As with extraction,
there are many choices of rotation method. In PROC FACTOR
,
24 methods are available. Each method uses a slightly different algorithm
to achieve the same broad goal—simplification of the factor
structure. In general, the rotation methods fall into two broad categories:
orthogonal and oblique (referring to the angle maintained between
the X- and Y-axes).
Orthogonal rotations produce factors
that are uncorrelated (i.e., maintain a 90° angle between axes),
and oblique methods allow the factors to correlate (i.e., allow the
X- and Y-axes to assume a different angle than 90°). Traditionally,
researchers have been guided to orthogonal rotation because (the argument
went) uncorrelated factors are more easily interpretable. In addition,
the mathematics is simpler, a fact that made a significant difference
during much of the 20th century when EFA was performed by hand calculations
or on early computers. Orthogonal rotations (which include quartimax,
equamax, and varimax) are generally the default setting in other statistical
software. However, in PROC FACTOR
, there
is no default method of rotation—the researcher must specify
the method in order for it to occur.
Let us review some of
the more common methods of rotation. The first three methods are orthogonal
methods (i.e., produce uncorrelated factors) and the latter two are
oblique methods (i.e., produce correlated factors).
Varimax rotation
seeks to maximize the variance within a factor (within a column of
factor loadings) such that larger loadings are increased and smaller
are minimized. In this way, this method simplifies the factors by
identifying a set of variables that have high loadings on one and
another set that have high loadings on another. However, some variables
might not load strongly on any of the factors.
Quartimax tends
to focus on rows, maximizing the differences between loadings across
factors for a particular variable—increasing high loadings
and minimizing small loadings. This method seeks to simplify interpretation
at the variable level—it wants each variable to load on one
factor, and only one.
Equamax is considered
a compromise between varimax and quartimax, in that it seeks to clarify
loadings in both directions. It attempts to find a solution where
factor loadings clearly distinguish the factors, and the variables
load on one primary factor.
Promax is recommended
by Thompson (2004) as the more desirable oblique rotation choice. This
method starts by actually conducting an orthogonal rotation to clarify
the pattern of loadings. It then slowly starts to allow the factors
to become correlated through a procrustean rotation (which is less
common and not discussed here). The initial method of rotation can
be specified through the PREROTATE
option.
If no option is specified, PROC FACTOR
will
default to using a varimax rotation.
Direct oblimin rotation
is another oblique rotation that can sometimes be problematic but
often gives very similar results to promax.
The two oblique methods, promax and oblimin, have parameters
that allow the researcher to limit how correlated factors can be.
The POWER
option raises the factor loadings
to a specified power, effectively increasing the correlation of the
factors as the power becomes larger. The default power for these methods
is 3, but users can adjust the value to any value greater than or
equal to 1. A low correlation can be specified as POWER=1
,
but a high can be specified as POWER=5
. Even
though researchers can adjust the possible correlation between factors,
they cannot force the factors to be correlated if they are not—in
other words, you can limit how strongly correlated the factors are,
but not the minimum correlation.
Each of the techniques described above
is specified on the ROTATE
option in the FACTOR
procedure.
The different rotation methods summarized above can be requested using
the following keywords.
rotate =
VARIMAX /*Varimax rotation*/
QUARTIMAX /*Quartimax rotation*/
EQUAMAX /*Equamax rotation*/
PROMAX /*Promax rotation*/
OBLIMIN /*Direct Oblimin rotation*/
Please note that the mathematical algorithms for each
rotation discussed above are different, and beyond the scope of this
brief introduction. However, the overall goal of each method is the
same: simplicity and clarity of factor loadings. For details about
how they achieve these goals, see SAS (2015). For a good overview
of the technical details of different versions of varimax rotation,
see Forina, Armanino, Lanteri, and Leardi (1989).