Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Multivariate Behavioral Research

ISSN: 0027-3171 (Print) 1532-7906 (Online) Journal homepage: https://www.tandfonline.com/loi/hmbr20

Promin: A Method for Oblique Factor Rotation

Urbano Lorenzo-Seva

To cite this article: Urbano Lorenzo-Seva (1999) Promin: A Method for Oblique Factor Rotation,
Multivariate Behavioral Research, 34:3, 347-365, DOI: 10.1207/S15327906MBR3403_3

To link to this article: https://doi.org/10.1207/S15327906MBR3403_3

Published online: 10 Jun 2010.

Submit your article to this journal

Article views: 515

View related articles

Citing articles: 42 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=hmbr20
Multivariate Behavioral Research, 34 (3), 347-365
Copyright © 1999, Lawrence Erlbaum Associates, Inc.

Promin: A Method for Oblique Factor Rotation


Urbano Lorenzo-Seva
Universitat Rovira i Virgili

A usual practice in factor analysis is the oblique rotation of the retained factors. For this
purpose, different rotation procedures have been proposed. The present article discusses
Promin, an alternative to Promaj (Trendafilov, 1994). It is easier to use than Simplimax
(Kiers, 1994) and the rotated solution obtained by Promin is very similar to the one
obtained by Simplimax. Promin also seems to perform better than other well-known
procedures.

After factors have been retained by factor analysis, they need to be


interpreted. Unfortunately, the raw orthogonal pattern is undetermined and
a theoretical interpretation cannot be made. To solve this indeterminacy, the
orthogonal pattern is rotated to simple structure (Thurstone, 1947). On the
transformation to a Simple Structure position, factors can be correlated using
an oblique rotation as long as the theory underlying the analysis allows
correlation between factors.
A popular kind of oblique rotation is the one that first computes a truly
simple matrix, and then uses this matrix as a target pattern in the rotation. In
this article, three of these oblique rotation procedures (Promax, Promaj and
Simplimax) are briefly reviewed and a fourth, Promin, is proposed.
The Promax method was developed by Hendrickson and White (1964)
and aims to build a truly simple pattern matrix that is the target for the oblique
rotation. In fact, Promax starts from an orthogonal simple pattern (usually
Varimax, Kaiser, 1958), the normalized elements of which are raised to a
high power (usually four), while the signs are left unaffected. The elements
are normalized by rows so that all the communalities of the variables are one.
After these transformations are computed, this simple pattern is taken as the
truly simple target matrix for a congruence rotation (Mosier, 1939). It should
be noted that Promax has three main drawbacks. First, it depends on the
previous Varimax rotation: if Varimax fails to find a good previous simple
structure pattern, Promax will fail as well. Second, the congruence rotation
does not seem to be optimal when a hypothetical target is used. A Procrustes

The author is indebted to Henk A. L. Kiers for his helpful reviews and comments of an
earlier version about this article, and to an anonymous reviewer for his useful suggestions.

MULTIVARIATE BEHAVIORAL RESEARCH 347


U. Lorenzo-Seva

rotation seems much more appropriate. Third, the final position of the rotated
factors depends on the power that is used. There is no rule to determine the
best value of the power for each pattern, although a power equal to four
seems to give the best results.
Promaj was proposed by Trendafilov (1994) to counter some of the
limitations in Promax. Promaj also starts from an orthogonal simple pattern
(Varimax is also advised), which is majorized to become the truly simple
target matrix required in the final Procrustes rotation. The main merit of
Promaj is the procedure for building the truly simple target matrix from the
prerotated orthogonal position. It is based on the principle that a simple
structure is one in which a large number of loadings have zero values. So the
main point is to find which loadings should be zero values and make them zero
in the truly simple target matrix. To do this, a majorization procedure is
applied to the orthogonal pattern. Trendafilov (1994) proposed a general
procedure to build a truly simple target matrix when the exact number of zero
values required in the final matrix is unknown.
A brief description of how to carry out majorization is given below, but
for a fuller version see Trendafilov (1994). Let A(n × 1) be the vector to be
majorized and H the majorized solution of A so that,

(1) H = f (A)

where f is the majorization function. If h i is an element of H, then h i is


defined by the expression,

(2) hi = sign( ai) max( ai2 − c, 0)

where ai is an element of A, sign() is the sign of the ai element, max(x1, x2) is


the maximum value of the values x1 and x2, c is a constant and i ranges from
1 to n. The value of the constant c is a2k for the ak element of A that satisfies
two conditions simultaneously. If m is the mean of the square elements of
the column vector A, then
1. the ak2 value is lower than m
2. the ak2 value is the closest ak2 value to m
The effect of this majorization procedure is that the values in A closest to
zero are zero values in H, while the larger values in A are non-zero values in
H but still high (even though not as high as they are in A).
Note that even though Promaj is more a procedure for building a target
than a rotation procedure, its author advises that the oblique Procrustes
rotation should be used after the target has been built. So the problem of

348 MULTIVARIATE BEHAVIORAL RESEARCH


U. Lorenzo-Seva

Promax which remains unsolved is that the final oblique rotation heavily
depends on the previous orthogonal rotation.
Simplimax was developed by Kiers (1994) also to solve the problems in
Promax. The truly simple target matrix and the best target rotation are found
simultaneously in this method. The method does not depend on the previous
orthogonal pattern, because a random number of orthogonal patterns with a
specified number of zeros are taken as the starting point (it is also advisable
to take the Varimax position as a starting point). When building the best
simple target, no subjective decisions are needed, but the target is objectively
the best simple target with a specified number of zeros. The rotation used
is the semi-specified oblique Procrustes rotation (Browne, 1972), where the
specified values are the values that are zero values in the target matrix.
It should be said that Simplimax can give good results even with complex
solutions, where many of the variables have a complexity bigger than one.
Kiers (1994) gives a rotated solution for Thurstone’s 26 Box data (1947),
which is a good rotated solution indeed.
Simplimax solves all the problems in Promax, but finds some difficulties.
Even though Kiers (1994) gives some procedures to help the user to
determine the best number of zeros in the truly simple target, it can
sometimes be difficult to decide. As a reviewer pointed out, the method
requires a number of solutions to be computed and the user to choose from
these solutions. This problem of choise is similar to the decision of how many
factors to retain when applying factor analysis. The conclusion is that
Simplimax is a powerful and efficient rotation method but difficult to use.
This article proposes a variation of Promaj. As has been seen above,
Promaj makes the user free to decide on the orthogonal prerotation method
and the last oblique rotation method. In fact, all the available oblique rotation
methods need some decisions to be taken (some parameters, and the
prerotation or the postrotation method need to be specified). The rotation
procedure presented here needs no specifications, as all the decisions have
already been taken in advance or as a function of the pattern to be rotated.
These decisions seem to lead to a useful rotation method, called Promin.

Promin: An Algorithm for Oblique Semi-Specified Procrustes Rotation

Let A(m × r) be an orthogonal pattern matrix, T(r × r) the oblique


transformation matrix and B(m × r) a simple semi-specified pattern. So
Promin minimizes the function f,

(3) a f
f T, B = AT − B
2

MULTIVARIATE BEHAVIORAL RESEARCH 349


U. Lorenzo-Seva

where T is subject to

(4) Diag(T −1T −1′ ) = I

The problem is solved in two steps. First, a truly simple semi-specified


target is obtained. Then, the semi-specified Procrustes rotation over this
target is computed.
To find the simple semi-specified target B, the Weighted Varimax rotation
(Cureton & Mulaik, 1975) of the given orthogonal pattern A is computed.
Note that the Varimax (Kaiser, 1958) solution which is computed in the
Weighted Varimax (Cureton & Mulaik, 1975) procedure should be computed
from random starts and care should be taken to prevent any possible
permutation effect (Ten Berge, 1995),

(5) P = AW

where W(r × r) is the Weighted Varimax transformation matrix and P(m × r)


is the orthogonal rotated pattern. Using this Weighted Varimax rotation
(suggested by a reviewer) allows Promin to achieve acceptable results even
with complex solutions.
Let V(m × m) be a diagonal matrix that is computed as,

(6) V = diag(PP9)-1/2

This allows the F(m × r) matrix to be computed as the product

(7) F = VP

where the F matrix is now the row-normalized matrix of the P matrix. Finally,
the mean and the standard deviation of the squared elements of each column
of the F matrix is computed. Let m( r ×1) be the vector with these means and
let s(r ×1) be the vector with these standard deviations. So the B(m × r) semi-
specified target matrix is built as follows:
1. if f2ij < (m j + –4sj) then the corresponding element of B is given by bij = 0.
2. if f 2ij > (m j + –4sj) then bij is unspecified.
The matrix B has two different kinds of value: values equal to zero and
values that are unspecified. The values equal to zero are the specified values,
while the unspeficied values will be free values in the rotation.
The constant value (m j + –4sj), used to decide what elements are to be taken
as specified zero values in the target, considers the mean and the standard
deviation of the square elements of each column. This mean can easily
350 MULTIVARIATE BEHAVIORAL RESEARCH
U. Lorenzo-Seva

decrease when a huge number of variables have very low values, as is the case
in large patterns with a clear Simple Structure. If the mean is very small,
some low values, that should be specified to be zero values, will be taken as
unspecified values, which will cause led Promin to give bad results when
computing with very large matrices. Considering the standard deviation helps
to correct this effect. As the standard deviation will be large (because some
of the values will be close to one) the cut point can be increased by
considering a proportion of this standard deviation. After some tests, it was
decided that the proportion 1/4 gave acceptable results. If the solution to be
rotated is a complex one (where variables have significant loadings on more
than a factor), this standard deviation will not be so large (as there will be few
values that are close to zero in the columns), so the cut point will not be
increased too much over the mean.
Once the simple semi-specified target is known, the f function becomes
the function g,

(8) g(T) = ||AT - B||2

where T is constrained to Equation 4. The solution to this semi-specified


Procrustes rotation has already been given by Browne (1972) and
Meredith (1977). Actually, Browne’s (1972) algorithm is a particular case
of Meredith’s (1977) hyperplane-fitting methods. However, as Simplimax
(Kiers, 1994) is based on Browne’s (1972) algorithm, in this article Promin
will be computed using the same algorithm to find the transformation matrix T.
The semi-specified rotation has some advantages over the fully specified
Procrustes rotation,
1. When defining the Simple Structure, all that is said is that some
loadings should be different from zero while others should be zero (Thurstone,
1947). Although it is difficult to know how high the loadings which are different
from zero should be, it is safe to say that the zero loadings should have zero
values. So it is safer to specify the values that should be as close as possible to
zero, while the others remain free values in the truly simple target.
2. This rotation criterion gives more freedom to the final oblique
rotation, because it uses a partly rather than fully specified target. So the
dependence on the orthogonal rotation computed when building the target
matrix is minimized in some way.

Studies Using Artificial Data

To study the performance of Promin three studies with artificial data


were carried out. The first study tests the performance of different rotation

MULTIVARIATE BEHAVIORAL RESEARCH 351


U. Lorenzo-Seva

methods as the correlation between factors gets progressively higher. The


second study tests the rotation methods when a perfect Simple Structure
solution (Thurstone, 1947) can be found. The last study tests the rotation
methods when a small proportion of the variables in the solution are variables
with a complexity bigger than one.
The rotation procedures tested in these three studies are: Promin, Promaj
(as proposed by Trendafilov, 1994), Promax, and Direct Oblimin (g = 0)
(Clarkson & Jennrich, 1988). The software used for these simulation studies
was Matlab 4.0 for a Silicon Graphics Computer.

First Study

The first study tests the performance of different rotation methods as the
correlation between factors gets progressively higher. To test the rotation
procedures, a sample of 250 artificial oblique solutions was built.
A sample of 250 oblique patterns (P) was built. Each of these patterns
had 5 factors and 100 variables (all of which had a perfect complexity of
one). The non-zero loadings were uniform randomly chosen between 0.40
and 1. Finally, the number of variables with non-zero loadings on each factor
was uniform randomly chosen, with the constraint that each factor had at
least five non-zero loadings.
Then 250 interfactor correlation matrices (F) were built. All the non-
diagonal elements in the first sample of 50 matrices were equal to 0.10, in the
second sample they were all equal to 0.20, etc. So the correlations between
factors rang from 0.1 to 0.5 in this study.
Once these True Oblique Patterns P(100 × 5) and True Interfactor
Correlation matrices F(5 × 5) had been built, an orthogonal pattern A was
computed for each pair of True Oblique Pattern and True Interfactor
Correlation matrix.
Let Q(100 × 5) and D(5 × 5) matrices be the eigendecomposition of F,

(9) F = QDQ9

with Q9Q = QQ9 = I and D diagonal. Let the T(5 × 5) matrix be a random
orthogonal transformation matrix (T9T = TT9 = I). So the orthogonal pattern
A(100 × 5) is computed as,

(10) A = PQD 1/2T

which ensures that

352 MULTIVARIATE BEHAVIORAL RESEARCH


U. Lorenzo-Seva

(11) AA ′ = PQD1/ 2 TT ′ D1/ 2 Q ′P ′ = PΦP ′ = R *

where R*(100 × 100) is the reduced correlation matrix.


This A orthogonal pattern computed for each P and F pair will now be
obliquely rotated using all the rotation methods tested in the experiment.

Oblique Rotation of the Solutions

Note that the parameters for each rotation procedure were the following:
1. In Promaj, the orthogonal prerotation was Normalized Varimax with
20 random starts and the oblique rotation was a fully specified Procrustes
rotation. For the Procrustes rotation, Browne’s (1972) method was used.
However, note that the target was a fully-specified target.
2. In Promax the orthogonal prerotation was Normalized Varimax with
20 random starts and the parameter k = 4.
3. In Direct Oblimin the parameter g was 0, so the method tested was in
fact Direct Quartimin (Clarkson & Jennrich, 1988).
4. Promin was computed in the way described above.
Finally, all the convergence criteria were set to 0.00001.

Results of the Study

Once the oblique patterns and the interfactor correlation matrices had
been computed by each rotation method from each A orthogonal pattern, the
factors were permuted and reflected (if necessary) to obtain optimal
congruence between the oblique patterns obtained and the True Oblique
Patterns (P). Next, the Root Mean Square of the Residual (RMSR) between
the rotated patterns and the True Oblique Patterns was computed. The
RMSR was also computed between the obtained interfactor correlation
matrices and the True Interfactor Correlation matrices. To summarize the
information, the mean of the RMSR for each set of 50 True Oblique Patterns
(each set with a different correlation between factors) was computed. The
mean of the RMSR for each set of 50 True Interfactor Correlation matrices
was also computed. Table 1 (next page) shows the results.
This table seems to show that:
1. Promax and Promaj have difficulties in recovering the oblique
solution as the true correlation between factors gets higher. Promax seems to
fail to recover mainly the True Oblique Pattern, whereas Promaj seems to fail
to recover mainly the interfactor correlation matrix. Table 2 shows one of the
obtained interfactor correlation matrices obtained where the True Interfactor

MULTIVARIATE BEHAVIORAL RESEARCH 353


U. Lorenzo-Seva

Table 1
Mean of the 50 RMSR Computed for True and Rotated Oblique Patterns and
for the True and Obtained Interfactor Correlation Matrices (first study)

Mean of the RMSR between true and rotated oblique patterns

METHOD F = 0.1 F = 0.2 F = 0.3 F = 0.4 F = 0.5

PROMAX 0.021872 0.041981 0.064294 0.092641 0.129190


PROMAJ 0.000138 0.001598 0.008369 0.032406 0.036820
D OBLIMIN 0.000050 0.000104 0.000165 0.000229 0.000273
PROMIN 0.000000 0.000001 0.000001 0.000001 0.000002

Mean of the RMSR between true and the obtained interfactor correlation
matrices

METHOD F = 0.1 F = 0.2 F = 0.3 F = 0.4 F = 0.5

PROMAX 0.000003 0.000083 0.000598 0.002513 0.007692


PROMAJ 0.000316 0.004158 0.019450 0.086557 0.122399
D OBLIMIN 0.000116 0.000250 0.000394 0.000543 0.000617
PROMIN 0.000001 0.000002 0.000002 0.000002 0.000004

Table 2
An Interfactor Correlation Matrix Obtained by Promaj When the True
Interfactor Correlations were 0.5 (first study)

1 2 3 4 5

1 1.000000 0.430107 0.379002 0.470712 0.319710


2 0.430107 1.000000 0.390711 0.482624 0.326299
3 0.379002 0.390711 1.000000 0.367803 -0.057171
4 0.470712 0.482624 0.367803 1.000000 0.362213
5 0.319710 0.326299 -0.057171 0.362213 1.000000

354 MULTIVARIATE BEHAVIORAL RESEARCH


U. Lorenzo-Seva

Correlation matrix had a constant value of 0.50. As can be observed in this


table, Promaj underestimates the correlation between factors. This effect on
Promax is minimized by the fact that the oblique rotation computed is a
congruence rotation, while the oblique Procrustes rotation applied in Promaj
constrains the rotated solution more. The conclusion is that Promax and,
particularly, Promaj underestimate the interfactor correlation matrix.
2. Direct Oblimin (g = 0) and Promin are similar although the rotated
solution given by Promin seems to recover the true solution better.

Second Study

The first study worked with solutions that are quite unnatural in the usual
application of the factor rotation: it is not usual for factors to have similar
correlation values with all the others in the solution. So a second study was
carried out, but this time the True Interfactor Correlation matrices had
random values.
For this study, a sample of 50 True Oblique Patterns P(100 × 5) were
computed in the same way as they were computed in the previous study.
Fifty True Interfactor Correlation matrices F(5 × 5) were also computed.
Now the values of the off-diagonal elements were uniform randomly chosen
between 0.00 and 0.50.
For each pair of True Oblique Patterns and True Interfactor Correlation
matrices, an orthogonal pattern A(100 × 5) was computed as in the previous
study. The matrix A wil now be obliquely rotated using all the rotation
methods tested in the experiment.

Oblique Rotation of the Solutions

The rotation methods tested in this second study are the same as in the
previous study, and they were computed in the same way.

Results of the Study

Once the 50 oblique rotated patterns and the 50 interfactor correlation


matrices were obtained using each rotation method, the factors were
permuted and reflected (if necessary) to obtain optimal congruence between
the oblique patterns obtained and the True Oblique Patterns (P). Next, the
RMSR between the oblique patterns obtained and the True Oblique Patterns
were computed. The RMSR between the interfactor correlation matrices
obtained and the True Interfactor Correlation matrices were also computed.

MULTIVARIATE BEHAVIORAL RESEARCH 355


U. Lorenzo-Seva

Table 3 shows the mean, the worst and the best of the RMSR for the patterns
and for the interfactor correlation matrices.
This table seems to show that:
1. Even though Promax and Promaj still have difficulties in recovering
the oblique solution, their performance seems better than in the previous
study. Promax still has more problems recovering the True Oblique Pattern,
while Promaj seems to have difficulties in recovering the True Interfactor
Correlation matrices. It should be said that, on average, Promaj is better than
in study 1, although it still gives some bad results.
2. Direct Oblimin (g = 0) and Promin are still similar, while the rotated
solution given by Promin also seems to recover the true solution better.

Third Study

In the two previous studies, the performance of the rotation methods was
tested by rotating solutions where all the variables are variables with a
complexity one. That is to say, solutions that have a Simple Structure

Table 3
Mean of the 50 RMSR Computed for True and Rotated Oblique Patterns and
for the True and Obtained Interfactor Correlation Matrices [The worst and
the best RMSR obtained are also shown (Second Study)]

RMSR between true and rotated oblique patterns

PROMAX PROMAJ D OBLIMIN PROMIN

MEAN 0.103934 0.032033 0.000159 0.000001


WORST 0.230261 0.104460 0.000371 0.000008
BEST 0.025531 0.003113 0.000010 0.000000

RMSR between true and the obtained interfactor correlation matrices

PROMAX PROMAJ D OBLIMIN PROMIN

MEAN 0.002522 0.059401 0.000353 0.000002


WORST 0.006606 0.115427 0.000728 0.000021
BEST 0.000596 0.007147 0.000073 0.000000
356 MULTIVARIATE BEHAVIORAL RESEARCH
U. Lorenzo-Seva

(Thurstone, 1947). However, these rotation methods will frequently have to


deal with solutions where most of the variables have a complexity one, but
some of them do not. While the solution can be considered to be mainly a
Simple Structure solution, some of its variables are in fact complex variables.
This third study will deal with this kind of solutions.
For this study, a sample of 50 True Oblique Patterns P(100 ×5) was
computed in the same way as they were computed in the first study. To each
pattern, loadings for 25 new variables were added. These variables had
complexities 3 (that is, they had non-zero loadings in 3 randomly chosen
factors) and their non-zero loadings were uniform randomly chosen between
0.40 and 0.50. So, the final True Oblique Patterns had 125 variables (100
variables with a complexity one and 25 variables with complexities three).
Fifty True Interfactor Correlation matrices F(5 × 5) were also
computed. Now the values of the off-diagonal elements were uniform
randomly chosen between 0.00 and 0.50.
Note that randomly choosing the loadings of the complex variables and
the correlation between factors values does not ensure that the diagonal
elements of R*(125 × 125) will have values smaller or equal to one (this is
the usual situation in factor analysis). So, each pair of True Oblique Patterns
and True interfactor Correlation matrices were checked to verify that it was
so. If a pair led to an improper R* diagonal value, the pair was discarded and
a new True Oblique Pattern and True Interfactor Correlation matrix built.
For each pair of True Oblique Patterns and True Interfactor Correlation
matrices, an orthogonal pattern A(125 × 5) was computed as in the previous
study. Each A was then obliquely rotated using all the rotation methods tested
in the experiment.

Oblique Rotation of the Solutions

The rotation methods tested in this third study are the same as in the
previous study, and they were computed in exactly the same way.

Results of the Study

Once the 50 oblique rotated patterns and the 50 interfactor correlation


matrices had been obtained using each rotation method, the factors were
permuted and reflected (if necessary) to obtain optimal congruence between
the oblique patterns obtained and the True Oblique Patterns (P). Next, the
RMSR between the obtained oblique patterns and the True Oblique Patterns
were computed. The RMSR between the obtained interfactor correlation
matrices and the True Interfactor Correlation matrices were computed.

MULTIVARIATE BEHAVIORAL RESEARCH 357


U. Lorenzo-Seva

Table 4 shows the mean, the worst and the best of the RMSR for the patterns
and for the interfactor correlation matrices.
This table seems to show that:
1. Whereas Promax has serious difficulty in recovering the True Oblique
Pattern and the True Interfactor Correlation matrix, the performance of
Promaj seems to be better than in study 2. However, they both go to worse
results than Promin.
2. Whereas Direct Oblimin (g = 0) and Promin give similar results in the
previous studies, here Direct Oblimin fails to recover the True solution with
the precision that Promin does.
3. Promin does not perform as well as in the previous studies, but it still
gave acceptable results. Note that the worst oblique pattern obtained by
Promin is much better than any of the patterns obtained by any of the other
rotation methods. Also note that only when using Promin can the True
solution sometimes be perfectly obtained by rotation.

Table 4
Mean of the 50 RMSR Computed for True and Rotated Oblique Patterns and
for the True and Obtained Interfactor Correlation Matrices [The worst and
the best RMSR obtained are also shown (Third Study)]

RMSR between true and rotated oblique patterns

PROMAX PROMAJ D OBLIMIN PROMIN

MEAN 0.167440 0.027112 0.012497 0.000732


WORST 0.279867 0.059734 0.020953 0.004621
BEST 0.086943 0.001678 0.010426 0.000000

RMSR between true and the obtained interfactor correlation matrices

PROMAX PROMAJ D OBLIMIN PROMIN

MEAN 0.125388 0.034653 0.032261 0.001006


WORST 0.163053 0.085252 0.051466 0.006946
BEST 0.080221 0.002071 0.022877 0.000000

358 MULTIVARIATE BEHAVIORAL RESEARCH


U. Lorenzo-Seva

Analysis of the Twenty-four Psychological Tests Data

The Twenty-four Psychological Tests Data (Holzinger & Swineford,


1939) were rotated using Promax, Promaj, Direct Oblimin (g = 0), Simplimax
and Promin. The correlation matrix for the 24 psychological tests was taken
from Harman (1976, p. 124). This matrix was analyzed by Principal
Component Analysis and four components were retained, thus obtaining the
orthogonal unrotated pattern. This orthogonal unrotated pattern was rotated
by Promax, Direct Oblimin (g = 0) and Promin, whereas the patterns rotated
by Promaj and Simplimax were taken directly from their authors
(Trendafilov, 1994; Kiers, 1994). The parameters for each of the rotation
procedures were the same as for the artificial data analysis discussed above.
The oblique patterns obtained by Promin and Direct Oblimin (g = 0) are
shown in Table 5 (next page). The absolute elements larger than.30 are set
in bold face. Note that the solutions are quite similar between them.
To compare all the rotated patterns with the others, the RMSR between
patterns was computed. (See Table 6).
It should be said that the differences between the rotated solutions in
Table 6 are small, whereas the more similar methods are Promin and Direct
Oblimin. Note that Simplimax was computed by Kiers (1994) fixing the
variables complexity a priori to one, what could explain why it is the rotation
procedure that differs more to the others with these data.

Analysis of the Twenty-six Box Problem

Finally, Thurstone’s (1947) 26 Box Data was analyzed. Thurstone built


this three-dimensional box problem by measuring a collection of 30 random
boxes. The three dimensions, x, y and z, were measured and recorded for
each box. Using these three dimensions, 26 variables were computed by
nonlinear combination formulas (see Table 7). As Kiers (1994) points out, a
simple pattern may not exist for these data, but a good approximation can be
expected. Rotation methods specifically developed to find simple structures
will fail to find such an approximation.
The rotation methods that showed a good performance with these data are
Weighted Varimax (Cureton & Mulaik, 1975) and Simplimax (Kiers, 1994).
These rotation methods can deal with solutions in which the variables in the
pattern have a complexity that is bigger than one. As Promin uses Weighted
Varimax rotation as a previous orthogonal rotation, it should be able to deal with
complex data as well. Promin was applied to the orthogonal pattern given by
Thurstone (1947). The oblique pattern obtained by Promin and the oblique

MULTIVARIATE BEHAVIORAL RESEARCH 359


U. Lorenzo-Seva

Table 5
Pattern Matrices by Promin and Direct Oblimin (g = 0) (24 Psychological
Tests) [Elements larger than 0.30 or smaller than -0.30 set in bold face]

Promin Direct Oblimin (g = 0)

FI FII FIII FIV FI FII FIII FIV

v1 0.64 -0.04 0.16 0.06 .65 .01 .12 .05


v2 0.47 -0.01 0.05 0.00 .47 .03 .03 -.01
v3 0.61 0.01 -0.09 0.06 .59 .04 -.10 .04
v4 0.56 0.11 0.03 -0.02 .55 .14 .01 -.03
v5 0.07 0.75 0.12 -0.04 .01 .78 .11 -.01
v6 0.07 0.78 -0.05 0.08 .00 .80 -.06 .09
v7 0.08 0.84 0.05 -0.12 .01 .86 .04 -.09
v8 0.25 0.52 0.15 -0.04 .21 .56 .13 -.02
v9 0.08 0.81 -0.09 0.08 .01 .83 -.10 .09
v10 -0.30 0.08 0.85 -0.03 -.24 .12 .78 .05
v11 -0.11 0.04 0.59 0.27 -.06 .09 .52 .33
v12 0.04 -0.14 0.80 -0.10 .10 -.08 .73 -.03
v13 0.32 0.00 0.59 -0.12 .35 .06 .53 -.07
v14 -0.10 0.11 -0.01 0.59 -.08 .12 -.04 .59
v15 -0.03 0.01 0.01 0.54 -.01 .03 -.02 .54
v16 0.33 -0.12 -0.03 0.54 .35 -.08 -.07 .51
v17 -0.12 0.01 0.17 0.60 -.09 .03 .11 .61
v18 0.14 -0.18 0.34 0.41 .19 -.13 .28 .43
v19 0.15 0.00 0.12 0.36 .16 .03 .07 .36
v20 0.35 0.27 0.00 0.21 .33 .31 -.02 .21
v21 0.30 0.00 0.44 0.07 .33 .05 .38 .10
v22 0.33 0.24 0.02 0.22 .31 .27 .00 .22
v23 0.45 0.23 0.15 0.09 .44 .27 .12 .10
v24 0.00 0.24 0.47 0.14 .02 .28 .41 .18

360 MULTIVARIATE BEHAVIORAL RESEARCH


U. Lorenzo-Seva

Table 6
RMSR Between the Oblique Patterns Obtained by Each Rotation Method for
the 24 Psychological Tests

PROMIN SIMPLIMAX PROMAJ D OBLIMIN PROMAX

PROMIN 0 0.0548 0.0490 0.0292 0.0322


SIMPLIMAX 0.0548 0 0.0773 0.0609 0.0457
PROMAJ 0.0490 0.0773 0 0.0376 0.0483
D OBLIMIN 0.0292 0.0609 0.0376 0 0.0319
PROMAX 0.0322 0.0457 0.0483 0.0319 0

pattern obtained by Kiers (1994) using Simplimax are shown in Table 7 (next
page).
Note that in Table 7 the elements of the pattern that were specified to be
zero values in the semi-specificed target are bold face. Clearly, Simplimax
built the best semi-specified target matrix. The Simplimax solution shown in
Table 7 was computed by Kiers (1994) who chose the values p = 25...30.
Because there is a large jump in Simplimax optimal function values after the
analysis for p = 27 (see Kiers, 1994, p. 575 for details), the number of zeros
in the semi-specified pattern was set to 27. So the number of zero values
expected for to Thurstone’s construction formulas was found. Promin sets
42 zero values. Obviously, Promin sets more zero values than the number of
zero values expected, and fails to find the proper unspecified loadings in two
kind of variables:
1. Variables with significant loadings different from, but still close to,
zero values or at least small when compared to the larger loading in the same
component. For example, variables such as (x2 + y2 + z2)1/2 or x2 y. This kind
of loadings is difficult to detect by Promin as unspecified loadings because
they are smaller than the means of the squared elements of each column of the
orthogonal pattern F, where the unspecified values are computed. However,
as these loadings are close to zero values and are specified to be zero values
in the semi-specified target matrix B, this kind of mistake is not very
significant.
2. Variables with two large loadings on both components. For example,
variables such as x/y or x/z. Promin fails to detect them because in the
orthogonal pattern F, where the unspecified loadings are computed, these

MULTIVARIATE BEHAVIORAL RESEARCH 361


U. Lorenzo-Seva

Table 7
Pattern Matrices Obtained by Promin and Simplimax for the 26 Box Problem
Elements Specified to be Zero in the Semi-specified Target Matrix Set in Bold Face

Promin Simplimax

FI FII FIII FI FII FIII

x 1.00 -0.06 -0.02 0.99 -0.01 -0.01


y 0.16 0.94 -0.03 0.06 0.94 0.04
z 0.11 -0.02 0.97 0.01 0.05 0.97
xy 0.71 0.63 -0.08 0.64 0.65 -0.02
xz 0.67 -0.08 0.64 0.60 0.00 0.65
yz 0.10 0.57 0.60 -0.02 0.61 0.64
x2 y 0.88 0.35 -0.03 0.83 0.39 0.01
xy 2 0.47 0.81 -0.04 0.38 0.82 0.03
x2z 0.84 -0.09 0.41 0.79 -0.02 0.42
xz 2 0.54 -0.07 0.85 0.45 0.02 0.86
y2z 0.10 0.73 0.39 -0.02 0.76 0.45
yz 2 0.10 0.37 0.75 -0.02 0.43 0.78
x/y 0.67 -0.87 0.07 0.75 -0.83 0.02
y/x -0.67 0.87 -0.07 -0.75 0.83 -0.02
x/z 0.74 0.04 -0.84 0.81 0.02 -0.82
z/x -0.74 -0.04 0.84 -0.81 -0.02 0.82
y/z -0.01 0.93 -0.88 -0.02 0.86 -0.81
z/y 0.01 -0.93 0.88 0.02 -0.86 0.81
2x + 2y 0.63 0.70 -0.08 0.55 0.72 -0.02
2x + 2z 0.63 -0.11 0.69 0.56 -0.03 0.69
2y + 2z 0.12 0.57 0.58 0.00 0.61 0.63
(x2 + y2)1/2 0.62 0.69 -0.07 0.54 0.71 -0.01
(x2 + z2)1/2 0.60 -0.09 0.67 0.53 -0.01 0.68
(y2 + z2)1/2 0.14 0.57 0.55 0.02 0.61 0.60
xyz 0.55 0.43 0.42 0.45 0.48 0.46
(x2 + y2 + z2)1/2 0.44 0.48 0.44 0.34 0.53 0.48

362 MULTIVARIATE BEHAVIORAL RESEARCH


U. Lorenzo-Seva

variables cannot have such large loadings, as the maximum communality is 1.


Promin does not always fail with this kind of variable, as y/z and z/y are
properly detected.
Despite the differences in the semi-specified target matrix, both methods
give good approximations to the Simple Structure that is expected from the
nonlinear combinations of the 26 variables. Bentler’s (1977) simplicity index
was computed with both solutions. The simplicity index for the pattern
obtained by Promin is 0.69, whereas it is 0.71 for the pattern obtained by
Simplimax. As Simplimax aims at minimizing the sum of squared small
loadings it can be considered as an ideal reference point for Promin. The
example shows that Promin, which is considerably simpler to use than
Simplimax, nevertheless gets very close to the Simplimax solution. While
other rotation procedures are as simple to use as Promin, Promin is the one
that gets closer to Simplimax with this complex data set.
The correlation between factors obtained by Promin are 0.16 for factors
I and II, 0.22 for factors I and III, and 0.38 for factors II and III; whereas the
correlations for Simplimax (Kiers, 1994) are 0.20, 0.27 and 0.28
respectively.

Discussion

Promin is an oblique rotation procedure which combines strategies from


Promax, Promaj and Simplimax. In fact, it can be considered as a rotation
procedure where all the steps in the rotation are perfectly defined in advance
and no freedom is given to the user.
Promin also could be regarded as a particular case of Simplimax.
Whereas Simplimax tests a number of random orthogonal starts and a
different number of zero values in the true simple target matrix, Promin only
tests one orthogonal start (Weighted Varimax), one number of zero values
and their position in the target matrix. Even though Simplimax is a much more
powerful rotation method, Promin seems to be good enough, and it is much
simpler to use.
The following conclusions have been drawn from the work in this article:
1. Working with artificial data, Promin rotated the solutions closer to
the true solutions than any other rotation procedure tested. Direct Oblimin
(g = 0 ) performed in a very similar way to Promin. However, when only a
few variables in the pattern are complex variables, Promin seems to perform
better than Direct Oblimin (g = 0).
2. Working with artificial data, Promaj and Promax proved that they
have difficulties when there is a high correlation between factors in the true
solution. These rotation procedures seem to depend heavily on the previous

MULTIVARIATE BEHAVIORAL RESEARCH 363


U. Lorenzo-Seva

orthogonal pattern. Note that Promin does not seem to have this dependence,
even though a previous Weighted Varimax rotation is performed.
3. When working with the real data tested here, Promin and Direct
Oblimin (g = 0) reach very similar solutions. As there seems to be an
interaction between the data and the oblique rotation procedures (Hakstian
and Abell, 1974), it is good that two methods reach similar oblique solutions
using different procedures if (a) the rotated solutions are indeed at least as
simple and interpretable as the solutions obtained using other rotation
procedures, and (b) both methods reach the same solutions using very
different procedures. Promin and Direct Oblimin (g = 0) have both
conditions.
4. When working with very complex data (such as the 26 Box Problem),
the rotated pattern obtained by Promin is similar to the one obtained by
Simplimax. Note that Promin uses Weighted Varimax as a pre-rotation
orthogonal method. This means that; (a) Promin can deal with solutions in
which variables have a complexity which is bigger than one, and; (b) Promin
can fail (as Weighted Varimax can fail, Kiers, 1994) if the solution has a very
high obtuse configuration, which is very unusual in real data (Cureton &
Mulaik, 1975). Note that Simplimax does not have this limitation. However,
Promin will usually obtain good results and it is simpler to use than
Simplimax.
To sum up, Promin reaches solutions which are as good as the solutions
given by using other well-known procedures, and is simpler to use. Although
Promin and Direct Oblimin can obtain very similar solutions when working
with simple data, Promin should be preferred if some variables in the solution
have a complexity bigger than one. Finally, even though Promin can deal with
very complex solutions, it would be advisable to use Simplimax for these
cases (specially if an obtuse configuration is expected).
A Windows 95 program (WPROMIN.EXE) for computing PROMIN
rotations and a short manual can be obtained from the author by sending an
e-mail to the author (uls@fcep.urv.es).

References

Bentler, P. M. (1977). Factor simplicity index and transformations. Psychometrika, 42, 277-295.
Browne, M. W. (1972). Oblique rotation to a partially specified target. British Journal of
Mathematical and Statistical Psychology, 25, 207-212.
Clarkson, D. B. & Jennrich, R. I. (1988). Quartic rotation criteria and algorithms.
Psychometrika, 53, 251-259.
Cureton, E. E. & Mulaik, S. (1975). The weighted varimax rotation and the promax rotation.
Psychometrika, 40, 183-195.

364 MULTIVARIATE BEHAVIORAL RESEARCH


U. Lorenzo-Seva
Hakstian, A. R. & Abell, R. A. (1974). A further comparison of oblique factor transformation
methods. Psychometrika, 39, 429-444.
Harman, H. H. (1976). Modern factor analysis (3rd ed.). Chicago: University of Chicago Press.
Hendrickson, A. E. & White, P. O. (1964). PROMAX: A quick method for rotation to oblique
simple structure. British Journal of Statistical Psychology, 17, 65-70.
Holzinger, K. J. & Swineford, F. (1939). A study in factor analysis: The stability of a bi-
factor solution (Supplementary Educational Monographs No 48). Chicago: University
of Chicago, Department of Education.
Kaiser, H. F. (1958). The varimax criterion for analytic rotation in factor analysis.
Psychometrika, 23, 187-200.
Kiers, H. A. L. (1994). Simplimax: Oblique rotation to an optimal target with simple structure.
Psychometrika, 59, 567-579.
Meredith, W. (1977). On weighted procrustes and hyperplane fitting in factor analytic
rotation. Psychometrika, 42, 491-522.
Ten Berge, J. M. F. (1995). Suppressing permutations or rigid planar rotations: A remedy
against nonoptimal varimax rotations. Psychometrika, 60, 437-446.
Thurstone, L. L. (1947). Multiple Factor Analysis. Chicago: University of Chicago Press.
Trendafilov, N. (1994). A simple method for procrustean rotation in factor analysis using
majorization theory. Multivariate Behavioral Research, 29, 385-408.

Accepted September, 1998.

MULTIVARIATE BEHAVIORAL RESEARCH 365

You might also like