In its final solution, factor analysis creates one new variable for each factor. These variables can be saved using methods namely Anderson-Rubin, Bartlett scores, regression, etc.
I prefer saving variables as Anderson-Rubin scores. Why would one choose other types for saving scores?
The purpose of saving factor scores is to compute it for other category with same factors. For example, you have similar variables for expectations and experience. Doing EFA for the both category would result different variables with different factors. In that situation you have to do EFA for one category and saving factor scores will help you for computing same factors for other category.
This function can be done with 'Compute Variable' in SPSS.
This is how I use the factor scores.
For calculating/estimating factor score two general classes of methods are known. The first class is often considered as "refined” methods producing approximately standardized factor score estimates. The second class of scoring procedures has been referred to as the “inexact” methods in which factor scores are computed by summing the responses of subsets of the factored items.
You can find the answer here:
http://psychology.okstate.edu/faculty/jgrice/factorscores/compute_q.html
Bartlett scores are strongly correlated with their own factors, but the scores between different factors may correlate, the factors are orthogonal but the scores may not be. Anderson-Rubin are strongly correlated with their own factors and uncorrelated with scores from other factors: both the factors and the scores are orthogonal.
See the link below for more detail.
http://pareonline.net/pdf/v14n20.pdf
The original methods of relating factors to variables not in the factor analysis is to compute factor scores. The method of that computation is non-critical as they all correlate highly. All of them have shrinkage going to a new sample and so the differences are not important; Horn has shown that unit weighting the salient variables generalizes better to a new sample than the "exact" methods unless you have a very large N. (Gorsuch, R. L. (2003). Factor analysis. In J. A. Schinka & W. F. Velicer (Vol. Eds.), Handbook of psychology. Vol.2: Research methods in psychology (pp. 143-164). Hoboken, NJ: John Wiley & Sons.)
The other method is that of "extension analysis". This determines the correlation of variables not in the factor analysis with the factors -- NOT with factor estimates. (Gorsuch, R. L. (1997). New procedure for extension analysis in exploratory factor analysis. Educational and Psychological Measurement, 57(5), 725-740. 10.1177/0013164497057005001. It means that all variables, including nominal, can be evaluated with the factors.
The only reason to compute factor scores is if transformed scores are needed, such as computing powers of a factor for polynomial analysis.
While extension analysis can be done with other programs, UniMult (a program I'm developing which is in beta testing; see UniMult.com) has extension analysis (as well as factor scores by regression or unit weighting) built in. Whenever a factor analysis is computed, the correlations of the factors with all non-factored variables are immediately shown and can be immediately tested against other variables. (See UniMult.com)
An article that helped me understanding the topic which I recommend , http://pareonline.net/getvn.asp?v=14&n=20
- Badges
- Science topic