In this chapter we examined
the concept of rotation — the purpose of rotation, what actually
rotates, and several methods for rotation. We also examined the rotation
results among each of our three example data sets. When the data was
clear with a strong factor structure (as in examples1 and 2, with
the engineering and SDQ data), almost any rotation will do a good
job of clarifying the factor structure. We argued that oblique rotations
performed slightly better given that the hypothetical factors were
correlated in each of our data sets. However, if the factors were
truly uncorrelated, the orthogonal and oblique rotations should identify
very similar solutions. Finally, although the pattern and structure
coefficients were reported, as this was an oblique rotation, the results
were clear and, thus, the oblique rotation did not overly complicate
the results.
In the third example
(GDS data), the way forward was less clear. We explored a single-factor
model, which might ultimately be the most desirable, given our preference
for parsimony. However, in this model, the communalities were lower
than ideal, and the overall variance accounted for was relatively
low. Guidelines that we previously explored recommended extracting
either three, five, or eight factors, but none of them seemed to make
more sense to us than the single-factor model (you will have to decide
whether the three-factor solution is the best or not). This scale
might need a larger sample, revision, or it might need to be examined
in the context of confirmatory methods in order to determine which
model might be superior. However, we cannot just be guided by empirical
data. The latent variables that we construct have to make sense.
At the end of the day,
EFA is about empirically constructing a narrative that makes sense.
Of all the models, see whether you can come up with one that makes
more sense, rather than simply saying that all the items measure depression.