logistic regression. However, I want to reduce the predictors to just a few variables with factor analysis. SPSS gives regression factor scores as variables for each of my cases. Do I replace my original 20 predictors with these 5 regression factor score variables, or do the factors need to be on another format before using them as predictors?

$\endgroup$
1
$\begingroup$

Provided that the factor analysis is itself valid, then you can replace your 20 predictors with the 5 factor score variables. That is, there is no special relationship between multinomial logistic regression and factor analysis that makes the application of factor analysis as a way of pre-processing the data invalid.

In terms of the validity of the factor analysis, you may want to use dimension reduction technique that are designed for categorical data. From memory, SPSS has one or two (HOMALS and Categorical Principal Components Analysis).

The other practical challenge is that it is possible that the best predictor variable may end up with relatively little weight in the factor analysis, so it may be useful to only do the factor analysis with variables that are known to have some predictive relationship with your dependent variable.

$\endgroup$
  • $\begingroup$ Thank you! But given that the factor scores used as predictors are "composite"' how one one properly interpret mlogit reg output? $\endgroup$ – JCV Dec 6 '12 at 2:25
  • $\begingroup$ This is what I was alluding to when I said "Provide that the factor analysis is valid". If all you are doing is factor analyzing a lot of variables without having a specific theoretic rationale for doing so there is no real way of interpreting the coefficients. Typically, when people use factor analysis in this context they either have a good reason to do so based on theory (e.g., a reason to believe that there are some underlying dimensions, as is often the case with survey ratings, as an example), or, they are just trying to predict and do not need to interpret the coefficients. $\endgroup$ – Tim Dec 13 '12 at 23:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.