Rotate factor loadings

`B = rotatefactors(A)`

B = rotatefactors(A,'Method','orthomax','Coeff',gamma)

B = rotatefactors(A,'Method','procrustes','Target',target)

B = rotatefactors(A,'Method','pattern','Target',target)

B = rotatefactors(A,'Method','promax')

[B,T] = rotatefactors(A,...)

`B = rotatefactors(A)`

rotates the *d*-by-*m* loadings
matrix `A`

to maximize the varimax criterion, and
returns the result in `B`

. Rows of `A`

and `B`

correspond
to variables and columns correspond to factors, for example, the (*i*, *j*)th
element of `A`

is the coefficient for the *i* th
variable on the *j* th factor. The matrix `A`

usually
contains principal component coefficients created with `pca`

or `pcacov`

,
or factor loadings estimated with `factoran`

.

`B = rotatefactors(A,'Method','orthomax','Coeff',gamma)`

rotates `A`

to
maximize the orthomax criterion with the coefficient `gamma`

,
i.e., `B`

is the orthogonal rotation of `A`

that
maximizes

sum(D*sum(B.^4,1) - GAMMA*sum(B.^2,1).^2)

The default value of 1 for `gamma`

corresponds
to varimax rotation. Other possibilities include `gamma`

=
0, *m*/2, and *d*(*m *-
1)/(*d *+ *m* - 2), corresponding
to quartimax, equamax, and parsimax. You can also supply `'varimax'`

, `'quartimax'`

, `'equamax'`

,
or `'parsimax'`

for the `'method'`

parameter
and omit the `'Coeff'`

parameter.

If `'Method'`

is `'orthomax'`

, `'varimax'`

, `'quartimax'`

, `'equamax'`

,
or `'parsimax'`

, then additional parameters are

`'Normalize'`

— Flag indicating whether the loadings matrix should be row-normalized for rotation. If`'on'`

(the default), rows of`A`

are normalized prior to rotation to have unit Euclidean norm, and unnormalized after rotation. If`'off'`

, the raw loadings are rotated and returned.`'Reltol'`

— Relative convergence tolerance in the iterative algorithm used to find`T`

. The default is`sqrt(eps)`

.`'Maxit'`

— Iteration limit in the iterative algorithm used to find`T`

. The default is`250`

.

`B = rotatefactors(A,'Method','procrustes','Target',target)`

performs
an oblique procrustes rotation of `A`

to the *d*-by-*m* target
loadings matrix `target`

.

`B = rotatefactors(A,'Method','pattern','Target',target)`

performs
an oblique rotation of the loadings matrix `A`

to
the *d*-by-*m* target pattern
matrix `target`

, and returns the result in `B`

. `target`

defines
the "restricted" elements of `B`

, i.e., elements
of `B`

corresponding to zero elements of `target`

are
constrained to have small magnitude, while elements of `B`

corresponding
to nonzero elements of `target`

are allowed to take
on any magnitude.

If `'Method'`

is `'procrustes'`

or `'pattern'`

,
an additional parameter is `'Type'`

, the type of
rotation. If `'Type'`

is `'orthogonal'`

,
the rotation is orthogonal, and the factors remain uncorrelated.
If `'Type'`

is `'oblique'`

(the
default), the rotation is oblique, and the rotated factors might be
correlated.

When `'Method'`

is `'pattern'`

,
there are restrictions on `target`

. If `A`

has *m* columns,
then for orthogonal rotation, the *j*th column of `target`

must
contain at least *m* - *j* zeros.
For oblique rotation, each column of `target`

must
contain at least *m* - 1 zeros.

`B = rotatefactors(A,'Method','promax')`

rotates
A to maximize the promax criterion, equivalent to an oblique Procrustes
rotation with a target created by an orthomax rotation. Use the four
orthomax parameters to control the orthomax rotation used internally
by promax.

An additional parameter for 'promax' is `'Power'`

,
the exponent for creating promax target matrix. `'Power'`

must
be `1`

or greater. The default is `4`

.

`[B,T] = rotatefactors(A,...)`

returns the
rotation matrix `T`

used to create `B`

,
that is, `B = A*T`

. You can find the correlation
matrix of the rotated factors by using `inv(T'*T)`

.
For orthogonal rotation, this is the identity matrix, while for oblique
rotation, it has unit diagonal elements but nonzero off-diagonal elements.

rng('default') % for reproducibility X = randn(100,10); % Default (normalized varimax) rotation: % first three principal components. LPC = pca(X); [L1,T] = rotatefactors(LPC(:,1:3)); % Equamax rotation: % first three principal components. [L2,T] = rotatefactors(LPC(:,1:3),... 'method','equamax'); % Promax rotation: % first three factors. LFA = factoran(X,3,'Rotate','none'); [L3,T] = rotatefactors(LFA(:,1:3),... 'method','promax',... 'power',2); % Pattern rotation: % first three factors. Tgt = [1 1 1 1 1 0 1 0 1 1; ... 0 0 0 1 1 1 0 0 0 0; ... 1 0 0 1 0 1 1 1 1 0]'; [L4,T] = rotatefactors(LFA(:,1:3),... 'method','pattern',... 'target',Tgt); inv(T'*T) % Correlation matrix of the rotated factors ans = 1.0000 -0.9593 -0.7098 -0.9593 1.0000 0.5938 -0.7098 0.5938 1.0000

[1] Harman, H. H. *Modern Factor
Analysis*. 3rd ed. Chicago: University of Chicago Press,
1976.

[2] Lawley, D. N., and A. E. Maxwell. *Factor
Analysis as a Statistical Method*. 2nd ed. New York: American
Elsevier Publishing, 1971.

`biplot`

| `factoran`

| `pca`

| `pcacov`

| `procrustes`

Was this topic helpful?