PSYCHOMETmKA--VOI~. 23~ NO.

3
S E ~ M B E I ~ , 1958

THE VARIMAX CRITERION FOR ANALYTIC ROTATION IN

F A C T O R ANALYSIS*
HENRY F. KAISER
UNIVERSITY OF ILLINOIS

An analytic criterion for rotation is defined. The scientific advantage

of analytic criteria over subjective (graphical) rotational procedures is dis-
cussed. Carroll's criterion and the quartimax criterion are briefly reviewed;
the varimax criterion is outlined in detail and contrasted both logically and
numerically with the quartimax criterion. It is shown that the normal varimax
solution probably coincides closely to the application of the principle of simple
structure. However, it is proposed that the ultimate criterion of a rotational
procedure is factorial invariance, not simple structure-although the two
notions appear to be highly related. The normal varimax criterion is shown
to be a two-dimensional generalization of the classic Spearman case, i.e., it
shows perfect factorial invariance for two pure clusters. An example is
given of the invariance of a normal varimax solution for more than two
factors. The oblique normal varimax criterion is stated. A computational out-
line for the orthogonal normal varimax is appended.

I n factor analysis, an analytic criterion for rotation is defined as one

t h a t imposes mathematical conditions beyond the fundamental factor
theorem, such that a factor matrix is uniquely determined. Historically,
the first such criterion was Thurstone's treatment of the principal axes
problem [10]: from a n y arbitrary factor matrix he suggested rotating under
the criterion that each factor successively accounts for the maximum variance.
But principal axes have seldom been accepted as psychologically very interest-
ing ([9], p. 139). The rotation problem for psychologically meaningful factors
is usually handled judgmentally. Scientifically, however, this procedure is
not very satisfactory: the ad hoc quality of subjective rotation makes
uniquely determined factors impossible; only factors t h a t are subject to the
uncertainties and controversies besetting a n y a posteriori reasoning can be
defined. I n contrast, an analytic criterion for rotation would allow factor
analysis to become a straightforward methodology stripped of its sub-
]ectivity and a proper tool for scientific inquiry.

The Quartimax Criterion

The first analytic criterion for determining psychologically interpretable
factors was presented in 1953 by Carroll [1]. I n an a t t e m p t to provide a
*Part of the material in this paper is from the writer's Ph.D. thesis. I am indebted
to my committee, Professors F. T. Tyler, R. C. Tryon, and It. D. Carter, chairman, for
many helpful suggestions and criticisms. Dr. John Caffrey suggested the name varimax,
and wrote the original IBM 602A computer program for this criterion.
I am also indebted to the staff of the University of California Computer Center for
help in programming the procedures described in the paper for their IBM 701 electronic
computer. Since their installation is partially supported by a grant from the National
Science Foundation, the assistance of this agency is acknowledged.
187
188 PSYCHOMETRIKA

mathematical explication of Thurstone's simple structure, tie suggested that

for a given factor matrix,
(1) / = ~_, ~_,a~a~
• <t i

should be a minimum, where j = 1, 2, . . . , n are tests, s, t = 1, 2, . . . , r

are factors, and a~. is the factor loading of the jth test on the sth factor.
It appears that Carroll was motivated in writing (1) primarily by a close
inspection of Thurstone's five formal rules for simple structure ([12], p. 335),
particularly the requirement that a large loading for one factor be opposite
a small loading for another factor.
In his original paper, Carroll provided two numerical examples of the
application of his method. Without the restriction of orthogonality, these
illustrations gave somewhat equivocal results--while the application of (1)
appears to bring one close to the desired simple structure, the criterion has
an obvious bias in being too strongly, influenced by factorially complex tests.
In the light of later developments, Carroll's criterion should probably
be relegated to the limbo of "near misses"; however, this does not detract
from the fact that it was the first attempt to break away from an inflexible
devotion to Thurstone's ambiguous, arbitrary, and mathematically un-
manageable qualitative rules for his intuitively compelling notion of simple
structure.
Almost simultaneously with Carroll's development, Neuhaus and
Wrigley [7], Saunders [8], and Ferguson [2] proposed what is usually called
the quartimax method for orthogonal simple structure. Neuhaus and Wrigley
suggest that a most easily interpretable factor matrix, in the simple structure
sense, may be found when the variance of all nr squared loadings of the
factor matrix is a maximum, i.e.,
(2) q, = fur ~_, ~ (a~,) 2 -- ( ~ , ~ a~,)']/n2r 2 -- maximum.
i * i .

Saunders' approach requires that the kurtosis (fourth moment over second
moment squared) of all loadings and their reflections be a maximum,
(3) q2 = nr ~_, ~_~ a~./(~_, ~E] a~.)2 = maximum.

While Ferguson, basing his rationale on certain parallels with information

theory, calls simply for
(4) qa = ~ ~ a~. = maximum.
i

All these investigators are concerned with attaining a factor matrix

with a maximum tendency to have both small and large loadings. While
less obviously related to Thurstone's rules than Carroll's criterion, the
emphasis on small loadings coincides with Thurstone's requirements of
 HENRY F. KAISER 189

zero loadings. For orthogonal factors, criteria (2), (3), and (4) are equivalent
because of tile invariance of the sum of the communalities, ~ ] ~ ] , a~ .
(This term, as well as other constants, disappear when differentiated in the
calculus problem involved in finding the required critical point.)
Indeed, it turns out that they are also equivalent to Carroll's criterion
in the orthogonal case. Minimizing (1) is equivalent to maximizing (4)
since the squared eommunality of a test is
constant = ( ~ ] a ~ ) 2 = ~]a~, + 2 ~a~,a~,
, s s<t

and tile sum of squared communalities over all tests is

constant = + 2
i s i s<~

= q3 + 2L
Thus, since the quartimax criterion plus twice Carroll's criterion is a con-
stant, maximizing q3 is equivalent to minimizing f.
Neuhaus and Wrigley- realized that none of these criteria can be real-
istically applied without the aid of an electronic computer--the calculations
involved are too extensive for a desk calculator or punched card mechanical
computers. Consequently, they programmed the quartimax method for the
Illiac* and provided a rather extensive numerical investigation of the empirical
properties of the quartimax method.
Their results were perhaps more encouraging than Carroll's. Under
the restriction of orthogonality Carroll's criterion (or the equivalent quartimax
method) does not show nearly so obvious a bias as does Carroll's criterion
when the restriction of orthogonality is removed. However, as an explication
of orthogona] simple structure, the quartimax method does have a systematic
bias which will be more fully examined in the next section.

The Varimax Criterion

ti'rom the outset, the above methods consider all nr loadings simul-
taneously. In every ease, however, these criteria m a y be applied separately
to each row of the factor matrix and summed over rows for the final criterion
because of the invariance of the comlnunalities. For example, Neuhaus and
Wrigley could have defined the simplicity, say, of tile factorial composition
of the jth test as the variance of the squared loadings for this test,
(5) q* = [r ~ ] (a~,) ~ -- ( ~ ] a~,)2l/r 2.
s

*The Illiac is the University of Illinois electronic computer. Subsequently, the quar ti-
max criterion has been programmed for the CRC-102A (Neuhaus), and the IBM 701
(Kaiser). The varimax criterion described in the next two sections has been programmed
for SWAC at UCLA (Comrey), the IBM 701 (Kaiser), Illiac (Dickman), and the IBM
650 (Vandenberg).
190 PSYCHOt~IETRIKA

To obtain the total criterion for the entire factor matrix, (5) could then be
summed over all tests to give
(6) q* = ~ {It E (a~.)~ - ( ~ a~.)2]/r21 -
t s s

Maximizing q* is equivalent to maximizing qa, again because constant terms

vanish when differentiated.
Equation (6) perhaps provides some insight into the quartimax
criterion--its aim is to simplify the description of each row, or test, of the
factor matrix. It is unconcerned with simplifying the columns, or factors, of
the factor matrix (probably the most fundamental of all requirements for
simple structure). The implication of this is that the quartimax criterion
will often give a general factor. Under requirement (5) there is no reason
why a large loading for each test may not occur on the same factor. In practice,
this tendency for the quartimax criterion to yield a general factor is most
pronounced when the unrotated factor matrix has a pronounced general
factor.
From the simple structure viewpoint, an immediate modification of the
quartimax criterion is apparent. Let us define the simplicity of a factor as
the variance of its squared loadings,
(7) v* = [n ~ , (a~) 2 -- ( ~_, a~,) 2]/n 2.
i i

And for the criterion for all factors, define the maximum simplicity of a
factor matrix as the maximization of
(S) v'* = ~ v * , = ~ {In ~--~.(a~,)2 -- (~,a~,)2]/n=},

t.he variance of squared loadings by columns rather than by rows.

Since a factor is a vector of correlation coefficients, the most interpret-
able factor is one based upon correlation coefficients which are maximally
interpretable. Those correlations which satisfy this condition are patently
obvious: correlations of 4- 1, which indicate a functional relationship, and
correlations of zero, which indicate no linear relationship. On the other hand,
middle-sized correlations are the most difficult to understand. Thus, it is
seen why v* in (7) could be maximized for the maximum interpretability or
simplicity of a factor, and more generally, why the interpretability of an
entire factor matrix could be considered best when (8) is a maximum.
Criterion (8) is the original raw varimax criterion [4]. In the original
proposal of this criterion, it was shown to be mathematically equivalent,
in the orthogonal case, to minimizing
(9) c* = ~ , { In ~_, a ;2, a ~;, _ ( ~ , a~,) ( ~., a~ ,) ]/n 2 }
• <t i i i

i.e., minimizing the covariance of pairs of columns of squared loadings and

 HENRY F. KAISER 191

summing over all possible pairs of columns for the criterion. Criterion (9)
then bears the analogous relationship to Carroll's criterion (1) that the
varimax criterion (8) does to the quartimax criterion (6).
Some distinctions between quartimax and varimax orthogonal solutions
can be illustrated numerically. In Table 1 solutions for Thurstone's eleven-
variable box problem ([12], pp. 373-375) are given. It will be noted that the
quartimax solution [7] could hardly be called a simple structure. There is
a large general factor, and the second factor seems only vaguely concerned

Table I

Thurstonels ll-Variable ~ox Froblem a

Subjective quartimax Raw Varimax

(oblique)

Test " X Y Z X Y Z X Y Z

x 9o o5 oo 68 65 05 91 19 16

y 04 85 O1 83 -47 O0 05 93 25

z 03 05 79 42 -08 79 ii 17 88

xy 62 63 -06 99 ii -OL 6h 74 2O

yz -o5 5h 57 71 -ho 56 o2 65 75
x2y 82 37 -01 92 41 03 8h 51 22

xy 2 35 76 02 96 -18 03 37 86 28

2x + 2y 53 71 -09 i00 O0 -07 5h 82 18

(x2 * 22) ~ 52 71 -08 99 -O1 -07 53 81 18

(x 2 * z2) ½ 52 -07 65 59 38 68 60 09 77
xyz 42 h3 L3 88 Oh L5 L8 58 65

aDecimal points omitted.

with dimensions of boxes. On the other hand, the raw varimax solution
closely parallels Thurstone's original subjective solution, given the restric-
tion of orthogonality.
In Table 2 are solutions for Holzinger and Harman's 24 psychological
tests ([3], pp. 229-233). Both the quartimax [7] and the raw varimax methods
seem to duplicate the subjectively rotated simple structure patterns. But
the respective variance contributions of the factors are perhaps more interest-
ing. It is seen that the dispersion of the ~i a~ for the subjective solution
is less than the corresponding figures for the two analytic methods. In other
words, Holzinger and Harman have made the factors a little more level or
 192 PSYCHOMETRIKA

Table 2

Holzinger and Harman's Twenty-four Psychological Tests a

Subjective Quartimax Raw Varimax Normal Varimax

Test A B C D A B C D A B C D A B C D

i i0 32 62 20 37 19 60 07 2~ 20 65 13 lh 19 67 17
2 07 15 41 13 24 07 38 04 17 07 42 08 IO 07 43 IO
3 iO 12 53 13 31 el 48 el 22 02 52 06 15 02 5h 08
4 15 18 53 12 36 07 h6 -oi 27 08 52 Oh 20 09 ~h 07
5 75 15 26 15 81 14 -02 -04 78 21 12 06 75 21 22 13
6 72 05 28 25 81 03 OO 06 78 i0 13 lh 75 IO 23 21
7 81 08 27 Ii 85 07 -Oh -IO 8h 15 IO O0 82 16 21 08
8 54 26 38 14 66 20 20 -04 60 25 31 05 54 26 38 12
9 76 -04 29 30 86 -06 -02 io 84 el 12 19 80 el 22 25
IO 28 66 -19 I~ 23 70 -12 Ii 17 71 -08 17 15 70 -06 24
II 27 61 -Oh 29 31 62 el 23 22 63 06 29 17 60 08 36
12 13 72 09 03 16 69 19 -oi 06 70 23 Oh 02 69 23 ii
13 2h 63 31 02 35 57 32 -08 2h 59 39 -01 18 59 41 06
14 23 19 -02 48 32 19 -03 42 26 20 el h6 22 16 04 50
15 ii l~ 08 50 25 ii I0 h5 17 ii 13 48 12 07 lh 50
16 05 22 34 h5 29 13 37 37 17 13 41 41 08 I0 41 43
17 15 24 -03 62 28 24 02 57 20 23 05 61 14 18 06 64
18 el 39 20 52 22 32 30 h7 08 31 32 51 09 26 32 5h
19 12 22 18 39 28 18 19 32 19 18 22 36 13 15 2h 39
20 31 18 46 29 52 09 35 14 42 12 43 21 35 Ii h7 25
21 17 46 33 2h 35 38 35 14 23 he 40 20 15 38 42 26
22 31 12 40 40 53 04 30 26 hh 06 37 32 36 oh 41 36
23 31 29 54 25 55 19 4h 09 43 21 52 16 35 21 57 22
2h 39 46 14 31 49 43 io 20 40 h6 18 27 34 h4 22 34
~a~ 343 292 268 236 559 242 196 142 431 260 26~ 186 350 244 308 236
j -js

aDecimal points o~itted.

even in their contribution to variance than the analytic criteria. Of the

two analytic criteria., the raw varimax solution has given a solution which is
closer in this respect to Holzinger and Harman's. It is also noteworthy that
as a result of these differences the large loadings of the factors with the larger
variance contributions for the analytic methods are larger than the large
loadings for the smaller factors, and similarly, the small loadings for the
larger factors are larger than the small loadings for the smaller factors.
Holzinger and Harman's subjective solution does not show this svstematie
bias; their solution gives a more equitable patterning of factor loadings.
How this bias may be removed is indicated in the next section. This
leads to a revision of the varimax criterion, which appears to have more
important characteristics than merely satisfying the rules of simple structure.
Factorial Invariance: Normal Varimax
It seems reasonable to attribute the systematic bias seen in both the
quartimax and varimax solutions of the Holzinger-Harman data and other
examples [4] to the divergent weights which implicitly are attached to the
tests by their communalities. When one deals with fourth-power functions
 HENRY 1% KAISER 193

of factor loadings, a test with comnmnality 0.6, for example, would tend to
influence the rotations four times as much as a test whose eommunality was
0.3. Thus, while the most obvious weights have been applied to the tests,
namely the square roots of their communalities, after the fact it seems that
there is probably a better set of weights--weights which would tend to
equalize to a greater extent the relative influence of each test during rotation.
There seems no rational basis for choosing among different weighting
schemes. Let us then make the agnostic confession of ignorance which pervades
any form of correlational analysis. For the purposes of rotation, weight
the tests equally, in the sense that the lengths of the common parts of the
test vectors have equal length. (The author is indebted to Dr. D. R. Saunders
for this suggestion.) The varimax criterion could then be rewritten as
(10) v = ~ {In ~ (a~Jh~) 2 - [~_~ (a~8/h~)]2]jn2},
, i

where h~ is the communality of the jth test. In contrast to (7) and (8), where
the variance of the squared correlations of the tests with a factor is maximized,
the variance of the squared correlations of the common parts of the tests
(the reflections of the tests onto the common-factor space) with a factor is
now being maximized. [Note from (10) that we are not advocating a permanent
weightia~g of the tests by a weight inversely as the square root of their com-
munalities. During rotation this weighting extends the common part of
each test vector to unit length, but after rotation each of these vectors is
shortened to its proper length by reweighting directly as the square root of
the test's communality.]
As will be seen in Table 2, under this modification the varimax criterio~
(the normal varimax, since rotation is with respect to normalized common
parts of tests) has effectively removed the small but disturbing bias in the
raw varimax solution of Holzinger and Harman's example. It also has been
shown in a number of other examples [61 that the normal varimax does not
seem to deviate systematically from what may be considered the best
orthogonal simple structure.*
Thus far, however, merely a numerical-intuitive basis for a weighting
procedure which leads to "prettier" results has been provided. Such a basis
is quite unsatisfactory theoretically. Indeed, this sort of ad hoc thinking
could conceivably lead to a different set of judgmentally determined weights
for any particular example--a situation as scientifically reprehensible as
the subjective graphical methods.
There is a more fundamental rationale for attempting to establish the
normal varimax criterion (10) as a mathematical definition for the rotation
*Professor Andrew Comrey has apparently reached the same conclusion in an exten-
sive application of the normal varimax criterion to interitem correlation matrices of the
MMPI (personal communication). A further example, available from the writer, is the
normal varimax solution of Thurstone's classic PMA study[ll] (dittoed).
194 PSYCHOMETRIK&

problem. Consider the situation illustrated in Fig. 1. There are two clusters
of tests, each of which is pure in the sense t h a t the reflections of the test
vectors of the cluster onto the two-dimensional, common-factor space are
eollinear. (While these clusters are drawn less t h a n 90 ° apart, the following
a r g u m e n t is perfectly general.)

"n"

='n-
Fzocm~ 1
Case for which a normal varimax solution is invariant under changes in the composition
of the test battery.

I t is shown below t h a t the angle of rotation in a plane which maximizes

(10) is

(11) ~b = ~ arctan + ' '

E (u~ - :~) - [(E u,): - ( E ~;):] '
i i i

where
u+ = ( a i l / h + ) ~ - (a~2/h+) 2,
and
vi = 2 ( a ~ l / h i ) ( a i J h i ) .

Let nA (n~ > 1) be the n u m b e r of tests in the first cluster and nB ( n B > 1)
be the n u m b e r of tests in the second cluster (n = nA =k nB). I t is readily
a p p a r e n t t h a t all tests of the first cluster have the same values for u, and v~ .
 HENRY F. KAISER 195

Let these values be UA and v , . Similarly let the values for the second cluster
bc uB and v, . In this case (11) reduces to
2nAns(u,VA + u~v8 -- naY, -- u,va)
(12) ¢ -- ~ a r c t a n n A n , ( u 2 + us2 -- vA2 -- vB2 -- 2 u a u , + 2vAv,) *

A most important result is shown in (12). The n A n , term may be

cancelled, indicating that the angle of rotation does not depend on the
number of tests in each cluster, i.e., Jot the case illustrated in Fig. 1, the normal
varimax solution is invariant under changes i n the composition of the test battery.
This invariance property would seem to be of greater significance than
the numerical tendencies of the normal varimax solution to define math-
ematically the doctrine of simple structure. Although factor analysis seems
to have many purposes, fundamentally it is addressed to the following
problem. Given an (infinite) domain of psychological content, infer the
internal structure of this domain on the basis of a sample of n tests drawn
from the domain. The possibility of success in such inferences is obviously
dependent upon the extent which a factor derived from a particular battery
or sample of tests approximates the corresponding unobservable factor in
the infinite domain. If a factor is invariant under changing samples of tests,
i.e., shows factorial invariance ([12], pp. 360-361), there is evidence that
inferences regarding domain factors are correct.
The normal varimax solution, according to the above result, allows such
inferences; regardless of the sampling of tests, for the problem shown in Fig.
1 it is possible to infer precisely the domain normal varimax factors. This is
not true for either the quartimax or raw varimax solutions since the angle of
rotation is a function of nA and ne .
Note that domain normal varimax factors are not said to be more
meaningJul than domain factors according to some different criterion; it is
suggested that observed normal varimax factors will have a greater likelihood
of portraying the corresponding domain factors.
Although one often gets the impression that simple structure is the
ultimate criterion of a rotational procedure, it is suggested here that the
ultimate criterion is factorial invariance. The normal varimax solution was
originally devised solely for the purpose of satisfying the simple structure
criteria. But the fact that it shows mathematically this sort of invariance
suggests that Thurstone's reasoning was basically directed toward factorial
invariance. The principle of simple structure may probably be considered
incidental to the more fundamental concept of factorial i::variance. This
viewpoint renders meaningless the arguments concerning "psychological
reality" of general factors, bipolar factors, simple structure factors, etc.
Admittedly, the result (12) is for a special case. The correlations among
the variables within each of the two pure clusters must form a perfect
Spearman matrix, and the reduced correlation matrix as a whole must be
 196 PSYOHOMETRIKA

of rank two. Normal wu'imax does however give invariant results for two
such Spearman clusters simultaneously, and consequently the normal varimax
criterion is a two-dimensional generalization of the classic Spearman case.
Obviously, the next step would be a generalization along the same lines to
the r-dimensional ease; thus far, however, work on this problem suggests
virtually insuperable mathematical difficulties.
To investigate numerically the tendency of the criterion to give factoriatIy
invariant solutions for r larger than two, again consider Holzinger and
Harman's empirical data. Taking their centroid loadings, the first five tests
were rotated then the first six, etc., systematically until the analysis of
all 24 tests was reached, as in Table 2. The results of this application of the
normal varimax criterion are given in Table 3. There, the normal varimax
loadings for the four factors as a function of the changing number of tests
are given.
Note that factors A and C are essentially invariant from the outset;
the loading changes, while somewhat systematic, are negligible--24 appears
to be essentially infinite. On the other hand, factors B and D show more
movement before they become stable. The reason for this is readily apparent.
For both B and D, this movement occurs only while they are underdetermined,
i.e., only while they contain no appreciable Ioadings. However, once they
pick up a test or two with high loadings, their ambiguous definition abruptly
stops, and they settle down to exhibit a degree of invnriance similar to the

Table 3A

Normal Varimax Loading Changes for Holzinger and Harman's Factor A (n = 5, 6, ..., 24) a

Test 5 6 7 8 9 I0 ii 12 13 lh 15 16 17 18 19 20 21 22 23 2}~

i 19 19 18 19 18 16 16 16 16 16 15 16 15 16 16 15 15 15 14 lh
2 12 12 12 13 12 12 12 12 12 ii ii Ii Ii 12 ii Ii II Ii i0 i0
3 15 16 15 16 16 17 17 17 17 16 16 16 16 17 17 16 16 16 15 15
4 21 21 21 22 21 21 21 21 21 21 21 21 21 21 21 20 20 20 20 20
5 79 78 78 79 78 76 75 76 76 76 75 75 75 76 76 75 75 75 75 75
6 78 77 77 78 77 76 77 77 76 75 75 75 75 75 75 75 75 75 75
7 83 8h 83 82 82 82 82 82 82 82 82 83 82 82 82 82 82 82
8 59 58 55 55 55 55 55 55 55 55 55 55 55 55 55 54 5h
9 82 83 82 83 83 81 80 80 80 80 80 80 80 80 80 80
iO 18 15 17 17 16 16 16 15 16 16 16 16 15 16 15
II 19 21 21 19 18 18 18 18 18 18 18 18 18 17
12 03 02 03 03 03 03 03 03 03 03 03 O2 O2
13 I~ 19 19 20 20 20 20 19 19 19 18 18
14 23 21 22 21 22 22 22 22 22 22 22
15 ii 12 ii 12 12 12 12 12 12 12
16 09 09 iO 09 09 09 09 O9 O8
17 13 14 lh 14 lh lh lh 14
18 Ol Ol OO OO oo OO OO
19 13 13 13 13 13 13
20 35 35 35 35 35
21 16 16 16 15
22 36 36 36
23 35 35
2h
34
 ~ 0

O~
II,
I I I 0
~0000
Co

~o 0~

~ o ~ o

P
~
o
i
~ 0 ~ 0 0 0 ~

~ o ~ o o o ~
¢0
I
~ ~ o ~ w ~ o o ~

i
~o ~ o o ~ ~
o ~ ~ o ~ o
i
~ o ~ o o ~ ~

cl
~
o
i c~ 1
~o ~o ~ o o ~ ~ o ~ 0 ~ ~ ~ 0 ~ 0 0 0 ~ c~
~ ~ o ~ o ~

~ o ~ o ~ o o ~ ~
5" !
H

~ o ~o ~ o o ~ ~ I
~ ~ ~ o ~ o ~ o ~ o ~ o ~ o ~ o o 0 ~
~ ~ ~ 0 ~

~ o ~o ~ o o ~ ~ I
~ ~ ~ 0 ~ 0 ~ ' ~ P ~ o o ~ ~ ~ o ~ o o o ~
~ o ~ ~ ~ 0 ~ P
~ o ~o ~ o o ~ ~ 0 ~ o ..... o g ~ o S ~ o o 0 ~
~ ~ ~ o ~

~ o ~ o o ~ ~ o ~ o~ ~ o~
o g o~
!
 198 FSYCHOMETRIKA

Table 3D

Nodal Varimm.x Loading Changes for Holzinger and Ha~man's Factor D (n = 5, 6 . . . . . 2L) a
;q

Test 5 6 7 8 9 i0 ii i2 13 lh 15 16 17 18 19 20 -~! 22 23 24

1 Ol -00 O0 Ol -07 Ol Ol Ol 02 08 13 i4 15 17 19 18 i~ i? 17 17
2 Ol Ol 02 Ol O0 02 02 02 Ol 05 08 09 09 l0 ii ii i0 lO i0 iO
3 -00 Ol Ol ~OO 05 -00 Ol -00 -01 02 06 07 07 08 09 09 O~ 08 08 08
-03 -03 -03 -Oh -01 -04 -03 -oh -Oh Ol 05 06 06 o7 09 08 o~ 08 o? o7
5 -00 -06 -0] -02 -09 -06 i06 -05 -05 07 12 12 lh lh 15 lh 1k ih lh 13
6 06 09 09 05 05 o6 06 05 17 21 21 22 21 22 22 22 22 22 21
7 -06 -06 -06 -i0 -I0 -09 -i0 03 07 O? O9 08 09 O9 O9 08 08 O8
8 -04 -lh -07 -08 -07 -07 05 iO iO 12 12 i4 13 13 i3 12 i2
9 15 II 12 12 Ii 22 26 26 27 25 26 26 26 26 25 25
I0 -00 -07 O0 02 15 19 19 23 24 25 25 25 2L 2h 2J,
L1 07 12 15 27 31 32 35 36 37 37 37 J? 37 36
12 -lh -ii -00 05 05 09 ii 13 12 i2 I_1 I_I ii
13 -17 -05 oo oi o4 07 08 08 07 07 07 o6
L,7 ~9 h9 50 50 50 50 50 50 50 50
15 L9 49 50 50 50 50 50 50 50 50
16 )~2 L2 L~ &5 L~4 L~ ~L /45 ~3
17 6h 65 6h 6£ 65 6h 6L 65
18 5L 55 55 5b 51~ ~ 94
19 ho &o }40 39 39 39
2O 26 26 26 25 25
21 27 26 26 26
22 36 36 36
23 22 22
2~ 3~

aDeci~al points omitted.

other two factors, which had high loadings from the beginning. For n = 24,
there appear to be good approximations to the domain normal varimax
factors.

The Oblique Case

If the restriction of orthogonality is relaxed, it is impossible to apply
directly the quartimax criterion (4) or the normal varimax criterion (t0).
This is because interfactor relationships are not considered when the criteria
are in this form, and when applied all factors will collapse into the same
factor--that one factor which best meets the criterion. However, Carroll's
version of the quartimax criterion explicitly considers interfaetor relation-
ships and an oblique solution is attainable. As suggested by (9), if
(13) c = E {In E (a~,/h~)(a~/h~) - ( E a~/h~)(E a~tlh~)]/n~},
• <t i i i

it may be shown that in the orthogonal ease v = - 2c. This alternative form
of the normal varimax may then be used to obtain oblique factors. The
mathematical problem of minimizing (13) is exactly analogous to Kaiser's
[5] treatment for Carroll's criterion. Comput~tionally, the (iterative) solution
involves finding the latent vector associated with the smallest latent root of
a constantly changing symmetric matrix of order r.
 HENRY F. KAISER 199

Computational Appendix
To compute an orthogonal normal varimax solution, the following
procedure is suggested. The first step is to normalize the rows of the arbitrary
reference factor matrix (e.g., principal axes or centroids) by dividing each
element by h;. Rotation to the direction of the normal varimax factors may
then be carried out with respect to these normalized loadings.
The criterion (10) will be applied to two factors at a time. For this purpose,
the following notation for an orthogonal rotation is convenient.

where x; and y~, the present normalized loadings, are constants, and X;
and Yi , the desired normalized loadings, are functions of ¢, the angle of
rotation.
It is immediately seen that
(14) X~ = x; cos¢ -b y~ sin ¢,
(15) Y; = - x t sin ¢ -{- yj cos ¢.
Thus,
(16) dXJd¢ = Y~,
(17) d Y~/d¢ = - X i .
According to (10), in this plane,
(18) n 2 v~ -- n ~ (X2) 2 -- ( ~ X2) 2 -~ n ~ (y2)2 _ ( ~ y~)2
should be a maximum. Differentiating (18) with respect to ¢, using (16) and
(17), and setting the derivative equal to zero,
(19) n ~ XY(X 2- y2) _ ~..Xy ~ (X ~ - Y~) = O.
To solve (19) for ¢ in terms of xi and y~, substitute the values of X; and
Y~ from (14) and (I5), consult a table of trigonometric identities, and, after
a good deal of algebraic manipulation,
(20) ¢ = ¼ arctan
2[~ ,, E (x ~ ~ y 2) ( x2 y ) - - E (x ~ y~) E (2xy)]
- -

~{ E [(2 _ y~)2 _ (2xy)~]} _ { [ E (~ - y~)]~ - [ E (2~y)]~}"

If ui = xi,2 -- Yi2 and vi = 2xiy~, (20) reduces to the form (i1) above.
Of course, (11)or (20) is only a necessary condition for a maximum.
By taking the second derivative of (18) sufficient conditions for a maximum
200 PSYCHOMETRIKA

m a y be found. These are summarized below.

sign of numerator
-}-

0 ° to 0 ° to
sign of "~22½° --22½°
denominator +22½ ° -22½ °
to ~-45 ° to --45 °

T h e sign of n u m e r a t o r a n d d e n o m i n a t o r refer to t h e r i g h t - h a n d m e m b e r
of (20); t h e values in the cells refer to ¢.
These single-plane rotations are m a d e on factors 1 with 2, i with 3, • • • , 1
w i t h r, 2 with 3, • .- , 2 with r, • • • , (r -- 1) with r, 1 with 2, - . . iteratively
until r(r - 1)/2 successive rotations of ¢ = 0 are obtained, i.e., until the
process converges. (It was shown [6] t h a t v in (10) c a n n o t be greater t h a n
(r - 1)/r, a n d since each successive application of (20) can result 0nly in
a non-decrease of v, this iterative procedure m u s t converge.) A f t e r con-
vergence, each normalized test v e c t o r is restored to its proper length b y
m u l t i p l y i n g b y ha .
Since this article was accepted for publication, t h e a u t h o r has p r e p a r e d
,~ detailed outline for coding an electronic c o m p u t e r p r o g r a m for tlle v a r i n ~ x
criterion. This (dittoed) p a p e r is available f r o m t h e writer.
REFERENCES
[1] Carroll, J. B. An analytical solution for approximating simple structure in factor
analysis. Psychometrika, 1953, 18, 23-38.
[2] Ferguson, G. A. The concept of parsimony in factor analysis. Psychometrika, 1954,
19, 281-290.
[3] tIolzinger, K. J. and Harman, H. H. Factor analysis. Chicago: Univ. Chicago Press,
1941.
[4] Kaiser, H. F. An analytic rotational criterion for factor analysis. Amcr. Psychol-
ogist, 1955, 10~ 438. (Abstract)
[5] Kaiser, H. F. Note on Carroll's analytic simple structure. Psychometrika, 1956, 21,
89-92.
[6] Kaiser, H. F. The v~rimax method of factor analysis. Unpublished doctoral disserta-
tion, Univ. California, 1956.
[7] Neuhaus, J. 0. and Wrigley, C. The quartimax method: an analytical approach to
orthogonM simple structure. Brit. J. statist. Psychol., 1954, 7, 81-91.
[8] Saunders, D. R. An analytic method for rotation to orthogonal simple structure.
Princeton: Educational Testing Service Research Bulletin 53-10, 1953.
[9] Thomson, G. H. The factorial analysis of human ability. (Sth ed.) New York: Houghton
Mifflin, 1951.
[10] Thurstone, L. L. Theory of multiple factors. Ann Arbor: Edwards Bros., 1932.
[11] Thurstone, L. L. Primary mental abilities. Psychometric Monogr. No. 1, 1938.
[12] Thurstone~ L. L. Multiple-factor analysis. Chicago: Univ. Chicago Press, 1947.
Manuscript received 11/15/57
Revised manuscript received 1/9/58