EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT

1974, 34, 111-117.

LITTLE JIFFY, MARK 1

IV

HENRY F. KAISER
University of California, Berkeley
and
United States Coast Guard Academy
JOHN RICE
University of California, San Diego

IN this paper three changes and one new development for the
method of exploratory factor analysis (a second generation Little
Jiffy) developed by Kaiser (1970) are described. Following this
short description a step-by-step computer algorithm of the revised
method-dubbed Little Jiffy, Mark IV-is presented.
Extensive empirical experience with &dquo;a second generation Little
Jiffy&dquo; (Kaiser, 1970) indicates that the method, for large matrices,
consistently mildly underfactors. A revision is called for. Thus, the
writers adopt as the answer for the crucially important question of
the &dquo;number of factors&dquo; Guttman’s (1954) classic weaker lower
bound, the index of the covariance matrix (with zeros in the diago-
nal) under consideration. This answer is the same as that given by
Kaiser’s (1956, 1960, 1970) extensively used &dquo;eigenvalues greater
than one of R.&dquo;
In a second generation Little Jiffy, for the transformation prob-
lem, the writers believe that Kaiser’s (1970) winsorizing procedure
is unnecessary, and that it would be better to allow possibly a little
distortion in the explication of Thurstone’s simple pattern and to
retain, undisturbed, the elegance (factor-analytic model-free, trans-
formation method-free, column orthogonality) of the Harris-Kaiser
(1964) independent cluster solution. The column orthogonality of
this solution renders legitimate the interpretation of factors as linear
1
This research was supported in part by the Office of Computing Ac-
tivities, National Science Foundation.
111
112 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT

combinations of the original variables (or their common parts, or

their images) given by the columns of the pattern matrix, in contrast
to the typical solution. This column orthogonality also allows one to
determine additive relative contributions of factors, in contrast to
the typical solution involving correlated factors.
To assess the simplicity of the transformed pattern matrix (and
its possible distortion from an ideal simple pattern), an &dquo;Index of
Factorial Simplicity&dquo;

has been developed (Kaiser, .1974), where q is the number of factors

and v,., is an element in the quartimax transformed Harris (1962)
(unit-length) eigenvectors. An analogous IFS(J) for each row fol-
lows by removing the 2~ in the numerator and the denominator of
the above formula. Any IFS must lie between zero and one, attain-
ing its maximum value of one only under perfect unifactoriality.
Subjective appraisal, based on extensive experience, suggests that
these Indices of Factorial Simplicity may be evaluated according to
the following table:

in the .90s marvelous,

in the .80s meritorious,
in the .70s middling, (2)
in the .60s mediocre,
in the .50s miserable,
below .50 unacceptable.
The above two changes and one new development constitute the
revision to a second generation Little Jiffy dubbed &dquo;Little Jiffy,
Mark III&dquo; in an unpublished paper. The following change estab-
lishes &dquo;Little Jiffy, Mark IV.&dquo;
Measures of Sampling Adequacy, MSA, as defined by Kaiser
(1970), have proved unstable for poor data. Professor Ingram
Olkin (personal communication) has suggested that this instability
may be corrected by taking a normalized MSA (so that it must lie
between zero and one) :
 KAISER AND RICE 113

overall, and

for each variable separately, where rik is an original correlation and

qjk is an anti-image correlation. Extensive numerical experience with
these measures suggests that they also may be evaluated according
to the above table (2). The borderline of acceptability here, again
.50, occurs when

where a careful reading of Guttman (1953) indicates that it

should be.
This paper appended to Kaiser’s (1970) original paper constitutes
the final version of Little Jiffy: Little Jiffy, Mark IV.

Computer Algorithm
1. Input and output the correlation matrix R of order p, or input
the score matrix Z of order n X p, where n is the number of in-
dividuals. R must be positive definite so that n must be greater than
p (and also no variable can be a linear combination of other vari-
ables). If scores are input, compute and output the vector of means
AVE, the diagonal matrix of standard deviations SD, and the cor-
relation matrix R of the variables, and write the original score
matrix on auxiliary storage for use at the end of the program in
computing factor scores. (If there are missing observations in the
score matrix, one wants to insert something in the missing slots to
insure that R is positive definite. It is recommended that column
means from the available scores be inserted; since most funda-

mentally below one is working with the image score matrix, one is
thus obtaining regression estimates of the missing observations.)
2. Find
114 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT

f or j 1, 2, ’’’, p, and output RMS (J), the root-mean-square cor-

=

relation of each variable with the remaining p - 1 variables and

RMS, the overall root-mean-square correlation.
3. Saving R, find R-1. If R is singular, go to the end of the pro-
gram ; the problem cannot be completed. Let 82 = (diag R-1) -1, the
diagonal matrix of anti-image variances. Compute and output
SMC (J) = 1 - si~2, j = 1,2, .. &dquo;p, the squared multiple correlation
of each variable on the remaining p - 1, &dquo;best possible&dquo; systematic
estimates of communalities (Guttman, 1956). Write R-1 on auxiliary
storage for use in computing the factor score weight matrix at the
end of the program.
4. Find the anti-image correlation matrix Q = SR-1S and then
determine

5. Now compute

for j 1, 2, p. MSA (J) is the revised (Mark IV) Measure of

=

Sampling Adequacy for the jth variable, and MSA is the revised
(Mark IV) overall Measure of Sampling Adequacy.
6. Find the eigenvalues of R and let q equal the number of
eigenvalues greater than one; q is the number of factors.
7. Form Harris’ (1962) covariance matrix, C B-1RB-t, and then
=

find the diagonal matrix M2 of the q largest eigenvalues (Harris

eigenvalues) of C and the associated q unit-length column (Harris)
eigenvectors. Let E be the p X q matrix of these eigenvectors.
8. Saving a copy of E, apply the (raw) quartimax criterion
(Wrigley and Neuhaus, 1954) to E, calling the result V. An algorithm
for the quartimax method is given by three changes in Kaiser’s
(1959) algorithm for the varimax criterion. In that write-up (a) omit
steps 2 and 17; (b) in step 5 let tan 4qJ A/B; and (c) tighten the
=

convergence criterion in step 11 to at least 8 < 10&dquo;~.

9. For each column of V find the sum of cubes and, if the sum is
negative, change the signs of all the entries in the column.
10. Find T = E’V, the quartimax transformation matrix.
11. Replace the diagonal elements of the diagonal matrix M2 of
Harris eigenvalues with
 KAISER AND RICE 115

for s = 1, 2, q. Call this new diagonal matrix M2, also. It con-

tains the Harris eigenvalues of the Harris rescaled image covariance
matrix.
12. Form LSTAR = T’M2T. LSTAR is the factor intercovariance
matrix. Let D2 = diag (LSTAR) be the diagonal matrix of &dquo;natural&dquo;
factor variances.
13. Reorder elements of D2in order of decreasing size, and re-
order columns of V and rows and columns of LSTAR accordingly.
14. From V find the IFS(J) and IFS

for j = 1,2,..., p.
15. Multiply each element of V by B/p&horbar;a step thus changing the
loadings from column-normalized to column-standardized. Output
this column-standardized factor pattern matrix. Return it to its
original column-normalized version V by dividing each element
by Vp. Now output IFS(J) and IFS.
16. For each factor compute

for 8 = 1, 2, ’’’, q, and output these relative variance contributions

of factors (percentages).
17. Find the factor intercorrelation matrix,

and output
18. a. Form A =
SVD, the factor pattern matrix (conventionally
scaled).
b. Form and output B =
AL, the factor structure matrix.
c. Output A, asterisking the salients-those elements of A which
116 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT

are larger than one in the column-standardized factor pattern

matrix V.
19. Calculate the standardized factor score weight matrix W :

and output. (Note that (S-2 - R-1) is just -R-1 with zeros in the
diagonal.)
20. Determine the domain validities (indices of reliability) of the
factors, DV (S) :

for s = 1, 2, &dquo; ’, q, and output.

If a correlation matrix was input at the beginning, the problem is
finished. If scores were input, the program goes on to compute
(image) factor scores. The formulas used below will be developed
in detail in a forthcoming paper.
21. Rescale the standardized factor score weight matrix W to be

where (SD) is the diagonal matrix of original standard deviations.

Call this new factor score weight matrix W, also. It has been re-
scaled to accept the original raw scores and to produce factor scores
with standard deviations equal to 100.
22. To calculate factor scores, one goes through the following loop
for each individual:

a. Read in a row vector z’ of p (raw) scores for a given individual.

b. Subtract the vector of original means AVE from z. Call these
centered scores z, also.
c. Let x W’z.
=

d. Find x + 500, the desired factor scores for one individual.

These factor scores have mean 500 for all factors.
e. Print and punch the vector of factor scores for one individual.
f. Go back to (a) for the next individual’s observation vector.

REFERENCES
Guttman, Image theory for the structure of quantitative variates.
L.
, 1953, 18, 277-296.
Psychometrika
Guttman, L. Some necessary conditions for common factor anal-
ysis. Psychometrika
, 1954, 19, 149-161.
 KAISER AND RICE 117

Guttman, L. Best possible systematic estimates of communalities.

, 1956, 21, 273-285.
Psychometrika
Harris, C. W. Some Rao-Guttman relationships. Psychometrika
,
1962, 27, 247-263.
Harris, C. W. and Kaiser, H. F. Oblique factor analytic solutions
by orthogonal transformations. Psychometrika, 1964, 29, 347-
362.
Kaiser, H. F. The varimax method of factor analysis. Unpublished
doctoral dissertation, University of California, Berkeley, 1956.
Kaiser, H. F. Computer program for varimax rotation in factor
analysis. EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT, 1959,
19, 413-420.
Kaiser, H. F. The application of electronic computers to factor
analysis. EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT, 1960,
20, 141-151.
Kaiser, H. F. A second generation Little Jiffy. Psychometrika
, 1970,
35,401-415.
, in
Kaiser, H. F. An index of factorial simplicity. Psychometrika
press, 1974.
Wrigley, C. W. and Neuhaus, J. O. The quartimax method: An
analytical approach to orthogonal simple structure. British Jour
nal of Statistical Psychology, 1954, 7, 81-91.