All extracted factors are initially orthogonal (Thompson,
2004), but remain so only as long as the rotation is orthogonal (we
discussed this briefly in the section above). In practice, this results
in interpretation of one matrix (a pattern matrix) when conducting
an orthogonal rotation, and two matrices (a pattern and structure
matrix) when conducting an oblique rotation. To further complicate
interpretation, an unrotated factor matrix is always output, and a
reference structure matrix is output when an oblique rotation is done
—neither of which is traditionally interpreted. The presence
of multiple matrices, whether they are to be interpreted or not, is
a source of much confusion in practice. Let’s
start with a few gentle conceptual definitions of each:
Unrotated
factor matrix coefficients are generally reported by
most statistical software regardless of rotation. They represent the
unrotated factor loadings, and are generally not of interest. In PROC FACTOR
,
they appear toward the top of the output in a table entitled “Factor
Pattern.” Despite the title, this should not be confused with
the pattern matrix that is described below.
Pattern matrix
coefficients are essentially a series of standardized
regression coefficients (betas or βs in the regression world)
expressing the variable as a function of factor loadings. These are
also output for all rotation methods. You can think of these as the
list of ingredients in the recipe (e.g., to make Item 13, add 0.70
of factor 1, 0.13 of factor 2, 0.02 of factor 3, etc. Mix well…
delicious!). Like regression coefficients, they hold all other variables
(factors) in the equation constant when estimating the pattern matrix
coefficients. So, if factors are uncorrelated, pattern and structure
coefficients are identical. As factors become more strongly correlated,
the two types of coefficients will become less alike. Thus, think
of pattern matrix coefficients as “row” statistics,
describing the individual item’s unique relationships to each
factor. In PROC FACTOR
, these appear toward
the bottom of the output in a table entitled “Rotated Factor
Pattern.”
Reference
structure matrix coefficients are a series of semipartial
correlations between the individual variables and the overall factors.
These are output by SAS when conducting an oblique rotation and are
not commonly output in other statistical software. This matrix provides
the same information as the pattern matrix, only in a different metric:
total variance vs units of standard deviation. Although the different
scale can be useful to those who think in terms of correlations, this
matrix is not traditionally analyzed.
Structure
matrix coefficients are simple correlations between an
individual variable and the composite or latent variable (factor).
These are output only when conducting an oblique rotation. In multiple
regression, these would be similar to correlations between an individual
variable and the predicted score derived from the regression equation.
The difference between structure and pattern coefficients is the difference
(again, returning to regression) between simple correlations and semipartial
(unique relationship only) correlations. In PROC FACTOR
,
these appear after the pattern matrix in a table entitled “Factor
Structure.”
Individuals are often
confused by the various matrices that are output as well as the different
matrices output by orthogonal and oblique rotation. As we mentioned
above, an orthogonal rotation will output 1) an unrotated factor matrix
and 2) a pattern matrix, but an oblique rotation will output 1) an
unrotated factor matrix, 2) a pattern matrix, 3) a reference structure
matrix, and 4) a structure matrix.
The pattern matrix is generally interpreted when conducting
an orthogonal rotation, and both the pattern and structure matrix
are interpreted when conducting an oblique rotation. The structure
matrix is provided only during an oblique rotation because it provides
information about the items under the assumption of correlated factors.
If all factors were perfectly uncorrelated with each other, the pattern
and structure coefficients would be identical. In this case, there
is no effect of holding other factors constant when computing the
pattern matrix, and the structure and pattern coefficients would be
the same, just like simple and semipartial correlations in multiple
regression with perfectly uncorrelated predictors. Thus, the structure
matrix is useful only when interpreting an oblique rotation.
Thompson (2004) and others (Gorsuch, 1983;
Nunnally & Bernstein, 1994) have argued that it is essential to
interpret both pattern and structure coefficients in an oblique rotation
in order to correctly and fully interpret the results of an EFA. In
practice, few do. Further, when rotations are oblique and factors
are correlated, they argue it is important to report the intercorrelations
of factors also. We will highlight this process when appropriate.