Let’s
review the syntax to produce an EFA on the engineering data using
oblimin rotation. We will use iterated PAF extraction, although any
method would provide similar results. (See Chapter 2.) We selected
oblimin rotation for this data because the factors are known to be
correlated. The syntax to run this analysis is presented below.
proc factor data = engdata nfactors = 2 method = PRINIT priors = SMC
rotate = OBLIMIN;
var EngProb: INTERESTeng: ;
run;
This syntax produces
a series of results—within which we can see the four factor
loading matrices (discussed above) of particular interest to us. These
matrices are presented in
Figure 4.2 Factor loading matrices output.
Notice how the coefficients in the unrotated factor matrix (in the
upper left of the figure) are generally larger in absolute magnitude
and more difficult to interpret than the other matrices. Although
we expected (and see evidence of) two very clear factors, the unrotated
factor loadings do not immediately identify these two separate factors.
All 14 items seem to have similar loadings on factor 1 and it is only
in combination with the loadings on factor 2 where the two factors
separate. If we plot each item in two-dimensional space (as done in
Figure 4.1 Example of unrotated and rotated factor loading plot from engineering data), we clearly
see the separation; however, rotation can help us interpret these
results without plotting.
The rotated pattern
matrix and the rotated structure matrix (in the upper right and lower
right of the figure) support the same two-factor solution proposed
in the plot of the unrotated results: the eight engineering problem-solving
items load on factor 1, and the six interest items load on factor
2. However, notice that the matrices have slight differences in their
factor loadings. The off-factor loadings (e.g., loading of EngProbSolv1
on factor 2) in the rotated structure matrix are higher than those
in the rotated pattern matrix. As we stated previously, this is because
the structure matrix represents the simple correlation between the
item and the factor, but the pattern matrix reflects the unique relationship
between the two, holding the other factor constant. The factors in
this example analysis are correlated, r = 0.37. Thus, holding one
constant removes the shared variance from the analysis and clarifies
the unique factor structure. It is important to interpret both in
order to diagnose how the correlation between factors might be affecting
the results.
The reference structure
matrix (in the lower left of the figure) presents a nearly identical
picture as the rotated pattern matrix. There are slight differences
in the loadings that can be traced to the difference in metric, but
it does not tell us anything new. This matrix is not traditionally
interpreted, and it is not output by other programs. We present the
results below only in order to provide an explanation of the results
that you might see. Since it does not provide any information beyond
what is provided in the pattern matrix, we will not include it in
subsequent examples.
Note: (a) unrotated factor matrix;
(b) pattern matrix; (c) reference structure matrix; and (d) structure
matrix.