bruce thompson-exploratory and confirmatory factor analysis_ understanding concepts and...
TRANSCRIPT
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 1/185
E X P L O R A T O R Y
AND
C O N F I R M A T O R Y
F C T O R
N L Y S IS
Understanding
oncepts
and
Applications
B R U C E T H O M P S O N
American Psychological Association • Washington DC
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 2/185
Copyr igh t © 2004 by the American Psychological Associat ion. A ll r ights reserved.
Except as pe rmi t t ed under th e U n i t e d States C o p y r i g ht Ac t of 1976, no part of th i s
pub l i c a t i on may be
reproduced
or
d i s t r i bu t e d
in any f o rm or by any
means ,
o r
stored
in a
database
or
re t r i eva l sys tem, wi tho u t
th e
p r io r wr i t t en pe rmiss ion
of the
publ i she r .
Publ ished by
American Psychological Associat ion
750 First Street, N E
Washing ton ,
DC
20002
www.apa .o rg
T o order In the U .K. , Europe , Af r i ca , and the Mi d d l e
A P A Order
Depa rtmen t East , copies
may be
ordered
from
P.O. Box
92984
Am erican Psychological Associat ion
Wash ing ton , D C
20090-2984
3 H e n r i e t ta Street
Tel: (800)
374-2721
Covent
Garden, London
Direct : (202) 336-5510
WC2E
8LU
Eng land
Fax:
(202) 336-5502
TDD/TTY:
(202) 336-6123
Online:
www.apa.org/books/
E-mai l : o rde [email protected] rg
Typeset
in
G o u d y
by
Wor ld
Composition
Services , Inc. , Sterl ing,
V A
Pr in te r: Data Rep roduct ions , Au burn Hi l l s , Ml
Cover
Designer: Veronica
Brcza,
Sterl ing,
V A
Techn ica l /Product ion Ed i to r :
D an
Brach tesende
T he
opinions
and
s tatements publ ished
are the
responsibi l i ty
ot the
au thor ,
and
such
op in ions and s t a t ements do no t necessari ly represent th e pol icies of the A me r i can
Psychological Associat ion.
Library
of ongress Cataloging in Publication Data
Thompson, Bruce , 9 5 —
Exploratory
and
conf i rmatory factor analysis
:
u n d e r s t an d i n g
concepts
and
ap p l i c a t i o n s
/
Bruce Thompson .—1s t
cd .
p. cm.
Includes bibl iographical references .
ISBN 1-59147-093-5
1.
Fac to r ana lys i s—Textbooks . 1. Ti t le .
BF3 9 . 2 . F3 2 T4 8 2 0 0 4
150 ' .1 '5195354—dc22
3 596
British L ibrary Cataloguing in Publication Data
A C 1P record is av ai la ble f rom the B ri t ish
Library .
Printed in
th e United States
o f
Am erica
First Edition
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 3/185
CONTENTS
Preface ix
Chapter 1 . Introduction to Factor Analysis 3
Purposes of
Factor Analysis
4
Two
M a j o r Classes
of
Factor Analysis
5
Organizat ion
of the
Book
7
Chapte r 2 . Foundational
Concepts
9
Heuris t ic
Example 9
Two
Wor lds
of
Statistics
11
Some Basic Matrix Algebra Concepts
12
More Heurist ic Results
15
Sample Size Considerat ions
in EFA 24
Major
Concepts
25
Chapte r
3.
Exp loratory Factor An alysis Decision Sequence
27
Matr ix
of
A ssociation
Coefficients 28
N u m b e r
of Factors 31
Factor Extraction Method
36
Factor
Rotation 38
Factor Scores 44
Major
Concepts
47
Ch apter 4. Com ponents Versus Factors Co ntroversies 49
Measurement
Error Versus Sampling Error
50
Pr incipal Components Versus Pr incipal
Ax es Factors 53
Major Concepts 55
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 4/185
Chapter 5. Comparisons of Some
Extraction, Rotation,
and
Score M ethods 57
Factor Scores in Different Ext rac t ion Methods 57
Structure Coeff ic ients
i n a
Components
Context
61
Communal i ty Coef f ic ien t s a s R
Values 61
Major
Concepts
64
Chap ter 6. Obl iqu e Ro tat ions and High er-Order Factors 67
M a x i m i z i n g Sample
F it
Versus Repl icabi l i ty
69
Higher-Order Factors 72
Major Concepts 81
Chapter 7. Six Two-Mode Techniques 83
Cattell 's "Data Box"
83
Q-Technique 86
Major Concepts
91
Chapte r 8 . Conf i rmatory Rota t ion and Factor
In te rpre ta t ion
Issues
93
"Best-Fit" Ro tat ion Ex am ple 94
Factor Interpre ta t ion Considerat ions 96
M a j o r
Concepts 98
Chap ter 9. Intern al Repl icabi l i ty Ana lyses 99
EFA Cross-Validation 101
EFA Boots t rap
102
Major
Concepts 106
Chapter 10. Co nfirm atory Factor Ana lysis Decision Sequence ... 109
CFA
Versus
EFA 110
Preanalysis
Decisions 114
Postanalysis Decisions
127
M aj o r
Concepts 131
Chapte r 11. Some Confi rmatory Factor Analys is
Interpre ta t ion Principles 133
Model Ident i f icat ion Choices 134
Different Es t imat ion
Theories
138
CFA Pat tern and Structure Coefficients 138
Comparisons
of
Model Fits
142
Higher -O rder M ode ls 145
M a j o r Concepts 147
vi CONTENTS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 5/185
Chapter 12. Testing Mode l Invar iance 153
Same M ode l , Different Est imates 153
S am e M ode l , Equa l Es t imates 155
Same M ode l , Test Invar iance of Param ete r s 159
M a jo r Concepts 159
A ppe nd ix
A:
L ibQU A L + ™ D a ta
163
A ppe nd ix
B :
SPSS S yn t ax
to
E xe cu t e
a
"Best-Fit"
Fac tor Rota t ion 169
Notation 177
Glossary
179
References
183
I nde x 191
About the A u t h o r 195
CONTENTS
vii
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 6/185
PREFACE
Investigation
of the
s tructure under lying var iables
(o r
people,
or
t ime)
has intrigued social scientists since th e early origins of psychology. Conduct -
in g one s
first
factor analysis can yield a sense of awe regarding th e power
of these methods
to
inform judgment regarding
the
dimensions underly-
ing constructs.
This book presents the impor tant concepts required for implement ing
two disciplines of factor analysis: exploratory factor analysis (EFA) and
confirmatory factor analysis (CF A). T he book may be uniq ue in its
effort
to present both analyses within the single rub ric of the general l inear model.
Throughout th e book canons of best factor analytic practice are presented
and
explained.
The book has
been
wr i t ten to str ike a happy medium between accuracy
and completeness versus overwhelm ing technical complexi ty .
A n
ac tua l
data set, randomly drawn
from a
large-scale international study (cf.
Cook
6k Thompson, 2001) involving faculty and graduate student perceptions of
academ ic libraries, is presented in Appe ndix A. Throughout the book differ-
ent combinations of
these
variables and participants are used to il lustrate
EFA
and CFA
applications.
Real u nde rstand ing of s tat is t ical concepts comes wit h applying these
methods
i n
one s
o w n
research.
The
next-best learning opportunity involves
analyzing
real data provided
by
others.
You are
encouraged
to
replicate
the
analyses
reported
in the
book
and to
conduct methodological var iat ion s
on
the reported analyses. You also may find it useful to conduct complementary
analyses to
those reported here (e.g.,
if 11
variables from
100
faculty
are
analyzed, conduct
th e
same analysis using data from
the 100
graduate
s tudent s ) .
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 7/185
There are also other data sets that have a long history of use in
heuristics
in the factor analytic context . For exam ple, the d ata reported hy
Holz inger
and Swineford (1939) are widely reported. Readers are encouraged
to access and analyze these and other data sets.
To improve the instruct iona l effect iveness of the book, some liberties
have been taken in using Am erican Psychological Association (A PA ) style.
For
example,
to
d is t inguish between score world
an d
area world statistics,
area world statistics are often presented with percentage signs, even across
the rows of a single table, for emphasis and cla rity. And more t han th e
usual num ber of decim al places may be reported, to help the reader com pare
reports in the book with the related reports in computer outputs , or to
dis t inguish
one block of results from ano ther by using different numbers of
decimal places
for
different results.
I
appreciate
in
this regard
th e
l a t i tude
afforded
me by the A PA
produc-
tion
staff.
I also app reciate the o pp ortu nit y to wr ite the book in my own
voice. This book has been wr itten as if I were speaking direc tly to yo u, the
student ,
in my own
office
or
classroom. Fortunately,
I
have
a
very
b ig
office
I
also apprec iate th e thoughtful suggestions of Randy A rnau , Univer s ity
of Southern M ississippi, Susan Crom well, Texas
A& M
Univers i ty ,
and
Xitao Fan, University of Virginia, on a
draft
of the book. Notwithstanding
their suggestions, of course I remain responsible for the book in its final form.
PREFACE
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 8/185
E X P L O R T O R Y
AND
O N F I R M T O R Y
F C T O R N L Y S I S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 9/185
1
INTRODUCTION TO
FACTOR ANALYS IS
The early 1900s were exciting times fo r social scientists. Psychology
was
emerging
as a
formal
discipline.
In
France, Alfred Binet created
a
measure
of in tel lectual performance that w as the forerun ner to modern
IQ tests. But from the very outs et there were heated controversies
regarding
definition s of intelligence and related questions abou t how to
measure intell igence.
Some psychologists took the vie w that in tel l igence was a s ingle general
abi l i ty .
Those taking th is
view
presumed
that
ind iv idua l s who w ere
very
bright in one area of performa nce tended to perform
wel l
across the full
range of in tel le ctu al tasks. This cam e to be called the general , or G, theory
of
intell igence . Other psychologists argued that individuals may be prof icient
at one
task,
bu t
that
expertise
in one
area
has
l i t t le
or no
predict ive abi l i ty
as regards other areas of cognitive performance. These latter psychologists
argued that intell igence involved specific factors that correlated min ima l ly
with
each
other.
However, in these early years psychologists who wished to test their
dive rgen t
views
had few
analy t ic tools w ith which
to do so. To
address this
void, Spearman (1904) conceptual ized methods that today we term factor
analysis.
In the intervening decades, numerous statistical developments
mean that today
the
analyst confronts
an
elaborate array
of
choices wh en
using
these methods.
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 10/185
Of
course,
for
many decades factor analysis
was
infrequently
used.
In
the era of statistical analysis via hand-cranked calculators, the computations
of factor analysis were u s u a l l y overwhelming. In part this was also because
responsible researchers conducting analyses prior
to the
advent
of
modern
computers felt obligated
to
calculate their results repeatedly, until
at
least
two
analyses
of the
data yielded
the
same results
Happily,
today the
v e r y ,
very complicated calculations of modern factor
analysis can be accomplished on a microcomputer in only moments, and
in
a
user-friendly
point-and-click environment.
As a
consequence, today
factor analytic reports comprise
as
many
as 18% to 27% of the
articles
in
some
journals (Fabrigar, Wegener, MacCallum, &
Strahan,
1999; Russell, 2002).
PURPOSES OF
FACTOR
ANALYSIS
Factor analytic methods can be used for various purposes, three of
which
are
noted
here.
First, when
a
measure
and its
associated scoring keys
have been developed,
factor
analysis can be used to inform
evaluations
of
score
validity.
V a l i d i t y questions focus
on
whether scores measure "the correct some-
thing." Measures producing scores that are completely random measure
"nothing" (Thompson, 2003).
Such
tools yield scores that are
p e r f ec t l y
unreliable. We never develop tests to measure nothing, because we have
readily available tools, such as dice and roulette wheels, that we can use to
obtain random scores. Obviously,
for
scores intended
to
measure something
(e.g.,
intelligence, self-concept),
the
scores must
be
reliable.
But r e l i a b i l i t y
is
only a necessary—not a sufficient—condition for validity.
V a l i d i t y questions involve issues such
as,
"Does
the
tool produce scores
that seem to measure the intended dimensions?" and "Are items intended
only
to
measure
a
given dimension actually measuring
and
only measuring
that dimension?" These are exactly the kinds of questions that factor analysis
w as created to confront. So it is not
s u r p r i s i n g
that historically "construct
v a l i d i t y
has
[even]
been
spoken
of as ...
'factorial validity'
"
(Nunnally,
1978, p.
111).
Joy Guilford's discussion some
50
years
ago
illustrates this application:
Val id i ty , in my opin ion is of two
kinds .
. . . The factorial validity
[italics
added] of a test is given by its . . . [factor weights on] m e a n i n g f u l ,
common,
reference
factors.
This
is the k ind of va l id i t y
t ha t
is really
m e a n t w h e n th e
ques t ion
is
asked
"Does
this test
measu re
w h a t
it is
supposed to measure? A mor e p e r t i n en t
ques t ion
should he "What
does this
test
measure?
The answer then should be in t e r m s of factors.
EXPLORATORY AND CONFIRMATORY
FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 11/185
... 1 pred ic t a t i me when a n y test a u t hor wil l b e expected to present
in forma t ion rega rd ing
t he
factor
compos i t ion of his [sic] tests. (Gui l fo rd ,
1946,
pp.
428, 437-438)
Although modern v a l i d i t y language does not recognize a type of v a l i d i t y
called "factorial validity," factor analysis nevertheless
is
h e l p f u l
in
addressing
construct v a l i d i t y questions.
Thus, Gorsuch (1983)
has
noted that
"a
prime
use of
factor analysis
has
heen
in the development of
both
the operational constructs for an area
and the
operational representatives
for the
theoretical constructs"
(p.
350).
And Nunnally (1978) suggested that "factor analysis is intimately involved
with questions
of
validity.
. . .
Factor
a n a l y s i s is at the
heart
of the
measure-
ment of psychological constructs" (pp. 1 1 2 — 1 1 3 ) .
Second, factor analysis
can be
used
to
develop theory
regarding th e
nature
of
constructs At least historically,
e m p i r i c a l l y
driven theory development
was sometimes held
in
high regard.
In
some
of
these applications, numerous
d i f f e r en t measures are administered to various samples, and the results of
factor analysis then are used to
specify
construct dimensions. This applica-
tion,
unlike the first one, is inductive.
A
classic example
is the
structure
of
intellect model developed
by
Guilford (1967). Guilford administered dozens
of
tests
of
cognitive abilities
over
years to
derive
his
theory,
in
which
he
posited that intelligence consists
of more than
100
d i f f e r en t abilities that
are
independent
of
each
other.
Third, factor analysis can be used to summarize relationships in the form
of
a
more parsimonious
set of
factor
scores
that
ca n
then
be
used
in
subsequent
analyses (e.g., analysis of variance, regression, or descriptive discriminant
analysis).
In this application, unlike the first two, the factor analysis is only
an intermediate step in i n q u i r y , and not the final analysis. Using
f e w e r
variables in substantive a n a l y s i s tends to conserve
degrees
of freedom and
improve power against Type
II
error. Furthermore, factor scores
can be
computed
in
such
a
manner that
the
unreliable variance
in the
original
variables
tends
to be
discarded
once the
original scores
are
reexpressed
in
a smaller set of factor scores.
TWO M A J O R CLASSES OF FACTOR
ANALYSIS
There are actually two discrete classes of factor analysis: exploratory
factor
analysis (EFA) and confirmatory factor analysis (CFA). Analyses such
as
those originally proposed
by
Spearman (1904) have
now come to be
called exploratory factor analysis.
In
EFA,
the
researcher
may not
have
any
specif ic
expectations
regarding
the
number
or the
nature
of
underlying
constructs
or
factors. Even
if the
researcher
has
such expectations
(as in a
INTRODUCTION TO
FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 12/185
val id i ty inv es t iga t ion ) ,
EFA
does
not
require
the
researcher
to
declare these
expectations, and the analysis is not influe nce d by these expectations.
Confi rmatory factor
analysis
methods were developed much more re-
cent ly (cf. Joreskog, 1969). These analyses requ ire that the researcher mu st
have
specific
expectat ions regarding
(a) the
n u m b e r
of
factors,
(b)
wh i ch
variables reflect g iven factors,
and (c)
whether
th e
factors
are
correlated .
CFA expl ic i t ly and directly tests the fit of factor models.
Researchers without theories cannot use CFA, but researchers with
theories usua l ly find CFA more
useful
tha n EFA. Confi rmatory factor
analysis
is more useful in the presence of theory because (a) the theory is di rect ly
tested
by the
analysis
and (b) the
degree
of
model
fit can be
quan t i f ied
in
var ious
ways.
Of course, some researchers begin an
inquiry
wi t h
specific
theoretical expectations, but after invoking CFA find that their exp ectat ions
were wi ldly wrong, and may
then
revert to E FA procedu res .
Alth ou gh factor analysis has been ch aracteriz ed as one of the m ost
powerful methods yet for reduc ing var iab le complex i ty to greater s implici ty
(Kerlinger, 1979, p. 180) and as the furthes t logical developm ent and
re igning queen of the correlat ion al methods (Ca t tel l , 1978, p . 4) , the
methods have been harshly crit icized by others (cf. Armstrong, 1967). For
example , with respect to research in the area of com m un ication , Cron khite
a n d L i s k a (1980) noted:
A p p a r e n t l y , it is so easy to f ind semantic scales which
seem
re leva n t to
[ information] sources , so easy to name or descr ibe potential/hypothetical
sources,
so
easy
to capture college
s tuden ts
to use the scales to rate the
sources , so easy to s u b m i t
those
ra t ings to fac tor ana lys is , so much fun
to name the
fac tors when
one's research
assistant re tu rns
with the
computer
p r i n t o u t , and so rewa rd ing to have a guaranteed publication
w i t h no
fe a r
of nonsignif icant resu l ts
that
researchers, once exposed to
the pleasures of the factor analytic
approach,
r a p i d l y become addicted
to it . (p. 1 0 2 )
However, these cr i t ic isms apply more
to EFA
than
to
CFA, because
the only danger in EFA would he that factors might not be in terpretable ,
in
which case
th e
resul ts
may be
regarded
as
nega t ive .
But CFA
p resumes
the invocation of specific expectations, w ith the attend an t p ossibili ty that
the predetermined theory in
fact
does not fi t , in which case the researcher
m ay be
unable
to
exp la in
the
relat ionships among
the
var iab les be ing ana-
lyzed. Both EFA and CFA remain useful today, and our selection between
the two classes of factor analysis general ly depends on wh ether w e have
specific theory regarding data s t ructure.
But even researchers who only conduct research in the presence of
such theory
mus t
unders tand EFA. Mastery
of CFA
requ i res und ers tand ing
of EFA, which
was the
his torical precursor
t o
CFA. Indeed,
a
f u n d a m e n t a l
EXPLORATORY AND
CONFIRMA TORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 13/185
purpose of this hook is to teach the reader what is common and what is
different across these tw o classes of factor analysis .
Understa nding EFA also may facili tate deeper insight
into
other statis-
t ica l
analyses .
A n
important idea
in
statistics
is the not ion of the
general
l inear model (GLM;
cf.
Bagozzi, Fornell,
&
Larcker, 1981; Cohen, 1968;
Kn a p p , 1978). Central
to the GLM are
realizations tha t a ll statistical analyses
are correlat ional and can yield effect sizes analogous t o r and apply weights
to measured variables to obtain scores on the composite variables tha t are
actual ly
t he
focus
o f all
these analyses (see Thom pson, 1991).
In
fact , because
all
parametric methods (e.g. , analysis of variance, regression, multivariate
analysis
of
var iance)
are
special cases
of
canonical correlation analysis,
and one kind of EFA is always im plic it ly invoked in canonical analysis
(Thompson, 1984,
pp.
11-13),
it can be
shown that
an EFA is an
impl ic i t
part of all parametric statistical tests. And as Gorsuch (2003) recently
pointed out , no factor analysis is ever completely exploratory (p . 143) .
ORGANIZATION OF THE BOOK
The book is organized to cover EFA first, followed by CFA. The
concepts und erlying EFA generally have c ounterpa rts in CFA. In fact, a
major
focus of the hook is com m unic ating the l inkages between EFA and
CFA.
In
general,
if a
statistical issue
is
important
in
EFA,
the
same issue
is
impor tan t
in
CFA.
For
example ,
jus t
as EFA
pat tern
and
s t ruc tu re
coeffi-
cients are
both
important in interpretation when factors a re correlated, the
same sets of
coefficients
are impo rtan t when factors are correlated in CFA.
Chapter 2 covers some fou nda tiona l statistica l topics, such as score
world
versus area
w orld
statistics,
and
basic m atrix algebra concepts. Chapters
3 through 9 elaborate EFA decisions and a naly tic choices. Finally , chapters
10 through
12
cover
CFA
methods.
Each of the chapters concludes with a section called M ajor Concepts."
These sections
are not
in tended
to summar ize all the key
points
of the
chapters. Instead, these sections emphasize major concerns tha t are i n s u f f i
ciently recognized in contemporary practice, or that otherwise might get
lost
in the
details
of
chapter coverage.
INTRODUCTION TO FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 14/185
2
FOUNDATIONAL CONCEPTS
A simple heurist ic example of exploratory factor analysis (EFA) may
m a k e the discussion of some founda tional concepts m ore concrete. Along
the way a few elem entary ideas involvin g a type of ma them atics called
matrix
algebra
will be introdu ced. It is som etimes jo kin gly said
that
matr ix
algebra invariably induces psychosis. It is t rue
that
matr ix a lgebra in some
respects
is
very diffe rent from
th e
simple algebra taught
in
high school.
A nd
knowing the mechanics of matrix algebra thankfully is
not
necessary for
unders tanding factor analys is .
But we do
need
to
know that some basic
concepts of matrix algebra (such as when matrices can be multiplied and
how
div i s ion
is
performed wi th matr ices)
are
important , even though
the
mechanics of the process are unimportant .
HEURISTIC EX AMPLE
Table
2 .1
presents hyp othetica l data in volv ing seven students ' rat ings
of
me reg ardin g how strongly th ey ag ree (9 = mo st agree) versus disagree
(1 =
most
disagree) that I am
Handsom e, Bea utiful , U gly, Bril l iant , Smart ,
and Dumb. Because these var iables a re di rect ly measured, w i thout applying
any
weights (e.g., regression
(3
weights)
to the
scores,
these
variables
are
called synonymously measured
or observed
variables.
As we
shall soon see, scores that
are
obtained
by
applying weights
to
th e
observed scores
are
synonymously called composite, latent,
or
synthetic
variables.
A ll
analyses within
the
general l inear model (G L M ) yield scores
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 15/185
TABLE 2.1
Heuristic
EFA
Data
Student/
Statistic
Barbara
Deborah
Jan
Kelly
Murray
Susan
Wendy
Mean
S
Measured variable
Handsome
6
8
9
5
4
7
3
6.00
2.16
Beautiful
5
7
8
4
3
6
2
5.00
2.16
Ugly
4
2
1
5
6
3
7
4.00
2.16
Brilliant
8
7
9
9
9
7
7
8.00
1.00
Smart
6
5
7
7
7
5
5
6.00
1.00
Dumb
2
3
1
1
1
3
3
2.00
1.00
fo r eve ry pa r t i c ipan t on at least one composi te var iable , and often on several
composi te var iables .
When we conduct a fac tor analys is , we are explor ing the relationships
among measured var iables and t r y ing to de te rmine w he the r these relation-
ships can be summar ized in a smal le r number of latent constructs. Several
different statistics
can be
used
to
summar ize
the
re la t ionships among
the
variables (e.g. , Pearson correlation coefficients, Spearman's rho coeff ic ients) .
Here we use Pearson product-moment correlation coefficients. Table 2.2
presents the ma t r ix of Pearson correlation coefficien ts for all the pairwise
com bination s of these six variables.
The Table 2.1 data set involves re la t ively few pa r t i c ipan ts (n = 7) and
re la t ive ly few va riable s ( i> = 6) . These features make the data set easy to
unders tand. However , the data and the resul ts are unreal is t ic . Real EFA
problems involve more par t ic ipants
and
more variables. Factor analysis
is
usually
less re levant
when w e
have
so few
variables, because
w e
gain li t t le
by sum ma r iz ing re la t ionships among so few measured var iables in a smal ler
set of la tent var iables .
TABLE 2.2
Bivariate Correlation Matrix
Variable
Variable
Handsome
Beautiful
Ugly
Brilliant
Smart
Dumb
Handsome
1.0
1.0
-1.0
.0
.0
.0
Beautiful
1.0
1.0
-1.0
.0
.0
.0
Ugly
-1.0
-1.0
1.0
.0
.0
.0
Brilliant
.0
.0
.0
1.0
1.0
-1.0
Smart
.0
.0
.0
1.0
1.0
-1.0
Dumb
.0
.0
.0
-1.0
-1.0
1.0
10
E XPLO RA TO RY A N D C O N F I R M A T O R Y F A CT O R
ANALY S I S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 16/185
Furthermore, we would rarely confront a correlation m atr ix, such as
that in Table
2 . 2 ,
in which the
correlation coefficients
are
only -1.0, +1 .0,
or 0.0. N or would we
tend
to see relationships fo r which th e latent constructs
are so
obvious
from
a
s imple examinat ion
of the
correlation matrix even
without performing any statistical analyses.
An examination of the Table 2.2 correlations suggests that two factors
(i.e.,
latent variab les) underlie the covariations in the Table 2 .1 data. Three
patterns
are
obvious
for the
heuristic example.
First, one
factor involves perceptions
of my
physical attractiveness.
Three
variables have r values
that are 1.0 (or
100%): Handsome,
Beautiful,
and Ugly, although scores on Ugly are scaled in the opposite direction
from
scores on Handsome and
Beautiful. Second,
another factor involv es
perceptions
of my
intellectual prowess.
Three
variables have
r
values
that
are
1.0 (or
100%): Brilliant, Smart,
and
Dumb, although scores
on
Dumb
are scaled in the opposite direction from scores on Brilliant and Smart.
Third, th e fact that all pairwise combinations of r values are 0.0 (or
0 % > )
for combinations involving one measured variable from one set (e.g., Hand-
some) and one measured variable from the other set (e.g., Brilliant) suggests
that
the two latent variables also are uncorrelated with each
other.
TWO W O R L D S OF STATISTICS
All statistics can be conceptualized as existing in or describing dynam ics
of
one of two "worlds": the
score
world or the
area
world. All statistics in
the score world are in an unsq uared me tric, as are the scores them selves.
For example, if we have d ata on how many dollars our classmates have on
their persons,
each
measured var iab le score is in the me tric, dollars. Stati stics
in the score world have implicit superscripts of 1. Example statistics in this
world are the mean, the median, the standard deviation (SD), and the
correlation
coefficient
( r ) .
If the mean is 7, the mean is 7
dollars.
If the SD is 3, the SD is 3
dollars.
The weights (e.g., regression (3 [beta] we ights , regression
b
weights)
applied
to
measu red variables
to
obtain scores
on
composite variables
are
also
in the score world (o r they couldn't be applied to the measured variables ).
The
statistics
in the
area world
are in a
squared metric.
For
example,
i f the scores are mea sured in dollars, and the v arian ce is 9, the variance
(SD
2
) is 9 squared dollars (not 9 dollars). M any of the statistics in this world
have explicit su perscripts of 2 (e.g., r
2
, R
2
) or names
that
communicate
that
these statistics
are
squared valu es (e.g.,
the sum of
squares,
the
mean square).
Other
statistics
in
this world
do not
have either superscripts
of 2 or
names
clearly
indicating
that
they
are in the
score world (e.g.,
th e
variance,
th e
covariance).
FOUNDATIONAL CONCEPTS 11
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 17/185
TABLE 2.3
Unsquared and Squared Factor Weights
Factor Squared values
Measured variable
Handsome
Beautiful
Ugly
Brilliant
Smart
Dumb
1.0
1.0
-1.0
.0
.0
.0
II
.0
.0
.0
1.0
1.0
-1.0
100
100
100
0
0
0
II
0
0
0
100
100
100
Any
statistics
we
compute
as
ratios
of one
area world stat ist ic
to
another
area
world stat ist ic
are
also themselves
in the
atea world .
For
example,
both
R
2
and the analysis of variance (ANOVA) effect size T |
2
(eta-squared ) are
computed
as:
sum of
squares E X P L A I N E D
su m
of squaresjoTA L
Rel iab i l i ty
coefficients (e.g., Cronbach's a) are also in the area world,
because they
are
computed
as
ratios
of
reliable
variance or
rel iable
sum of
squares divided by total score v arian ce or the total sum of squares. R eliab i l i ty
statistics
are in the
area world ,
even
though
they
can be
neg ative (e.g.,
-.3), and indeed even though they can be less than -1.0 (e.g., -7.0;
Thomp-
son, 2003).
Table
2.3
presents
the
unsquared factor weights (analogous
to
regres-
sion
(3
weigh t s )
for the
Table
2 .2
correlat ion mat r ix (and
the
Table
2 .1
data used
to
create
the correlation
ma t r i x ) .
Table 2.3
also presents
the
squared valu es associated wi th this analysis .
To
help readers keep straight
which
stat ist ics are in the area versus the score world, in many cases I
present these statistics
as
percen tages (e.g.,
r = 1.0 =
100%).
The Table 2.3 results from the factor analysis of the Table 2.2 correla-
tion
mat r ix confi rm
our
expec tat ions.
The
Table
2.3
mat r ix (and
the
Table
2.1 data
and the
Table
2.2 correlation
matrix) suggests that
two
uncorrelated
factors
(i .e.,
my
perceived physical at t ract iveness
and my
perceived intel lec-
tua l
abi l i t ies) underl ie the six measured variables.
S OME BASIC M ATRIX ALGEBRA CONCEPTS
As noted
previously ,
it is not
important
to
know
how to do
mat r ix
algebra (which
is
for tuna te, because
as
noted previously this knowledge
can
12
EXPLORATORY
A N D
C O N F I R MA T O R Y F A C TO R ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 18/185
cause severe psychological distress). Readers who desire an understanding
of the mechanics of matrix algebra can consult any number of hooks (e.g.,
Cooley & Lohnes, 1971, pp. 15-20; Tatsuoka, 1971, chap. 2). But it is
important
to know some matrix algebra
concepts,
so that one can
have
some understanding
of
what
is
being done
in
certain analyses
Different books
use di f ferent
notations
to
indicate
that a
matrix
is
being discussed. In this book, capital letters in bold are used to represent
a matrix (e.g.,
X, R).
There
are six
matrix algebra concepts that must
be
mastered.
These
are discussed below.
Rank
In this book we
p r imar i l y
consider matrices
that
are two-dimensional
(i.e., number of rows > 1, and number of columns > 1). Matrix rank character-
izes how
many rows
and
columns
a
particular matrix has.
When we specify
rank, we always report the number of rows first, and the number of columns
second, using subscripts.
For
example,
the
measured variable scores reported
in Table 2.1 could be represented as: X
7x6
. And the bivariate correlation
mat r ix
presented in Table 2 . 2 i s
R(
1X
6.
Transpose
For
some purposes
we can
reorganize
the
data
in a
matrix such that
the
columns
in the
original matrix become
the
rows
in a
second matrix,
and the
rows
in the
original matrix become
the
columns
in the
second
m a t r i x . Such a new matrix is called a transpose. We designate a transpose
by putting a single quotation or tick mark on the matrix. For example, if
w e have a matrix, Y ^
X
2 , the transpose of this m a t r i x is ¥ 2 x 3 ' -
Consider
the following matrix, ¥3
x
i\
1
4
2 5
3
6.
The
transpose
of
this matrix,
¥2
x
/, is:
1 2 3
4 5 6 .
Conformability
In regular (nonmatrix) algebra, any two numbers can be multiplied
t imes each other
(e.g., 4x9 = 36; 2 x 8 = 16). In matrix algebra, however,
matrices can be multiplied times each other only if a certain condition is
met: The number of columns in the
left
of two matrices being mu l t ip l i ed
FOUNDATIONAL
CONCEPTS
13
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 19/185
mus t equa l
t he
num b e r
of
rows
in the
right matrix being multiplied. Matrices
me eting this condit ion (arid eligible for m ultip licatio n) are said to be con-
formable.
F or
example ,
the Table 2 .1
m a t r i x ,
X y
X
5 can be
pos tmul t ipl ied
by the Table 2.2
correlation
matrix, R 6
X
6 , because 6 (columns in X) equals
6
( the number
of
rows
in
R .
When w e
m ul t iply matr ices ,
th e
rank
of the
result
o r
product matrices
wil l equal the num ber of rows in the lef tmost matr ix , and the num ber of
columns
in the
produc t ma t r ix will equa l
th e
num b e r
of
columns
in the
r ightmost m atr ix . For example , the matr ices Zy
x 6
and W
6
x
2
are conform able.
The product of these two matrices would have a rank of 7 by 2:
Z y
x 6
W
6 x 2
= 0
7 x 2
.
Symmetric Matrix
Some "square" matr ices (number of rows equal number of columns)
are symmetric, meaning
that
every
i,j
e lement
of the
matr ix equals
th e
corresponding j,i e lement of the same matr ix. For example , in the Table
2 .2
corre la t ion matr ix,
the
3 1 e l ement
in R
equ als -1.0 eq uals
the
1 3
e lement
in R.
Correlation m atr ices
a re
a lways sym metr ic .
A
related property
of symmetric matrices, which are a lways conformable to themselves, is that
a
sym metr ic m atr ix a lways equals
i ts own
transpose,
by
definit ion (i .e. ,
Rfi x 6
=
R-6 x 6 •
Identity Matrix
In regular algebra,
a ny
number t imes
1
equals
th e
o r ig ina l num ber
that
w as m ul t ip l i ed times
1
(e.g.,
9x1 =9 ; 15x 1 =
15) .
In
matrix algebra
there is a
m a t r i x , ca l led
the
identity
matr ix
( I) , which when mul t ipl ied t imes
any m atrix , equ als the init ial matrix . For exam ple, let 's say we have ¥ 3 x 2 :
1
4
2 5
3
6.
For this example, to be conformable the ident i ty matr ix must
have
two rows. The ide nti ty ma trix is always sym me tric. Th e identi ty m atrix
always has
ones
on the
"diagonal" running
from the
upper
left
corner down
to the lower right corner, and zeroes in all
off-diagonal
locations. So for
this
example I
2 x 2
would be:
1 0
0 1.
For
this example
¥ 3
x
71
2 x 2
(i.e.,
Y
pos tmul t ipl ied
by I)
yields
a
product
matrix that contains exactly
the
same numbers that
are in
Y }
x 2
itself.
T he
J 4
EXPLORATORY
A N D C O N F I R M A T O R Y F A CT O R ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 20/185
identity matrix
is a
critical matrix algebra concept, because
as we see
next
I is invoked in order to be able to do matrix division.
Matrix Inverse
For symmetric matrices,
one
could state
a
matrix equation such
as
R ( j
x
6 R 6 x 6
=
l e x 6 . The matrix, R < s
x(
," ' ( i .e . , a matrix with a-1 superscript),
is called an inverse. Not every symmetric matrix has an inverse. If a matrix
eligible
for
inversion (e.g.,
a
symmetric matrix) does
not
have
an
inverse
fo r a
given data set,
the
matrix
is
said
to be
"ill conditioned."
Some symmetric matrices, such as some correlation matrices, do not
have an inverse, because the numbers in the matrix cannot yield an R"
1
such that
R R' = I. If this
happens
in factor analysis,
this
may be due to
sample
size
being too small.
Some factor analyses
require
that
an
inverse
be
computed
in
order
to
compute
the
factor solution.
When
using factor analyses requiring matrix
inversion, if a matr ix is ill conditioned, the solution simply cannot be
computed using
the
particular technique requiring
the
inversion. Another
technique must be selected, or sample
size
must be increased.
In regular algebra, division is straightforward. For example, to compute
the
F C A L C U L A T E D
m
ANOVA
or in
multiple regression,
we
compute
F =
[ S u m of Squares E X P L A I N E D / ^ / E X P L A I N E D ] / (Sum of Squares U N E X P L A I N E D /
^ /U N E X P L A I N E D -
But in
matrix algebra
the way
that we do
division
is by
postmultiplying
the
numerator matrix times the inverse of the matrix with which we wish t o divide.
For example, th e division of 62
x
2 by W 2
x
2 is accomplished using the form ula :
62
x
2 W 2
x
2
=
^2 x 2 . (This is actually on e basis fo r obtaining F C A L C U L A T E D
fo r a multivariate
ana lys i s
of variance [MANOVA] and is directly analogous
to the
univariate
ANOVA
computation, except
for the use of
matrix
algebra.)
Division is used a lot in univariate statistics. Division is also used a
lot in multivariate statistics, including factor analysis. Clearly, access to an
inverse is usually critical for most factor analytic applications.
M O R E
HEURISTIC
RESULTS
Throughout the GLM, weights are applied to the scores on the mea-
sured variables
to
obtain scores
on
composite variables.
In the
GLM, weights
are
invoked (a) to compute scores on the
latent
variables or (b) to interpret
what
the
composite variables represent. Logically, because
the
weights
are
used
to
obtain
the
scores
on the
composite variable,
the
weights must
be
FOUNDAT1ONAL
CONCEPTS
15
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 21/185
of
some assistance
in
understanding
the
nature
of the
latent variables
(i.e.,
help in interpreting what latent variables represent).
For
example,
in
mul t ip le regression these weights
are
often called
j3
(beta)
weights. A set of weights for the predictors in regression is called the
regression
equation.
These
P weights in regression are also sometimes called
standardized regression
c o e f f i c i e n t s because they are applied to the measured
predictor variables
in
their t-score form (i.e.,
th e
measured variables have
been
transformed so
that
th eir means equ al zero and the ir standard deviations
and variance s equal one). It is actually a misn om er to call the weigh ts
"standardized,"
because
th e
weights
are a
constant
for a
given data set. Thus,
th e
weight
for the
f i r s t predictor
itself
cannot
be
standardized.
If p = .5, this weight is applied to the f i r s t predictor z scores of
everybody in the sample. We
cannot
apply the formula ^ = ( X j - X ) / SD to
a given weight. If everybody is assigned a weight of Pi = .5, SDpi = 0,
then
th e division is m athem at ical ly impermiss ible . To be technically correct, we
should call such weights "weights applied
to the
standardized measured
variables" rather
than
"standardized weights."
Perhaps
we are
performing
a
regression using
tw o
predictor variables.
Steve has z\ = -1 and ^2 = .5. The form of the regression equation is:
For a given data set , Pi might equal 2.0, and P? might equal — 1 . 2 .
Note
that ,
as is the case thr oug hou t the G LM , weights are not necessarily
correlation coefficients (see Courville & Thompson, 2001).
Thus,
as in this
example,
it is
possible that none
of the
weights will
lie in the
range between
-1 and +1. For Steve, Y
S T
E V E = 2 .0 (0 .7 ) + -1.2(0.5) = 1.4 + -0.6 = 0.8.
In
regression,
these Y
scores
are the
composite, latent,
or
synthetic variables.
Pattern
Coefficients
In
factor analysis,
pattern
c o e f f i c i e n t s
are the
weights applied
to the
measured variables
to
obtain scores
on the
factor analysis latent variables
(called
factor
scores) . As n oted , these w eights are analog ous to the P weights
in
multiple regression. These weights
are
also analogous
to the
standardized
discriminant function coefficients in descriptive discriminant analysis (Hu-
berty,
1994)
and the
stan dardized canonical function
coefficients in
canoni-
cal correlation analysis (Thompson, 1984, 2 00 0a).
T he
factor pattern coefficients (P y
x
F )
a re
computed part ly
to
reexpress
th e variance represented in the correlation matrix that is being analyzed,
and from wh ich the factors (i.e., the sets of pattern coefficients) are extracted.
T he
factors
are
extracted
so
that
the first
factor
can
reproduce
th e
most
variance
in the
matr ix being analyzed,
th e
second factor reproduces
th e
16 E XPLO RA TO RY A N D C O N F I R M A T O R Y
FACTOR
A N A L Y S I S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 22/185
second most variance, and so forth. The ability of one or more factors to
reproduce the ma trix being analyzed is qu antified by the reproduced m atrix
of associations, such as the reproduced inter var iable correlation m atr ix
( R y x V
+
) - This
is the "how f u l l is the glass" perspective.
The
ability
of the
factors
to
reproduce
th e
matrix being analyzed
can
also be qu antified by com puting the ma trix
that
is left after a given number
of
facto rs have been extracted . This is called the resid ual m atr ix of associa-
tions, such as the residu al inter var iab le correlation ma trix ( R
V x V
~ ) -
This
is
the
converse, "how empty
is the
glass" perspective.
If the factor pat tern coefficients perfectly reproduced the matrix of
associations (here
an
intervariable Pearson correlation coefficient m atri x) ,
then R y
x
v ~ would consist entirely
of
zeroes, indicating that
there is no
information or variance
left
in this m atrix. An d if the factor pat tern
coeffi-
cients perfectly reproduced th e matr ix of associations,
then
th e entries in
RV x v
+
would exactly match the entries in Ry
x
v.
The reproduced intervariable correlation matrix (Ryxv
+
) is com-
puted as:
Note
that
th e
pattern coefficient matrix
and its
transpose
are
conformable,
and the product matrix has the rank
v
by
v
(the same rank as the original
intervariable correlation matrix).
Possible Numbers of
Factors
In
regression there
is
only
a
single "equation" (set
o f
weights,
[3
weights)
in a given analysis. In fac tor analy sis the sets of weights (i.e., patte rn
coefficients) are called /actors rather than equations, mainly just to be incon-
sistent and to confuse th e graduate students. The number of factors worth
keeping can range between one and the number of variables.
Let's
go
back
to our
example
of
ratings
by
seven participants
of me as
regards six measured variables. If every entry in the interv ariab le correlation
matrix was either a +1 or a -1, the r
2
between every pair of measured
variables
would
be
100%. This would imply that
a
single factor underlay
th e ratings. W e would extract one factor (Pg
x
|) consisting exclusively of
negative
or
plus ones.
A nd
this single factor w ould perfectly reproduce
th e
original
R
6 x
6 mat r ix .
Technically,
there
would be five addit ional factors, each consisting
exclusively of zeroes, meaning that they each contain and reproduce no
information. But we would never bother with such factors. Indeed, we
usually do not bother even w ith factors that reproduce some, but only a
FOUNDAT10NAL CONCEPTS I 7
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 23/185
l i t t le information,
if we
deem
th e
amount
of
informat ion
or
variabil i ty
reproduced to not be noteworthy.
At the other extreme,
if
somehow all the ra tings were perfectly uncorre-
lated, then e very off-diagonal element
of
R
6 x 6
would
be a
zero (and R ^x
5
=
I
6x6
).
Now,
no two
variables could
be
combined
to
create
a
factor (i.e.,
each
measured variable would define
its own
factor).
There
would
be six
factors.
Each factor would have one +1 value, and the rema ining five entries
would be
zeroes.
A t this second extrem e (per fectly uncorrelated measured variab les),
RV x v
+
would exactly match th e entries in Ry
x
v- In fact , whenever we
extract all possible factors (i.e., the number of factors equals the number
of measured var iab les) , the pattern coefficients will perfectly reproduce the
original matrix of associations being analyzed.
Factor
Structure
Coefficients
All GLM analyses are correlational and yield weights (e.g., factor
pattern coefficie nts) that are applied to m easured variables to obtain scores
on composite va riables.
These
weights are sometim es correlation coefficients,
and
sometimes they are not (cf. Courville & Thompson, 2001;
Thompson,
1984, pp . 2 2 -23 ) .
Researchers can also comp ute biva riate correlation
coefficients
between
measured variables an d their composite variables (e.g., regression Y scores,
factor
scores, discriminant function scores). These correlations across various
GLM analyses are always called structure coefficients.
Throughout the G LM consult ing the struc ture coefficients is almost al-
ways essential to correctly interpreting results. For example, in regression
Courvil le a nd
Thompson
(2001) documented the misinterpretations that ca n
occur when only p weights a re consulted in interpretation (also see Cooley
&Lohnes , 1971, p. 55; D unla p& La ndis , 1998; Thompson &B orrello , 1985).
For
descriptive disc rimina nt analysis, Hu berty (1994) noted that "con-
struct definit ion and struc ture dime nsion [and not hit rates] constitute th e
focus [italics added]
of a
descriptive discriminant analysis"
(p .
206).
The
importance
of
structure
coefficients
in
canonical correlation analysis
is
widely
recognized (cf. Cohen
&
Cohen, 1983,
p.
456; Levine, 1977,
p.
20) .
For
exam ple, M eredith (1964) noted, "If the variables within each set are
moderately intercorrelated th e p ossibil i ty of interpreting th e canonical vari-
ates by inspection of the appropriate regression weights [func tion coefficients]
is
practically nil"
(p .
55) .
Structure coefficients
are
equal ly important
in
factor analysis (cf. Gra-
ham, Guthrie , & Thompson , 2003; Thompson , 1997). Thus, Gorsuch (1983)
emphasized
that a
basic
[italics added]
matrix
fo r
interpreting
th e
factors
is th e
factor s t ructure"
(p .
2 0 7 ) .
18
EXPLORATORY A N D
C O N F I R M A T O R Y F A CT O R ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 24/185
In the GLM, sometimes the weights are correlation coefficients. F or
example , in regression, if the predictor variables are perfectly uncorrelated,
the P weight for a given predictor equals that variable 's correlation with
the outcome var iab le (Yi) .
In
factor analysis,
the
s t ruc ture
coefficients ( S v
x
p )
can
be
computed
using the equation
P v x F
Rp 'xF
S vxFi
where R
F x F
is the
interfactor corre la t ion mat r ix (not
the v by v
intervar iab le
correlation matrix). When
the
factors
are
perfect ly uncorre la ted,
by defini-
tion R p x F
=
I p x F -
an
J therefore
P y x F
=
S V X F -
In other words, whenever the factors are perfectly uncorrelated the
pattern coefficients are also structure coefficients (and therefore must fall
with in
th e
range — 1
to
+ 1 ) ,
and in
this case
are
more appropriately termed
pattern/structure coefficients.
And wh en factors are first extrac ted, the factors
are
always
perfectly uncorre la ted.
"Loadings"
It is not uncom mon in factor analytic reports to see authors referring
to coefficients as "loadings." The problem is that this term is inherently
ambiguous. In some reports authors
refer
to pattern
coefficients
as "loadings."
In
other
reports authors
refer
to structure coefficients as "loadings." In some
reports
it is
impossible
to
discern wh at coefficients
th e
authors
are
descr ibing.
And in
some reports
th e
pa t te rn
and the
s tructure coefficients
are not
equal but authors
refer
to both sets of coefficients as "loadings." These
ambigui t ies
cause confusion tha t impedes clear
scientific
d ialogue. Therefore,
some journals (see Thompson
&
D aniel, 1996) form ally discourage
the use
of this vag ue colloq uialism . For the same reasons, the term loading is not
used
here.
Communality Coefficients
For the Table 2 .3 factors, because the factors are perf ectly u ncorrelated ,
the unsquared coefficients reproduced in Table 2 .4 are pat tern/structu re
coefficients.
Recall that Pearson correlation coefficients (a) characterize
the linear relationship between two in terv ally scaled varia bles but (b) the
coefficients
themselves are not inte rva lly scaled.
That
is, an r of 1.0 is no t
twice as
large
as an r of
0.5.
T o
m a ke
th e
correlations intervally scaled,
so
that
we can say how
much larger one r is versus another r, we m us t first square the rs before w e
compare them. Thus, the r of 1.0 is four times larger than the r of 0.5,
FOUNDATIONAL
CONCEPTS
19
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 25/185
TABLE
2.4
Pattern/Structure Communality Coefficients
and
Eigenvalues
Factor
Measured variable
Handsome
Beautiful
Ugly
Brilliant
Smart
Dumb
Sum of
squared column values
1.0
1.0
-1.0
.0
.0
.0
3.0
II
.0
.0
.0
1.0
1.0
-1.0
3.0
1
00
100
100
100
100
1 00
Note
The eigenvalues (3.0 and 3.0) are in a squared, area-world metric, even though they are not pre-
sented as percentages.
because
r
2
=
l.O
2
=
100%
is fou r
times larger than
r" =
0.5""
=
25%. Because
the Table 2.4 structure coefficients are correlation coefficients, they too
must
be squared before they are compared.
One set of manipulations involves taking the squared structure coeffi-
cients for the uncorrelated factors (the formula is d i f fe rent for correlated
factors) and
adding
the
squared values along
the
rows across
the
columns
(e.g., 100%
+ 0% =
100%).
The
resulting coefficients, called
communality
coefficients h
2
) , are area-world statistics. Because these factors are uncorre-
lated, the variance each measured variable shares with a factor is unique
or nonoverlapping, and thus the
h
value for a measured variable indicates
ho w much of th e variance in a measured variable the
factors
as a se t can reproduce .
If a measured variable had a communality coefficient close to 0%, this
would
mean
that this variable is not being represented within the factors.
If the
researcher desires
the
variable
to be
represented
in the
factors,
it may
be
necessary
to
extract additional factors.
An
equally valuable perspective takes
a
converse view
of the
interpreta-
tion of the communality coefficients. The h value for a measured variable
reflects
ho w
much
of the
variance
of a
given measured
variable was
useful
in
delineating the
factors
a s a
set.
There
is yet another
perspective from which communality coefficients
can be viewed. The communality coefficient for a variable is a "lower bound"
estimate of the re l iabi l i ty of the scores on the variable. This means that if
a variable has an h
2
of 50%, we believe that the reliability of the scores on
the variable is no lower than .5 (and, of course, may be higher).
In
some unusual cases communality coefficients greater than 100%
occur. These are called Heywood cases (Heywood, 1931). Such solutions
are
statistically "inadmissible,"
and
they indicate fundamental problems
s imi la r
to
estimating variances
to be
negative.
20
E XPLO RA TO RY A N D C O N F I R M A T O R Y
FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 26/185
Table 2 .4
presents
th e
communal i ty coeff ic ients
fo r each of the six
measured
variables in the heurist ic example. Each of these is 100%. This
reflects th e fact that for the present example, if we reexpress th e data
presented in Table 2.1 in the form of scores on the two factors presented
in
Table
2.3,
we
lose
no
informat ion whatsoever
Eigenvalues
For the Table 2 .3
factors,
if we add the
squared s t ruc ture coeff ic ients
down
th e
columns across
th e
rows,
we get a
very important
set of
squared,
area-world stat is t ics called e igenva lues (X ). As is the case e lsewhere in statis-
tics, important concepts are given multiple names, to confuse everybody;
natural ly it is the best use of effort to be confusin g to focus such atten tion
primari ly on the most important concepts, by giving them the most synony-
mous names . Another synony m ous term for eigenvalues is characteristic roots.
T he eigenvalues fo r this s imple heurist ic example are 3.0 and 3.0, as
reported in Table 2 .4 . (N orm al ly e igenvalues descend in valu e , but for this
heurist ic example
the first two
eigenvalues were exactly equal.) Eigenvalues
are always indices of the amount of information represented in some multi-
variate result . In some m ul t iva r ia te analyses e igenvalues are mul t iva r i a t e
squared correlation coefficients . Clearly, 'given that
the first two
e igenvalues
associated with
th e
factor an alysis were
both
greater than 1.0,
in EFA
eigenvalues
are
not
squared correlation coefficients .
T he fo l lowing
four
s ta tements apply to e igenvalues in an EFA:
1. The n u m b e r of e igenvalues equals th e n u m b e r of measured
variables being analyzed.
2. The sum of the
e igenvalues equals
th e
n u m b e r
of
m easured
variables.
3. An
e igenvalue divided
by the
n u m b e r
of
measured var iables
indicates th e proport ion of informat ion in the m a t r ix of associ-
ations being analyzed that
a
given factor reproduces.
4. The sum of the
e igenvalues
for the
ext racted factors divid ed
by th e num b e r of measured variables indicates th e proportion
of the inform at ion in the m atr ix being analyzed that the factors
as
a set
reproduce.
For th e current example th e number of measured variables is six. The refore,
there are six eigenvalues associated with the Table 2 .2 corre la t ion m atr ix.
Because the sum ot the e igenv alues for the cu rren t exam ple is 6.0, give n
that
the first two
e igenvalues
are 3.0 and
3.0,
the
remaining e igenvalues
mus t be 0.0, 0.0, 0.0, and 0.0.
Because 3.0 / 6 equals 0.5, the first e igenvalue indicates
that
Factor I
reproduces
0 .5 (or
5 0 % )
of the
informat ion conta ined
in the
Table
2 .2
FOUNDATJONAL CONCEPTS 21
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 27/185
HANDSOME
BEAUTIFUL
UGLY
BRILLIANT
SMART
DUMB
DSOME
BEAUTIFUL
UGLY
BRILLIANT SMART
DUMB
0
0
0
0
0
0
0
0
0
0
0
0 0
0 0 0
This result reflects the fact
that
fo r
these
heurist ic data the two factors
alone reproduce
all the
information contained within
the
original
Tahle
2 .2
correlation matr ix.
SAMPLE SIZE
CONSIDERATIONS
IN
EXPLORATORY FACTOR A N A L Y S IS
Sample size
affects
the precision of all statistical estimates, inc lud ing
those
made in EFA. V arious researchers have proposed rules of t hum h fo r
sample size m i n i m u m s that
are a
funct ion
of the
ra t io
of the
num b e r
of
people
to the
num b e r
of
measured var iables .
T he
recommended ratios often
are wi th in th e range of 10 to 20 people per measured variable. Gorsuch
(1983) suggested
that
"an absolute minimum ratio is five individuals to
every variab le, but not less t h an 100 in div idu als for any an alys is" (p. 332) .
However, some
Monte
Carlo
s imulat ion research (Gua dagnol i
&
Ve l i -
cer, 1988) suggests tha t th e most cri t ical issue is how saturated th e factors
are
by the
measured var iables .
They
suggested
that
replicable factors tend
to be
es t imated
if :
1. factors
are
each defined
by
four
or
more measured variables
with structure coefficients each greater than |.6|, regardless o f
sample size; or
2 .
factor s are each defined with 10 or more s tru ctu re coefficients
each around |.4|,
if
sample size
is
greater
than
150;
or
3.
sample
size
is at
least 300.
MacCal lum, Widaman, Zhang,
and
Hong (1999) found that sample
sizes
as low as 60 accu rately reproduced population pattern coefficients if comm u-
nalities were all .60 or greater. If h values were arou nd .50, sample
sizes
of
100 to 200 were required.
O f
course, when
it comes to
m athe ma t ical ly complex analyses such
as EFA, m ore
is
always better. This pr inciple
is
par t icular ly re levant , because
one may not know prior to conduc t ing the analysis what the s a tura t ion
pat tern
of the
factors will
be, and so
rules revolving aroun d saturation
2 4
E X P L O R A T O R Y
A N D
C O N F I R M A T O R Y
FACTOR
A N A LYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 28/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 29/185
3
EXPLORATORY FACTOR ANALYSIS
DECISION SEQUENCE
Exploratory
factor analysis (EFA) in actuality involves a linear se-
quence of decisions each involving a menu of several available choices.
Given that there are five major decisions, with many choices at each decision
point,
the
number
of different analysis
combinations that
are available to
the researcher is quite large. These five decisions, discussed in more detail
in the remainder of the present chapter, address the following questions:
1. Which matrix of association coefficients should be analyzed ?
2. How
many factors should
be
extracted?
3.
Which method should be used to extract the
factors?
4 -
H ow
should
th e
factors
be
rotated?
5. How
should factor scores
be
computed
if
factor scores
are
of
interest?
In EFA, using more than one set of analytic choices is usually best practice.
A s Gorsuch (1983)
wisely
suggested, Factor th e data by several
different
analytic procedures and hold sacred only those factors that appear across
all the
procedures used
(p.
330).
This chapter presents the EFA decision sequence. Not
every
choice
at each decision point is described. But the most common selections are
presented.
And in
addition
to
covering
the
most commonly encountered
choices, th e presentation is
sufficiently
broad to give th e reader a
flavor
of
27
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 30/185
TABLE 3.1
SPSS Variable Names and Variable Labels for the Appendix A Data
Variable SP SS variable label
PER1 I
Willingness
to
help users'
PER2
'1
Giving users
individual
attention'
PER 3 '1 Employees who deal with users in a caring fashion'
PER4
'1
Employees
who are
consistently courteous'
PER5 '2 A haven fo r quiet an d solitude'
PER 6 '2 A meditative place'
PER7 '2 A
contemplative environment'
PER8
'2
Space that facilitates quiet study'
PER9 '3 Com prehensive print collections'
PER10 '3 Complete runs of journal titles'
PER11
'3
Interdisciplinary library needs being addressed'
PER12
'3
Timely document delivery/interlibrary loan'
th e
complexi ty
of the choices.
Exploratory factor analysis ,
to pu t it
mi ld ly ,
is
not any one
s ingle analys is—EFA encompasses
a
broad
family of choices.
The sm a l l , hut real , data set presented in A p p e n d i x A wi l l he used to
make this discussion
concrete. The
da ta were randomly sampled
from the
L i h Q U A L + ™
s tu dy
of
user perceptions
of
serv ice qu al i ty
at
academic l ibrar-
ie s
in the
Uni t ed
States and Canada
(cf.
Cook &. Thompson,
2001;
Thomp-
son,
Cook, &
Heath, 2001; Thompson,
Cook, & Thompson,
2002) . Appen-
d ix
A presents the ra t ings provided by 100 g ra d u a t e s tu den t s and 100
faculty
from
one LibQUAL+ s t u dy . Table 3.1 presents the 12 measured var iables
sampled from
the
data set.
MATRIX
OF A S S OC IATION COEFFICIENTS
The scores on measured var iables are not the actual basis for E F A .
Ins tead , these da t a are used to compute some matr ix of biv ariate associat ions
am ong the m easured v ariables. I t is the m atrix of associat ions that is analyzed
within EFA. Indeed, g iven
the
m a t r ix
o f
associations (e.g.,
R H
X | I
) for a
data set ,
all the
steps
of the
factor analysis (except
for the
c om pu t a t ion
of
factor
scores) can be
accomplished
even
without
access to
original data
(e.g.,
X i o c x i i ) .
Because
th e
factors
are
ext rac ted
from
som e m a t r ix
of
associat ions,
the factors themselves are sensit ive only to
those
dynamics captured in a
given s ta t is t ic m easur ing b ivar ia te re lat ionships .
If the
chosen
stat ist ic only
evalua tes
the
orderings created
by the
m easured var iables , ignor ing d is tances ,
then the
factors
wi l l
also
be
exc lus ively order-dr iven.
If the
statistic chosen
to characterize associat ion is in f l u enc ed b y
both
re la t ionships and score
variabili t ies,
then
the
factors
in
turn wil l also
be a
fu nc t ion
of all
these
28 E XPL OR AT OR Y AND
C O N F I R M A T O R Y F A CT O R ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 31/185
dynamics . And for the
sam e data set,
different
factors
m ay
emerge depending
on the selection of the matr ix of associations to he analyzed.
The
Pearson produ ct-m om ent hiv aria te correlation m atrix
is the ma-
trix
of associations most
commonly used
in
EFA.
In fact, in
most statist ical
packages
th e
default
(used unless
th e
user overrides
th e
default choice)
matr ix
of associations used in EFA is the Pearson correlation matrix.
But
there
ar e
quite
a
n u m b e r
of
other available choices.
Different
statistics
that characterize relationships are sensi tive to different aspects of
the data. Different relationship s tatistics also presu m e that different levels
of
scale underlie
th e
data.
For example, th e Pearson r requires that da ta mus t he interv ally scaled.
Spearman's rho,
on the other hand,
presumes only that data
are at
least
ordinally scaled.
But if
data
are
intervally scaled,
the
researcher could
use
either Pearson or Spearman
coefficients
as the basis of the ana lysis. In fact,
Spearman's
rho is
Pearson's
r
between
tw o
variables
once
interval data have
been converted into ranks with
no
ties.
For example, presume a data set with interval scores of three people
on two variab les:
Part ic ipant X Y
Barbara 1 1
Deborah
2 2
Jan
3 96
The
Pearson correlation coefficient (r
x
y )
can be
c o mp u t e d
as:
r
XY
=
COVxv
/
(S D
X
*
S D
Y
) ,
where
the
covariance
of X and Y is
c o mp u t e d
as :
C O V
X Y
=
( 2 ( X , - X ) ( Y
i
- Y ) ) / n - l ,
or
equiv alent ly (because
x, = X; — X and y, = Y, — Y) as:
COVxv
=
( I x y , ) / n
-1.
For these init ial data, we obtain:
Part ic ipant X X x Y Y y xy
Barbara 1 2.0
-1.0
1
33.0 -32.0 32.0
Deborah
2 2.0 0.0 2
33.0 -31.0
0.0
Jan
3 2.0 1.0 96 33.0 63.0 63.0
Sum
6.00
99.00
95.00
Mean 2.00
33.00
SD 1.00 54 .56.
DECISION SEQUENCE 9
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 32/185
So the covariance is
95.00
/ 2 = 47.50. And the Pearson r equals:
47.50
/
( 1 . 0 0 x 5 4 . 5 6 )
47.50 /
54.56
= .87.
If
these interv ally scaled data are conv erted into ranks, the only change
is that Jan's score
of 96 on Y
hecomes
3
w h e n
the
scores
are all
converted
into ranks. Now we can c o mp u t e th e Spearman's rh o using th e Pearson
r
formula:
Part icipant
Barbara
Deborah
J a n
Sum
M e a n
SD
X
1
2
3
6.00
2.00
1.00
X
2.0
2.0
2.0
X
-1.0
0.0
1.0
Y
1
2
3
6.00
2.00
1.00
Y
2.0
2.0
2.0
y
-1.0
0.0
1.0
xy
1.0
0.0
1.0
2.00
The co va rian ce is 2.00 / 2 = 1.00. A nd the Pearson r equals:
L O O /
(1.00
x
1.00)
l . O O /
1.00
=
1.00.
In
effect,
w hether computed on ordinal or on interv al data, Spearman's
rh o
addresses
the
qu est ion,
"Do the two
v ariables order
th e
people
in
exact ly
th e
same order?" Pearson's r evaluates this question also,
but
takes into
account
as
wel l
th e
distances between
th e
ordered scores. Spearman's
rh o
presumes
that
no
such informat ion
is
present
in
da ta
(or
ignores this
informa-
t ion) , which is w hy rho can be used w h e n both variables are ordin ally scaled.
These
two m atrices of association are two of the
choices
that researchers
can use in
extracting factors.
But a
third choice
is the
mat r ix
of
c ovariances
itself. In fact, th e covariance matr ix is the most common choice as the
matrix of
associations investigated
in
confirmatory factor
analysis
(CFA).
From the
formula
T
X
Y
= CO VX Y / ($ DX * S DY ), we can see
that
rX Y
will
equal
CO V
X
Y
if either (a) r (or the covar iance) is zero, or (b) the two
standard
deviat ions are reciprocals of each other (e.g., 1 and 1, .5 and 2).
In all
other cases,
r
x
y ?t /
CO V
XY
In many contexts th e covariance is used primari ly as an intermediate
calculat ion in obtaining th e correlation coefficient, and not to describe
relationship or
association.
The
covariance
is used infrequently
because,
un l i ke r, the covariance does not have a defini t ive range of possible
values.
For
example, consider
th e
following combinat ions
of r and SD
values:
30 E X P L O R A T O R Y AND C O NFI RM A TO RY FA CTO R A NA LY SI S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 33/185
r
XY
* SD
X
* SD
Y
=
CO V
X Y
1.00
* 100 * 100 =
10000
0.50
* 100 * 200 = 10000
0.01
*
1000
*
1000
= 10000.
Clearly the covariance can equal a giv en valu e und er a wid e array of
circumstances
The covariance is joi ntl y influenced by three aspects of the two vari-
ables: th e correlation between the two variables, th e var iabi l i ty of the
first variable, and the variability of the second variable. Therefore, when
exploratory factors are extracted from a covariance matrix, some factors
m ay be a function of correlations, whereas others may be more a function
of score spreadoutness. Sometimes we want our factors to be sensitive to
an array of aspects of the scores. But at
other
times we may prefer all the
factors
to be sensitive only to a single aspect of our data.
N U M B E R O F FACTORS
A
critical decision
in any EFA is
determining
how
many factors
to
retain. There are numerous strategies for making this decision. In general,
several strategies should be used with the hope that different approaches
to making this decision will corroborate each other. Although empirical
evidence
can
inform
this
jud gm ent, these decisions
are in the final
analysis
matters of exactly that: judg m ent. Zw ick and V elicer (1986) presented what
is
the most authoritative empirical evaluation of how well these various
strategies work.
Statistical Significance Tests
Statistical significanc e tests
d ue to
Bartlett (1950)
can be
used
in
either
of tw o
ways.
First, the correlation matrix can be tested to evaluate whether
the matrix is an identity matrix. If this null hypothesis that the correlation
matrix
is an identity ma trix
cannot
be rejected,
then
factors
cannot
sensibly
be extracted
from
th e matr ix . Second,
after
each factor is extracted, with
factors being extracted successively so as to contain progressively less and
less inform ation or variance, the residu al correlation m atrix can b e analyze d
to evaluate whether any information remains (i .e. , whether the residual
matrix
is essentially a n identi ty ma tr ix) .
The problem with these applications is the same general problem that
applies
in all
statis tica l significance tests (see Thompson, 1996). Statistical
significance
is driven in large part by sample
size.
Because researchers gener-
ally
use EFA
only with reasonably large samples, even trivial correlations
DECISION SEQUENCE 31
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 34/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 35/185
Series'
r-
o>
i-
Factor
Figure 3.1. Scree plot
for
first
100
cases
and 11
variables presented
in
Appendix
A.
recognize and retain. Trivial factors, however, are analogous to scree, and
should he left hehind in the extraction process.
A scree plot graphs eigenvalue magnitudes on the vertical access, with
eigenvalue numbers const i tut ing th e horizontal axis. T he eigenvalues are
plotted as asterisks w ithin the graph, and suc cessive valu es are connec ted
by a line. Factor extraction should be stopped at the point where there is
an "elbow," or leveling of the plot. This visual approach, not invoking
statistical significance, is sometimes called a "pencil test," because a pencil
can be laid on the rightmost portion of the relevant graphic to determine
where th e elbow or flattening seems to occur.
Statistical
pack ages prov ide the scree plot on requ est. Of c ou rse, be-
cause
this strategy requires judgment, different researchers given the same
plot m ay d isagree regarding what number of factors should be retained.
Figure
3.1 presents the scree plot
associated with
the
analysis
of the
first 100 cases of data for the first 11 va riab les listed in A p p e n d ix A . These
data are perceptions of 100 grad ua te studen ts regarding
library
service quality
on the first 11
variables listed
in
Table 3.1.
This
plot suggests that three
factors shou ld be extracted. Em piric al varia tions on this "visual scree" test
use regression analyses to evaluate objectively where th e elbow in the plot
occurs (Nasser, Benson, & Wisenbaker, 2002).
nspe tion
of the Residual Correlation Matrix
A s noted previously, as more factors are extracted th e entries in the
res idual correlation matrix
(R
v x
v)
approach zeroes.
If all
possible factors
are extracted, then the resid ual m atrix
wi l l
always consist only of zeroes.
DECISION
SEQUENCE
33
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 36/185
Logically, another approach to de te rmin ing the n u m b e r of notewor thy
factors involves examination
of the
res idual matr ix
as
successive factors
are
extracted. Compu ter packages provide the res idual matr ix on request . And
some packages count th e n u m h e r of entries in one triangle of the res idua l
correla t ion matr ix that
are
greater
than
|.05|
to
help inform this eva luation .
N o te
that
this criterion has nothing to do wi th p < .05 and that the cutoff
va lue is somew hat arbi trary. Researchers m ay invoke other cr i ter ia , and
m ight elect to use larger cutoffs when sample sizes are sm all and the standard
errors of the
correlation coefficients
a re
larger.
Parallel
nalysis
Horn (1965) proposed
a
strategy called parallel analysis that appears
to be among the best methods for deciding how many factors to extract or
reta in (Zwick
&
Velicer,
1986).
The
strategy involves taking
the
actual
scores on the m easured variab les and creating a random score m atrix of
exactly th e same rank as the actual data with scores of the same type
represented in the data set.
There
are various approaches to this s tra tegy.
The
focus
of the method is to create eigenv alues
that
take into account
th e sampling error that influences a g iven set of measured v ar iables . If scores are
random ly ordered , the related correlation m atrix wi l l approximate an id enti ty
matr i x , and the e igenva lues of that m atr ix
wi l l
f luc tua te around 1.0 as a
func-
tion
of sampling
error.
The range of fluctuations will tend to be narrower as
th e sample size is larger and the n u m b e r of measured var iables is sm aller.
If
one presumes
that
actual d ata are approximately normally dis tr ib uted ,
then the short
SPSS
syntax presented by Thompson and Daniel (1996)
m ay
be
used
to
conduc t
th e
analysis.
This
program creates randomly ordered
scores consisting of the same intege r valu es (e.g., 1, 2, 3, 4, or 5) represented
in some data sets such
as
Likert scale data.
Other approaches use regression equation s pu blish ed in prior sim ulatio n
studies to
pred ic t
th e
eigenvalues that would
be
produced
from th e
correla-
t ion m atr ices for random ly ordered d ata having a cer ta in rank. M ontanell i
and Humphreys (1976) produced one such equation that uses as inpu t only
the ac tual sample
size
and the nu m ber of measured var iables . Lautenschlager ,
Lance ,
and
F laher ty (1989) pub lished
an
al ternative formula that uses these
tw o
predictors
bu t
also uses
the
rat io
of
var iables-to-par t ic ipants
as another
predic tor var iable .
A more sophisticated approach exactly honors not only the types of
possible
values contained in the da ta set but also the exact values (and
their d escriptive statistics includ ing shape statistic s such as skewness) in
th e real data set. Table 3.2 presents the first 25 scores from A p p e n d i x A
on the
variables PERI, PER2, PER5,
and
PER6.
The data on these
variables
34 E XPL OR AT OR Y AND C O N F I R M A T O R Y FA C T O R ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 37/185
TABLE 3 2
Variables Randomly Sorted for Parallel Analysis
Original data
ID
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Mean
SD
PER1
8
5
6
5
5
7
8
7
1
9
9
4
6
7
9
9
8
8
9
8
9
7
7
8
8
7.08
1.93
PER2
7
7
5
5
5
7
8
7
3
9
9
3
6
7
9
9
8
5
9
7
9
7
8
9
8
7.04
1.84
PER5
3
4
5
4
5
7
6
5
1
5
7
6
7
2
7
5
7
4
6
5
9
7
5
5
6
5.32
1.75
PER6
2
5
3
4
4
8
7
3
1
7
9
5
7
3
5
5
6
8
7
6
9
6
6
3
6
5.40
2.14
Randomly sorted data
RAN1
9
8
7
5
5
5
9
9
8
8
8
7
1
8
7
9
9
7
6
4
8
7
6
8
9
7.08
1.93
RAN2
5
9
5
7
5
5
9
7
9
3
8
3
7
9
9
7
7
7
6
8
8
7
9
8
9
7.04
1.84
RAN5
3
7
5
4
4
5
5
6
2
5
1
7
7
5
7
9
6
4
7
6
7
5
6
5
5.32
1.75
RAN6
6
4
5
4
7
3
3
5
8
7
9
6
1
3
5
7
3
6
6
2
6
5
9
8
7
5.40
2.14
were independen t ly randomly ordered to create th e scores on va r iab les
RANI,
R A N 2 , RAN5, and RAN6.
N o te that the means and s tandard deviat ions on rela ted actual and
randomly ordered data match exactly (e.g., X P H R I = X R A N I
=
7.08; SDpgRj =
SD^AMi
=
1-93). Other d escriptive statistics, such as coefficients of skewness
and kur tos is ,
wou l d
also match fo r each ac tua l and randomly ordered vari-
able pair .
However, randomly ordered scores should have bivariate correlations
approaching zero with eigenvalues all f luc tua t ing around one. Table 3 .3
presents the correlation matrix for the actual data presented in Table 3.2,
and correlation
coefficients
for the same data after independent random
ordering of the scores on each var iable .
Note that
the correlation
coefficients
for
the
randomly ordered var iables
are
closer
to
zero
than the
correlations
among
the
actual scores.
DECISION
S E Q U E N C E
35
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 38/185
TABLE
3.3
Bivariate Correlations For Actual and Randomly Ordered Data
PER1/ PER2/ PER5/ PER6/
Variable
RAN1 RAN2 RAN5 RAN6
PER1/RAN1
.15
-.13
.41
PER2/RAN2 .87 .01
-.02
PER5/RAN5 .51 .46 -.55
PER 6/RAN6 .54 .43 .73
Note. Th e
coefficients above
t he
diagonal
are
correlations among
the
randomly ordered data.
The co-
efficients
below th e diagonal are correlations among the actual data in its original order as presented in
Table 3.2.
TABLE 3.4
Eigenvalues for Actual and Randomly Ordered Data
Number
1
2
3
4
Sum
Actual
data eigenvalue
2.776
0.828
0.276
0.119
4.000
Random
order eigenvalue
1.756
1.082
0.797
0.364
4.000
In
parallel analysis
the
eigenvalues
in
successive pairs from
the
actual
and
the randomly ordered data are compared. Factors are retained whenever
the
eigenvalue
for the
actual data
for a
given factor exceeds
the
eigenvalue
for
the
related factor
for the
randomly ordered scores.
These
eigenvalues
for
the
Table
3.2 data and the Table 3.3 correlation matrices are presented
in Table 3.4.
For this example, because 2.776 is greater than 1.756, the first factor
from
the
actual data would
be
extracted. Because 0.828
is
less than 1.082,
the second factor would not be extracted.
Brian
O'Connor
(2000)
has
written several
SPSS and SAS
syntax
files
to execute variations on
these
analyses. The programs can also be down-
loaded from http://flash.lakeheadu.ca/~boconno2/nfactors.html
FACTOR EXTRACTION METHOD
There
are
numerous statistical theories
that can be
used
to
compute
factor
pattern coefficients. Probably the most frequently used EFA extraction
method, perhaps because it is the default analysis in most statistical packages,
is called pr incipal components analysis (see Russell, 2002).
Principal components analysis assumes that the scores on measured
variables have perfect reliability.
Of
course, scores
are
never perfectly reliable
36 E X P L O R A T O R Y A N D CONFIRMATORY FACTOR
N LYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 39/185
(see Thompson,
2003).
Furthermore, th e analysis attem pts to reproduce th e
variance
or information in the sample data, rather than the population. Of
course, if the sample is reasonably representative of the population data,
th e sample factors should tend to match th e population factors.
Principal
components analysis, which actually
was the
analysis
d e-
scrihed in
chapter
2,
uses ones
on the
diagonal
of the
correlation
mat r ix .
This seems reasonable to the extent that scores o n a given m easured variable
should correlate perfectly with the same scores on the same variable.
However, if scores on a variable are not perfectly reliable, then cor-
relations
of
scores
on a
measured var iable with itself will
not be
perfect
(i.e., +1.0)
if
measurement error
is
taken into account. Logically,
it
might
be reasonable to alter the d iag on al entries of the c orrelation m atri x to take
measurement error
into
account.
It was
noted previously that co m m un ality coefficients
are
lower-bound
(or
conservative) estimates of score reliability. Another common factor
extraction method, principal axes factor analysis uses
the
communal i ty coeffi-
cients to replace th e ones on the diagonal of the correlation matrix.
Principal axes factor analysis often starts w ith
a
principal components
analysis.
Then the
c o m m un a l ity coefficients
from
this analysis
are
su bst i tu ted
respectively for the ones on the diagonal of the original correlation matrix.
It
is
im portant
to
emphasize that
no
off-diagonal data
are
altered
in
this
process. Then
a new set of
factors
and
their corresponding com m unali ty
coefficients are
estimated.
This
process
is
successively repeated un til
th e
communal i ty es t imates
stabilize. This process is called i terat ion. Iteration is a statistical process of
estimating a set of results, an d then
successively
tweaking results un t i l tw o
consecutive sets
of
estimates
are
su fficiently close
that the
iteration
is
deemed
to
have converged. Iteration
is
used
in
var ious ways throughout mult ivar ia te
statistics,
inc luding
in
other aspects
of
EFA, such
as
factor rotation, which
is discussed in the next section.
The principal axes communality iteration process does not always
converge. Estimates may sequentially bounce u p, bounce dow n, bounce up,
and so forth. If this happens, often the sample size is too small for the
num ber of m easured variables and the m odel being estimated.
Always carefully check whether i teration
has
converged. Some com-
puter programs print only innocuous-sounding warnings about failure to
converge, and then still
pr in t full results.
So the unwary
analyst
may not
notice
a
convergence problem. Programs also have b ui l t- in maxim um nu m-
bers of
iterations (e.g., 25) .
If a
solution does
not
converge, override
th e
default n um ber of i terations. If a solu tion does no t converge in 100 or 125
iterations, u sually the solution will not converge at any num be r of iterations.
Either
a different factor extract ion method must be selected or sample size
must
be
increased.
DECISION SEQUENCE 37
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 40/185
Principal
components
analysis
does
not
have
to be the f i r s t step of the
principal axes factor an alysis. An y reliab ility estimates that the researcher
deems reasonable can he used as initial estimates. For example, it the mea-
sured
va riab les are test scores, the relia bil ity coeffic ients reported either in
test ma nu als or in a prior "reliability generalization" (R G ) an alysis (V acha -
Haase, 1998) can be used as initial estimates placed on the diagonal of the
correlation matrix.
Another strategy for creating initial estimates of com munal i ty coeffi-
cients is available in most statistical packages. The multiple
R
2
of a given
measured variable with all the other measured variab les is used as the initia l
estimate
for that
variable's
h
2
(e.g.,
h
}
2
=
R
2
P E
R i
x
P E R 2 , P E R 5 , P E R 6 , ^ =
R
2
p E R 2 x
P E R I , P E R 5 , P E R 6 ) -
In theory, what one u ses as the ini tial com m un ality coefficients in this
iteration process
will
affect
how
many iterations must
be
performed (with
princip al components inv okin g ones on the diagonal of the correlation
matrix obviously taking
th e
most iterations)
but not
what
are the final
estimates.
If
this
is so,
given that
the
only downside
of
numerous iterations
is
wasting a little electricity, it doesn't matter too much which initial
estimation procedure
is
used.
Other factor extraction method s includ e alpha factor analysis, maxi-
m um
likelihood factor analysis, image factor analysis,
a nd
canonical factor
analysis.
Alpha factor analysis focuses on creating factors with maximum
reliability. Max imum likelihood
analysis focuses
on
c reating factors that repro-
duce
th e
correlation
or
covariance matrix
in the
population, versus
in the
sample. Image factor analysis focuses
on
creating factors
of the
latent variables
that exclude
or
m inim ize "unique factors" consisting
of
essentially only
on e
measured variable. Canonical
factor
analysis seeks to identify factors that are
maximally related to the measured variables.
FACTOR ROTATION
Factor rotation involves mov ing
th e
factor axes measuring
the
locations
of the measured variables in the factor space so
that
the nature of the
underlying constructs becomes more obvious to the researcher. Factor rota-
tion
can be
performed either graphically, guided only
by the
researcher's
visual jud gm ent, or em pirically using one of the d ozens of
different
algorithms
that have
been
developed over th e years. Analytical rotation has some
advantages over graphical rotation, because
th e
procedure
is
automated,
and
because
any two
researchers analyzing
the
same data will arrive
at the
same factor struc ture when using the same em pirical rotation method.
38
EXP LOR ATOR Y
A N D
C O N F I R M A T O R Y
F ACTOR
AN ALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 41/185
TABLE 3.5
Correlation
and
Squared Correlation Coefficients
for 6
Variables
and
100 Participants
Variable
PER1
PER2
PER3
PER5
PER6
PER7
PER1
1.00
.85
.78
.40
.32
.48
PER2
72.3%
1.00
.72
.45
.32
.50
PER3
60.8%
51.8
1.00
.51
.40
.50
PER5
16.0%
20.3%
26.0%
1.00
.66
.80
PER6
10.2%
10.2%
16.0%
43.6%
1.00
.71
PER7
23.0%
25.0%
25.0%
64.0
50.4
1.00
Note.
Th e (unsquared)
correlation coefficients
are
presented
in the
bottom
triangle. T he squared
correla-
tion
coefficients are presented as
percentages
in the top
triangle;
values
above
43% are italicized.
Rotation
is not
possible when only
one
factor
is
extracted.
But in
vir tual ly
all cases involving two or m ore factors, rotation is u sually essential
to
interpretation.
T he first 100 scores on six LihQUAL+™ va r iab les (PERI , PER2,
PER3, PER5, PER6,
and
PER?) presented
in
Appendix
A can be
used
to
demonstrate
w hy
rotation
is so
essential
in
EFA. Table
3.5
presents
th e
bivar iate r and r
valu es associated w ith
these
data .
A n examination of the r va lues in Table 3.5 ( r e m e m b e r
that
r values
mus t
be squared before their magnitudes can be compared in an in terval
metric) suggests (a) the existence of two factors, consisting respectively of
PE RI , PE R2, and PER3, and of PER5, PER6, and PER 7, because a ll pairs
of
th e
variables
in these tw o
sets have
r
values greater
than
4 3 % ,
and (b)
the two
factors probably
are not
highly correlated, because
the r
va lues
between measured variables across the two sets are all less than 2 7 %. T he
unrotated pattern /s tructure coeff ic ients
and the
eigenvalues presented
in
Table
3.6 confirm th e expectat ion of two factors. The two factors reproduc e
84.2% (3.81 + 1 .23 = 5.05 / 6 = .842) of the var iance in the correlation
matrix involving the six measured variables.
However ,
th e
u nrotated factor pattern /s tructu re coefficients presented
in Table
3.6 do
n o t match
ou r
expectations regarding
th e
composition
of
th e factors. For example , as reported in the table, all six variables are
correlated
.70 or
more with Factor
I Yet the
F igure
3 .2
plot
of the
factor
space, based on the p attern/ struc ture coefficients reported in Table 3.6, d o
match ou r expectat ions regarding factor s tructure.
This
suggests
that
the
patterns in our unrotated results are exactly what w e expect, b u t that th e
patterns
cannot
be readily discerned
from
statistical results in the form of
unrotated results .
The idea of rotation is that th e locations of the axes in the factor
space are com pletely arbi trary. W e can move (i.e., rotate) these axes to any
DECISION S E Q U E N C E 39
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 42/185
TABLE 3.6
Unrotated Principal Components
Pattern/Structure Coe fficients
Unrelated factors
Squared
structure
Variable
PER1
PER2
PER3
PER5
PER6
PER7
X
I
0.81
0.81
0.82
0.80
0.70
0.83
II
-0.50
-0.46
-0.35
0.43
0.54
0.41
I
65.9%
65.6%
67.9%
63.6%
49.2%
69.3%
3.81
II
25.3%
20.9%
12.1%
18.8%
29.7%
16.4%
1.23
h
2
91 .2%
86.5%
80.0%
82.4%
78.9%
85.7%
Note. Pattern/structure coefficients
>
|.35|
are
italicized.
For the
principal components model
the
eigenval-
ue s
equal
the
sums
of the squared unrelated
pattern/structure coefficients
for a
given factor
(e.g.,
.659
+
.656 + .679 + .636 + .492 + .693 = 3.81).
O
1
0.5 -
n
1 -0.5
-0.5
1
0.5
A
Series
1
Factor
I
Figure 3.2. Table
3.6
variables presented
in
factor space before rotation.
location, and in any manner we wish. However, we may not move the
location
of any
variable
in the
factor space.
Persons first learning of rotation are often squeamish about the ethics
of this proced ure. Ce rtainly ethics have
an
important bearing
on the
exercise
of statistical choice. If w e have several mice in a d rug experiment, a nd Elwood
in the e xperim ental group is not responding w ell to the new intervention, i t
is unethical to take Elwood out for a li ttle walk a t 3:00 a.m. that results in
Elwood's "accidental" demise. That
is
unethical
scientific
conduct (and
cruelty to Elwood) . B ut factor rotation is not in any respect unethical (a s
long as we only move fac tor axes and not measu red va riable s).
Table 3.7 presents an exam ple of an em pirical rotation of the Table
3.6
pattern/s tructure
coefficient
matr ix .
Note
that (a) the communality
coefficients fo r each of the variables rem ain unaltere d by rotation and (b)
the total variance reproduced in the factors as a set (84.2% = [2.59 + 2 .46 =
5.05
/ 6 =
.842]) remains unaltered.
40
E XPL OR AT OR Y A N D C O N F I R M A T O R Y FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 43/185
TABLE
3.7
Rotated Principal Components Pattern/Structure Coe fficients
Variable
PER1
PER2
PER3
PER5
PER6
PER7
Trace
Rotated factors
I II
0.94 0.19
0.90 0.23
0.84 0.31
0.28 0.86
0.13 0.88
0.32 0.87
Squared structure
I II
87.5% 3.7%
81.4% 5.1%
70.1% 9.9%
7.9% 74.5%
1.8% 77.1%
10.6% 75.2%
2.59 2.46
/i
2
91 .2%
86.5%
80.0%
82.4%
78.9%
85.7%
Note. Pattern/structure coefficients
>
|.35|
are
italicized.
For the
principal components
model
th e
trace vari-
ance
reproduced by the
rotated factors equals
the
sums
of the
squared rotated pattern/structure coeffi-
cients for a
given
factor (e.g., .875 + .814 + .701 + .079 + .018 + .106 = 2.59).
However,
it is
true that rotation tends
to
redist r ibute
th e
location
of
variance within th e solut ion. In the unrotated solution the first factor
reproduces th e b u l k of the observed variance (3.81 / 6 = 63.5%). In the
rotated solution the two rotated factors reproduce 43.2% (2.59 / 6) and
41.0% (2.46 / 6) of the observed v ariance. The rotation has almost equalized
the location of the informat ion within the solution. This tends to happen
to the extent that rotations are performed in pursu i t o f "simple structure."
Thurstone (1935,
p.
156, 1947,
p.
335) proposed
the
basic concept
of
simple structure
and the
associated criteria
fo r
evaluating simple structure.
Think about what set of factor pattern coefficients would maximize our
abil i ty
to
interpret
th e
na ture
of the
latent construc ts und erlyin g scores
on
measured variables.
T he un rotated results presented in Table 3.6 w ere virtu ally impossible
to
interpret because
all of the
m easured v ariab les were very highly correlated
with
both
factors. A more interp retation -frien dly solution w ould strike a
balance between hav ing enoug h m easured variables correlated w ith
a
factor
that some measured variables reveal the underlying nature of a factor, but
not so many that th e factors cannot be d is t inguished from each other or
are so complex
that
their nature
cannot
easily he apprehended.
A s imple s t ructure would involve
from
a column perspective roughly
an equal number of measured variables being appreciably correlated with
each factor (e.g., 6/2 = 3), bu t from a row perspective m ost or all measured
variables being appreciably correlated with only one factor.
These
criteria
are met
very well
for the
rotated factor solution presented
in Table
3.7.
Empir ical (i .e., nongraphical) rotation methods perform
the
rotation
by applying an algorithm to derive a transformation matrix ( T p x p ) b y which
the unrotated pattern coefficient mat r ix ( P v x p ) is postm ul t ipl ied to yield
the rotated factors. There are dozens of formulas for deriv ing T p
x
p matrices,
DECISION SEQUENCE
41
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 44/185
several of w hich are availab le in various statistical packages. These c o m p u te r
programs wil l
pr in t the transformation matr ices on request. The more a
t ransformation matr ix deviates
from
be ing an identi ty matr ix, the more the
unrotated versus th e rotated factor pattern coefficients
wi l l
b e different .
Orthogonal Rotation
Figure
3.2 and Table 3.7 involve a case in which th e original 90-degree
angle between the two factors was m ain ta ined in the rotation process. This
means that the rotated factors wi l l remain orthogonal or uncorrela ted.
Factors are always orthogonal upon e xtraction, and rem ain unc orrelated if
orthogonal rotation is used.
The correlation
coefficient
ac tu ally is a fun ctio n of the cosine of the
angle between ei ther measured or composi te var iables .
When
two var iables
have been standardized to have a length or var iance of 1.0, as factors are,
the cosine of the angle b etw een tw o factors is the correlation o f the
factors
(see Gorsuch, 1983, chap. 4).
The
most common or thogonal rota t ion method,
and
indeed
the
most
common rotat ion of a n y ki nd , is the v a r imax rota t ion method developed by
Kaiser
(1958) .
This is the
default rotation
in
most statistical packages.
V a r i m a x tends to focus on m a x i m i z i n g the
differences
between the squared
pattern /s tructure coefficients on a factor ( i .e. , focuses on a column perspec-
tive ). In my experience, in about 85 % of exploratory factor analyses va r i max
will yield s imple s tructure.
Another
orthogonal rotation method
is
quar t imax
which tends
to
focus on s imple s t ruc tu re
pr imar i ly
from th e perspective of rows (i .e. , fo r
variables across
th e
factors) . Quart imax
is
most appropriate w hen
the first
factor is expected to be a large "G" or general factor that is saturate d by
disproportionately many of the variables.
E q u a m a x
is a compromise betwe en
var i max an d quar t imax methods .
Oblique Rotation
In some cases s imple s truc ture cannot be obtained us ing or thogonal
rotation. Typically this is indicated by var iables having pattern /s tructure
coefficients that are
large
in
absolute value
on two or
more factors (which
is sometimes cal led mult ivocal v s. un ivoca l ) . Such factors are difficult to
interpret. The researcher must then tu rn to oblique rota t ion in the pursu i t
of s imple s t ruc tu re .
However , as noted in chapter 2, w hen factors are rotated orthogonally,
the
pa t te rn coefficient
and the
s tructu re coeff ic ient m atr ices , which always
have the same ran k, also
wil l
contain exactly the same numbers . This occurs
because when factors are uncorrelated the interfactor correlation matrix
42 EXPLORATORY A N D C O N F I R M A T O R Y FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 45/185
p )
is an identity matrix, and given that P y
x
p R F X F
=
S v x F i if R p x F
=
I p
x F
, then necessarily P
V x
F
=
S y
x
p .
But once
factors
are
rotated obliquely,
th e
factors
are
correlated,
and
P\/ x F Sy
x
F -
Correlations among the factors that deviate further
from
zero
lead
to greater
differences
between the pattern and the structure
coefficients.
A nd
once factors
are
rotated o bliqu ely,
th e
pattern coefficients
are no
longer
correlation
coefficients.
On ob lique ly rotated factors, any or even all of the
pattern coefficients can be outs ide th e range
—1
to +1.
When orthogonal rotation is performed in computer packages,
R F
X
F
will not be printed, because the interfactor correlation is by d efini t ion known
to be the identify matr ix . And because PV x F equals Sy
x
F. only one factor
matr ix
wi l l
be
printed, which
may be
somewhat ubiquitously labeled some-
thing like "varimax-rotated factor matrix." When
an
oblique rotation
is
performed, all
three matrices
(Py
x
F >
RF x
F )
and Sy
x
F)
will always
be
printed,
so
that
the
results
can be
correctly interpreted
by
consult ing
both
Py
x
p
a n d S
V x F
.
When
oblique rotation is necessary, p r om a x rotation (Hendrickson &
White, 1964) is almost always a good choice. Promax is actually conducted
as a series of rotations. The first rotatio n is an orthogonal rotation, such
as var imax . These pa t tern /s t ruc ture coefficients are then raised to some
exponential power,
k
which is sometimes called a pivot power. If an even-
numbered pivot power (e.g.,
k =
4) is used, the signs of the resulting
coefficients
are restored after the
exponential operation
is performed .
Raising
var imax pattern/s tructure coefficients
to an
exponential power
makes
all the
resulting coefficients closer
to
zero.
But the effect is
differential
across
origin al value s that
ar e
larger
in
magnitude versus smaller
in
magni-
tude. Fo r example, .9
4
is .656, whereas ,2
4
is .002. The resulting coefficients
tend to have simple structure.
N e xt , a "Procrustean" rotation is invok ed . Procru stean rotation can
also be useful as a rotation method on its own merits and not as part of
promax rotation,
as
when
a
researcher wants
to
rotate factors
to
best
fit a
theoretical structure (Thompson & Pitts, 1982), or when conducting an
EFA
meta-analysis (Thompson, 1989). Indeed , some test distrib utio ns have
been developed
fo r
s tatis tically
evaluating
the fit of
actual
and
theoret ical ly
expected structures,
or of
actual structures
from
d ifferent studies (Thomp-
son, 1992b).
Procrustean rotation derives its name from the Greek myth about an
innkee per w ith only one bed . If gu ests we re too short for the bed, he stretched
the guests on a rack. If guests were too tall for the bed , he cut off parts of
their legs.
Procrustean rotation finds the best fit of an actual matrix to some
target matrix (e.g., a hypothetically expected pattern matrix, the factor
solution for the same variables in a prior
s tudy) .
In the promax use of
DECISION SEQUENCE
43
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 46/185
Procrustean rotation,
th e
varim ax pattern/struc ture coefficients each raised
to the pivot power hecome th e target. T he resu lting Procrustean rotation
i s the promax solution.
The researcher can control the degree of correlation am ong th e factors
( R p
x
p )
hy
changing
th e
pivot power.
T he
higher
the
exponent,
the
greater
will be the
degree
of
correlation among
the
factors.
Another oblique rotation method
is oblimin.
Ob limin invokes
a
va lue
called delta
to
control
the
degree
ot
correlation among
th e
rotated factors.
Delta values
of
zero yield factors
that are
more highly correlated, whereas
large negative valu es produ ce factors that
are
more uncorrelated. However,
in cases where oblique solutions are required, promax is
usually
a very
good choice.
FACTOR SCORES
As
noted
in c hapter 1, sometim es factor analysis is used not to
identify
an d
interpret factors,
hut to
obtain
a
more parsimonious
set of
composite
scores (i.e., factor scores)
that are then
used
in
subsequent analyses (e.g.,
regression) instead
of the
measured variable scores. There
are
four methods
of obtaining factor scores (i.e., composite variable scores for each person
on
each
factor)
that
are briefly summarized here.
Chapter
4 includes more
detailed
comparisons of these methods.
The
most commonly used method
of
obtaining factor scores
(F^
x
y)
is called the regression
method.
As a first step, the measured variables are
each converted into z scores with means of zero and SDs of 1.0.
Then
the
following
algorithm
is
applied:
F N x F
=
Z N
x
v R V x v P V x
F -
The rightmost portion of the formula can be reexpressed as :
W v x F
R-VxV PvxF-
This is the "factor score matrix output" by
SPSS.
Second, Bartlett (1 937 ) proposed m ethods for estim ating factor scores
intended
to
minimize
the
influence
of the
un iqu e factors consisting
of
single
measured
variables not usu ally extracted in the analysis. Third, Anderson
and R ub in (1956) proposed methods sim ilar to Bartlett 's methods, b ut w ith
an added constraint that th e resu lting factor scores m ust be orthogonal.
All of
these factor score estimation procedures yield factor scores that
are themselves z scores,
if the
extraction method
is
principal components,
or
the
estimation method
is
Anderson—R ubin. This does
n ot
create
difficulties if
44
EXPLORATORY A N D CON FIRMATO RY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 47/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 48/185
TABLE 3.9
z Scores and Standardized Noncentered Scores for the
Four
Measured Variables
z
scores
ID
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Mean
SD
ZPER1
.48
-1.08
-.56
-1.08
-1.08
-.04
.48
-.04
-3.14
.99
.99
-1.59
-.56
-.04
.99
.99
.48
.48
.99
.48
.99
-.04
-.04
.48
.48
0.00
1.00
ZPER2
-.02
-.02
-1.11
-1.11
-1.11
-.02
.52
-.02
-2.20
1.07
1.07
-2.20
-.57
-.02
1.07
1.07
.52
-1.11
1.07
-.02
1.07
-.02
.52
1.07
.52
0.00
1.00
ZPER5
-1.33
-.75
-.18
-.75
-.18
.96
.39
-.18
-2.47
-.18
.96
.39
.96
-1.90
.96
-.18
.96
-.75
.39
-.18
2.10
.96
-.18
-.18
.39
0.00
1.00
ZPER6
-1.59
-.19
-1.12
-.65
-.65
1.21
.75
-1.12
-2.06
.75
1.68
-.19
.75
-1.12
-.19
-.19
.28
1.21
.75
.28
1.68
.28
.28
-1.12
.28
0.00
1.00
Standardized
CPER1
7.56
6.00
6.52
6.00
6.00
7.04
7.56
7.04
3.94
8.07
8.07
5.49
6.52
7.04
8.07
8.07
7.56
7.56
8.07
7.56
8.07
7.04
7.04
7.56
7.56
7.08
1.00
CPER2
7.02
7.02
5.93
5.93
5.93
7.02
7.56
7.02
4.84
8.11
8.11
4.84
6.47
7.02
8.11
8.11
7.56
5.93
8.11
7.02
8.11
7.02
7.56
8.11
7.56
7.04
1.00
noncentered
CPER5
3.99
4.57
5.14
4.57
5.14
6.28
5.71
5.14
2.85
5.14
6.28
5.71
6.28
3.42
6.28
5.14
6.28
4.57
5.71
5.14
7.42
6.28
5.14
5.14
5.71
5.32
1.00
CPER6
3.81
5.21
4.28
4.75
4.75
6.61
6.15
4.28
3.34
6.15
7.08
5.21
6.15
4.28
5.21
5.21
5.68
6.61
6.15
5.68
7.08
5.68
5.68
4.28
5.68
5.40
1.00
ote
Th e
means
of the
standardized noncentered
scores are the
same
as the
means
of
these
variables
in their
original
form, a s
reported
in Table
3 2
descriptives variables=cperl
to
cper6
.
compute btscrl= .55080 * cperl) + .62861 * cper2)
+
-.17846
* cperB) +
-.18121
*
cper6)
.
compute btscr2= -.11583 * cperl) +
-.22316
* cper2)
+
.60668 * cperS) + .61055 * cper6) .
print formats fscrl fscr2 btscrl btscr2 F8.3) .
Table
3.10
presents
both
th e regular
regression
factor scores and the
standardized
noncentered
factor
scores
for
these data.
Both sets of factor
scores
are
compared
here
only
fo r
heuristic purposes.
In
ac tua l research
th e
score es t imation method would be selected on the basis of the research
questions and the related preferences for
different
factor
score
properties.
EXPLORATORY
AND CONFJRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 49/185
TABLE
3 10
Conventional Regression Factor Scores
F
2 5 X
2 )
and Standardized
Noncentered Factor
Scores for the EFA of the Four Measured Variables
Regression Standardized noncentered
ID
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Mean
SO
FSCR1
.773
-.437
-.770
-1.037
-1.139
-.428
.386
.199
-2.300
1.115
.741
-2.295
-.970
.505
1.080
1.284
.368
-.522
1.013
.230
.537
-.259
.288
1.169
.470
0.000
1.000
FSCR2
-1 .824
-.442
-.483
-.485
-.138
1.334
.520
-.786
-1
.898
-.008
1.256
.797
1.230
-1 .826
.115
-.578
.582
.476
.339
.010
1.950
.763
-.052
-1
.089
.235
0.000
1.000
BTSCR1
7.170
5.960
5.627
5.360
5.258
5.969
6.783
6.596
4.097
7.512
7.138
4.102
5.427
6.903
7.477
7.681
6.765
5.875
7.410
6.627
6.934
6.138
6.685
7.566
6.867
6.397
1.000
BTSCR2
2.309
3.691
3.650
3.649
3.996
5.467
4.654
3.348
2.235
4.126
5.390
4.930
5.363
2.307
4.249
3.555
4.715
4.610
4.472
4.143
6.083
4.897
4.082
3.045
4.369
4.133
1.000
T he
standardized noncentered factor scores
are perfectly
correlated
w i t h their regular regression factor score counterparts (i.e., r^scRi x B T S C R I
=
1.0 and
f p s c R Z
x
R T S C R Z
=
1-0)i
anc
l
the
correlations across
the two
factor
scores also match across the two score types (i.e., T F S C R I x F S C R 2
=
r
B T S C R i x
R T S C R 2 *
which here
is
0.0).
For the
Tahle
3.10
results,
a
comparison
of the
means on the
s tandardized
noncentered
factor scores suggests
that the 25
participants on the
average rated their
l ibraries
higher
on
Factor
I
(6 .397)
than on
Factor
II
( 4 -133) .
M A J O R
CONCEPTS
The sequence of EFA involves num erous decisions (i.e., what associa-
tion
matrix
to
analyze,
h ow ma n y
factors
to
extract ,
factor
extraction method,
DECISION SEQUENCE
47
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 50/185
factor rotation method, factor score estimation method). For each of the
steps
of
analysis, there
are
myriad possihle decisions
or
ways
of
informing
judgment . Thus, across all the steps and all the choices fo r
each
step, th e
possihle number
of
com hinations
of
analytic
choices is
huge
It is
also important
not to
l imit
the
analysis
at any
stage
to a
single
analytic
choice.
For example, in deciding how many factors to extract,
researchers do well to consult several procedures (e.g., scree plot, parallel
analysis) and
then
evaluate
whether
various strategies corroborate
each
other. The most confidence can be vested in solutions fo r w hich factors are
stable
or
invariant over various analytic
choices.
The
purpose
of EFA is to
isolate factors that have simple structure
or ,
in
other words, are interpre table . Any thoug htful ana lytic choices that yield
clear factors ar e justif ied, including analyzing th e data in lots of different
ways
to "let th e data speak."
In practice, factor rotation is necessary. The only exception is when
a
single factor
is
extracted,
in
which case
no
rotation
is
possible.
Nor in
such a case is
rotation
needed, b ecause presum ably every variab le will saturate
the single factor, and the factor is w hatev er explains all the v ariable s bein g
highly
associated with
each other.
Usually when m u ltip le factors
are ex-
tracted, reasonable simple structure
is
realized with varimax rotation.
A nd
i f
these
factors
ar e
interpretable,
th e
correct rotation method
has
typically
been identif ied. However, even when
an
orthogonal rotation yield s simple
structure,
when
th e
purpose
of the
analysis
is to
obtain factor scores
for use
in subsequent statistical analy ses, if the researcher be liev es
tha t
the latent
constructs
have
notew orthy correlations,
an
oblique rotation (e.g., promax)
may be
sui table.
4 8 E XPL OR AT OR Y A N D C O N F I R M A T O R Y FACT OR A N A L Y S / S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 51/185
4
COMPONENTS VERSUS
FACTORS CONTROVERSIES
In the previous chapter it was noted that there are myriad methods
wi th
which
to
extract factors, some
of
wh ich
are
avai lab le
in
commonly
used statistics packages.
It is
impor tant
to
understand
th e
extent
to
which
extract ion choices
m ay
affect factor interpretations, and, especially,
to
under-
stand why and when differences m ay arise across method s for a given data set.
For reasons that
wil l b e
explained mom entari ly ,
i t
m ight
be
argued that
wh e n
th e
n u m b e r
o f
var iables
is [at
least] moderately large,
fo r
example,
greater than
30, and the
analysis contains
virtually no
variab les expected
to have low com m unalit ie s (e .g. , .4) , practically any of the exploratory
[extraction] proced ures . . .
wil l
lead to the same interpretations.
(Gorsuch,
1983,
p.
123)
Similarly , Cliff (1987) suggested that "the choice of common factors or
components methods often makes
vir tual ly
no
difference
to the conclusions
of
a
s tudy"
(p.
349). These d y namics will
be
i l lustrated here with data
for
different variables
and
participants from
the
Appendix
A
scores.
Table 4.1 compares
three
of the more elegant extraction methods that,
unlike principal components analysis, each invokes iteration
as
par t
of the
extract ion process: alpha, m ax im um l ikel ihood, and image factor analysis.
As an
inspection
of the ta ble d results suggests, for these d ata the three
extraction methods yield reasonably comparable factor structures.
49
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 52/185
TABLE 4.1
Varimax-Rotated Pattern/Structure Coefficient Matrices
fo r
Three Factor
Extraction Methods
n =
100 Faculty)
Variable
PER1
PER2
PER3
PER4
PER5
PER6
PER?
PER8
PER9
PER10
PER11
I
.20
.12
.23
.16
.89
.87
.92
.72
.34
.21
.17
Alpha
II
.90
.82
.62
.86
.13
.14
.18
.31
.18
.23
.34
III
.18
.24
.22
.26
.22
.22
.20
.18
.44
.72
.56
I
.20
.11
.26
.19
.90
.87
.92
.72
.39
.20
.23
ML
II
.91
.86
.64
.84
.13
.15
.20
.32
.22
.23
.37
III
.16
.19
.12
.20
.18
.19
.09
.16
.32
.95
.40
I
.21
.12
.24
.18
.86
.84
.86
.71
.37
.25
.21
Image
II
.84
.82
.61
.80
.14
.16
.19
.32
.21
.31
.36
III
.19
.20
.19
.24
.20
.20
.18
.16
.35
.45
.42
Note.
Pattern/structure coefficients greater than |.4| are italicized. ML = maximum likelihood.
The largest difference occurs on the third factor, which involves only
2 or 3 of the 11
measured variables.
And the
results from
the
image factor
analysis differ somewhat more from the
first
tw o sets of results than these
tw o
sets
differ
from each other.
This is
partly because im age factor analysis
focuses on m inim izing the influence of factors involving single variables not
reflected
via
covariations
in the
"images"
of the
other
measured variables.
However,
the
d ifferences across extraction m ethods tend
to
decrease
as there are more measured variables. For example, regarding image analysis,
as
more measured variables are added
there
is an increasing likelihood
that
most measured variables will be correlated with another measured variable
or a combination of the other measured variables, so the influences of
"specific"
factors consisting
of a
single measured variable,
and not
represent-
ing true latent variables,
will
tend to become less as there are more m easured
variables.
Bear
in
mind that
th e
present small heuristic analysis involving
only
11
measured variables
is
unrea listic . Researchers would rarely
(i f
ever)
factor analyze
so few
measured variables.
MEASUREMENT ERROR
VERSUS
SAMPL ING ERROR
As
noted
in chapter 3, principal axes factor analysis is the same as
principal components analy sis, except that in principal axes factoring
only
the
diagonal
of the
correlation matrix
is
modified
by
removing
the
1.0's
originally alway s residing there and iteratively estimating the true commu-
50 EXPLORATORY A ND CONFIRMATORY
FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 53/185
TABLE 4 2
Communality Coefficients If ) Through Four Iterations
n - 100 Graduate Students)
Iteration
a r i a b l e
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
83714
74596
67464
65751
75740
55652
73583
65091
89374
76327
69098
67103
79961
58434
77671
67977
2
92230
76468
69167
67066
81378
58885
78818
68493
3
93880
76277
68996
66857
81939
58870
79155
68521
4
94904
76061
68843
66692
82197
58798
79257
68479
nality coefficients. This is
done
to take into
consideration
the fact that all
scores have some measurement error (i.e., scores are never perfectly relia ble).
Table 4-2
illustrates
th e iteration
process
for a
principal axes factor
analysis involving th e first eight measured variables presented in Appendix
A, and
l imi t ing
the
analysis
to the 100
graduate students
(i.e., the first 100
scores in Appendix A). In this particular analysis, before iteration w as begun,
R
values were substituted on the diagonal of the correlation matrix.
These
results
involve
the R between a given measured variable and the seven
remaining measured variables in the analysis.
For
example,
as
reported
in
Table
4.2,
the
R
2
fo r
predicting
the 100
scores
on the
measured variable, PERI, with
the
scores
on the other
seven
measured variables in this analysis (i.e., PER2, P ER 3,. . . PER8) w as .83714.
The R for
predicting
the 100
scores
on the
measured variable, PER2, with
th e
scores
on the other
seven measured variables
in
this analysis (i.e., PERI,
PER3, . . . PER8) was .74596.
Using the R value s in this man ner is sensible, although other ini t ia l
estimates of communal i ty m ay also be plausible, because the reliabilities of
th e
scores
do impact th e
multiple regression effect size (Thompson, 2003).
If
a particular variable was perfectly unreliable (i.e., measured
nothing,
or
only
random
fluctuations
in
scores),
then
the R"
would theoretically
be zero.
Table 4.2 then
presents
the /T
est imates
fo r
each
of four
successive
steps
of
iteration. These
can be
obtained
in
SPSS
by
invoking
the
CRITERIA
subcommand:
factor variables=perl to per 2/
analysis=perl
to
per8/
criteria=factors(2)
iterate(1)
COMPONENTS
VERSUS
FACTORS CONTROVERSIES 51
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 54/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 55/185
PRINCIPAL
COMPONENTS VERSUS
PRINCIPAL AXES
FACTORS
By
far the two
most commonly used estimation methods
are
principal
components an d principal axes factors analyses. Principal
com ponents
analy-
sis
is the
default
method
in
commonly used statistical packages,
and is
used
with
considerable freq uenc y whe never exploratory factor analyses (EFA s)
are perform ed (Fabrig ar et a l., 1999; Russell, 2002). The
only
difference
between these tw o methods is that 1.0's are used on the d iagonal of the
correlation matrix
in
principal components analysis.
In
prin cipal axes factor-
ing,
th e
communali ty est imates
on the
diagonal
of the
correlation matrix
are iterativ ely estimated un til the iterations converge (i.e., rema in
fairly
stable
across a contiguous pair of estimates).
It would be a major understatement to say
that
"analysts differ quite
heatedly over the uti l i ty of princip al components as compared to comm on
or
principal factor analysis"
(Thompson
&
Daniel, 1996,
p.
201).
As
Cliff
(1987) noted,
Some author i t ies
insist
that componen t analysis is the only suitable
approach,
and that the c o m m o n
factors methods just superimpose
a lot
of ext raneous m umbo jum bo, dea l ing w i th fundamen ta l l y immeasurable
things, th e c o m m o n
factors. Feelings are,
if
anything,
even
stronger
on
th e o ther
side.
. . .
Some even insist that
th e
term "factor analysis" must
not
even
be
used
when a
components analysis
i s
performed,
(p .
349)
Indeed,
an
entire special
issue on
this controversy
was
published
in
the j ou rna l ,
M ult ivariate
Behavioral
Research
(Mulaik ,
1992). And Gorsuch
(2003) recently noted that "in following th e dialogue for the past 30 years,
th e only major change seems to be that th e heat of the debate has in-
creased" (p. 155).
A series of com parisons of princ ipal com ponents versus principal axes
factors for the App endix A d ata m akes this discussion concrete. Table 4.3
presents a principal com ponents
versus
a prin cip al axes factor analysis of
TABLE
4.3
Components
Versus
Principal
Axes
Factors for
our
Measured
Variables
n= 100 Graduate Students)
Varimax-rotated
components
Variable
PER1
PER2
PER5
PER6
.94
.94
.28
.13
II
.19
.22
.8 6
.92
If
92.3%
92.7%
82.1%
85.3%
Varimax-rotated factors
.8 8
.90
.27
.16
II
.22
.25
.7 9
.7 8
tf
83.0%
87.2%
70.4%
62.9%
Note.
Pattern/structure coefficients
greater
than |.4|
are
italicized. Given
tha
small number
of
variables,
the
convergence criterion fo r principal axes factor extraction was raised from the default of .001 to .003.
COMPONENTS VERSUS FACTORS CONTROVERSIES 53
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 56/185
TABLE
4 .4
Varimax-Rotated Components Versus Factors fo r Eight Measured
Variables With Factors Estimated with Ft
2
Starting Values
and 1.0 Starting
Values n =
100
Graduate Students)
Components
Principal-axes
factors
3
Principal-axes
factors
Variable
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
I
.94
.8 8
.8 2
.8 2
.27
.15
.33
.27
II
.19
.23
.32
.29
.8 8
.8 4
.8 5
.8 3
tf
91
.5%
82.8%
77.2%
76.1%
84.0%
73.0%
82.6%
76.7%
I
.96333
.82979
.75473
. 74967
.26440
.18993
.32802
.28089
II
.18636
.26108
.34148
.32049
.86862
.74214
.82784
. 77795
tf
96.3%
75.7%
68.6%
66.5%
82.4%
58.7%
79.3%
68.4%
I
.96362
.82971
.75466
.74959
.26447
.18985
.32802
.28089
II
.18625
.26112
.34151
.32054
.86841
. 74235
.82782
.77796
tf
96.3%
75.7%
68.6%
66.5%
82.4%
58.7%
79.3%
68.4%
Note. Pattern/structure coefficients greater than |.4|
are
italicized.
Communality
estimation
was
begun using
fl
2
starting values.
Com munality e stimation
was begun
using
1. 0
starting values.
four m easured variables. Altho ugh the p atterns of the results (two m easured
variables
saturate each
of the two
latent
variables) are
similar,
tw o differ-
ences appear.
First,
th e
pattern/structure coefficients tend
to be
smaller
for the
pr inci-
pal axes factors. Second,
th e
co mmu na l i t y
coefficients
tend
to be
smaller
for
the principal axes factors. These tw o outcomes are a function of changing
only
the diagonal of the correlation matrix from 1.0s to sm aller positive
numbers w i t h in the prin cipa l axes analyses.
Table 4-4 presents a p rincipal com ponents
analysis
versus tw o principal
axes factor analyses of eight m easured variables. W ith regard to the pr incipal
axes factor analyses,
the first
analysis again began with h est imation using
R
values of each measured variable with the rem aining seven m easured
variables, and estim ation continued until convergence was achieved in seven
iterations.
The second prin cipa l axes factor
analysis
began
h
est imation
using
1.0 values (i.e., the unaltered correlation matrix was
used
as a starting
poi nt), and estim ation continued un til convergence w as achieved in eight
iterations. Clearly
the two
sets
of
results
are
comparable, even
fo r
this
example w ith only eight m easured variables, which is why these results had
to be
reported
to
more than
tw o
decim al values.
This is the general pattern in regard to selecting starting points for
iteration. Theoretically, roughly
th e
same values
will be
realized regardless
of
the
initial estim ates. H owever, using
less
reasonable initial estim ates (i.e.,
1.0's vs. R or rel iabil i ty
coefficients) will
require m ore iterations. On a
modern microcomputer ,
th e
add it ional computat ions
from
more iteration
will
usually
go unnoticed.
54
EXPLORATORY AND
CONFIRMATORY FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 57/185
The interpretations of the Ta ble 4-4 com ponents versus factors wou ld
be the same, even though the h values tend to be larger for the component
analysis.
And the two sets of results are more com parable for the Table 4-4
com parisons involv ing eight m easured variables than for the Table 4-3
comparisons involving
four
measured variables. Why?
Two factors affect the equivalence of the principal components an d
the princ ipal axes factors for a given d ata set. First, if the relia biliti es of the
scores being
analyzed
are quite high
(i.e.,
approach 1.0), then th e
h
values
will tend to each be roughly 1.0, which is the same
h
val ue used for each
variable in the principal components analysis. This m ight occur, for exam ple,
i
the researcher was analy zing m easured variables that were each scores on
tests each consisting of a large number of items.
Second, there is also a less obvious dyn am ic at work . The only difference
between princip al com ponents and princip al axes factor analyses involves
changing
the
diagonal
of the
correlation matrix when principal axes factor
analyses are conducted. But the impact of altering only the diagonal of the
correlation matrix
differs
depending on how many measured variables are
being analyzed
As
Snook
and
Gorsuch (1989,
p.
149) explained,
"As the
number
of
variables decreases, the ratio of diagonal to
off-diagonal
elements also de-
creases, and therefore the value of the communality has an increasing effect
on the analysis." For example, with four measured variables, there are 16
(4 X 4) entries in the correlation m atr ix, of which
four
are on the diagonal
(4 /16 = 25 .0%) . When there are eight m easured v ariables , there are 64
(8 X 8) entr ies in the c orre lation m atr ix, of wh ich eight are on the diagon al
( 8 / 6 4 = 12.5%).
When there are 100 measured variables, there are 10,000 entries in
the correlation matrix, of which 100 are on the diagonal (100 / 10,000 =
1.0%). And it is more common to factor analyze 10 0 than 4 measured
variables As there ar e m ore m easured variables, even it their reliabilities
are modest, principal components
and
principal axes factors tend
to be
increasingly similar (Ogasawara, 2000). And principal components have
some add itional desirable properties, as dem onstrated in chapter 5, that
makes
these
analyses attractive,
at
least
to
some researchers.
MAJOR CONCEPTS
There
are
numerous
EFA factor extraction methods
(e.g.,
alpha, canon-
ical,
image, m a xi m u m likelihood). However,
most f requent ly
seen in pub-
lished
research are principal components and principal axes methods. Princi-
pal
components
is the
default
method in
widely
used
statistical packages.
COMPONENTS VERSUS FACTORS CONTR OVERSIES 55
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 58/185
Principal axes m ethods acknowledge measurem ent error
by
iteratively
approximating
the
com m unali ty estimates
on the
diagonal
o f the
correlation
matrix.
When
score re liabilit y is high for the m easured variab les, there will
be
little difference between
th e
factors extracted using these
tw o
methods.
In any case, the differences in the two methods
will
also be smaller
as the number of factored variables increases.
This
is because iterative
estimation of comm onalities changes
only
the diagonal entries of the correla-
tion
matr ix ,
and the
diagonal
of the
matr ix
is
progressively
a
smaller con-
stituent of the correlation matrix as more measured variables are analyzed
(e.g.,
4 diagonal entrie s w ith 4 m easured variables is 4 out of 16 [25.0% ]
matrix entries;
8
diagonal entries with
8
measured variables
is 8 out of 64
[12.5%] ma trix entries). And in practice we usually factor
fairly
large num bers
of measured variables.
There is also a trade-off in considering measurement error variances.
Each sample data
se t
also contains sampling error variance.
The
more steps
we iterate to take into account measurement error, th e more statistical
estimates we are making, and the more opportunity we are creating for
sampling
error to influence our results.
56
EXPLORATORY AND CONFIRMATORY
FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 59/185
5
COMPARISONS OF SOME
EXTRACTION, ROTATION,
AND SCORE METHODS
The exploratory factor analysis (EF A) statistics called structure coeffi-
cients were
first
introduced
in
chapter
2, and the
statistics called communal-
ity
coefficients were
first
introduced
in
chapter
3 .
Both
of
these statistics merit
further discussion.
This
chapter further elaborates
the
concepts und er ly ing
s tructure
and communali ty (h ) coefficients hy invoking factor score statis-
tics.
First,
however, some comparisons of factor score estimation methods
may be helpful.
FACTOR SCORES IN DIFFERENT EXTRACTION METHODS
As
noted
in
chapter
4, the two
most commonly used
EFA
approaches
are principal components and principal axes factor analyses. And, as noted
in
chapter
3,
there
are
four different methods with which
to
estimate
factor
scores.
Here three
factor score methods (regression, Bartlett,
and
Anderson-
Rubin)
are
compared
to
faci l i ta te deeper understanding
of the
factor score
estimation choices
and to
elaborate addit ional differences
in the two
most
common
extraction methods.
In real
research,
factor scores are typically only estimated
when
the
researcher elects to use these scores in further substantive analyses (e.g., a
57
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 60/185
mul t iva r i a t e analysis of var iance comparing m ean
differences
on three factors
across men and women) . In these applications th e researcher will use only
one
factor score estimation method. Here several methods
are
compared
s imultaneously only
for
heuristic purposes.
Factor scores
can he
derived
hy
invoking
SPSS
COMPUTE
sta tements
that apply the factor score coefficient matrix (Wy
x
F) to the measured
variahles in either z
score
or
standardized, noncentered form. This
has the
heuristic value
of
making this compu tation process concrete
and
explic i t .
However , in actual research it is
both
easier and more accurate to obtain
the factor scores by using the
SPSS
SAVE subcommand within the
FACTOR procedure.
For ex am ple , to obtain regression factor scores
from
a principal compo-
nents analysis involving
11
measured var iables (PERI through P ER U ),
th e
syntax would
be:
factor variables=perl to per 2/
analysis=perl
to
perll/criteria=factors 3)/
extraction=pc/rotation=varimax/
print=all/save=reg all
reg_pc)/ .
The execution of this commands yields three factor scores for each partici-
pant f rom a pr incipal
components
analysis that are labeled REG _PC1,
REG_PC2, and
REG_PC3.
Factor scores from a principal axes factor analysis estimated using th e
Bart le t t method would be produced with the syntax:
factor variables=perl
to per 2/
analysis=perl to perll/criteria=factors 3)/
extraction=paf/rotation=varimax/
print=all/save=bart all bar_paf)/ .
Factor scores were computed for all 100
faculty
using the first 11 measured
variables presented
in
Appendix
A fo r all six
combinations
of two
extraction
methods (principal components vs. principal axes factor analysis) and three
factor score estim ation m ethods (regression, Ba rtlett ,
an d
Anderson—Rubin) .
Table
5.1 presents th e
three
sets of principal components factor scores fo r
the first 5 and
last
5
participants
in the
data set. Notice that when principal
components methods a re used, all three factor score methods
yield
identical factor
scores
for a given participant (e.g., fo r participant 101, REG_PC1 =
BAR_PC1 = AR_PC1 = +.35) .
Table 5.2 presents the three sets of principal axes analysis factor scores
for the first 5 and last 5 participants in the data set. Notice that when
58 EXPLORATORY AND CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 61/185
C O
_0)
.CO
C O
Q
Q)
C O
C O
>
£
o
C O
Q)
C O
lo
^
£ 2
o
Q.
£
,_
o
•
c _ )
^ * j
UJ CO
DO
5
1— £
C D
c S
X
C O
rrt
,0
C O
o
o
C O
£ 2o
0
C O
LL
0)
O
0)
CD
CO
£•
O
C O
II
c
'
- ^
1
CL
LU
T T ^
^
O
jE
r:
li
t
C
Q
D C
O
CO
C D
T3
C
5?
CO
C O
c
o
'co
CO
Q)
o>
0>
Q C
CO
o
Q.
DC
O
CL
DC
O
CL
1
D C
CO
O
C L
oj
C O
(M
o
Q.
D C
m
o
i
D C
<£
C O
CO
o
CL
j
O
LU
DC
OJ
O
Q.
C D
LU
D C
T—
O
o _
d
LU
DC
Q
C O
LO
in
CO
00
in
f
in
CO
00
in
CO
o
§
in
CO
00
s
in
CO
r
00
in
i
o
in
CO
CO
m
l
o
h- oo
CD t-
OJ CO
OO
CO
•*
CD
|
C- CO
CD i-
CM CO
C» CO
•^- ^—
'
•* CO
I
s
- OO
CD i-
1
CM CO
CO CO
1
52
i
CO
•t
0O
CO
CM
CO
O)
li
CO
CM
CO
O5
^
-.
l
CO
CM
CO
0>
£
1
§
CO
^
03
in
l
CO
in
in
CO
••-
00
in
T—
1
CO
m
in
CO
*-
00
in
CO
m
. m
.
O
00 O>
CO C*-
T — 1
^c^
O
CO
i-
O>
00 O)
CO f~
T 1
s ^ -
1— ^
1
O
CO
- CT>
00
O)
CO ("-
^ - 1
1
•*
CM
O
CO
T O
CD r~ -
CT> O)
5 >
CO
CM
00
5
CD
CM
i—
^
O
l
CD
OJ
1
r*-
00
l
CD CO
m oo
T
|
in h-
• < f r m
CM
i-
i i
r-- co
CO S-
T
_ _
^
l
CO CO
in oo
T
|
in r~-
^T
i
h- 00
C O
h-
••- -r-
CO CO
m oo
r- \
m r —
-I- in
CM t-
l i
K
CO
CO K
*
T
en
o
en
o
T- CM
COMPARISONS OF METHODS
59
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 62/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 63/185
princ ipal axes factor analyses method s
are
used,
all
three factor score methods
yield
different
fac tor scores
for a
given partic ipan t (e.g. ,
fo r
participan t 102,
RG_PAF1 = -.76; BR_ PAF1 = -.86; ARJPAF1 =
-.83).
Table 5.3 presents th e Pearson correlation coefficients among all pair-
wise
combinat ions of the 18 factor scores for the 100 participants in this
analysis.
Recal l
that all six
analyses invoked rotation
to the
varimax cri ter ion,
which
means that the three
factors
in all six
analyses were
perfectly
uncorre lated .
As the
correlation
coefficients in Table 5.3
indicate,
the
correlations
of th e
factor scores from
th e
principal components analyses exactly matched
th e
correlations (here
all
zero)
of the
factors themselves.
This
always (and
only always) happens
in
prin cipa l components analysis.
The
exact matching
of
factor score correlations with factor correlations is
another
appealing
feature
of principal components analyses.
STRUCTU RE COEFFICIENTS IN A COMPONENTS CONTEXT
As
noted
in chapter 2, correlations between measured variables and
composite variables
are
always called structure coefficients (r ) throughou t
the general l inear model (GLM; Courville Thompson, 2001). Table 5.4
presents
th e
Pearson
correlation
coefficients
for the 11
measured variables
with
the
three factor scores
from the
principal components analysis.
These
were
obtained
wi th
the
SPSS
syntax:
correlations variables=perl
to
peril with
reg_pcl to reg_pc3 .
Note that th e Table 5.4 structure coefficients
literally
computed as
bivar ia te correlation
coefficients
exact ly match
the
pat tern/s t ructure coeffi-
cients produced by the SPSS principal components analysis of the same 11
measured variables.
These
varima x-rota ted pat tern/s t ructure coefficients are
presented in Table 5.5.
COMMUNALITY
COEFFICIENTS AS
R
1
VALUES
In chapter 2 it was noted that communality coefficients indicate th e
proportion of variance of a measured variable that the factors as a set can
reproduce. Conversely, h
can be
viewed
as
characterizing
how
much
of the
variance of a measured variable was
useful
in delineating the extracted
factors.
COMPARISONS OF
METHODS 61
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 64/185
TABLE
5 .3
Factor
Score C or r e la tions
n = 100
Faculty)
Principal components
Variable
REG.PC1
REG-PC2
REG-PC3
BAR_PC1
BAR_PC2
B A R _ PC3
AR-PC1
AR_PC2
AR-PC3
RG-PAF1
RG-PAF2
RG-PAF3
BR-PAF1
BR_PAF2
BR_PAF3
AR_PAF1
AR_PAF2
AR-PAF3
REG_PC1
1 0000
.0000
.0000
1
0000
.0000
.0000
1 0000
.0000
.0000
.9903
.0054
.0485
.9904
.0061
-.0476
.9917
.0044
-.0023
Regression
REG_PC2
1 .0000
.0000
.0000
1 .0000
.0000
.0000
1 0000
.0000
-.0017
.9779
.1007
-.0064
.9748
-.0116
-.0050
.9793
.0383
REG-PC3
1 .0000
. o o o o l
.ooool
1 . O O O O l
. o o o o l
. o o o o l
1. OOOOl
.0596
.0543
.95711
-.03141
-.0577]
.95011
.0208)
.00351
.95351
BAR_PC1
1.0000
.0000
.0000
1 .0000
.0000
.0000
.9903
.0054
.0485
.9904
.0061
-.0476
.9917
.0044
-.0023
Bartlett
BAR-.PC2
1 .0000
.0000
.0000
1 .0000
.0000
-.0017
.9779
.1007
-.0064
.9748
-.0116
-.0050
.9793
.0383
Anderson-Rubin
BAR_PC3
1
.0000
. o o o o l
. o o o o l
1 .OOOOl
.0596
.0543
.95711
-.03141
-.05771
.95011
.02081
.00351
.95351
A R _ P C 1
1 .0000
.0000
.0000
.9903
.0054
.0485
.9904
.0061
-.0476
.9917
.0044
-.0023
AR_PC2
1
.0000
.0000
-.0017
.9779
.1007
-.0064
.9748
-.0116
-.0050
.9793
.0383
AR_PC3
1.0000
.0596
.0543
.95711
-.03141
-.05771
.95011
.02081
.00351
.95351
One
heur is t ic
to
mak e
these
concepts concrete is to
conduct
a
m u l t i p le
regression analysis in which the measured variables in tu rn are predicted
by the factor scores
from
the pr inc ipa l components analysis.
These
regression
analyses can be executed using the SPSS syntax:
regression variables=reg_pcl to reg_pc3
perl to perll/dependent=perl /
enter reg_pcl
to
reg_pc3
.
regression variables=reg_pcl to reg_pc3
perl
to
perll/dependent=per2
/
enter reg_pcl
to
reg_pc3
.
regression variables=reg_pcl
to
reg_pc3
perl
to
perll/dependent=perll
/
enter reg_pcl
to
reg_pc3
.
6
EXPLORATORY
AND C ONFIRM ATOR Y FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 65/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 66/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 67/185
Dependent Var iab le.
.
PER1
1
Wi ll in gne ss
to
help users
M ethod: Ente r REG _PC1 REG _PC2 REG_PC3
M ultiple
R
.93803
R Square .87990
Variables
in the
Equation
Variable B SE B Beta T Sig T
REG-PC1 .277091 .048319 .202837 5.735 .0000
R EG_ P C 2
1.224812 .048319 .896592 25.349 .0000
REG _PC 3 .255 119 .048319 .186753 5.280 .0000
(Con stan t) 7.650000 .048077 159.121 .0000
E n d
Block Number
1 Al l
r eques ted var iab les en t e r ed .
De pen de n t Variable. . PER2 1 G iving use rs in dividual att
M ethod: Ente r REG _PC1 REG_PC2 REG_PC3
M ultiple
R
.90666
R
Square .82204
Variables
in the
Equation
Variable
R E G PC 1
RE G
PC 2
REG_PC3
(Constant)
B
.167676
1 .353442
.333508
7.310000
SE B
.066672
.066672
.066672
.066338
Beta
.108283
.874029
.215374
T
2.515
20.300
5.002
110.194
S i g T
.0136
.0000
.0000
.0000
Figure 5.1.
M ultip le re gres s ion predictin g
pleasured
variables using
the
three factor
s co r es
n =
100 faculty).
Factor scores can also be used fo r heuristic purposes, to better grasp
the nature of both structure coefficients and communal i ty
coefficients.
Throughout
th e G L M , a structure coefficient, score-world statistic, can
be computed (l i teral ly) as the bivariate
product-moment
correlation of a
measured variable (e.g. , th e scores of n people on a factored variable) with
a
latent variable (e.g. ,
th e
principal components factor scores
of n
people).
And a
com muna lity coefficient,
an
area-world statistic,
in the
orthogonal
principal components case
can be
computed
as the R
2
between
th e
factor
scores
of n
people with
th e
scores
of n
people
on the
measured variables.
COMPARISONS OF METHODS 65
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 68/185
6
OBLIQUE ROTATIONS
AND
HIGHER-ORDER
FACTORS
In chapter
3 the
notion
of
factor rotation
was
introduced.
It was
noted
t h a t
ro tat ion is a lways poss ib le in exploratory factor analys is (EFA) whe never
there
is more
than
one factor , and
that rotat ion
is almost always necessary
in these cases
to
obtain s imple s t ructure .
O ne
class
of
methods involves
or thogonal
rotation
and
leaves
the
factors s til l perfectly uncorrelated.
A
second class o f me thods involves obl ique ro tat ion and allows t he unro ta ted
factors ( that a re
ini t ia l ly a lways uncorrelated)
to
become correlated. Ob lique
rotations present some interpretation challenges
tha t
mer i t fu r ther
discussion.
Table 5.5 (see chap. 5) presented the v ar im ax- ro ta ted p r inc ipa l compo-
nents
f rom
an
analysis
o f the first 11
measured var iab les presented
in Appen -
dix
A for
data provided
by the 100
facu l ty members .
This
solution
had
reasonably s imple s t ructure .
A few
measured var iab les were
no t
q u i t e
as
clear
in
regard
to
what factors they reflected.
For
example , PERU
had a
pa ttern /stru ctu re c oeffic ient on F actor III of .72, but on Factor II had a
pat tern/s tructure coeff ic ient
of
.34.
How ever , i t is im por t ant to reme mb er that a l l cor relat ion coeff ic ients ,
including structure coefficients , must be squared before their magnitudes
can be
compared. Thus, because
.72 is
51.8%,
and .34 is
11.6%,
the
larger
of these tw o s tructure coeff ic ients w as actual ly 4 .5 times larger than the
smaller value.
67
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 69/185
TABLE 6.1
Promax-Rotated Components for 11 Measured Variables n = 100 Faculty)
Pattern Structure
Variable
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
PER9
PER10
PER11
I
.04
-.07
.10
-.01
.95
.93
.94
.81
.20
-.03
-.10
II
.94
.92
.75
.90
-.06
-.05
.00
.20
-.12
.01
.17
III
-.03
.03
-.03
.05
.04
.04
.00
-.07
.70
.85
.75
I
.40
.31
.39
.37
.94
.93
.95
.85
.48
.38
.33
II
.94
.91
.78
.92
.34
.34
.38
.49
.31
.43
.51
III
.46
.46
.40
.50
.46
.45
.45
.41
.73
.84
.79
If
88.0%
82.2%
61 .5
84.7%
88.8%
86.9%
89.7%
75.5%
57.0%
71 .2%
64.7%
Note.
Pattern coefficients greater
than
|.4| are italicized.
Although the v a r ima x - r o t a t e d solution had s imp le s t r uc tu r e , for he u -
r is t ic purposes the solution was also rotated to the p r o ma x
criterion.
Because
S
V x
p equa ls
P
V X
H ( R F
X
H ) .
a n
d
R p x F
n o w
wi l l
no t equa l
I P
X
F .
the s t r u c t u r e
and the
pattern coef f ic ient matr ices wi l l
no longer be
equa l . Indeed , hecause
the
pa t te rn coef f ic ients
are no
longer cor re la t ion coef f ic ients , some, m a n y ,
or all of the pa t te rn coef f ic ients may be less than -1 or grea te r
than
+1.
Table 6.1 presents the pattern and the s t r uc tu r e ma t r i c e s f rom this
ana lys is .
Table
6.2
presents
the
fac tor
correlation
ma t r ix f r om
the
ana lys is .
In
chapter
2 the
notion
of the
c o mmu na l i t y c o e f f i c i e n t
was first
intro-
duced. In ac tua l i ty , the examp le invo lved a v a r ima x - r o t a t e d p r i n c i p a l c o m-
ponents so lu t ion. In an
orthogonal
ro ta t ion the
h
2
can be c o mp u t e d by
s u m m i n g the squared s t ruc ture coef f ic ients for a var ia b le ac ross the factors.
This
y ie lds
the total
var iance
of the
measu r ed va r i ab le common
to the
fac tors
as a set. The computation is
s implif ied
because the fac tor s are unco r r e -
la ted, and so
there
is zero over lap be tw een the common var iances of a g iven
measu r ed va r i ab le on any two fac tor s .
B ut
when factors are cor re la ted , the h is der ived by first computing
the
p roduc t s
of a
g iven measured var iab le ' s
pattern
coef f ic ient
on
each
TABLE 6.2
Factor Correlation Matrix
(R
3x 3
) for the
Table
6.1
Factor Pattern Matrix
Factor I II III
1.000
.398 1.000
III
.473 .502 1.000
68 E X P L O R A T O R Y AND CO N F I R M A TO R Y F A CTO R ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 70/185
factor with the corresponding s truc ture coefficient on each factor , and then
s u mming
these
products .
For
example ,
for the
measu red var iah le PERI
in
the
Table
6.1 solution, h for this measured var iahle is computed as :
[0.040
x
0.399]
+
[0.936
x
0.937]
+
[-0.029
x
0.460]
=
0.016 + 0.877 + -0.013 = 87.96%.
It
should he
noted that
th e algorithm for a given measured var iab le ,
ZP
K
S« across the k factors , is a more general
formula
fo r h
2
. When factors
are uncorrelated, £P
K
SK across the k factors can be reexpressed as ZP
K
2
,
because
in
this case
the
pat tern
coefficients are
also s tructure
coeff icients .
In
o ther
words , fo r pr incipal components and pr incipal axes analyses th e
algorithm ZP^Sx across the fe factors
always
works , whether or not the
factors
are
correlated.
M A X I M I Z I NG S A M P LE
F IT VERSUS
RE PLI CA BI LITY
In th e present comparison the Tab le 5 .5 var im ax-ro tate d so lu t ion and
th e Table 6 .1
promax-ro tated so lu t ion would general ly lead
to the
same
interpretat ions regarding
the
m a k e u p
of the three
unde rlying constructs .
Where the so lu t ions d i f f e r is in the representat ions of the correlat ions among
the
latent constructs .
The first componen t in
Table
5.5
wou ld
be
interpreted
as
measur ing
th e
Library
as
Place.
This component w as
saturated w ith percept ions
of
l ibraries inc lud ing r espec t ive ly
the
i tems
7 . A
co n tem p la t i v e env i ro nmen t ,
5. A
haven
fo r
qu ie t
and
solitude,
6. A
meditat ive p lace,
and 8.
Space
tha t f ac i l i t a tes qu ie t s tudy .
The
second
com ponent: might
be
labeled Service Affe ct.
This compo-
n en t w as
saturated respect ive ly with
the
i tems
1.
W ill ingness
to help
users ,
2 . Giving users indiv idual a t ten t ion , 4 . Employees who are consistently
cour teous ,
and 3.
Employees
w ho
deal with users
in a
caring fashion.
T he third com ponent might be labeled Information
Access. This
component
w as
saturated respect ively with
the
item s 10.
Comple t e
runs
of
jour nal t i t les , 11. Interdiscip l inary l ibrary needs being addressed,
and
9.
C omprehens ive pr in t collections.
But
fo r
these data
th e
p rom ax- ro ta ted r esu l t s r epo rted
in
Table
6.1
would lead
to the
same factor labels
as those
produced
in the
var imax-
rotated solution. T he pat tern coeff ic ients in Table 6.1, like the
pat tern /
s tructure coeff ic ients presented
in
Table 5.5,
had
s imple s t ructure .
T he s tructure coeff ic ients in Table 6.1 ref lect a dynamic invo lv ing th e
components
being
fairly
highly correlated,
as
repor ted
in
Table 6.2. Most
of the s tructure c oeff ic ients in
Table
6 .1 are qui te notew or thy for a ll var iab le s
O BL I Q U E
ROTATIONS AND H I G H ER - O R D ER FACTORS 69
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 71/185
on a l l components . This is mere ly indica t ive of the fac tors having corre la-
t ions of .398, .473, and .502.
Of course , the com ponen ts a re sufficiently discre te tha t i t
st i l l
seems
reasonahle to recogn ize the ex i s t ence of three la ten t va r iab le s , because even
the two
mos t h igh ly cor re la ted com ponen ts on ly have 25 .2% com mo n
var iance
( . 502
2
=
. 2 5 2 ) .
But for a
case such
as
th is ,
how w ou l d one
se lec t
a
f a c t o r an a ly t i c so l u t i on ?
Three
re la ted
issues
seem
pa r t i cu la r ly
re levan t .
Analytic Purpose
In
som e fac to r ana ly t i c app l i ca t ions
the
researcher
is p r ima r i ly
in ter-
es ted in in terpre t ing the fac tors . In other appl ica t ions , fac tor scores are
genera ted fo r use in subs equ en t s ta t i s t i c a l analyses . As the present exam ple
suggests, when fac to r in te rp re ta t ion i s the focus, rota t ion choice may be
less
c r i t i c a l a s long as a g iven a na lys i s y ie lds a r e a son ab le s imp l e s t r u c t u r e .
B ut
w hen com pos i t e scores
are
going
to be
crea ted
for use in
sub sequen t
analyses ,
and w hen the ob l iqu e fac to rs a re h igh ly cor re la ted , some pre fe rence
migh t
be
given
to the
ob l ique so lu tion ,
so
th a t
the
cor re la t ions
of the
la tent
var iables wil l b e honored in
these subsequen t su bs tan t ive ana lyses . Ho wever ,
there is one o ther a ppro ach to es t im at ing fac tor scores , som et ime s ca l led
unit
weighting, tha t may finesse these concerns (see G orsu ch, 1983, pp.
268—276) .
As noted in chapter 4, eve ry sample con ta ins some id iosync ra t i c and
nonreproduc ib le va r iance
as a
func t ion
of
samp ling error.
The
more pa r ame -
ters w e e s t ima t e , the more po ten t i a l the re is for our re su l t s to cap i t a l i ze
excess ively on
u n i q u e
fea tures of the
s amp l e , w i th
the
consequence
that
resul ts wil l
genera l ize less
w e l l to new
samp le s .
One way to capi ta l ize
less
on id iosync ra t i c
fea tures
of the sam ple is to com pute fac tor score es t im ates
w i t h o u t
us ing
the fac tor
score we igh t ma t r ix ( W v
x
p ) -
One approach
s imply
ident i f ies those m easured va r iab le s tha t
def in i -
t ively
mea su re
a
single fac tor .
Then
composi te scores
on
each fac tor
are
com puted m ere ly by su m m ing the scores on the re leva n t measured va r iab le s ,
by av era ging these scores , by av eraging the
z
scores of only the re levant
var iab les ,
or by do ing wha teve r e l se the re sea rche r deems though t fu l and
reasonab le . For ex am ple , ac ross both the va r imax - ro t a t ed and the pro 'max-
rota ted resul ts
for the
p re s en t e xamp le ,
the
s ame m ea su red va r i ab le s wo u ld
be used
to
ob ta in un i t -we igh ted compos i t e scores
on the
three construc ts .
Parsimony
Centur ie s ago a ph i losopher named Wi l l i am of Occam a r t i c u l a t ed the
concept w e recognize today as the law of pars im ony. Thus, th is precep t i s
somet imes ca l led
Occam's
razor. Bas ically,
the law of
pars im ony says that
w h e n
tw o
exp l an a t i on s
fi t a set of
facts rough ly
equa l l y
we l l ,
a ll
things
70 EXPLORATORY
AND
CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 72/185
equal, the
simpler explanation
is
more
l i ke l y to be
true.
We
tend
to p re f e r
parsimonious explanations because things that are true are also more l i ke l y
to replicate in
fu tu r e
research, and researchers
u s u a l l y
do not want to be
embarrassed by
making discoveries
that no
one else
can
replicate.
In
an EFA context, orthogonal solutions are more parsimonious
than
their oblique counterparts because f e w e r parameters are estimated when an
orthogonal rotation
is
implemented. This
is
because when
R
F x
F
equals
Ip
x F
,
then PV
x
F equals Sy
x
F-
In any s t a t i s t i c a l ana ly s i s ,
including EFA, what
we are
trying
to do is
fit a
model
to our
data.
The
more parameters
we
estimate,
the
greater
is
the likelihood that our model
w i l l
fit the sample data. For example, if n is
2, for any two measured
v a r i a b l e s ,
r
2
is
a l w a y s
+1. For a regression ana ly s i s ,
when
the
number
of
predictor variables equals
n — 1, the
degrees
of
freedom
error
is
zero,
and the
resulting
R is
always
+1,
regardless
of
what
the
predictor
variables
are.
The
same situation applies across
d i f f e r e n t
analyses.
For
example,
in
predictive
discriminant analysis (PDA), when
less
parsimonious equations
called quadratic versus linear rules
are
used,
the
prediction
hit
rate tends
to be higher in the original sample, but the prediction eff icacy of the same
prediction equation across other samples tends to be lower than in the
original sample. Thus, Huberty (1994, p. 64) noted a general preference for
the more parsimonious PDA equations, even though the fit of the model
to
sample data tends
to be
better
for the
less parsimonious solution. Elsewhere
he noted
that
the linear rule is suggested because of presumed greater
stability
[across samples]
than
that
for the quadrat ic rule, in general (p. 260).
Interpretation Difficulty
It must also be
sa id that
factor interpretation
i tself
is more complex
when
factors are correlated.
This
is because more parameter estimates must
he
simultaneously interpreted. Specifically, throughout
the
general linear
model (GLM), whenever weights
are not
structure coefficients,
both
stan-
da rd iz ed weights and structure
coe f f i c i en t s
must be examined to arrive at
correct interpretations.
This
is true with reference to analyses including mul tiple regression
(cf. Courville &
Thompson,
2001;
Thompson
& Borrello, 1985), canonical
correlation analysis (Meredith, 1964;
Thompson,
1984), and confirmatory
factor
analysis (cf. Graham
et
al.,
2003; Thompson,
1997).
This is
also true
in
EFA.
As
Gorsuch (1983) emphasized, proper
interpretation of a set of
factors
can probably only occur if at least S and P are both examined (p . 208).
A variable does not contribute to a factor only if the variable's pattern
and
structure
coe f f i c i en t s are both
zero.
If a
structure coefficient
is
near zero,
but the pattern coefficient is not, the measured variable contributes to the
OBLIQUE ROTATIONS
AND
HIGHER-ORDER
FACTORS
71
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 73/185
solut ion indi rec t ly
by m odify ing the
re la t ionship
of
other measured var iables
w ith the
factor. Across
the GL M
this
is
called
a
suppressor
e f f e t
(see Co urv il le
& Thompson, 2001; Horst, 1966).
For the Table 6.1 resul t s , the analysis does not present overwhelming
difficulties,
because the s t ruc tu re is so clear. The three factors
( Library
as
Place, Service Affect , and Inform ation Access ) w ere highly correlated
but
sti l l
sufficiently
discre te
to be
recognized
a s the
constructs described
by
the pattern coefficients. But the structure coefficients in this case clear ly
indica te tha t all 11 variable s are correlated w ith al l three c onstructs. Such
findings
suggest
the
presence
of
overarching superordinate const ruc ts ,
or
higher-order factors.
H I G H E R - O R D E R
FACTORS
When
variables are correlated (i .e . , R y
x
v ^ l v \ v ) >
then
factors can
be
computed
as
weighted aggregates
of the
measured va r i ab le s .
By the
same
token, when factors are correlated (i.e., R F
X
F IF
X
F ) >
then factors can be
extracted
from
this matr ix
as
wel l .
The factors extracted from interv ariab le correlat ions (o r other stat ist ics
measu r ing associat ions)
are
called
first-orde r
factors.
T he
factors
then ex-
t rac ted
from
the inter fac tor correlat ion s among the first-order factors are
called
second-order
factors.
If the
second-order factors
are
correlated,
then
third-order
factors
can be
ext rac ted. However ,
as
Kerl inger (1984) noted,
whi le ordinary fac tor
analysis is
probably
wel l
unders tood, second-order
factor
analysis,
a
vi ta l ly impor t an t pa r t
of the
analysis, seems
not to be
widely k n o w n
and
understood
(p .
x i v v ) .
It is argued here that these higher-order factors
should
be ext rac ted
whenever fac tors are corre la ted. As Gorsuch (1983) em phasized:
Rotating
ob l i que ly
in
factor analys is
impl ies
that
the
factors
do overlap
and that there are,
there fo re ,
broader areas of generalizabil i ty than j u s t
a
p r i m a r y fac to r . I m p l i c it
in all
oblique
rotations
are
h ig he r - o r de r factors .
It
is recommended that these [always] be
extracted
and examined so
that
the
inves t iga tor
may
ga in
the
ful les t poss ible
understanding of the
data.
(
P
. 2 5 5 )
Too few
researchers report ing correlated
first-order
factors conduct these
needed higher-order analyses.
The different levels of factor analysis give
different
perspec t ives on
the
da ta . Thompson (1990b)
offered th e
analogy
o f
hiking
in the
m oun t a in s
versus flying over them. When you are hiking you can see rocks and trees
and fish in
streams. When
you fly
over
the
mountains,
you can see the
pat terns made
by the
moun ta ins
as a
range.
The
perspect ives complement
72 EX P LO R A TO R Y A ND CONFIRMATORY
FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 74/185
TABLE 6.3
Second-Order
Factor Pattern/Structure Coefficient Matrix
First-order factor
A
If
FACTJ
.773 59.8
FACTJI
.792 62.7
FACTJII
.832 69.2
each other, and each is needed to see patterns at a given level of specif ic i ty
ver sus
generality.
The literature includes various examples of higher-order factor analyses
(cf . Borrello & Thompson, 1990; Cook, Heath, & Thompson, 2001; Cook
&
Thompson,
2000; Thompson & Borrello, 1986). Here
both
the implemen-
tation
of a
higher-oider analysis using SPSS
and the
interpretation
of
such
an analysis are illustrated.
Example 1
For the Table 6 .2 interfactor correlation matrix (R ^
x
5) the second-
order factors can be eas i ly computed in
SPSS
by analyzing the correlations
among
the first-order
principal components scores.
The
syntax
is:
subtitle '5
n=100
Faculty
AAA
First-order
A A A A A A
.
execute .
temporary
.
select
if (ranktype eq 3) .
factor
variables=perl
to per 2/
analysis=perl to erll/criteria=factors(3)/
extraction=pc/rotation=promax/
print=all/save=reg(all
reg_pc}/ .
subtitle
6 Second-order $$$$$$$$$$$$$$$$$$$$$$$$' .
execute
.
factor
variables=reg_pcl to reg_pc3/
print=all .
This works only fo r principal components analysis, because only in
principal
components analysis
wi l l
the factor correlations and the factor
score correlations match exactly.
The
second-order factor derived
in
this
manner is presented in Table
6.3.
Higher-Order Interpretation Strategies
The interpretation of higher-order factors poses some special dif f icul -
ties. Factors
are
abstractions
of
measured variables. Second-order factors,
OBLIQUE
ROTATIONS AND
HIGHER-ORDER
FACTORS 73
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 75/185
then,
are abstractions of abstractions even more removed f rom the measured
va r iab le s .
Somehow
we
would like
to
interpret
the
second-order factors
in
terms of the measured variables, rather than as a manifestation of the factors
of
the
measured variables.
As Gorsuch (1983) suggested,
T o avoid basing interpretat ions upon interpretat ions of interpretat ions,
the relationships of the original variables to each level of the higher-
order factors are
de te rmined .
. . . Interpret ing from th e variables
should
improve
th e
theore t i ca l unders tanding
of the data and
produce
a
better
ident i f icat ion
of
each
higher-order
factor, (pp. 245-246)
There are three strategies for doing so.
Firs t ,
Gorsuch (1983) suggested
that the
second-order factors directly
ref lec t ing the measured variables could be evaluated by computing the
product matrix of the first-order
pattern
matrix times the second-order
pattern matrix ( i .e . , here
PH
x 5
P}
X
] = PH
x
i) . Second,
Thompson (1990b)
reasoned
that
an orthogonally rotated product matrix might be interpreted.
He
reasoned
that if
rotation
is
sensible
in
both
the first and the
second
order,
then
rotation of the product matrix to achieve simple structure might
also be
usefu l .
Of course, in the present example involving only one second-order
fac tor ,
rotation
is not possible. Rotation is only possible when there are two
or more factors.
Third, Schmid and Leimaii (1957) proposed an elegant method for
expressing
both the first-order and the
second-order factors
in
terms
of the
me a su r e d variables, but also res idua l iz ing (removing) all variance in the
first-order factors
that
is
also present
in the
second-order factors.
This
allows
the researcher to determine what, if any, variance is unique to a given level
o f analysis or perspective. Schmid and Leiman (1957) suggested that their
Schmid-Leiman solution not only preserves
the
desired interpretation char-
acteristics
of the
oblique solution,
but also
discloses
the
hierarchical structur-
in g of the variables (p. 53).
This analysis
can be
implemented using SPSS matrix commands.
The
input into the analysis includes (a) the obliquely rotated first-order pattern
matrix (here
P j i x ? )
an
d (b) the second-order pattern matrix (here
P ?
x
i )
augmented
by a
symmetrical (here 3 x 3 ) matrix containing zeroes off-
diagonal
and the
square roots
of
uniquenesses
on the
diagonal.
This augmented matrix, A, is of rank / by s + /, where / is the number
of f irst-order factors and s is the number of second-order factors. In the
current example, this augmented matrix
is
A3
x4
.
R e c a l l
that
a communality coefficient
(h
) indicates in an area-world
squared metric the percentage of variance in a measured variable
that
is
74
EXPLORATORY
A ND
CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 76/185
TABLE
6.4
Augmented Matrix A
3 x 4
SECOND
0.773
0.792
0.832
U|
0.634
0.000
0.000
SORT of uniquenesses
Un
0.000
0.611
0.000
U
0.000
0.000
0.554
Note.
SORT (0.402) = 0.634.
containe d in the fa ctors as a set. The uniqueness ( i t ) i s an area-w orld
s tat i s t ic
ind ica t ing the percen tage o f var i ance in a measu red var i ab le
n o t
con ta ined in the fac tor s as a set (e.g., 100% - h
2
= u
2
, or 100% - 59.8% =
4 0 . 2 % ) .
Table
6.4 presents the m a t r i x , A^^ for this analysis .
The SPSS
sy n t ax
file for the
p resen t example
is:
MATRIX .
COMPUTE
PI =
{0.040, 0.936,
-0.029
0.917,
0 031
0.751, -0.026
0.900, 0.050
0.037
0.037
0 . 944
0.805
0.198
-0.070,
0.103,
-0.014,
0.947,
-0.061,
0.933, -0.049,
0 . 003, 0.003
0.202,
-0.070
-0.121,
0.700
;
-0.026,
0.010,
0 851
;
-0.096, 0.165,
0.753}
.
COMPUTE
P2 =
{0.773 ;
0.792
;
832}
COMPUTE
Pvh = PI * P2 .
COMPUTE
A =
{0.773, 0.634, 0.000, 0.000 ;
0.792, 0.000, 0.611, 0.000 ;
0.832, 0.000, 0.000, 0.554} .
COMPUTE SL = PI * A .
PI /
FORMAT='F8.3' /
TITLED First-order Promax-rotated Pattern
Matrix
/
O B L I Q U E ROTATIONS
AND
H I G H E R - O R D E R
FACTORS 75
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 77/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 78/185
TABLE
6.5
Pattern Product Matrix (Pn
x 1
)
Variable SECOND
PER1 .748
PER2 .698
PER3 .653
PER4 .744
PER5 .715
PER6
.713
PER7 .735
PER8
.724
PER9 .640
PER10 .696
PER11 .683
TABLE 6.6
Schmid-Leiman Solution for Example 1
First-order
Variable
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
PER9
PER10
PER11
SECOND
.748
.698
.653
.744
.715
.713
.735
.724
.640
.696
.683
FACTJ
.025
-.044
.065
-.009
.600
.592
.598
.510
.126
-.016
-.061
FACTJ I
.572
.560
.459
.550
-.037
-.030
.002
.123
-.074
.006
.101
FACTJ
II
-.016
.017
-.014
.028
.020
.020
.002
-.039
.388
.471
.417
x mple 2
The second
example invo lves perceptions
of
l i b ra ry s e rv i ce qua l i t y
provided by
61,316
u n i v e r s i t y
students
and facul ty us ing the 25
LibQUAL+ i tems presen ted in
Table
6.7. Four first-order fac tors were
extracted. For heur is t ic purposes on ly , in this analysis two
second-order
fac tors
were extracted. The
Schmid-Leiman
so lu t ion is reported in Table 6.8.
The richness of higher-order fac tor analy s is can be communicated by
i n t e r p r e t i n g these results .
The two
second-order fac tors measure Lib rary
as
Place ( SEC_B )
and all
other perceptions
as
SEC_A.
H o w e v e r , the
nine
Service Af fec t and the six
Personal
Control i t ems pa r t i cu l a r ly
dominate the first second-order
fac tor .
O B L I Q U E
ROTATIONS
AND H I G H E R - O R D E R
FACTORS
77
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 79/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 80/185
TABLE 6.8
Schmid-Leiman
Solution
fo r Example 2
(n
61,316 Students
and
Faculty)
Second-order
First-order
Variable
PR1
PR4
PR11
PR14
PR15
PR17
PR18
PR20
PR24
PR2
PR10
PR13
PR21
PR23
PR 3
PR8
PR9
PR19
PR22
PR5
PR6
PR7
PR12
PR16
PR25
Sum of
squares
SEC_A
.655
.624
.690
.635
.660
.704
.713
.610
.689
.200
.206
.262
.328
.260
.526
.566
.559
.416
.512
.628
.685
.708
.736
.761
.625
8.531
SECJ3
.153
.201
.260
.265
.256
.234
.216
.294
.310
.754
.820
.771
.705
.824
.281
.169
.274
.324
.406
.106
.239
.129
.202
.231
.344
4.312
FACTJ
.422
.450
.291
.403
.469
.366
.430
.382
.378
-.025
-.028
.028
.035
.005
-.060
.048
.033
.037
-.052
-.041
.000
-.039
.061
.103
.079
1.494
FACTJ
I
-.011
-.001
.003
.009
.005
-.006
-.008
.006
.004
.124
.132
.117
.095
.120
-.009
-.018
-.005
.014
.009
-.002
.012
-.001
.003
.003
.010
0.071
FACTJII
.024
-.012
.072
.005
-.022
.020
.004
-.057
-.029
.045
.025
.001
-.023
-.043
.010
.072
.027
-.021
-.044
.364
.306
.399
.306
.260
.068
0.576
FACTJV
-.045
-.062
.041
-.038
-.053
.047
.007
.054
.077
-.064
-.042
-.014
.086
.067
.419
.254
.308
.247
.449
-.019
.033
-.016
.012
.044
.228
0.694
set printback=listing .
DATA LIST
records=3 /
ROWTYPE_ (A8) FACTOR— (F2.0)
Prl Pr4 Prll Prl4 Prl5 Prl7 Prl8 Pr20 Pr24
Pr2 PrlO Prl3 Pr21 Pr23
Pr3 Pr8 Pr9
Prl9 Pr22
Pr5 Pr6 Pr7 Prl2 Prl6 Pr25
(1X,9F5.3/11F5.3/5F5.3).
BEGIN
DATA
FACTOR 1 .655 .624 .690 .635 .660 .704 .713 .610 .689
.200 .206 .262 .328 .260 .526 .566 .559 .416 .512 .628
.685 .708 .736 .761 .625
OBLIQUE ROTATIONS
AND HIGHER-ORDER FACTORS
79
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 81/185
TABLE 6.9
Equamax-Rotated Product Matrix
n =
61,316 Students and Faculty)
Variable
PR1
PR4
PR11
PR14
PR15
PR17
PR18
PR20
PR24
PR2
PR10
PR13
PR21
PR23
PR3
PR8
PR9
PR19
PR22
PR5
PR6
PR7
PR12
PR16
PR25
A
.653
.622
.688
.632
.658
.702
.711
.607
.686
.193
.198
.255
.321
.252
.523
.564
.556
.413
.508
.627
.683
.707
.734
.759
.622
Second-order factor
B
.159
.207
.266
.271
.262
.241
.223
.300
.316
.756
.822
.773
.708
.826
.286
.174
.279
.328
.411
.112
.245
.136
.209
.238
.350
FACTOR 2
.153 .201 .260 .265 .256 .234 .216 .294 .310
.754 .820 .771 .705 .824 .281 .169 .274 .324 .406 .106
.239
.129 .202 .231 .344
END DATA.
FACTOR
MATRIX=IN(FAC=*)
/
ANALYSIS=prl to pr25 /
PRINT=ROTATION /
CRITERIA=ITERATE(75) /
ROTATION=equamax
.
T he
equam ax- ro ta t ed pa t t e rn / s t ruc tu re coe fficients
for the
analysis
are
reported in Table 6.9. In the present example fac tor in terpre ta t ions would
be l i t t l e a l te red by this rotat ion.
O f course , th e produc t m a t r ix would neve r be obliquely rotated. That
would
imply the exis tence of higher-order fac tors . In conduc t ing ob l iqu e
80 E X P L O R A T O R Y AND C O N F I R M A T O R Y FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 82/185
rotations, always extract
the
progressively higher-order factors imp l ied
hy
nonzero inte rfacto r correla t ions, unt i l (a) there is only a single higher-orde r
factor o r (b) the factors at the highest level of analysis have s imp le stru ctu re
when ro ta ted or thogonal ly .
M A J O R
CONCEPTS
If
s imple s t ruc ture cannot be obtained with orthogonal rotation, re-
flected in
m any var iab les having p a t t e rn / s t ruc ture coef fi c ien t s tha t
are
large
in
magni tude
on two or
more factors, oblique rotation
may be
needed
to
derive interpretable factors . However, more parameters must
be
est imated
in
obl ique rotat ion. As a consequence, obl iq ue resul ts not o nly (a) tend to
fit sample data better but also (b)
tend
to yield factors that replicate less
well .
As a consequence, unless obl iqu e resul ts are substa nt ia l ly
be t te r
than
orthogonal resul ts , or thogonal resul t may be pre fe r red .
Whenever obl ique
ro ta t ion is
perform ed, higher-order factors
are im-
plied
and should be derived and interpreted . The Schmid-Leiman so lu t ion
is
part icular ly useful in such cases. T he
SPSS
syntax p resented in the chapter
invokes
SPSS
MATRIX p rocedure s and can be used as a m odel to perform
these analyses using this widely ava i l ab le software.
OBLIQUE ROTATIONS AND HIGH ER-O RDER FACTORS 81
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 83/185
7
SIX TWO-MODE TECHNIQUES
Psychologists of ten think
in
terms
of
different types
of
people, such
as wo rkaholics or Type A personalities. Yet the factor analyses discussed in
previous chapters d o
n ot
address question s regard ing types of people Instead,
all the previous discussion
implicitly
involved investig ations of the structures
under lying variables.
The
previous analyses involved
a raw
data matrix
in
which people
were represented as data rows and the variables were organized as colum n
fields wi th in the data matr ix . When da ta are analyzed in this manner ,
intervariable relationships and factors of the measured variables are the
focus of the analysis.
But
data matrices
do not
h av e
to be
organized
in
this manner.
For
example, data could be organized such that variables define the rows and
people define the columns. When da ta are organized in this ma nner, interper-
son relationships are the focus o f the analy sis, and people factors are the
result. Indeed, there
are
other data structures
that are
possible
as
well.
The
mathemat ics
of the
analysis remain
the
same across these varia-
tions.
The
computer package need
not
know what data features
are
being
correlated and analyzed
CATTELL'S "DATA BOX "
Raymond Cattell (1966a) sum ma rized six prim ary exploratory factor
analysis (EFA) choices in the form of his "data box." Cattell suggested that
83
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 84/185
data could involve (a) either one or more than one
variable,
(h) either one
or
more than
one person, and (c)
either
one or
more than
one occasion of
measurement.
These features
of the data are termed
modes.
Some data sets might involve several variables, several people,
and
several
occasions
of
measurement.
It is
theoretically possible
to do a
factor
analysis that simultaneously considers all three modes. Tucker (1966) pro-
vided some guidance on performing these analyses.
However, three-mode analyses
are
extraordinarily complex,
and
com-
monly used
software
does not support these methods. Thus, only two-mode
analyses are routinely seen in the li terature. Within each two-mode combina-
tion (e.g., variables and people, measured at only one occasion), there are
two
possible organizations
of the
data matrix (e.g., people
define the rows
and
variables
define the
columns,
or
variables
def ine th e
rows
and
people
define
the columns). Thus, there are six two-mode techniques. Each two-
mode technique
was
labeled
by
Cattell using
a
d i fferent
capital letter (e.g.,
R Q).
Six Two-Mode Techniques
R-Technique
When
people
define the
rows
of the
data matrix,
and
columns
are
used to represent different variables, the analysis examines the structure
underlying variables. This analysis is termed R-technique factor analysis. All
of
the
previous examples
in
this book were R-technique examples. This
is
the most commonly used two-mode EFA technique, and is so common
that people invoking this analysis often
do not
even label
the
analysis
R-technique.
Q-Technique
When
variables
define
the rows of the data matrix, and columns are
used to
represent
different
people,
the
analysis
identif ies
people factors
(rather than factors of variables).
This
analysis is termed Q technique factor
analysis. Q-technique is the
second
most
widely
used EFA method.
P-
Technique
When
variables
are
factored,
and
occasions constitute
the
rows
of the
data matrix,
th e
analysis
is
called
P-technique
factor analysis. P-technique
m ay involve a single-person study. However, the analysis might also involve
means, or medians, summarizing for each variable and each occasion all the
scores as a single number.
84 EXPLORATORY
AND
CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 85/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 86/185
of categories with each category co ntain ing exact ly a p rede termined num ber
of objects. The sorted o bjec ts can be anyth ing (e.g., ind ex cards conta ining
statements, photographs
of
ar t ) .
The
sorting criteria
can vary
(e.g.,
how
much participants agree or disagree with the statements, aesthetic valuing
of
the
ar t) .
Typically
the fixed single distributio n is developed to be sym me trical
and
quasi-normal
in
shape.
An
exam ple might involve sorting index cards
containing statements into categories
1
(m ost agree) through
7 most disagree)
using the
following
dis t r ibut ion:
Category 1 2 3 4 5 6 7
n
o f
cards
2 4 6 9 6 4 2
In th is example
every
column of data (i .e. ,
every
person's data) would
involve exactly two 1's, exactly
four
2's, and so forth. The mean of
every
column would be 4-00
SD =
1.58) with a coefficient of skewness of 0.00
and a
coefficient
of
kurtosis
of
— 0.50 .
Ranking
In a ranking strategy par ticipa nts are instructed to ra nk order the
objects
(e.g., index cards with statements) using some criterion (e.g.,
from
most agreem ent to most di sag reem ent). The appe al of this strategy is that
the method yields grea ter score va rian ce. Because more in form atio n is being
collected, this,
in
tu rn , theore t ica l ly
will
yield more sensi t ive
or
stable
interperson correlations,
and
thus better person factors.
The
downs ide
of
the strategy is that participants may find rank ordering 60 or 100 objects
to be
tedious
and
painful.
Mediated Ranking
Thompson (1980a) suggested
an
alternative two-stage strategy
fo r
col-
lecting ranked data that makes the ranking process more manageable for
participants. First, participants are asked to perform a conventional Q sort.
Second,
the
part icipants
are
then asked
to
rank order
the
objects with in
each
category. This means th at fewer objects must
be
simultaneously considered
by
the part icipants when performing a given ranking task.
euristi
Example
The
hypothetical data presented
in Table 7.2 can be
used
to
make
concrete the discussion of Q-technique interpretation issues. The example
presumes that five students and four faculty rank ordered 20
features
of
l ibrary
service quality
from 1
(most important)
to 20 least
important).
Table
88 EXPLORATORY AND CONFIRMATORY FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 87/185
TABLE
7.2
Q Technique
Heuristic
Data
S t u d e n t
F a c u l t y
I t e m
1 2 3 4 5 1 2 3 4 I t e m c o r e
1 1 1 1 6 1 3 1 1 1 1 2 0 1 0
2 0 S A E m p l o y e e s w h o i n s t i l l c o n f i d e n c e i n u s e r s
2 1 6 1 3 1 6 1 6 1 6 7 7 9 1 I A C o n v e n i e n t b u s i n e s s hours
3
1 3 1 1 1 1 1 3 8 1 1 1 7 I P S p a ce t h a t f a c i l i t a t e s q u i e t s tu d y
4 8 8 7 8 1 3 9 9 2 0 9 S A E m p lo y e e s d e a l u s e rs i n a c ar in g f a s h i o n
5 1 2 1 2 1 1 2 1 2 1 0 2 0 1 0 1 0 S A E m p l o y e e s u n d e r s t a n d n e e d s o f t h e i r u s e rs
6 7 7 8 1 7 8 8 1 6 8 I A
C o m p l e t e r u ns
o f
j o u r n al t i t l e s
7 1 1 1 2 9 1 1 6 1 6 8 1 7 S A G i v i n g u s e r s i n d i v i d u a l a t t e n t i on
8 1 5 1 7 1 5 1 5 9 3 3 3 1 6 S A
E m p l o y e e s
w h o a r e
c o n s i st e n t l y c o u r te o u s
9 9 9 9 7 1 5 1 7 1 3 1 7 3 I A
C o m p r e h e n s i v e p r i n t c o l l e c t i o n s
1 0 4 4 4 4 4 1 2 1 2 1 3 1 2 S A
E m p l o y e e s k n o w l e d g e
t o
an s w e r u s e r q u e s t i o n s
1 1
1 7 1 5 1 7 1 7 6 1 3 1 7 1 2 1 3 P C li b we b
s i t e e n a b l i n g
m e
l o c a t e i n f o
o n m y o w n
12 6 6 3 6 17 4 4 6 4 LP A
contemplative environment
1 3
3 3 6 3 3 6 6 4 6
L P A
p l a c e
f o r r e f l e c t i o n a n d
c r e a t i v i t y
14 14 10 14 19 14 15 15 11 14 1A Timely document delivery interlibrary loan
1 5 1 0 1 4 1 0 1 8 1 0 1 4 1 1 1 4 1 5 S A R e a d i n e s s t o r e s p o n d t o users q u e s t i o n s
1 6
2 0 2 0 18 10 1 9 1 1 1 4 1 5 1 1 P C
C o n v e n i e n t a c c e s s
t o
l i b r ar y c o l l e c t i o n s
1 7
1 9 1 9 2 0 1 4 2 0 5 5 2 5 L P A h a v e n f o r q u i e t a n d s o l i t u d e
1 8 2 5 2 2 5 2 2 5 2
L P A
c o m f o r t a b l e
a n d
i n v i t i n g l o c a t i o n
1 9
5 2 5 5 2 1 9 1 8 1 9 1 8 S A W i l l i n g n e ss t o h e l p u s e rs
2 0
1 8 1 8 19 2 0 18 1 8 1 9 18 1 9 I A
I n t e r d i s c l i b r a r y n e e d s b e i n g a d d r e s s e d
7.3 presents
th e
v arima x-rotated principal components pattern/structure
coefficients from
this
analysis.
The f i r s t question addressed by the analysis is "How many types of
people
are
there?" Here
th e
answer
was
two. Presumahly
th e
decision
to
extract two factors was guided hy various evalua tion criter ia (e.g., eigenvalues
greater than 1.0, scree test, parallel
analysis)
as it would he in any
other
EFA.
TABLE 7.3
Varimax Rotated Pattern/Structure Coefficients
Person
STUD1
STUD2
STUD3
STUD4
STUDS
FAC1
FAC2
FAC3
FAC4
Factor
I
96
94
85
8
75
03
07
.09
13
II
02
.01
12
23
.14
93
9
74
76
ote
Values greater than |.40|
are
presented
in
italics.
SI X TWO-MODE TECHNIQUES 89
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 88/185
TABLE 7.4
Salient Factor Scores
for
Factor
I
Sorted
by
Descending Levels
of Agreement
Factor score
Item
I II
SPSS variable label
19
18
13
10
7
16
20
17
-1.37
-1.35
-1.27
-1.27
-1.12
1.32
1.48
1.60
1.66
-1.51
-.90
39
96
26
1.52
-1.33
SA
Willingness to help users
LP A
comfortable
and
inviting location
LP A
place
for
reflection
and
creativity
SA Employees knowledge to answer user
questions
SA
Giving users individual attention
PC
Convenient access to library collections
IA Interdisc library needs being addressed
LP A haven for quiet and solitude
The
second
question addressed is
Which
people belong to the different
types?"
Here the five students saturate Factor I, and the four faculty saturate
Factor II. The pattern/structure matrix has simple structure. Would that
real results were always so
def in i t ive )
The third question addressed is Which variables were the basis for
delineating
the
different person
factors?" To
address this question, factor
scores are computed for each variable on both the person factors. These
factor scores are in ?-score
form,
as are all factor scores except standardized,
noncentered
factor scores.
The
variables with factor scores less
than
-1.0
or
greater
than
+1.0
are more than one standard deviation in distance from the factor score
means
of
zero.
These are the
items
of
greatest
importance or
least importance
primari ly to the persons defining the two factors. Table 7.4 presents all the
factor scores for Factor I, the student person factor, for
which
factor scores
were greater than |1.0|. Table 7.5 presents all the factor scores for Factor
I I , the
faculty
person factor, for which factor scores were greater than
|1.0|.
TABLE 7 .5
Salient Factor
Scores for
Factor
II
Sorted
by
Descending Levels
of Agreement
Factor score
Item I II SPSS variable label
3
18
17
12
20
19
26
-1.35
1.60
-.55
1.48
-1.37
-1.55
-1.51
-1.33
-1.28
1.52
1.66
LP Space that facilitates quiet study
LP A comfortable and inviting location
LP A haven for quiet and solitude
LP A contemplative environment
IA Interdisc
library needs being
addressed
SA Willingness to help users
90
EXPLORATORY
AND
CONFIRMATORY FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 89/185
In the present exam ple, lower rankings (e.g., 1, 2) indicated items
ranked most important.
And all the
noteworthy pattern/structure coefficients
in
Table
7.3 are
positive. This means that factor scores
in
Tables
7.4 and
7.5 that
are
smallest (i .e. , biggest nega tive value s) were items deemed most
important, whereas factor scores
that
are
largest were items deemed least
important.
The
factor scores
for
Factor
I
presented
in Table 7.4
suggest that
the
s tudents were most concerned that library staff are service oriented and tha t
the
library
is a comfortable place for reflection. At the other extreme,
the
students were least concerned about
the
library being quiet
and
about
library collections.
The
factor scores
fo r
Factor
I I
presented
in
Table
7.5
suggest
that th e
faculty
were most concerned about
the
library being quiet
and
least con-
cerned about
the
helpfulness
of
library
staff and the
interdisciplinary needs
of users being ad dressed.
These
hypothe t ica l
faculty
only care about w hether
the library coverage in their own d isciplines is ad equate
Variables with large factor scores and different signs across
factors
reflect
w hat v iews
part icularly differentiate
the two person
factors.
For exam-
ple,
s tudents want l ibrary
staff to be
helpful
(F[ = -1.37),
whereas this
is
less
important
to
faculty
(Fu =
+1.66). However,
both
s tudents
and faculty
w an t
th e
library
to be
com fortable
(Fj =
-1.35
and FH = — 1.51),
even though
they disagree wh ethe r this in volv es quie t
(Fj =
+1.60
and FU = -1.33).
MAJOR
CONCEPTS
In
many research situations, factoring variables
via
R-technique
ad-
dresses im port ant issues, such
as
bet ter unders tand ing
th e
nature
of
intelli-
gence or
self-concept.
But both
people
and
t ime
can
also
be
factored.
Factoring people
is
often
of
interest
in
psychology, because psychological
theory often is abou t type s of people (e.g., Type A perso nalitie s).
Q-technique factor analysis is the second-most comm only used EFA
method , after R-technique.
In
Q-technique, d eciding
how
many people
factors
to
extract addresses
the
question "How many types
of
people
are
there?" across
th e
variables used
in a
given analy sis. Persons' pa ttern
or
s t ructure
coefficients large
in
m a g n i t u d e
(plus or
min u s )
on a
given facto r
can be
examined
to
de te rmine
Which
factored people belong
to a
given
type?" And the
factor scores that
are
large
in
magnitude
can be
examined
to
de te rmine
Which
variables were
the
basis
for
del ineat ing
the different
person
factors?"
SIX
TWO-MODE TECHNIQUES
91
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 90/185
8
CONFIRMATORY ROTATION A N D
FACTOR INTER PRETA TION ISSUES
There are
var ious s i tua t ions
in
which
the
researcher
m ay
w a n t
to
evaluate
how
we l l
tw o
di f feren t
factor structures
f it
each other.
J u s t
as
unrelated factors c a n n o t be readi ly in terpre ted, as reflected in the Table
3.6
results (see chap.
3),
factors across samples often can n o t
be
accu ra te ly
compared unless the two structures are compared analytically using Procrus-
tean rota t ion methods.
These m eth o ds rotate one st ruc ture to a best-fit posi t ion w ith a
second target structure.
If the two
factor matrices
do n o t
match under these
condi t ions, they can n o t
f it
u n d e r
an y
other condit ions,
and the
s t ruc tures
s imply
are
incompat ib le .
Some
example applications
of
these methods
in -
c l u d e
empir ica l ly evaluat ing
1. w h e th e r the factor pattern /struc ture matrices for a given set of
variables across di f feren t samples
are
invariant across samples;
2 .
whether there
are
substantively based differences
in
s t ruc ture
fo r a set of
i tems across groups (e.g. , graduate students
vs.
facul ty) ;
3. whether the st ruc ture is replicable using in ternal replicabil i ty
( T h o m p s o n , 1994, 1996) cross-validation methods (i .e . ,
r andomly sp l i t t i ng
the
sample
to
in form judgment s abou t
replicabil i ty); and
93
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 91/185
4 - the factor adequ acy of an obtained pa ttern /stru ctu re m atrix
by
eva lua t ing
its fit to a
theoretical ly expected target stru ctu re
consist ing of
ones
(o r negative ones) and zeroes (Thompson
&
Pit ts, 1982).
Thompson (19 92 b) provid ed a part ial stat ist ical significan ce test d istr ibu tion
for evaluating these factor correlations across samples.
BEST-FIT
ROTATION
E X A M P L E
For heur is t ic purposes,
th e
invariance
of the
pr inc ipal components
var imax-rota ted pa t tern /s t ruc ture coeff ic ients involv ing percept ions of li-
brary
service qual i ty w i th regard to 11 var iables (PE RI through P E R U )
wil l
be evaluated. T he comparison involves th e st ruc ture for 100
facul ty
reported in T ab le 5.5 (see chap. 5) being compared wi th the corresponding
s t ruc tu re
of 100 graduate s tudents (see Append ix A). The gradu a te s tuden t s'
s t ruc tu re wi l l be
used
as the
target matrix
for
this analysis.
Ve ldman (1967 , pp .
2 38 — 2 4 4)
expla ined the mathemat ica l theory
under ly ing
the method . H e
noted that
th e cosines among the factors across
the two
solut ions equal
the
factor correlat ions across samples.
H e
also no ted
that this matrix of factor cosines can be used to rotate a second mat r ix to
its best-fi t posi t ion with a target matr ix .
The cosine m atri x is comp uted using the m atrix algebra fo rmu la:
C j
x
i = ( A ' i
x
K B
K x
j ) ' j
x
|
V j
x
j
E
A x
j ~ • V ' j
x
| ,
where I = the n u mb e r of fac tors in the A target pat tern matr ix; J = t he
num ber of fac tors in the B pat tern matr ix; and K in the nu mber of var iables
in the two analyses. This algori thm requires first the computa t ion of
Q i x i =
A ' |
x K
B
K x
j ( A '
l x K
B K
X
J ) J
X
I .
The matr ix V |
x
j is the mat r ix of pat tern coefficients extracted from Q [
x
j The
matr ix E
A x
j has
zeroes
off the
d i a g o n a l
and on the
diagonal
th e
e igenvalues
of
Q i x i .
This analysis can be conducted in SPSS using th e syntax presented
in
Ap p e n d i x
B. By the
way, this syntax invokes
th e
MATRIX procedures
avai lable in SPSS, i f one learns how to type syntax rather than execut ing
SPSS via point and c l ick. In App end ix B, a l l the syntax between the
MAT R I X c o mma n d a nd t he EN D M AT RIX command execu tes ma t r ix
algebra. T he adventurous reader can access an SPSS ma n u a l to discover
how to command
SPSS
to execute any matr ix procedure (e.g., transpose,
inversion, mul t ip l ica t ion)
o r
matr ix formula .
9 E X P L O R A T O R Y A N D CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 92/185
TABLE
8 1
Factor Correlations (Cosines)
Student
target
F CTJ
F CTJI
F CTJII
F CTJ
-.007
1.000
.022
Faculty matrix
F CTJI
1.000
.006
.024
F CTJII
-.024
-.022
.999
Table 8.1
presents
the
factor correlat ions across
the two
samples .
The
resu l t s indicate that the factors were h ighly congruent or invariant across
samples (e.g. , r^
x
u g = 1.00), al though the order of the factors was
d i f f e r en t
in the two
samples (e.g. , Factor
I in the
s tuden t sample
w as
Fac to r
II in
th e facu l ty
s a m p l e ) .
Table 8 .2
presents
th e
fac to r pa t t e rn / s t ruc tu re m at r ix
of the 100 f acu l ty
after
rotation to
best
fit
w i t h
the
s t ruc tu re
of the 100
gradu ate s tudents .
The s t ruc tu res are very similar across the two groups in th is data set .
The p rogram also produces cosines ( i .e . , here correlations) betw een
the
var iables
in the
target pat tern /s t ructure matr ix
and the
var iables
in the
best-f i t ro ta t ed pa t t e rn / s t ruc tu re c oe f f i c i en t ma t r ix . These cosines among
variables should all be reasonably large to ind ica te
that
the two solutions
fit reasonably wel l at the level of each of the var iables , and not jus t at the
factor level . If some var i ab les ' cosines are smaller , these are the var i ab les
that
could
not be
related wel l
in the
best- f i t
rotation.
For
th e
Table
8 .2 d a t a , th e lowest of the 11 va r iables ' cosines was .94
for variable PER9, and the 10 remaining
cosines
were all .98 or larger.
These
resul t s con f i rm fac to r f it even across the 11 measu red var i ab les .
TABLE
8 2
Pattern/Structure Matrix
for 100
Faculty Rotated
to
Best-Fit Position With
th e
Pattern/Structure Matrix
for 100
Graduate Students
Factor
Variable
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
PER9
PER10
PER11
FACT_I
.890
.868
.726
.866
.115
.124
.167
.315
.087
.209
.321
FACT_II
.204
.109
.231
.164
.9 04
.892
.905
.792
.323
.166
.105
F CTJII
.213
.239
.185
.266
.240
.239
.223
.170
.677
.801
.730
C O N F I R M A T O R Y R O T A T IO N
A N D
F A C T O R I N T E R P R E T A T I O N
95
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 93/185
F A C T O R
I NTER P R ETATI ON CONSIDERATIONS
These
results afford
an
opportunity
to
state some general principles
of
resul t
interpretation. The statement of these principles requires speci fy ing
which
aspects
of
resul t s
are and are not
relevant
to the
interpretation process.
Factor Order
Both
hefore and
af ter
rotation, factors are ordered by the amount of
information they
contain.
Prior to rotation, the eigenvalues characterize
the amount of information
contained
in the unrot ted factors. For example,
if the first eigenvalue associated with the first principal
component
was 3.0,
and 11 variables were included in the R-technique analysis, the result
indicates that 27.2% (3.0 / 11) of the variance in the variables can be
reproduced
by the first
component.
Factor rotation redistributes the variance across the factor in service
of pursuing simple structure. The eigenvalues are not relevant to evaluating
the
information contained
in
particular rotated factors. Computer packages
print sums
of
squares
to
address
the
issue
of the
information
contained
within specif ic rotated factors.
In
general,
the
order
of the
factors
is not of
interest
in
evaluating
the
solution.
For
example,
here the
analyst expected
three
factors
in both the
student and the f acu l ty samples: Service A f f ec t , Information
Access,
and Library as
Place.
These dimensions did emerge in
both
samples. No
particular factor ordering
was
theoretically expected
in
this analysis.
It
does
not
matter that
the
factors emerged
in a d i f f e r en t
sequence
in
these analyses.
In
some analyses
there
might
be an
expectation
that one
factor w i l l
dominate
the
factor space (i.e., contain
a
hugely disproportionate amount
of
variance).
Only
in such cases would a theoretical expectation regarding
f ac t o r
ordering
be relevant.
Reflection of Factors
Some factors involve pattern
an d
structure coefficients with
d i f f e r e n t
signs. Because most people
find it
easier
to think in
positive terms,
if the
larger
pattern
and
structure coefficients have negative signs
on a
given
fac tor , it is completely appropriate for the analyst to reflect the factor by
reve rs i ng the signs of the coefficients on any given factor (Gorsuch, 1983,
p. 181).
This
is
appropriate because psychological variables typically have
no
intrinsic scaling
direction. For
example,
an
achievement test could
be
scored
by
counting either
th e
number
o f
right answers
or the
number
o f
wrong
answers.
96
E X P L O R A T O R Y
AND CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 94/185
How ever , w hen a g iven fac to r i s r e f l ec t ed , al l the signs of al l the patte rn
and the s t r u c t u r e coef f ic ien ts on the fac to r mus t be rever sed . I t i s not
necessary to a c k n o w l e d g e the reflect ion process in descr ibing resu l t s . And
in a g iven so lu t ion , any c o m b i n a t i o n o f d i f f e r en t fac tors m ay b e r e f l e c t e d , '
i nc lud ing , o f course,
none.
Salience of Variables
It is common to see researchers invok ing some c r i ter ion to de te rmine
w hat va r iables are sal ien t to the in terp retat ion o f a g ive n facto r . V alues o f
pat tern /s t ructure coeff ic ien ts
of
|.3|, |-35|,
and
|.4|
are
used wi th some f r e -
q u e n c y . Such v a l u e s
m a y h e useful in
se t t i ng
th e
m i d p o i n t
o f
gaps b e tween
smal ler an d larger coeff ic ien ts o n a g iven fac to r .
Researchers
m ay
u n d e r li n e
o r
ita lic ize
th e
ent r ies
in
pa t t e rn / s t ruc tu re
matr ices that meet thei r cr i ter ia , to help readers see pat terns . Or researchers
m ay sort th e rows of the mat r i ces by the va lues of the biggest abso lu te
coef f ic ien ts on
d i f f e r en t
factors .
Some researchers blank ou t nonsa l i en t coe f f ic i en ts to make s imp le
st ructure more apparent . Unl ike i ta l ic iz ing
coef f ic ien ts
o r sor t ing factor rows,
omi t t i ng nonsa l i en t coe f f i c i en t s is n ot good pract ice . Doing so dep r ives
other
researchers of the oppor tun i ty to conduct secondary analyses , such as
i nvok ing d i f f e re n t ro ta t ion methods ,
o r
c o n d u c t in g best- f i t ro tat ion
o f
their
resul t s
w i t h yours .
It
is impor tan t not to ge t too rigid in determining sal ience cr i ter ia .
Use a sal ience cr i ter ion in a g i v e n s t u d y that does a good job for the g iven
resul t s in h igh l igh t ing s imp le s t ruc tu re .
Naming
actors
An impor tan t part of the interpretation process is n a m i n g the factors .
This
is
where
the art of the
fac tor analy t ic process
is
exercised . Factor names
are usual ly one- o r two-word phrases .
Factors should
never
be named af ter mea sured var iables . A fac tor i s
inherent ly an aggregat ion o f several measured factors . A factor ref lected in
only a s ing le measured var iable ( i .e . , a speci f ic factor ) is inheren t ly n o t
a t rue factor . Thus, a ll real fac tors invo lve mu l t ip l e var i ab les , and to th is
extent m u s t
be
n a m e d
in a
man ner r e f l ec t ing
th e
o veral l pat tern
o f
c o n t r i b u -
tion
o f di f fe ren t
variables
to the
factor ' s def in i t ion
(o r
conversely ref lect ing
th e unde r ly ing cons t ruc t 's ma n i fes t a t ion v ia the measu red var i ab les ) .
Factor naming also must be in fo rmed by the
d i f f e r en t i a l
sa tu ra t ion o f
a
g iven factor
by
g iven factored en t i t ies .
F or
example ,
in an
orthogonal
R -
technique analys is , i f one var iab le has a pat te rn /s t ru ctu re coeff ic ien t o f +.7
o n a
factor ;
and another
var i ab le
has a
pa t t e rn / s t ruc tu re coe f f i c i en t
o f
+ .5 ,
C O N F I R M A T O R Y R O T A T I O N A N D F A C T O R IN T E R P R E T A T I O N 97
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 95/185
do not forget tha t the first va r iab le has 49% of its va r iance in c o m m o n w i t h
the f ac to r , whereas the second measured var iab le has on ly 25% c o m m o n
variance . Thus, the
first
var iab le should have rough ly doub le in f luence on
selection of the factor label.
Fac to r names mus t be
suff iciently
broad to cap ture all of the p r i m a r y
relationships suggested
by the
pa t te rn
and
s t ruc ture coef f ic ients .
But there
is
ano the r cavea t
that
m u s t
be
m e n t i o n e d .
When
conduct ing fac tor analyses
and
com paring resul ts w i th those
in
pr ior re la ted s tud ies ,
it is
impor tan t
to not
p lace undue re l i ance
on
construct names assigned
by
previous researchers .
Resea rche rs may have nam ed the i r f ac to r s in ways tha t d o not seem ins igh t fu l
or even reasonable. If the fa ctor labels from pr ior s tud ies are accepted w i tho ut
scru t iny ,
inva l id compar i sons wi th re la ted p r io r f i n d i n g s m a y re su l t .
M A J O R C O N C E P T S
In
va r ious s i tua t ion s re sea rche rs may want to d e te rmine how we l l two
d i f f e ren t
f ac to r s t ruc tu res
for a
g iven
set of
fac tored ent i t ies (e .g ., var iab les
if
R - t ec h n i q u e
is
u s e d )
f it
each other. Empi r ica l me thods usua l ly m u s t
be
used for
this purpose , because ap parent
d i f ferences may be
method ar t i fac ts ,
and emp i r ica l me thods p rov id e a quan t i f i ca t ion o f degree o f s t ruc tu re cor re -
spondence.
The
A p p e n d i x
B
SPSS syntax
can be
used
as a
mode l
to
conduct
such best-f i t rota t ions .
In ev aluat ing correspondence of fac tors across samples or subsam ples ,
factor
order
is usual ly not a
cons idera t ion. Sim ilar ly ,
a
given fac tor across
samples
may be
or ien ted
in
oppos i te d i rec t ions , such that
the
coef f ic ients
for
a
fac to r
in one
sample mus t
be
ref lec ted
by
changing
a ll th e
signs
o f
a ll coefficients
on the fac tor . Factor or ienta t ions are comple te ly arbi t ra ry ,
but we
usually prefer
the
predominance
o f
l a r g e m a g n i t u d e
coefficients to
be
posit ive, because most people
find it
eas ier
to
th ink pos i t ive ly .
When
present ing fac tor s t ruc tures , researchers of ten highl ight
coeffi-
cients that they deem sa l ien t . Such salience criteria
are
usua l ly
set by
looking
for gaps be tween near-zero and no nzero co ef f ic ients (e .g . , absolute .4) in a
given solut ion. Sal ience cr i ter ia should not be used wi th excess ive
r ig id i ty .
Factors are synthet ic var iab les re f lec t ing un de r lying la tent cons tructs ,
and therefore should never be named af ter measu red var iab les . Eq ual ly
impor tan t , when read ing p r io r re sea rch ,
do not
accep t
the
factor labels
of
others at
face
va lue . When conduc t ing either sub jec t ive or empir ica l fac tor
meta-analyses (Thompson, 1989), focus on the inva r iance of the f ac to r
meanings across s tud ies , notwiths tand ing
th e
labels assigned
by
d i f f e r en t
researchers.
98 E X P L O R A T O R Y AND C O N F I R M A T O R Y F A C TO R
A N A L Y SIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 96/185
9
I N T E R N A L REPLICABILITY ANALYSES
In
quantitative research,
all
statist ical methods, including factor analy-
sis,
are susceptible to the influences of
three
kinds of errors: sampling error,
measurement error,
and
model specification error.
As
noted
in
chapter
4,
sampling error
is
that portion
of
variance
in
sample data
that is
unique
to
a given sample, and in
effect
reflects the personality of a given sample.
B y
definition, the sampling error in a given sample does not exist in the
population.
e surement
error is the
variance
in
data that
is due to
inherent
imperfec t ions in measurement (Thompson, 2003). We estimate these inf l u -
ences by computing ratios of reliable variance to the total observed score
variance; these ratios
are
called reliability
coefficients. In
every study
it is
important to estimate the reliability of the data in
hand
in the study and
not
merely cite reliability coefficients
from
the
manual
or
prior studies
(Vacha-Haase, Kogan, &
Thompson,
2000). As Wilkinson and the Task
Force on Statistical Inference (1999) emphasized:
It
is important to remember that a test is not
reliable
or unrel iable. . . .
Thus ,
authors should provide
re l iabi l i ty
coefficients
of the scores fo r
th e data being analyzed even when th e focus of their research is not
psychometric , (p.
596)
Model specification error
is the
failure
to use all and
only
the correct
variables,
or the f a i l u r e to use the correct statistical analysis, or
both,
thus obtaining
99
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 97/185
incorrect resul ts . Of course, as Ped haz ur (1982 ) has noted, The rub , how -
ever,
is
that
the
t rue model
is
seldom,
if
ever, known
(p.
229 ) .
And as
Du ncan (19 75 ) has noted, Indeed it wou ld req u ire no elaborate sophistry
to show that we
wi l l
never have the 'r ight ' model in any absolute sense
(p.
101).
Sampling error is par t icular ly pernic ious in mu l t ivar iate analyses , be-
cause more complex analyses afford more opportunit ies fo r these influences
to be
m anife s ted . With respect
to
exploratory factor analys is (EFA ), Gorsuch
(1983) noted, In factor analysis , one has nu merou s poss ibi l i t ies f or
capital iz-
in g on chance. Most extraction procedu res, inclu ding principal factor
solu t ions ,
reach the ir criterion by such capital iz at ion. The same is tru e
of rotat ional procedures, including those that rotate for simple structure
(p. 330) .
There
are two
major ways
of
empir ical ly address ing resul t repl icabi l i ty
quest ions: external repl icabi l i ty analyses an d internal repl icabi l i ty analyses
( H u b e r ty , 1994;
Thompson,
1996). Ex ternal repl icab i l i ty s tudies involve
collect ion of independent data from different par t ic ipants . These are true
replicat ion studies and provide the best evidence of resul t repl icabi l i ty by
direc t ly
and exp l ic i t ly address ing repl icabi l i ty concerns .
However, many researchers
find
themselves unable
to
a lways conduc t
external
repl icabi l i ty s tudies . In these situat ions researchers may use some
combination of three major logics for internal repl ica bi l ity analyses: cross-
val idat ion, th e jackkn i fe , and the bootstrap (Thompson, 1993b, 1994).
These
three methods combine the
subjects
in hand in
different ways
to
dete rmine wh ether resul ts are stable across sample variat ions, that is, across
th e idiosyncracies of in d iv id u a l s that make general izat ion in social science
so challeng ing (Thomps on, 1996,
p.
29 ) .
Here both EFA cross-validation and bootstrap methods wil l be il lus-
trated. These are, respectively, the simplest and the most complex internal
replicabil i ty methods. Jac kk nife applicat ions have been discussed elsewhere
( H u b e r t y ,
1994).
Across the three major internal re plic ab il ity analys is choices it is
important to remember that none of the methods is as useful as true replica-
t ion
s tudies .
A s
Thompson
(1993b) noted, because
all three strategies are typical ly based on a single sample of subjec ts in
w h i c h th e sub jec t s u s u a l l y
h a v e
m u c h in common (e .g. , point in t ime
of
measu rement , geographic or igin) re lat ive
to
what they wou ld have
in common
w i t h
a separate
sample ,
the three
methods
a l l
y ie ld somewhat
inf lated
es t imates of repl ica bi l i ty . Because inf lated es t imates of repl ica-
b i l i t y
provide a bet ter es t imate of rep l icab i l i t y than no estimate at all
( i .e . , s t a t i s t i ca l s ign i ficance t e s t ing) ,
these
procedures can st il l b e usefu l
in
focusing on the sine
qua non
of science, (pp. 368-369)
00 EXPLORATORY A ND
C O N F I R M A T O R Y F A C T O R ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 98/185
Cross-validation and bootstrap applicat ions wil l be il lustrated here using
for heu r is t ic purposes the first 30 cases of data and var iab le s PE RI through
PER8 from Appendix A .
EFA
CROSS-VALIDATION
Cross-validation methods executed
as
internal repl icabi l i ty analyses
randomly split the data into tw o subgroups. Analyse s are conducted sepa-
rately
w ithin each sub group, and then an empirical analys is must be execu ted
to quant i fy the degree of fit across the two analyses. These methods have
been used across the general l inear model ,
from
reg ression to canonical
correlation analyses (cf. Grossman, 1996;
Thompson,
1994). The methods
are
logical ly straightforward
and do not
requ ire specialized sof tware .
Table
9.1
presents
th e
cross-validation su bgrou p assignments ( 1
or
2 )
for the 30
cases produced
by
flipping
a
coin.
For
logistical purposes
as
is typical ly done, becau se keeping t rack
of the
subgroups
is
thereby faci l i ta ted,
here the 30 cases were split 16/14 rather
than
15/15. Table 9.2 presents th e
pat tern/s t ruc ture matr ices
from
principal components analyses rotated to
the varimax criterion for both subgroups.
T he best-f i t rotat ion procedure must then be invoked to al ign th e
factor s t ruc tures
to
opt imal
fit.
E ither so lut ion
can be
used
as the
target
f or
this analysis, al though some preference migh t be afforded to using the larger
sample as the target .
Table
9.2 also presents th e
second
solut ion rotated to
best-f i t
posi t ion w ith
th e
so lu t ion
for the f irs t
subgroup .
In
this
smal l
heur is t ic
example
th e
solut ions
in the two
subgroups
are
reasonably congru ent.
This is
reflected in the
cosines
of the two
solu t ions
vir tual ly
being
an
ident i ty matr ix.
The biggest disparity across the subgroups involves the variable PER8,
which in the second subgroup had pat tern/s t ruc ture coeff ic ients on Factor
of
.667
and on
Factor
II of
.596.
In the first
sub group, PER8
had a
t r iv ial
pat tern/s t ruc ture coefficient
on
Factor
I
( .158
2
=
2 . 5 % )
but a
huge va lue
on
Factor
II
(.840
=
70 .6%) .
This
dispari ty
is
also reflected
in the
cosines
for
the eight variable pairs
each
being greater than .94, except for PER8,
which had a cosine of .808.
Cross-validation can g ive some insight regarding potential resul t
repli -
cabi l i ty .
If
solu t ions wil l
n ot
replicate across sub groups
in an
inte rnal replica-
bi l i ty analys is involv ing a single sample,
then
replicat ion with a
t ru ly
inde-
pendent second sample seems part icularly un like ly.
And i t
should
be
emphasized that
the
fu l l
sample resul t s are a lways the basis for our interpre tat ions . T he
subsample resul t s are
only
employed as a vehic le for
exploring
th e
repl icabi l i ty
of the fu l l
sample results .
T he f u l l
sample results
are
used
INTERNAL
REPLICABILITY ANALYSES
101
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 99/185
TABLE
9.1
Cross-Validation
Random
Subgroup Assignments and Bootstrap
Resampling Assignments
C r o s s v a l i d a t i o n
Case
3
4
5
6
7
8
9
2
3
4
5
6
7
8
9
2
2
22
23
24
25
26
7
28
29
3
S u b g r o u p
Draw
3
4
5
6
7
8
9
2
3
4
5
6
7
8
9
2
2
22
23
24
25
26
7
28
29
3
29
8
3
24
3
7
5
6
22
2
26
3
4
5
2
22
5
26
4
5
4
8
24
22
25
R e s a m p l e
2
3
4
2
7
3
25
7
7
28
6
4
9
3
5
7
2
24
6
28
5
2
5
28
25
23
3
2
27
26
2
7
4
23
8
29
6
3
3
5
5
28
4
9
3
24
26
9
9
4
6
9
8
4
25
9
7
25
4
4
23
8
2
3
8
7
7
3
8
22
5
9
3
7
28
5
fo r interpretation, because we tend to vest more confidence in resu l t s
as sample
size
increases, and the fu l l s ampl e
analysis
involves
more
subjec ts
than do the s u bgr ou p ana l ys es . (Thompson,
1994,
p.
1 7 2 )
The subgroups are only used a s a vehic le to evaluate the possible replicability
of th e factor solution for the complete sample.
EFA
BOOTSTRAP
The bootstrap is an elegant computer-intensive internal replicab il ity
analysis conceptualized fo r several analyses by Efron and his col leagues .
Diaconis
an d
Efron (1983) provide
an
accessible art icle su mm ariz ing this
1 02 E X P L O R AT O R Y
A N D
C O N F f R M A T O R Y FA CTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 100/185
TABLE 9.2
Varimax-Rotated Factor Pattern/Structure Matrix for Tw o Subgroups
(n , = 16; n
2
= 14)
Subgroup 1
(target)
Variable
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
I
.896
.954
.784
.877
.376
.326
.261
.158
II
.340
.110
.474
.354
.792
.845
.906
.840
Subgroup
2
.929
.942
.809
.596
.345
.356
.228
.667
II
.214
.250
.496
.487
.898
.890
.926
.596
Subgroup 2
after
best-fit
I
.923
.935
.796
.583
.322
.333
.205
.657
II
.238
.274
.576
.502
.906
.899
.932
.613
Note. Pattern/structure coefficients greater than |.5| are presented in italics.
logic
in layperson's terms. Lunneborg's (2000) book provides a more techni-
cal and comprehensive view.
The
bootstrap
can be
conceptualized
as a
method
of
using
the
sample
data
in
such
a
manner that
the
sample data
are
themselves used
to
estimate
the distribution of a large number of sample results from the population
(i.e., the sampling distribution).
This
process of the estimation pulling itself
up by its own
bootstraps explains
why the
approach
is
named
the
bootstrap.
One way of
explaining
the
technique, although somewhat
s impl i f ied ,
begins
by describing the creat ion of a mega-file created by copying the sample
data set of size n
onto
itself ( i .e . , concatenating) a ve ry large number of times.
Then repeated resamples
of
cases,
each of
exactly
size
n
are
drawn
from the
mega- f i l e ,
and in each resample the results are analyzed independently.
The
number
of
resamples
is u su a l ly at
least
1,000, although 2,000 to
5,000 resamples
are
recommended.
In a
given resample,
a
given case might
be
drawn
not at
all, once,
or
several times.
And
this case might
be
drawn
not at all, once, or several times in the next resample. After all the resamples
are drawn, mean (or median) parameter estimates (e.g., R
2
, factor pattern/
structure coef f i c ients ) are computed across the resamples.
The
standard deviations
of the
estimates
are
also computed. These
are
empirical (as
against theoretically based) estimates
of the
standard errors
o f
each of the
parameter estimates.
There
is
only
one diff icul ty in
applying
the
bootstrap
in
multivariate
statistics such as EFA: When more than one set of weights (e.g., factors)
can be
computed,
the
weights across
the
resamples must
be
rotated
to
best
fit
w i t h
a common target matrix prior to computing averages and standard
errors. This is because in the first resample a construct might emerge as
Factor I, but in the next resample might emerge as Factor II. Factors must
I N T E R N A L
KEPLJCABILfTY
ANALYSES
103
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 101/185
TABLE
9.3
First Four Eigenvalues From Four Resamples n =
30)
Factor
Resample
1
2
3
4
Mean
SE/SD
I
3.867
4.798
5.357
5.190
4.803
0.577
II
2.523
1.776
1.312
1.553
1.791
0.454
III
0.600
0.675
0.548
0.605
0.607
0.045
be rotated to
bes t - f i t
position so that averages of apples and apples are being
computed across
the
resamples.
Of
course, these mechanics
are not d i f f i cu l t
when they
are
automated,
and bootstrap applications across various mult ivariate analyses have been
reported
and recommended (Thompson, 1988, 1992a, 1995). In EFA the
bootstrap can be used to
inform
the decision of how many factors to extract,
to help evaluate the replicahility of the structure, or both.
Factor Extraction Bootstrap Procedures
The
bootstrap
can be
used
as
another procedure
to
decide
how
many
factors to extract, in addition to the methods described in chapter 3. In this
heuristic example only four resamples are drawn. This makes the s u m m a r y
manageable, but of course is not a realistic description of an actual boot-
strap analysis.
As reported in Table 9.1, in the first resample, the first case drawn
w as
Case #1. In each draw all the data for a given case are extracted into
the
resample.
Case #1 was
also drawn again
in
Draw
21.
In
each
resample the eigenvalues are computed.
Then
means of the
eigenvalues
and the standard deviations can be derived. Because the resam-
ples are being used to estimate sampling distributions, these standard devia-
tions are actually empirically estimated standard errors (i.e., SDs of the
parameter estimates
in the
sampling distribution).
Table
9.3 presents the first three of the eight eigenvalues
from
principal
components analyses
of the
data from
the four
resamples.
The
table also
presents
the
mean eigenvalues across resamples
and the
standard errors
of
these estimates.
For
example,
the
mean eigenvalue across
four
resamples
for
the second factor without rotation was 1.791 (SEj. = 0.454). The mean
eigenvalue across four resamples
for the
third factor
was
0.607 (SE^
=
0.045).
F or these results the extraction of two factors is suggested. If the third
eigenvalue mean
had
been 0.999,
and the
standard error
had
been small
104
E XP L ORA T ORY
A N D
C O N F I R M A T O R Y
FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 102/185
bu t
nontr iv ial , the extraction of the third factor might have been warranted.
In other cases, even
if an
e igenvalue mean
is
slightly greater than 1.0,
a
relat ively
large standard error would
suggest
that not much fai th can be
vested
in
this eigenvalue consistently being greater than 1.0;
here
some
thought might he g iven to not extrac t ing the factor associated with this
re la t ively
smal l and unstable e igenvalue .
Pattern Structure Coefficients and Standard Errors
Bootstrap estimation of factor pattern/stru ctu re coefficients and their
stand ard errors requ ires the spe cifi catio n of a common factor space or targe t
to wh ich all resampled so lut ions can be rotated to a position of best fit.
The
s t r u c t u r e
in the
actual sample
can be
used
as the
target
for
this purpose.
Another
choice is a
theoretically expected factor structure.
In the
present example,
the
target matrix reflected
an
expectation that variables
P E R I
throug h PER 4 wou ld have +1.0 coefficients on Factor I , and zeroes
on Factor II. The target matrix also reflected an expectation that var iables
PER5 through PER8 would have +1.0 coefficients on Factor II, and zeroes
on
Factor
I.
Table 9 .4 presents the v ar imax-rotated pat tern/s t ruc tu re matr ices from
principal components analyses
in the
four
resamples. Note that
the first
factor
inc lu des var iables PE RI through PER4
in
Resamples 2
and #3 but
not
#1 and #4.
Table 9 .5 presents the four s t ruc tures each rotated to best-f i t posit ion
w i t h the target matrix. Table 9.6 summarizes the means of the paramete r
est imates
and
their standard errors.
Table 9 .6
also reports
th e
mean parameter
est imates d iv ided
by
their respective standard errors.
The ratio of a paramete r estimate to i ts standard error (SE ) is very
important in statist ics, and so this ratio is given m any names. These names
include critical ratio, Wald statistic, an d t. W e usual ly expect these t val ue s
to be
greater
than
roughly |2.0|
for
parameters that
w e
wish
to
interpret
as
being nonzero.
Overal l ,
the
heurist ic results,
if
they
had
been real , would have sup-
ported
a
conclusion that
the
fac tor s t ruc tu re
w as
reasonably replicable .
This
is reflected
in the fact that the em pirica lly estimated standard errors for
large
va lues
in
Table
9.6 are
quite small (e.g., .039, .009).
These
standard
errors indicate
that th e
noteworthy parameters
in the
model
do not
vary
mu ch across various configu rations of the people considered in the bootstrap
resamples. If we mix up our
sample cases
in
many
different
configurat ions
and
always
sti l l get the same basic factor structure, this would give some
hope that the
s t ruc ture
may
also replic ate across
the
configurat ion
of
par t ic i -
pants in a new sample .
I N T E RN A L RE P L I C A B I L J T Y A N A L YSE S 1 05
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 103/185
TABLE
9.4
Varimax-Rotated Pattern/Structure Matrix From Four Resamples n
K
= 30)
Factor
Variable
~i ll~~
Resample 1
PER1 -.066 .921
PER2 -.067 .958
PER3 .347 .836
PE R4 .251 .808
PER5 .847 .219
PER6 .904 .031
PER7 .917 -.021
PER8 .797 .135
Resample 2
PER1 .966 .136
PER2 .964 .130
PER3 .927 .266
PER4 .759 .431
PER5 .128 .927
PER6 .344 .784
PER7 .182 .845
PER8 .159 .775
Resample 3
PER1 .869 .297
PER2
.947 .127
PER3 .810 .493
PER4 .859 .292
PER5 .555 .670
PER6 .226 .899
PER7 .251 .882
PER8
.228 .825
Resample 4
PER1 .067 .947
PER2
.215 .933
PE R3 .485 .836
PER4 .541 .632
PER5
.922 .203
PER6 .911 .228
PER7 .878 .148
PER8 .788 .329
MAJOR
CONCEPTS
A ll statis tical methods, inclu ding factor analysis , are susceptible to the
influences of three kinds of errors: sampling error, measurement error, and
model specification error.
To
address
th e
influences
of
sampling error,
re -
searchers replicate resu lts wi th independ ent samples (i .e. , external replica-
/ 06 EXPLORATORY A ND CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 104/185
TABLE 9.5
Four
Resample Solutions Rotated to Best Fit With Target
Factor
Variable
I II
Resample
1
PER1 .920 -.079
PER2 .957 -.080
PER3 .841 .335
PER4 .812 .240
PER 5 .231 .844
PER 6 .044 .904
PER7 -.008 .917
PER8 .146 .795
Resample
2
PER1 .968 .125
PER2 .966 .119
PER 3 .930 .255
PER4 .764 .422
PER5 .139 .926
PER 6 .354 .779
PER7 .192 .843
PER8
.168 .773
Resample
3
PER1 .864 .312
PER2 .944 .143
PER3 .802 .507
PER 4 .854 .306
PER5 .544 .679
PER 6 .211 .903
PER7 .236 .886
PER8 .214 .828
Resample
4
PER1 .950 .012
PER2 .944 .159
PER3 .863 .435
PER4
.663 .503
PER5
.256 .909
PER6 .282 .896
PER7 .199 .868
PER8
.375 .767
tion). When external
replication
is
impossible
or
impract ica l , internal rep l i -
cab i l i ty methods , such as cross-validation or the bootstrap, may be usefu l .
Exploratory factor analysis cross-validation methods can be conducted
by randomly spli t t ing the sample and using SPSS to factor analyze both
data sets, and
then
us ing the Appendix B bes t - f i t rotation to q u a n t i f y degree
of factor
fit
across
the two
subsamples . However ,
the
basis
for
actua l fac tor
interpretation is a lways the s t r u c tu r e from the total sample. The subsamples
INTERNAL REPUCABiL/TY ANALYSES
10 7
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 105/185
TABLE 9.6
Bootstrapped
Estimates of Pattern/Structure Coefficients and Their
Standard Errors
Variable
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
Mean
.926
.953
.859
.773
.293
.223
.155
.226
Factor
I
SD/SE
.039
.009
.046
.071
.152
.115
.095
.090
23.47
102.33
18.49
10.86
1.93
1.94
1.62
2.52
Mean
.093
.085
.383
.368
.840
.871
.879
.791
Factor
II
SD/SE
.146
.096
.096
.102
.098
.053
.027
.024
0.63
0.88
3.99
3.62
8.60
16.45
32.58
33.09
Note. Pattern/structure coefficients greater than |.4|
are
italicized.
are created only
to
gain insight regarding
the
character
of the
total data
set and not for factor interpretation.
The EFA bootstrap (Thompson, 1988) is a particularly powerful inter-
nal replicability method. The bootstrap is so p o w er f u l because it resamples
the data in so many
d i f fe rent
combinations. However, as a practical matter,
specia l i zed software must be used to conduct these analyses.
10 8 EXPLORATORY AND CO NFIRMATO RY FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 106/185
10
CONFIRMATORY FACTOR AN ALYSIS
DECISION SEQUENCE
Although the basic concepts of factor analysis are roughly a century
old (see Spearman, 1904), more recent extensions of the exploratory factor
analysis (EFA)
concepts
have created
the
basic methods
for
confirmatory
factor analysis (CFA;
see
Joreskog, 1969).
Both EFA and CFA are
part
of
the
general linear model (GLM).
Thus,
many
of the
concepts introduced
previously
with regard
to EFA
apply
in CFA as
well. Indeed,
an
important
emphasis
in
this book
is
elaborating
the
similarities between
EFA and CFA
as
part
of a
single underlying GLM, while also highlighting
th e
differences
in the two sets of methods.
Of course, the basic difference in CFA is that one or more underlying
models (i.e., how many factors, which measured variables are
thought
to
reflect
w hich la tent v ariab les, whe ther the factors are correlated )
must
be
specified
even
to run the
analysis.
There
are
also statistical analyses
that
are possible
in CFA but
impossible
in EFA
(e.g., allowing error variances
to be
correlated).
And some procedures that are routine in EFA, such as factor rotation,
are irrelevant in CFA. This is because the a priori mode ls thems elves typic ally
specify simple structure,
by
constraining certain factor pattern
coefficients
to be
zero while
freeing
other pattern
coefficients to be
estimated.
Confirmatory
factor analysis
is a
very important component within
a
broader class
of
methods called
structural equation modeling
( S EM) ,
or
some-
109
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 107/185
t imes
covariance structure analysis.
CFA specifies th e
"measurement models"
delineating how measured variahles reflect certain latent variables. Once
these measurement models a re deemed satisfactory, then th e researcher can
explore path models (called structural models) that link th e latent variables.
Structural
equation modeling is beyond the scope of the present vol-
ume.
Excellent book-length treatm ents of SEM are availab le elsewhere (see
Byrne, 1994, 1998, 20 01 ). Cha pter-length expositions
are
also available
(see Thompson, 2000c). Schumacker
and
Marcoulides (1998)
and
Marcou-
lides and Schumacker (2001) describe recent extensions of
these
methods.
Suffice it to say that even w hen the researcher is app lyin g the wid er
range of analyses possible with SEM , CFA mod els are often a critical start ing
place for such analyses (Thompson, 2000c). It makes little sense to relate
constructs w ith in an SEM m odel if the factors
specified
as part of the model
are not worthy of further attention.
CONFIRMATORY FACTOR ANALY SIS VERSUS EXPLORATORY
FACTOR ANALYSIS
In EF A, all parameters im plicit in a factor model m ust be estimated .
For example, if there are 9 measured variables, and 3 princip al components
are
extracted and then rotated to the promax cr i ter ion, then 27 ( 9 x 3 )
pattern coeff icients , 27 ( 9 x 3 ) s tructure coeff icients , 9 comm unali ty
coeffi-
cients
h ; or
their corresponding uniq uenes ses estimated
as u = 1
—
h" ) ,
and 3 factor correlation
coefficients ( r ]
X
n , r\x\\\, i i i x i n )
are estimated.
Of course, because Pg
x
3 R^ x 3
=
S
y x
3 given P, if either the factor
correlation matrix (R ^
x
3) or the factor structure m atrix (8 9
x
5) is m a themati -
cally
determined, so is either R or S, and thus in actua lity only 30 parameters
(2 7
+ 3) are
estimated. N ote
that th e
uniquenesses
are
also mathematically
determined once
th e
pat tern
and
structure matrices
are
estimated.
In CFA, th e researcher can "constrain" or "fix" certain param eters to
mathematically "permissible" values (e.g., a variance m ay be constrained to
equal any positive number; a correlation, r, may be constrained to equal
-1, +1, or any
number
in
between), while
at the
same time "freeing"
the
use of the input data to derive estimates of other model parameters (e.g.,
factor p attern coefficien ts, factor varian ces) . For example, the researcher
might constrain variables VI, V2, and V3 to reflect only Factor I, variables
V4, V5, and V6 to reflect
only Factor
II, and
variables
V7, V8, and V9 to
reflect only Factor II I, w ith the th ree facto rs allowed to be correlated.
In EFA, th e researcher m ay expect certain coefficients, but these expec-
tations cannot be incorporated into the analysis. In CFA a researcher must
110 EXPLORATORY AND
CONFIRMATORY
FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 108/185
declare as input into the analysis one or more
specific
models, each containing
some "fixed" and some "freed" parameters.
The original computer software for CFA (and S EM ), such as LISRE L
(i.e., analysis of Linear Stru ctura l RELationships; Joreskog & Sorbom, 1984),
required
th e researcher to input these model expectations using ma trix
algebra. A thorough knowledge of both th e uppercase and the lowercase
Greek alphabet was also required
Modern software, such as AMOS (Arbuckle & Wothke, 1999) an d
EQS (Bentler, 1995), instead uses a graphics-oriented user interface. The
researcher's expec tations are declared using the software to dr aw a diag ram
representing
th e
input model.
The
results
are
then reported
by the software
in text form (i.e., sets of coefficients printed fo r different matrices) , but also
as
an
output diagram with estimates printed
on the
input model diagram.
By
tradition, in these v arious compu ter programs measured variables
are
represented
by
squares
or
rectangles. Latent variables
are
represented
as
circles or ovals. Regression paths or, in the factor analysis case, pattern
coefficients, are
represented
by
one-headed arrows. These
are
drawn
from
factors
to measured variab les to reflect the premise that the latent variables
are
in
fact
an
underlying influence
on the
manifestation
of the
factors
in
the form of scores on the measured variables. Correlations (and cova riances)
are represented as two-headed arrows draw n to connect either m easured or
latent variables in pairs for which correlations or covariances are freed to
be nonzero and are estimated in the analysis.
Figure
10.1 presents the inp ut graphic model declaration implied by
an EFA involving nine measured variables and three correlated factors.
Table 10.1 presents
a
sequential count
of the 30
param eters that
a re
estimated
as
well as a sequential count of the parameters mathematically determined
once th e estimates are in hand.
Figure 10.2 presents th e related CFA inpu t model presum ing that each
of the
nine measured variables
is
associated with only
one
factor, three
measured
variab les per factor. The absence of paths between selected pa irs
of
measured and latent variables (e.g., between V I and Factor II, between
V2 and Factor II) implies that these factor pattern coefficients have been
constrained
to be
zero.
This
model declares simple structure
on its face,
because
no
measured variable
is
allowed
to
function
as an
indicator
fo r
more
than on e
factor. Thus,
if
model
fit is
adequate ,
no
rotation
will be
necessary.
Table 10.2 presents
a
sequential count
of
parameters estimated,
no t
estimated, and implied by this model. However, one difference between
EFA
and CFA arises in this presentation. Note from Table 10.2 that th e
nine
uniquenesses
u
) are counted as estimated in the CFA analysis. In
the EFA
analyses these were taken
as
implied, because once
th e
pat tern
an d
structure
coefficients are in
hand,
the
communal i ty coefficients h
) are
CONFIRMATORY FACTOR
ANALYSIS
DECISION SEQUENCE
I 11
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 109/185
factor \Y
Figure
10 1
Exploratory factor analysis factor model with nine measured variables
an d three correlated factors
computed by simple multiplications and additions along a given row of the
factor matrix, an d
then
uniquenesses ar e computed as u
2
= 1 —
h
2
.
Figure
10.3 presents yet another factor model for this research problem.
This
figure
illustrates
some
of the
model elements that
are not
possible
in
EFA.
First, the uniquenesses or error variances involving el and e2 have
been allowed
to
covary
or be
correlated
in
this model.
No
error variances
can be correlated in an EFA model, but in CFA the correlations of various
112
EXPLORATORY
AND
CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 110/185
TABLE
10.1
Counts of Parameters Estimated
and
Implied) by the Figure
10.1
Factor Model
Pattern Structure
Variable
V1
V 2
V 3
V4
V5
V 6
V7
V8
V9
Factor r
I
II
III
1
2
3
4
5
6
7
8
9
—
28
29
II
10
11
12
13
14
15
16
17
18
—
30
M l
19
20
21
22
23
24
25
26
27
—
I
31)
32)
33)
34)
35)
36)
37)
38)
39)
II
40)
41)
42)
43)
44)
45)
46)
47)
48)
III
49)
50 )
51 )
52)
53)
54)
55)
56)
57)
u
2
58 )
59 )
60)
61 )
62)
63)
64)
65)
66)
Note. Coefficients mathematically
determined
and thus
implied
once estimates are in
hand
are
noted
in
parentheses.
pairs
of
error variances
can be
estimated,
as
deemed necessary
by the
researcher.
Second,
only tw o factors are allowed to be correlated, and this correla-
tion
is
estimated
as part of the
model.
In EFA, factors must be either all
correlated or all uncorrelated.
Third,
an "equality" constraint, designated by the same letter (here
a)
b eing placed on the paths for the factor pat ter n coeff icients for three
measured variables on Factor III, h as been imposed.
This
means
that
the
input data will be consulted to estimate
these three
pa ttern coefficients, but
th e estimate has been constrained such
that
th e same single number must
be used for all three of these pa ttern coef ficients. Eq ual ity constraints are
not possible within EFA.
Equali ty constraints have several possible uses
in
EFA,
but one use is
to
test
a
theoretical exp ectation
that all
measured variables
on a
given factor
measure
that
factor equa lly well.
Fo r
exam ple, when scoring
on
test
subscales
is done
by summing item scores on a given subscale, the implication is that
th e items on the subscale measure th e underlying
construct
equally well.
This premise can be directly tested in CFA.
In CFA one may also (and should ) test rival models. Equality con-
straints might
be
imposed
in one
model; however,
in a
riv al model
all freed
pattern coefficients might be estimated with no equality constraints.
This
rival model might reflect a belief that the test items do reflect the intended
CONFIRMATORY FACTOR ANALYSIS
DECISION SEQUENC E
113
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 111/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 112/185
TABLE
10.2
Counts
of Parameters Estimated, Not Estimated, (and Implied) by the
Figure 10.2
Factor
Model
Variable
V1
V2
V3
V4
V5
V6
V7
V8
V9
Factor r
I
II
III
1
2
3
—
—
—
—
—
—
—
10
11
Pattern
II
III I
(22)
(23)
(24)
4 (25)
5
(26)
6
(27)
7
(28)
— 8 (29)
— 9 (30)
—
12 —
Structure
II
31 )
32 )
(33)
(34)
(35)
(36)
(37)
38 )
(39)
III
40 )
41 )
42 )
(43)
(44)
45 )
46 )
47)
48 )
u
2
13
14
15
16
17
18
19
20
21
Note. Coefficients mathematically determined and thus implied once estimates are in hand are noted in pa-
rentheses. Fixed parameters are designated by dashes ( — ). Here 18 factor pattern coefficients are con-
strained to be zeroes. And the three factor variances are fixed to equal 1
.Os,
thus making the factor covari-
ances equal factor correlation coefficients, because for example r . „ = COV,
x
„/ (1.0 x 1.0).
ival Models
Like SEM, CFA involves testing the fit of models to data. When
models fit well, as reflected in various statistical fit indices, some researchers
erroneously infer that
the
correct model
has
been specified,
and
that
in
some sense
the
model
is
proven.
In fact, it is conceivable that several models may fit a given data set.
As Thompson (ZOOOc) noted, "the fit of a single tested model may always
be an
artifact
of
having tested
too few
models"
(p.
278).
Thus it is
desirable
to test the fit of several r iva l models when conducting a CFA.
The fit of a preferred model is more impressive when that fit occurs
in the context of testing several r ival models, especially when some of the
rival models are theoretically plausible and are not articulated so as to create
straw person arguments. The fit of a model is also more impressive when
the
tested model
is
strongly
disconfirmable.
Disconfirmability
is partially a function of the degrees of freedom fo r
the model test. Degrees of freedom indicate, given the presence of one or
more sample statistics (e.g., the mean of the sample data, the r between
two measured variables), how many pieces of information in the analyses
remain
free to
vary
after the
estimate
is in hand. For
example,
in the
presence
of knowledge that the sample mean is 3, given any two scores in the data
set
(1, 3, 5), the researcher can always accurately determine the third score.
CONFIRMATORY FACTOR
ANALYSIS
DECISION
SEQUENCE
115
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 113/185
Figure 10.3. Confirmatory factor analysis CFA) factor model with three model
features unique
to
CFA.
For
this problem
th e
degrees
of
freedom
is a
function
of the
sample
size, n,
and
equals
n — 1.
Tahle 10.3 presents the c om puta tions for stat istic al significan ce testing
in any r or
m ult ipl e regression
(R )
analysis involving
p
pre dictor variables
(e.g., for the
bivar iate correlation,
p = 1). There are two
related situations
in which
^ C A L C U L A T E D
(and the corresponding
^ C A L C U L A T E D
also) is unde-
116
EX PLORAT ORY
AN D CONFIRMATORY FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 114/185
TABLE
10.3
Statistical
Significance Calculations
for r
2
and R
2
Problems
Source
Model
Unexplained
Total
Sum-of-
squares
A
B
C
Degrees of
freedom
D
E
F
Mean
square ^ C L C U L T E D
G J
H
I
r
2
/
ft
2
K
Note. D = p (i.e., the number of
predictor
variables). F = n -1 .E = / ? - 1 -p. G = A / D . H = B / E . I = C/
F. J = G/ H. K = A / C .
fined
and a statistical significance test cannot be conducted. First, regardless
of
sample
size, if the sum-of-squaresuN^xpLAiNEO '
s
zero
(and correspondingly
the r
2
or R
2
is
100%),
the F
cannot
be
calculated, because
the
mean
s q u a r e u N E X P L A i N E D
w
'"
De zero
>
an
J division by zero is mathematically imper-
missible.
Second,
whenever
the
degrees
of freedomMODEL
equals
the
available
degrees of freedom (i.e., fo r these analyses
n —
1), the
sum-of-squaresuNEX-
P L A I N E D
W
'1I always
be
zero.
F or
example,
if X and Y are
both
variables (i.e.,
SD
X
> 0 and SD
Y
> 0), and n = 2, r
2
must be 100%, and the sum-of-
s q u a r e s u N E x p L A i N E D
must
^Y
definition both
be
zero. Similarly,
in a
regression
analysis
involving an y tw o predictors, the R
2
wil l always be 100% whenever
n = 3.
Clearly, an R" of 100% is entirely unimpressive when 10 predictor
variables are used and n is 11 However, an R of 100% would be very
impressive (even though the result still could not be tested for statistical
significance ) if n = 1,000, because this regression model would be highly
disconfirmable given
that
th e residual degrees of freedom
UNEX
pLAiNEn
equaled 989.
The
residual degrees
of
freedom
are an
index
of the
potential discon-
firmability of any model. The good fit of models is more impressive when
degrees of freedom
UN E
xpLAiNED
ar
e
barge,
because as a function of the residual
degrees
of
freedom
the
model
is
parsimonious
and yet has
good explana-
tory ability.
In CFA (and SEM), however,
total
degrees of freedom for tests of
model fit are not a function of the sample
size,
n. Instead, the total degrees
of
freedom
in CFA are a
function
of the
number
of the
unique entries
in
the covariance matrix, which is a function of the number of measured
variables (v) in the analysis. The total degrees of freedom for
these
tests
equals [ < v v + 1)] / 2). For example, in a study involving i> = 9 measured
variables,
th e
available degrees
of
freedom equal
([9 (9 + 1)] / 2) =
([9 (10)]
/ 2) = 45. This
corresponds
to
evaluating
9
variances
and 36
covariances (i.e., there are 9 + 36 = 45 unique entries in the associated
variance/covariance matrix).
CONFIRMATORY
FACTOR ANA LYSIS DECISION
SEQUENCE 117
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 115/185
In CFA (and SEM), each parameter e.g., factor pat tern coefficient,
model error variance,
factor
co rrelation) that is estimated
"costs"
one degree
of
freedom.
This
means that each a nd every model involving nine measured
variables and
est imation
of
an y
4 5
parameters will always perfectly reproduce
or
fit the
data.
The disconfirmable and/ or plausible models for a CFA are context-
specific, and
some such models
can
only
be
developed
by
consulting theory
and related em piric al research. How ever, several rival CF A models may
often be useful
across
a fairly
wide array
of
problems:
1 . Independence model. This
model
specifies
that
the
measured
variables
are all perfectly uncorrelated with each other, and
thus
tha t
no
factors
are
present.
The
model estimates only
the mo del specifica tion varianc es of each of the meas ured
variables.
This model
is
d isconf irmable
but usually not plausi-
ble.
H owever, the stat ist ics qua ntifying the fi t of this model
are
very useful
in
serving
a s a
baseline
fo r
evaluat ing
the fits
of plausible rival models.
2. One-factor model.
This model
is
disconfirmable,
but usually
not p lausib le for most research situations. Aga in, how ever,
even
if
this m odel
is not
theoretically plausible,
the fit
stat ist ics
from this model test can be
useful
in cha racte rizing the degree
of
sup eriori ty
of a rival
mul t i fac tor model .
And it is
desirable
to
rule
out the fit of this model even when m ultifa ctor models
are
preferred,
so
that support
for the
mul t i factor model
is
then
stronger.
3.
Uncorrelated
factors model. In EFA, as noted in chapter 3,
orthogonal rotat ions
are the
default
an d usually
provide simple
structure.
However, in CFA, correlated factors are usually
expected
and
almost
always
provide
a
better
fit to the
data.
However , it is
helpful
to quantify th e degree of superiority of
model
fit for a
correlated versus uncorrelated factors model.
Estimating
the factor correlations for /fac tors costs (/ [f - 1]) /
2 degrees of freedom, for exam ple, (3 [3 - 1] ) / 2 = (3 [2] ) /
2 = 3). Uncorrelated factor models ar e more parsimonious
when there are two or more
factors.
The quest ion is whether
the sacrifice of
pars imony (and
of
related
disconfirmability) is
worthwhile in service of securing a better model fit .
Model Identification
A
model
is said to be
"identified" when,
for a
given research problem
and data set,
sufficient
constraints are imposed such that there is a single
118
EXPLORATORY
AN D
CON FIRMATORY
FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 116/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 117/185
apples to apples as to how well each m easured varia ble reflects th e underly-
in g
construct.
The
selection
of
scaling
for the
latent variables
will not affect th e
model fit statistics. The scaling decisions are basically arbi trary and can
reasonably be made as a function of the researcher's s tyli stic preferences.
These are not the only conditions necessary to achieve model identifi-
cation.
When a single measured varia ble is posited to reflect multip le un de rly-
in g
factors, evaluating model identification becomes technically
difficult.
Bollen (1989)
and
Davis ( 1993 ) provided more details.
Fortunately,
modern computer software
is
relatively sophisticated
at
dete cting model und erid ent ific atio n. Most packages (e.g., AM OS) provide
diagnostics such
as
suggestions
of how
m any more param eters need
to be
"fixed" or constrained , and some guidan ce as to wh ich para me ters in particu-
lar
are good candidates for achieving model identification.
Matrix of Association Coefficients to Analyze
As noted in chapter 3, in EFA there are num ero us choices that can
be made regarding the matr ix of association coefficients to be analyzed.
Popular EFA
choices include
th e
Pearson
produc t— moment
r matr ix ,
th e
Spearman's
rho
matr ix ,
and the
variance/covariance matrix.
A s
noted
in
chapter 3, the
default
choice in statistical packages is the Pearson r matr ix ,
and
th is
is the
selection used
in the
vast preponderance
of
published
EFA
studies.
The
statistical
software
for CFA
supports
an
even larger,
dizzying
array
of
choices.
These
include coefficients with which many readers
may be
unfamiliar, such a s the po lyserial or the polychoric cor rela tion coefficients .
But
th e default choice in most CFA software is the covariance m atrix .
This
matr ix
has
covariances
off the
main diagonal. Remember that
T X Y
=
COVxv / [(SD
X
) (SD
Y
)], and that therefore COV
XY
= r
X Y
(SD
X
) (SD
Y
). The
correlation
is a
"standardized" covariance from which
the two
standard
deviations have been removed
by
div is ion .
The variance/covariance matrix has variances on the main diagonal.
Because the covariance of a var iable with itself equals the var iance of the
variable, it is
more common
to
hear researchers refer
to
th is matrix more
simply as the covariance matrix.
The computation of
both
th e Pearson r and the covariance matrices
presumes that th e data are inte rval ly scaled. This means that the researcher
believes that
th e
underlying measurement tool yielding
the
scores
has
equal
units
throughout. For example, on a ruler every inch (e.g.,
from
1" to 2",
from
11" to
12")
is the
same amount
of
distance
(or
would
be if the
ruler
were
perfect). Note tha t the a ssumption is that th e ruler has equa l intervals,
and not that the scores of different people are all eq uid ist an t.
120
EXPLORATORY
A N D
CONFIRMATORY FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 118/185
In psychology sometimes the interval scaling of the measurement is
more debatable when scores on abstract constructs are being estimated. For
example,
if
Likert scales
are
used
to
measure attitudes,
and 1 = never,
2 =
often, 3 = very
often,
and 4 = always, most researchers would doubt that
th e
data were intervally (continuously) scaled. However,
if 1 =
strongly
disagree, 2 = disagree, .3 = agree, and 4 = strongly agree, many (not all)
researchers would deem the scores intervally scaled, or at least approximately
intervally scaled.
Of course, whenever
there
is some doubt regarding the scaling of data,
or regarding the
selection
of matrix of associations to analyze, it is thoughtful
practice
to use
several reasonable choices reflecting
different
premises.
When
factors are invariant across analytic decisions, the researcher can vest greater
confidence in a view
that
results are not methodological artifacts.
In CFA, tor some complex statistical reasons, it is almost always prefera-
ble
to use the
coyariance matrix, rather
than the
Pearson
r
matrix,
as the
matr ix to be
analyzed. There
are
some conditions
in
which
the
Pearson
p r o du c t— m o m e n t matrix may he correctly used (see Cudeck, 1989). How-
ever, the covariance matrix will always yield correct results, as long as the
data are appropriately scaled and distributed.
Multivariate
Normality
As
noted
in
chapter
7, the
shapes
of the
data distributions being
correlated
affect the
matrices
of
association actually analyzed
in
EFA. Tradi-
tionally, considerable
attention
is paid to
these
issues in
Q-technique
EFA
applications, while less
attention
is focused on these
concerns
in R-technique
EFA applications, even when the Pearson product-moment r matrix is the
basis for analysis.
However, many CFA results are particularly affected by data distribu-
tion shapes. Certain statistical estimation theories presume multivar iate
normality, especially with regard to the standard errors for the estimates of
model parameters.
And
some
CFA
(and SEM) model
fit
statistics also
presume normality.
Even univariate normality is not as straightforward a s many researchers
erroneously believe. Misconceptions regarding univariate normality occur,
in part, because many textbooks only present a single picture of a univariate
normal distribution:
z
scores
that
are normally distributed (i.e., the "standard
normal" distribution).
There is only one univariate normal distribution for a given mean and
a given standard deviation. But for every different
mean
or every different
standard deviation, the appearance of the normal distr ibution changes.
Some
normal distributions when plotted in a histogram have the classic bell shape
that many associate with the normal or Gaussian curve.
C O N F I R M A T O R Y FACTOR
ANALYSIS
DECISION SEQUENCE 121
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 119/185
But bells come
in
infinitely many shapes. Some bells appear fairly flat
and wide, whereas others appear tall and narrow. Of course, one universally
true statement about normal dis tribution s with different appearances is that
they all are unimodal an d symmetrical (i.e., coefficient of skewness = 0)
and
have
a
coefficient
of
kurtosis
of
zero.
Bivariate
normality is
even m ore complica ted. Because there
are now
scores on two va riables for each person, a three-dimen sional space is requ ired
to represent th e distribution, an d three-dimensional objects must be located
in this space to represent the scores of different people. The objects used
might
be
little
BBs or
ballbearings,
one for
each person.
An X-axis and a perpendicular Y-axis might be drawn on a piece of
paper. A
vertical axis extending through
the X,Y
intercept might
be
used
to measure the frequencies at different score pairs. For example, if
C olleen
and Deborah both had scores of 1,1, their BBs would be stacked on top of
each
other
at
this particular score pair coordinate.
If
both
variables
are
z scores
and
normally distributed,
the BB s in
bivari-
ate data w ould create a literal three-dim ensio nal b ell shape. If we traced around
th e outer edges of the BB s on the paper, a circle wou ld be drawn on the paper.
If
we made a slice through the BBs anywhere up the frequenc y axis, parallel
to but moving away
from
the piece of paper, concentric and n arrow ing circles
would be
defined un til
we
reached
the top of the
bell.
If we
sliced th rou gh
the
BBs with a
knife,
in any location only with the restriction that the knife cut
al l
the way
down through
th e
vertical frequenc y axis, every slice would
yield
a univariate standard normal distribution.
If
the Y scores were more spread out than the X scores, the signa ture
or foot of the bell defined by trac ing on the paper a round the BB s wou ld
be an ellipse rather than a circle. If we sliced upward on the vertical axis,
but par allel to the paper, every slice would yield a concentric ellipse until
we reached the top of the bell. If we sliced throug h the BB s w ith a kn ife,
in any
location only with
t he
restriction
that th e
knife
cut all the w ay
down
through
th e
vertical frequency axis, every slice would yield
a
univariate
normal distr ibu tion , albeit potentially slices that were narrower or wider
than each other.
Univariate normality of both variables is a necessary but not sufficient
condition
fo r
b ivariate normality. B ivariate normality
of all
possible pairs
of
variables
is a
necessary
but not
sufficient condition
fo r
mult ivariate normali ty.
Of course, multivariate normality requires that more than three dimensions
are conceptualized. These are not easy dynam ics to picture or to understand.
Henson (1999) provided a very
useful
discussion of these concepts.
There a re primari ly tw o ways
that
researchers can test th e multivariate
normality of their data. First, if more than 25 or so measured variables are
involved, the grap hical techniqu e (not involving a statistical significance
test) described
by Thompson
(1990a)
can be
used.
J
2 2
EXPLORATORY
AND
CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 120/185
Second,
a statistical test of the mu ltiva riate c oefficient of kurto sis,
using the null hypothesis that kurtosis equals
zero,
is available in com monly
used CFA/SEM computer programs. These computer programs also provide
a "critical ratio" t
or
Wald statistic)
of
this p aram eter estimate
to its
standard
error.
Researchers
hope that
th e mu ltivaria te kurtosis
coefficient
is close to
zero, and that the null hypothesis is not r ejec ted.
What can he
done
if the
assumptions
of
m ultivaria te normality cannot
he met? One choice is to use a parameter estimation theory, described in
a subsequent section of this chapter, that does not require this assumption.
H owever, "distribution-free" estimation theory does require even larger sam -
ple
sizes than
do
more commonly used theories.
A second alter nat ive is to create item "parcels" by bund ling together
measured variables.
For
example,
if
eight measured variables
are
presumed
to measure Factor I, the data can be reexpressed as four measured variables.
For example,
each
person's score on the measured variable
that
is most
skewed left is
added
to the
person's score
on the
measured variable that
is
most skewed right,
the
score
on the
measured variable that
is
second-most
skewed left is
added
to the
score
on the
measured variable
that is
second-
most skewed right, and so forth. However, this strategy is subject to the
general limitations
of
parc eling methods (cf. Bandalos
&
Finney, 2001;
Nasser
&
Wisenbaker, 2003).
Outlier Screening
It is well-known that
outliers
can
grossly
influence statistical results for
a data set, even if the number of outliers is small and the sample size is
large.
Outliers can influence CFA results (see Bollen, 1987)
just
as they
can other analyses.
Yet misconceptions regarding outliers abound. Some researchers incor-
rectly believe that outliers have anomalous scores on all possible variables
with respect
to all
p ossible statistics.
One can
almost visualize evil, aberrant
people whose foreheads are (or should be) branded with large capital Os.
Outliers are persons with aberrant or anomalous scores on one or more
variables
with respect
to one or
more statistics.
But
most outliers
are not
outliers
on all variables o r w ith respect to all statistics. Consider th e scores
of
Tom, Dick,
and
Harry
on the
variables
X and
Y :
Person
X Y
Tom 1 2
Dick
2 4
Harry
499 998
CONFIRMATORY FACTOR ANALYSIS DECISION
SEQUENCE 1 23
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 121/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 122/185
TABLE
10.5
Covariance
and Correlation Matrices for Eight
Measured Variables
n= 100
Faculty)
Measured variable
Variable
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
PER1
1.847
1.768
1.340
1.681
0.921
0.962
0.986
1.344
PER2
0.844
2.374
1.278
1.787
0.816
0.797
0.835
1.468
PER3
0.653
0.549
2.280
1.488
0.992
1.026
1.132
1.176
PER4
0.825
0.774
0.657
2.248
1.043
1.019
0.967
1.338
PER5
0.328
0.256
0.318
0.336
4.274
3.543
3.375
3.322
PER6
0.349
0.255
0.335
0.335
0.844
4.122
3.231
3.150
PER7
0.388
0.289
0.400
0.344
0.872
0.850
3.506
2.935
PER8
0.461
0.444
0.363
0.416
0.749
0.723
0.730
4.605
Note. Covariances are not italicized; Pearson
rs
i.e.,
standardized covariances) are italicized. Pearson
r
C O V x v
/
SDx
x SCM
e .g. , 0.844
=
1.768
/
1.847
5
x
2.374
5
)). COV
XY
=
r
m
x SD
*
x S
f <
e
9 >
1
-
768
=
0.844
x
1.847
s
x
2.374
s
).
distances
can
also
he
obtained within SPSS
by
performing
a
regression
listing the measured variables as the predictor variables in a mult iple regres-
sion analysis and using any other variable as the dependent variable. For
this
example,
if
participant
ID
number
was
used
as the
pseudo-dependent
variable, the SPSS
syntax would
be:
regression
variables=id perl
to
per8/
descriptive=mean stddev corr/
dependent = id/eiiter perl to per8/
save=mahal mahal)
sort cases by mahal D) .
print formats mahal F8.4)
list variables=id perl to per8 mahal/cases=999
However, even when we know that a given person is an o u t l i e r on a
given set of variables in regard to
p a r t i c u l a r
s ta t is t ic s , it is not
a lways
clear
what should be
done
with the person's scores. If the researcher interviews
the outlier and determines that b izar re scores were intentionally generated,
then
deleting
the
person's scores
from
the
data
set
seems only reasonable.
But when
we
cannot
determine
the
reason
for the
score aberrance,
it
seems
less reasonable
to
delete scores only because
we are
unhappy about their
atypicality
with respect to deviations
f rom
the means.
One resolution refuses to represent all analytic
choices
as being either /
or. The analysis might be conducted both with the outlier candidates in the
data and with
them
omitted. It interpretations are robust across var ious
decisions,
at
least
the
researcher
can be
confident
that the resu l ts are not
artifacts
of
methodological choices.
C:ONFIRMATORY
FACTOR
ANALYSIS DECISION SEQUENCE
125
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 123/185
It is
also critica l
to
take
into account th e
analytic context when making
these decisions. The context is very different if one is doing research in a
previously
unexplored area versus an arena in which n ume rous studies have
been conducted. When
other
studies
have
been
done
th e researcher can
feel
more comfortable
that
outliers are not c omp romising gene ralizab ility,
i f results replicate across stud ies. By the same
token,
researchers may be
somew hat less concerned about outliers when they
are
also conducting
interna l replicability analyses.
A s Cohen, Cohen, Wes t , and Aiken (2003) argued,
The
extent
of
sc ru t iny
of these
statistics
depends on the
nature
of the
study. If the s tudy is one of a series of replications, inspection of graphical
displays for any
obvious
outl iers
is
normally
sufficient. Consistency of
the findings across replications provides assurance that the presence of an
outlier is not responsible for the results
[italics added]. On the
other hand,
i f
the
data
set is
un ique
and
un l ike ly
to he
replica ted (e.g.,
a
s tudy
of
40 ind iv idua ls wi th a rare
medical
disorder), very careful scrutiny of
th e
data
is in
order, (pp. 410-411)
Parameter stimation
Theory
Conventional graduate courses in applied statistics teach ana lyses such
as
t
tests, analysis
of
variance, multiple regression, descriptive discriminant
analysis,
and cano nical correlation analysis. All of these an alyses use sample
data to compute sample statistics (e.g., sample means, correlation coeffi-
cients) that
are
then taken
as
estimates
or
corresponding population param e-
ters
(e.g.,
p., a, p).
All of
these analyses invoke
a
single statistical theory
fo r
parameter
estimation called
ordinary
least squares
(OLS).
These estimates seek to max-
imize
the
sum-of-squaresMonEL,
or its
mu ltivariate equivalent,
and the
corres-
ponding r
l
effect size analog (e.g., R
2
, r\
2
) in the sample data.
However,
there
are (many) other parameter estimation theories that
can also be used. Indeed, it can be argued
that
classical statistical analyses
are not
very
useful, and may
even routinely lead
to
incorrect research
conclusions, because classic al methods are so hea vily influe nced by outlie rs
(see
Wikox, 1998).
Two alternative parameter estimation theories are
fairly
commonly
used
in CFA (and SEM). Hayduk (1987) and Joreskog and Sorbom
(1993)
provided more detail on these and other
choices.
When sample size is large,
and if the model is correc tly specified, the va rious estimatio n theories
tend
to
yield consistent
and
reasonably accurate estimates
of
population parame-
ters (e.g., population factor pattern coefficients).
First, maximum likelihood theory is the de fau lt parameter estimation
theory in most statistical packages. This estimation theory attempts to
/ 2 6
EXPLORATORY
AND CONFIRMATORY FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 124/185
estimate th e
true population parameters,
and not to
reproduce
th e
sample
data or minimize the sample
s u m - o f - s q u a r e s u N E x r L A i N E n
while maxim izing
th e
related sample
effect
size (e.g.,
R ',
Tp).
This is
accomplished
by
applying
maximum
likelihood
theory
to the
sample
covariance
matrix
to estimate
the population covariance ma trix, 2
V
x v, and then derivin g factors to repro-
duce that matrix rather than the sample covariance matrix.
Obviously, estimating population parameters with a statistical theory
that optim izes this estimation (rather than features
of the
sample data)
is
very
appealing. Better estimation of population parameters should yield
more replicab le results. This e stima tion theory also has the appeal
that
most
of
th e
statistics
that
evaluate model
fit are
appropriate
to use
when this
theory has
been
invoked. H owev er, m ax im um likelihood estimation theory
does require that
th e
data have
at
least
an
approximately mu ltivaria te
normal distribution.
Second, asymptotically
distribution-free
(AD F) estimation theory may
be invoked. As the name of this theory imp lies, these estimates tend to he
accurate
p aram eter estima tes regardless of distribution shapes as the sam ple
size is increasingly large. This is an appealing featu re of the theory. H owe ver,
very
large sample sizes
of
more th an 1,000 cases
are
required
to
invoke this
theory for some analytic problems.
Of course, even CFA m ax im um likelihood estima tion often requires
larger sample
sizes than
would EFA modeling for the same data set (MacC al-
l u m ,
Browne, (StSugawara,
1996). W est,
Finch,
and
Curran
(1995)
reviewed
some of the relevant issues and choices regarding th e distributional assump-
tions of variou s para me ter estimation theories.
Jackson (2001) found
that
sample
size per se, and not the
number
of
parameters being estimated,
is the
critical ingredient
in
deriving accurate
CFA estimates. And sample size is less of a
concern
when the average
pattern
coefficient is
|.80|
or
greater.
POSTANALYS1S DECISIONS
Once
the analysis is in hand, various postanalysis decisions arise. First,
the goodness of fit tor
each
rival model must be evaluated.
Second, specific
errors
in
model specification must
be
considered.
Model Fit Statistics
A unique feature of CFA is the capacity to quantify the degree of
model fit to the data. However, there are dozens of fit
statistics,
and the
properties of the fit
statistics
are not
com plete ly understood.
As
Arbuckle
and W othke
(1999) observed, "Model evalu ation
is one of the
most unsettled
CONFIRMATORY FACTOR
ANALYSIS DECISION SEQUENCE
\ 27
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 125/185
and
difficult
issues connected with s tru ctu ral modeling [and CFA]. . . .
Dozens of statistics . . . have heen proposed as measures of the mer it of a
model" (p. 395).
Classically,
Monte
Carlo
or simulation studies are used to explore th e
behaviors of statistics under variou s condition s (e.g., differe nt sample
sizes,
various
degrees of violations of distributio nal assumptions). In Monte Carlo
research, populations with known parameters
are
created (e.g., with
an
exactly known degree
of
deviation from normal i ty) ,
and
numerous random
samples (e.g.,
1,000,
5,000) from
these
populations
are
drawn
to
evaluate
how the statistics perform.
M any of the pre vious sim ulatio n studies of
CFA/SEM
model fit statistic s
specified
populations associated with known models,
and then
drew samples
and tested the fit of the known models. Such simulations studies are not
realistic
as regards real research situa tions , because in real research the
investigator does not usually know the true model. Indeed, if we knew the
true model,
we
would
not be
doing
th e
modeling study
Better simulation studies of fit statistics have created populations fo r
certain m odels, and
then
intentionally evaluated the behavior of
fit
statistics
fo r va rious conditions and v ariou s model m isspecifications. Recent s imula-
tion stud ies in this venue have pro vided im portant insights re garding how
the fit statistics perform (cf. H u & Rentier, 1999; Fan,
Thompson,
& Wang,
1999; Fan,
Wang,
& Thompson, 1997), but there is still a lot to learn
(Rentier,
1994).
Four
fi t
statistics
are
most commonly used today.
In
general,
it is
best
to consult several fi t statistics when evaluating model fit and to evaluate
whether th e results corroborate each other.
First, the
Jj statistical
significance
test
has
been
applied since the origins
of CFA. As noted in chapter 2, in EFA the reproduced correlation matrix
( R v x V
+
)
c a n
°e computed by mul t ip lying P V X F by Ppxv ' - T he residual
correlation matrix (Ry
x
v )
l s
computed by subtract ing R y
x
v
+
from R y x X -
The
same basic idea
can be
applied
in
CFA.
One can
estimate
how
much of a covariance matrix is reproduced by the parameter estimates, and
how m uch is not. In CFA, the n ul l hypothesis t ha t is tested is essentially
whether th e residua l covariance estim ate equals a matrix consisting only
of
zeroes.
Un like most sub stantive applications of null hypothesis statistical sig-
nificance
testing,
in CFA one
does
not
w an t
to
reject this null hypothesis
(i.e., does not wa nt statistical sign ifica nce ), at least for the preferred model.
The degrees of freedom for this test statistic are a function of the number
of measured variables and of the number of estimated param eters. For exam-
ple, if there are
eight m easur ed variables,
the
d/Toj^i.
»
re
36 ([8 / (8 + 1)] /
2). If eight factor pat ter n coeffic ients, eight error variances, and on e factor
1 28 EXPLORATORY A N D CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 126/185
covariance are estimated, the degrees-of-freedorriMODEi equal 19 (i.e., 36
17).
How ever, as with conventional statistical significa nce tests, though
the sample size is not used in computing the degrees of freedom for the test,
the sample size is used in comput ing the %
2
. Even for the same input
covariance matrix and the same degree of model fi t , Jj is larger for a larger
sample size. This makes this
fi t
statistic
not
very useful when sample size
is
large, as it
usually should
he in
CFA, because with large sample sizes
al l
models will tend to be rejected as not fitting. A s B entler and
B onnett
(1980)
observed, "in ve ry large samples vir tua lly all models that one might consider
would
have to be rejected as untenable. . . .
This
procedure
cannot
generally
be justified,
since
the
chi-square value
. . . can be
made small
by
simply
reducing sample size" (p. 591).
The % test is not very useful in evaluat ing the fit of a single model.
But the test is very
useful
in evaluating the comparative f i t s of nested models.
Models are nested when a second model is the same as the first model,
except that one or more additional parameters (e.g., factor covariances) are
estimated or freed in the second model.
The % "s and
their degrees
of
freedom
are
addi t ive.
For
example,
if for
the first
model positing
tw o
uncorrelated factors
the %
2
is
49.01 df = 2 0 ) ,
and for the second model estimating th e same parameters except that th e
factor
covariances
are freed to be
estimated
the yj is
31.36 df
=
19),
then
the differential fit of the two
models
can be
statistically tested using
%
2
=
17.65 (i.e., 49.01 - 31.36) with df = 1 (20 - 19). The ^ C A L C U L A T E D of this
difference
can be
determined using
th e
Excel spreadsheet function
= c h i d i s t ( 1 7 . 6 5 , 1 ) ,
which here yields f) = .000027.
When
ma ximu m likelihood estimation theory is used, and the mu ltiva r-
iate normality assumption
is not
perfectly met,
the J j
statistic will
be
biased.
H owe ver, Satorra and B entler (1994) developed a correction for this bias
that may be
invoked
in
such cases.
Second,
th e
normedfit index (NFI; Bentler
&
B onn ett, 1980) compares
the / for the tested model against the % for the baseline model presuming
that the measured variables are completely independent. For exam ple, if
the %
2
for the tested C FA model was 31.36, and th e %
2
for the baseline
independence model pr esum ing the measured variables were perfectly uncor-
related
was
724.51,
the NFI
would
be
(724.51
-
31,37)
/
724-51
=
693.15
/
724.51
= 0.96. The NFI can be as
high
as
1.0.
Generally, models with NFIs
of
.95 or more may be deemed to fit reasonably well.
CONFIRM ATORY FACTOR ANALY SIS DECISION SEQUENC E 1 29
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 127/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 128/185
When
maxim um likelihood estimation theory is used, and the multivar-
iate normality assumption is not
perfectly
met, the standard errors for the
parameter estimates
will
also
be
inaccurate. Satorra
and
Rentier (1994)
proposed methods fo r estimating th e correct standard errors, an d conse-
quently correct critical ratios, using
these
standard errors as denominators,
by taking into account data distribution shapes. Fouladi (2000) found that
these
corrections work reasonably we ll if sample
sizes
are 250 or more. Most
modern software provides these corrections as they are needed.
Parameters Erroneously
Fixed
For
each
parameter
fixed to not be
estimated that could
be
estimated,
th e CFA/SEM
software
can
estimate
th e
decrease (improvement)
in the
model fit
jj
resulting from
freeing
a given previously fixed parameter instead
to be estimated. In some packages (e.g., A MO S )
these modification
indices
are "lower-bound" or conservative estimates, and if a fixed pa ram eter is
freed the ~ f c reduction predicted by the software m ay be better than predicted.
Obviously, fixed parameters with the largest m odific ation indices are among
th e
leading candidates
fo r
model respecification.
Model
Specification
Searches
Confirmatory factor analysis is intended to be, as the name implies,
confirmatory. Respecifying
a CFA
model based
on
consultation
of
crit ical
ratio and modification index statistics is a dicey business, i f the same sample
is
being used
to
generate
these
statistics
an d
then
to
test
the fit of the
respecified model.
Using
th e
same sample
to
respecify
th e
model,
and then
test
th e
respecified model, increases capitalization on sampling error variance an d
decreases the replicability of results. And using the same sample in this
manner also turns
th e
analysis
into an
explanatory one, albeit using
CFA
algorithms.
As
Thompson (2 00 0c) emphasized, model respec ification should
never
"be based on b lind, dust-bow l empiricism . Models should only be respecified
in
those
cases in which th e researcher can articulate a persuasive, rationa le
as
to why the
modification
is
theoretically
and
practically defensible"
(p .
272) . Model respecification
is
reasonable when
each change can be
justified
theoretically, and when the fit of the respecified model can be evaluated
in a
holdout
or
independent sample.
MAJOR CONCEPTS
Exploratory factor analysis and CFA are part of the same general linear
model
(Bagozzi
et al., 1981). However, CFA models can estimate some
CONFIRMATORY
FACTOR ANALYSIS DECISION SEQUENCE
13 \
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 129/185
parameters that
cannot
he evaluated within EFA. For exam ple, only in CFA
can the covariances of error variances be nonzero and estimated. In EFA
factors can he
either correlated
or
uncorrelated,
hut in CFA
addit ionally
some factors may be allowed to he correlated whereas
other
pairs of factors
in
th e
same model
are
constrained
to he
uncorrelated.
O r,
uniquely
in
CFA,
factor correlations can he
freed
hut constrained to be equal. And in CFA,
unlike
EFA, all the pat tern coeff icients for a factor can he freed to be
estimated
but
constrained
to be
equal.
Confirmatory factor
analysis
m odels req uire the specification of one
or more hypothesized factor structures, and the analyses directly test and
quantify
the
degrees
of fit of rival
models. Because more
than on e
model
may
fit in a given
study,
it is important
to
test the fit of multiple plausible
models, so
that
any case for fit of a given model is more persuasive.
Only
"identified" factor models
can be
tested
in
CFA.
In
CFA, degrees
of freedom are a function of number of measured variables and not of sample
size. A
necessary
hut not sufficient
condition
fo r
model identification
is
that
th e
number
of
estimated parameters (e.g., pattern coefficients, factor
correlations, error varia nce s) cannot exceed th e ava ilable degrees of freedom.
Ideally,
considerably fewer parameters
will be
estimated
than
there
are
degrees
of
freedom,
so
that
the
model
is
more
disconfirmable. Fit
when more
degrees of freedom are unspent is inherently more impressive.
Some
factor
pattern coefficients or factor variances must also he constrained to obtain
model identification.
Like
EFA,
CFA
requires various decisions, such
as
what matr ix
of
associations (e.g., covariance m atr ix)
to
analyze,
an d
what estimation theory
(e.g., maximum likelihood) to use.
Some
estimation theories require that
data are m ultivariate normal. And screening data fo r outliers can be part icu-
larly
important when v ery elegant methods, such
a s
CFA,
are
brought
to
bear.
Confirmatory factor analysis model tests yield various statistics that
quantify degree of model fit. It is generally best practice to consult an array
of
f it
statistics (Fan
et
al., 1999;
H u &
Rentier, 1999)
in the
hope that they
will
confirm
each other.
The
analysis also yields criti cal ratios
that
evaluate whether freed
parameters were indeed necessary, and modification indices that indicate
whether fixed parameters instead should have been estimated. Respe cifying
models based
on
such results
is a
somewhat dicey proposition, unless theory
can be
cited
to j u s t i f y
model revisions.
Such
"specification searches" are,
however, completely reasonable and appropriate when th e respecified model
is
tested with
a
new, independent sample.
132 EXPLORATORY A N D CONFIRMATORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 130/185
1 1
SOME CONFIRMATORY
FACTOR
ANALYS IS
INTERPRETATION PRINCIPLES
Many princ iples of conf i rmatory fac tor analys is (CFA) mirror
those
in explora tory fac tor analys is (EFA ; e .g ., the impo rtance of con sul t ing both
pa t t e rn and s t ru c tu r e coe f f i c ien t s ) , g ive n the re la t ionsh ips o f these tw o
analyses wi thin
th e
genera l l inear model (GLM). However ,
there are
also
aspects
of CFA t h a t are
u n i q u e
to CFA
model tes ts .
For
e x a m p l e ,
in EFA we
a lmos t a lways
use
on ly s t anda rd ized f ac tor
pa t te rn
coef f ic ien ts . How ever , in CFA both uns tanda rd ized pa t t e rn coeffi-
c ien t s (ana logo us to regress ion uns tan da rd ized b we igh t s ) and s t anda rd ized
pat tern coef f ic ients (analogous to regress ion s tand ardized (3 we ights ) a re
c o m p u t e d .
The s ta t is t ic a l s ignif ican ce tests of
i n d i v i d u a l
C F A p a t t e r n
coeff ic ients
are conducted us ing the uns tandardized coef f ic ients . These coeff icients a re
c o m p u t e d
b y
d i v i d i n g
a
g iven pa rame te r
by i ts
re spec t ive st anda rd e r ror
to
compute a t (or Wald or c r i t ica l ra t io) s ta t is t ic . Noteworthy uns tandardized
coefficients us ing
these
tests also indicate the noteworthiness of the corres-
pond ing s t anda rd ized pa rame te rs , a s in mul t ip le reg res s ion .
H o w e v e r , w h e n p a r a m e t e r s a re compared w ith each
o ther ,
usua l ly th e
s t andard ized
coef f ic ients a re reported, again as in m ul t ip le reg res sion . T he
standa rdization allows apples- to-apples
comparisons of related parameters
with in a single model.
133
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 131/185
Five aspects of CFA p rac t ice are elaborated here . These inc lude (a)
a l t e r n a t i v e
wa y s
to
create
ident if ied
models ,
(b )
some es t imat ion theory
choices,
(c) the
impor tance
of
consu l t ing
b o th
p a t t e r n
and
s t ruc tu re coeffi-
cients when C FA factors are cor rela ted , (d) com par ing
f i t s
of nes ted models ,
and (e) ex tract ion o t h igher -order factors .
In
these i l lus tra t iv e analyses , th e Append ix A da ta p rov ided by the
100 facu l ty m e m b e r s for the f i r s t eight i tems are used . Readers may f ind i t
useful
to conduct paral le l analyses us ing the append ed data provided by the
100 g radua te s tuden ts .
MODEL IDENTIFICATION
CHOICES
A s
explained
in
chap ter
1 0 ,
parameters cannot
b e
est imated unless
th e
m o d e l
is
" iden t i f ied . "
A
m o d e l
is
u n id e n t i f i e d
o r
under iden t i f i ed
if for
t h a t model an inf in i te num ber o f es t imates fo r the paramete r s a r e equa l ly
p laus ib l e .
A necessary bu t insuff ic ient condit ion for ide nt if ic at io n is that the
number o f paramete r s be ing es t imated mus t be equa l to ( j u s t ident i f ied) or
less t han
(ove r iden t i f ied ) the ava i lab le deg rees o f f r eedom. In th i s exam ple
the
n u m b e r
of
m e a s u r e d v a r i a b l e s
is
eight . T herefore ,
the
ava i lab le deg rees
of f reedom
are 36 ([8 x (8 + 1)] / 2).
T h e
models tes ted here presum e that i tems PE R I through PER4
reflect
only
an
under ly ing f ac to r named "Service Af fec t"
and
tha t i t em s PER5
through PER8 reflect only an u nde r ly in g factor named "Lib rary as Place."
Thus ,
a total of up to
eight factor pat tern coefficients
are
es t imated . Also,
th e er rors var iances for
each
of the e igh t measu red var iab les are es t imated .
I t is assumed here tha t the factors are unc orrela ted .
H o w e v e r ,
for
th is model
w e
c a n n o t es t im ate these
16 (8 + 8)
paramete r s
because th e scales of the two la ten t factors have no t
been
indicated . These
scales
of
m e a s u r e m e n t
of
la ten t constructs
are
a r b i t r a r y . T h a t
is,
there
is
no
in tr ins ic reason
to
th ink tha t "L ib ra ry
as
Place" scores
in the
p o p u l a t i o n
should have
a
p ar t icu la r va r iance
or
s tandard dev ia t ion .
Bu t we
m us t dec la re
some scale for these late nt va riab les, or w ith
each
of
infini tely
m a n y
di f ferent
possible var iances there w ould
b e
inf in i te ly m any corresponding sets
of
o ther
parameters being es t imated .
There are two
bas ic ways
to
iden t i fy
C FA
models.
First , for a
g iven
fac tor we can "f ix" or constrain
any
g iven un stand ard ize d factor pa t tern
coefficient
to be
a ny m a th e m a t i c a ll y p l a u s i b l e n u m b e r . H o w e v e r , usua l ly
the number 1 .0 is used for th is purpose. In doing so , we are scal ing the
var iance of the la ten t construct to be measured as some func t ion of the
measured
v a r i a b l e
fo r
w h ic h
we fix the
pat tern coeff ic ien t .
Some
researchers
134 E X P LOR A TOR Y A ND CONFIRMATO RY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 132/185
prefer to f ix the
pa t t e rn
coef f i c i en t for the measu red var i ab l e be l i eved
to be
mos t p sy chom et r i ca l ly r e l i ab l e . Ho wev er ,
the
se l ec t ion
is
bas i ca l ly
arb i t r a ry .
Second, w e c a n
es t imate
all the
pa t t e rn coef f i c i en t s
and f ix the
fac tor
var iances to any m athe m at i ca l ly p l aus ib l e o r "admiss i b l e" num b er .
Because v ar i ance s canno t be nega t ive , adm iss i b l e num be rs fo r var i an ce
es t imates
m ay no t be
n e g a t i v e .
And the
factor var iance
c a n n o t b e
zero ,
o r we wou ld be eva lua t ing the
inf luences
of a var ia b le tha t was no longer
a
v a r i a b l e .
When
factor var iances
are f ixed, usual ly th e
var i ance paramete r s
a re
f ixed to be
1 .0 .
This
approach
has a
s ingu la r advan tage :
When
we a re
es t imat ing fac to r covar i ances ,
th e
covar iances
of the
fac to r s wi th
each
o t h e r now
become fac to r co r re l a t ions , because g iven tha t
r\
x
n =
COV[
x
n /
(SDj
x SD j i ) , when the fac to r var i ances a re both 1.0, w e o b t ai n C O V ]
x l
j /
( 1 x 1 ) .
This
makes in t e rp re t a t ion more s t r a igh t fo rward .
In add i t ion ,
t
s ta t i s t ics a re o n l y o b t a i n e d fo r es t imated fac to r pa t t e rn
coefficients .
So i t i s help fu l to kno w the
s t a t i s t i c a l
s ign i f i cance , o r W ald
rat ios,
fo r
each factor
p a t t e r n
coef f i c i en t ,
a n d
these
w o u l d
no t be
a v a i l a b l e
i f th e
model were
ident i f ied b y
cons t r a in ing some fac to r pa t t e rn coef f i c ien t s .
T h e
approach also
has the
a d v a n t a g e t h a t
a l l the
pa t t e rn
coeff ic ients
fo r a
g iven factor
can be
com pared w i th each o ther , b ecause they were
a ll
freed. In
genera l ,
w e m a y
expec t
t h a t
pa t t e rn
coeff ic ients on a
g iven fac to r
wil l b e
s i m i l a r
in
m a g n i tu d e s .
Table
11.1 compares three analyses
o f
these
da t a wi th th ree
d if fe ren t
model i den t i f i ca t ion s t r a t eg ies hav ing been invoked . In Model 1
a rb i t r a r i l y
the
uns t and ard ized pa t t e rn
coeff ic ient for
v a r i a b l e P E R I
on
Factor
I
("Service
Affect") was f ixed to
equal 1.0,
and the
uns t andard ized pa t t e rn coef f i c i en t
tor var iab le PER8 on Factor I I ("L ib rary as Place") was f ixed to equ al 1 .0 .
T he
fac to r var i ances were
b o t h
"freed"
to be
es t imated .
In
Model
2
a rb i t r a r i ly
the
uns t andard ized pa t t e rn
coeff ic ient fo r
v a r i a b l e
P E R 2
on
Factor
I
( "Serv ice A f fec t " )
w as f ixed to
equal 1.0,
and the
u n s t a n -
dardized
pa t t e rn coef f i c i en t
fo r
v a r i a b l e P E R 5
on
Fac to r
II ( " L ib ra ry as
Place")
was f ixed to
equal 1.0.
T he
fac to r var i ances were
b o t h
"freed"
to
b e
es t imated .
In
M o d e l
3 ,
po r t r ayed g raph ica l ly
in
Figure 11.1 ,
a ll
e igh t pa t t e rn
coefficients
were es t imated . How ever ,
to identify th e
model ,
b o t h
fac to r
var iances were const rained
to be
1.0.
N o t e
f rom
T a b l e
11.1 that
th e
s t andard ized pa t t e rn coef f i c ien t s
and the
fac tor
s t ru ctu re coeff ic ien ts are ide nt ica l across the three m odels, reg ard less of
how the
model s were
iden t i f i ed . And the f i t
s t a t i s t i cs
a re
i d e n t i c a l .
S o m y
preferen ce for ident i fy ing CFA models by f ix ing factor var ianc es i s arb i t rary
bu t perm iss i b l e as a mat t e r o f s ty l i s t i c
choice .
SOME INTERPRETATION PRINCIPLES 135
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 133/185
C O
C O
Q)
2 _
x:
1 —
73
C D
CD
0
*
C
o
C O
O
?±I
^
D
~
—
5•*""
V
^
>1
0
0
C D LL
Q o
O '
r
T—
"tjj
|l
T
-C
uj 55 "o
_J •* — o
DQ
LL .c
p* (_j fl)
"~ ~
C _ ;
c o i
IE
E
m J
_
it *
C D
n|
O
O
'
CD
S-_
0
-I—
C O
.£
"co
CL
O
O
C O
C D
N
~O
C O
c
C O
5 5
CO
'C D
TJ
o
-
CM
*
C D
T3
o
^
2
*
CD
T3
0
C D
f5
5 5
c
C D
~
C O
Q.
C D
3
3
C O
c
C D
e
CO
Q.
C D
p
0
£
C
C D
to
Q.
~ s
.c
c t
(j
>
—
—
—
—
—
—
—
—
_
-
o
0 0
CO
t/5
o c
o c
0C
0C
0C*
in a
CT) a
0C
00
L O
a
a> a
d c
0C
0C
o c
0C
o o
in a
o a
0C
o
in g
2;
o c
0C
0C
o c
o o
m a
en
a
o c
1
II
I-
C
D C D C
LU LL
C L Q
3 C
3 C
3 C
3 C
J
(i
3 0
D C l
3 C
J C l
3
0
3 C
3 C
3
C
3 C
3 C
3 C
J C
3
0
3 C
3 C
j
u
C
is
3 C
3 C
3 C
3 C
J C
3 a
3 C
3 C
J
U
3 C
3 <i
) C
J O
: Q
J U
. Q
3 O •* CO CO CO
3 O
CO
O
CO
CT)
3 O O3 O3 CT) h~
D O O O O O
D r-- o o o o
D
r- o o o o
D
CO O O O O
3
O0000
• CO CO CO £?•
i CO O CO CT) S
| (33 O)CT)h~ P
ddddO
D r*- ^r* r^ o
• * —
c o o > C D
3 h - i i
S CMCM^l-COini-
3 C O | | R oo
^03(331-
3
O O •* OOO
3
O •* 00CO CO
3 O
CO
O
CO
CT)
3 O CTi
C33
O3 f .
3 O O O O O
3
r- o o o o
3 h~ O O O O
3 CO
O O O O
3
O0000
JT
CO
CO CO O
i S O CO
C33
CO
o-d do co
3t^ U3
h^Oi—
C O O 3 C 3 3
3 r ~ 1 • *
CMoa^coini-
}°°
1 00 CO
CM
0
?
0
?"":
3 O *- f O O O
3 O "^ 00 CO CO
3
O CO O CO O3
3 O C33 O3 O3 1
3
O O O O O
3 r- o o o o
3 K O O O O
3 000000
3 O O O O O
^ oo co
z?
ro
i CO O CO £? CT)
|
CT)
CT>
CT)
gj 00
dddo c\i
3T 1 1 O
•»—
COCT)CT)
3 r - i 1 C D C M C M ^ r c o i n i -
3C O | |C D 0- j
w
-C33C33i-
3 o i- •* d o d
s <
^ • ^ L O C O h - O O o ^
^
10:0:0:0:0: ^3 •- % ^
JUJLULULU LLJ JcJo
N
^ LLLL
_
Q_ Q_ Q_
Q_
Q_
i > x"O x2 OOC
0)
C/}
£
CO
CL
_ c
03
t:
o
CL
03
0)
to
C D
•a
o
"O
N
(C
TJ
c
_ c o
^
3
1
0
T-^
rO
.
13 c:
°E
"g o
x CL
3- CL
«
*- o
E
2
t o Q J
*u
0
Q) CD
a> g-
c <p
2
c
C O C O
p
Jl
w "
w
* * • *
-Si
w DZ
-D
?
1-g
O.S
C L —
Ii
if
s i
> |,
o 01
r u
-—
"D
"w
~
C D ~
ra
-a
E g
-5
^
Q) O
^ c
03
_
C
ll
|z
^1
o "°
s ffl
J 3 6 E X P L O R A T O R Y ' AND CONFIRMATORY FACTOR A N A L Y S I S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 134/185
e1
e 2 )
(
e 3 j ( e 4
Perl iPer2J P e r 3
Per4i
.
tt
Per5
T
J
/-"
-\
( e 5 )
Per6
V
i
i
' e 6 )
Per?
r
Per?
V T
A A
( e7 ) ( e8
Figure
11.1. C on f irmatory fac to r ana lys is fac to r model wi th e ight meas ured variables
and two
co r re la ted fac to rs wi th fac to r var iances cons t ra ined
to
equal 1.0. ServAf f
=
"S erv ice
Af f ec t " ; LibPlace = "Library as P lace. "
SOM E
I N T E R P R E T A T I O N
P R I N C I P L E S
137
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 135/185
D I F F E R E N T
ES TIM A TIO N TH EO R IES
As noted in chapter 10, classical s tatis t ic s (e.g. , analysis of variance,
regress ion, canonica l corre la t ion analys is ) invoke
a
s ta t is t ica l es t imat ion
theory called ordinary
(east
squares
(OLS).
However , the re are n u m e r o u s
o ther s ta t is t ica l es t imat ion theories
t h a t
c a n b e invoked to e s t ima te mode l
pa rame te rs . T he theory most commonly used in CFA (and in s t r u c t u r a l
equa t ion m odel ing [SEM]) is m axim um l ike l ihood theory. Most of the f it
s tatis t ics
have
heen developed for this theory.
How ever , max im um like l ihood e s t ima t ion does p re sume that data a re
mul t ivar i a t e n o r m a l l y d i s t r i b u t e d .
A n o t he r
es t imat ion theory, asymptot i -
cally d i s t r i b u t ion - f ree (AD F) theory , can b e invoked i f the d i s t r i bu t iona l
assumptions
o f
m a x i m u m li k e li h o o d
c a n n o t b e
me t . However ,
A D F
es t ima-
t ion
does presume large sample s ize.
For com para t ive purposes ,
Ta b l e
1 1.2 presents
the
s tandardized pat tern
and the s t ruc tu re coeff ic ients for the Mode l 3 f it to the facul ty da ta us ing
three different es t imat ion theories .
N o t e t ha t th e
pa t t e rn
an d
s t ruc tu re
coefficients es t imated in a l l three models are reasonably comparable .
These analyses were compared pr im ari ly for heuris t ic purposes . How -
ever, i t is not entirely a bad th ing to inv oke d if fe ren t analytic s trategies for
the
same da ta , j u s t
to
conf i rm t ha t results
are not
ar t i fac ts
o f
ana lyt ic choices .
In this ins tance , probably only the maximum l ike l ihood e s t ima tes would
b e
reported,
but the
researcher would sleep easier knowing
t ha t
s imilar
factors
emerged for the other analyses.
CONFIRMATORY
FACTOR
ANALYSIS PATTERN
A N D
S T R U C T U R E C O EF F IC IEN TS
In EFA, or thogonal solut ion s a lmost a lways provide s im ple s t ruc ture ,
are the defaul t models in s ta t is t ica l packages, and d o m i n a t e the l i t e ra tu re .
How ever , in CFA, i t i s more usua l to test either correlated factors models
or b o t h
uncorre la ted
and
correlated factors models .
That is , usua l ly b o t h
C FA models are reasonably plaus ib le , and we w a n t to tes t p laus ib le
r iva l
models
t o
have
the
strongest possible supp ort
for our
preferred model , presum -
in g that
this model u l t imate ly
has
best
fit .
When factors a re u ncor re la ted , the CFA s tandardized fac tor pa t te rn
and the s t ruc tu re coeff ic ients for g iven v ar iab les on given fac tors exact ly
equa l each other , as ref lected in the i l lus t ra t ive resul ts reported in
Ta b l e
11.2. Bu t wh en CFA factors are correlated , exa ctly as is the case in EFA ,
factor pa t t e rn and fac tor s t ruc ture coef f ic ients are no
longer equal.
For mul t ip le regress ion (Courvi l le &
T h o m p s o n ,
2001; Th o mps o n
& B orre llo , 1985) , desc r ipt ive disc r im inan t analys is (H ub erty , 1 994) , and
/
3 8
E X P L O R A T O R Y
A N D
C O N F I R M A T O R Y
FACTOR
A N A L Y S I S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 136/185
C O
C D
b
CD
H
c
o
C O
.E
"5 5
UJ
Q)
C D
il
-C
1
—
E
o
LL
C D
>
CD
'
Q
co
=tt:
C D ,
E
\
^ «=T
T — ^ O
^~" A 1
O LL
—
w S
C D
.0 0
^
C O I
ro c
C O
"-"
^
LL
C
C O
tn
"c
CD
C D
o
O
H _
, .
o
2
55
£
03
•+-rf
03
C L
0
o
c o
u _
-o
0
o
CD
E
^ E
C O
2
LL
Q
C O
T3
Q
£
cn
2^
>
c
—
»
j
2
G
5 5
c
S
to
CL
£?
3
D
5 5
c
CD
C O
D L
CD
.•
O
5 5
^
—
-S -
c o
Q _
* p j -
^
n
CO
>
Z
—
—
—
—
—
~~
o
to
c
c
c
c
u
o
c
c
u
c
c
c
c
c
c
a
c
c
c
a
o
c
c
c
c
a
T;
0
C
a
^
a
c
a
u
Q
3 C
3 C
3 C
3 C
3
r
3
a
3
a
3 C
3
0
a
> >
a
3 C
3 C
3 C
3 C
3 C
3 CI
3 C
15 C
3 C
3 C
3 C
3 a
3 C
?E
3 C
3 C
15 U
r u
1 3 a
3 C
3 U
- u
3
a
3 C
: a
J LL
.Q
3 E
3 C
3 C
vJ CI
3
a
3 Ct
3 C
U C I
3 C
3 C I
3 C
3 C
3 C
3 C
3 C
3 T-
J T
D a
3 C
3 T
U
T
3
a
3 C
3 C
3 C
3 C
3 C
3 C£
3 0
3 U
3 C
3 c r
3 C
3 CI
3 C
: a
J U
.Q
3 C
3 C
3 C
3 C
3 r
3 r-
3
a
3 C
3
r
3 l~
3 a
3 C
3 C
3 C
3 C
3 C
- 0
- C
3 C
3 C
- a
- C
3 O
3 C
? E
3 C
3 C
3 •<;
D
C
3 a
3 C
3
^
3 C
3 a
3 C
•\
-
: a
J U
.Q
3 <;
3
0
3 O
3 C
- C
- C
3 C
3 C
•^
P
a
c
-
3
3
3 h
3 r-
D
a
3 C
0 C
£ £
3 C
h
h
a
c
3
3
3
3 ?
3
0
3 C
J-
C
S E
3 C
C
0
a
c
f
3
15
" L
: a
J U
- Q
5 ?
3 O
3 C
3 C
3 C
3 C
3 C
f a
3 C
1 5 a
3 C
- CI
- c
D
a
3 C
3 C
3 C
3 C
3 C
- U
- O
1 5 a
3 C
3 g
15
0
3 C
3 C
3 C
3 C
3 C
3
a
3 C
i a
3 C
\
^|
:
a
J U
.Q
3 O
3
C
3 a
3 C
3 C
3 C
3 C
3 C
3
P
3 C
3
a
3 C
3
a
3 C
3
a
3 C
3 C
3 C
3 C
3 C
3 a
D
0
3
a
3 C
3 O
3
0
•5 C
3 C
3 C
3 C
3 C
3 C
3
a
3 C
3
a
3 C
5 h
: a
J U
.D
3 P
3
0
3 h
3 C
3 E
3 C
3 C
3 C
3 C
3 t^
3 C
3 P
3 C
3
0
3 C
3 C
3 C
3 C
3 C
15
P
1 5 C I
3 a
3 C
3 T
j a
3 r~
3 C
3 C
It
3 C
3
^
0 C
3
f
3 C
: c
J U
.D
3
13
3
3
3
3
3
(5
3
h * " »
^D ^
—
CO O5 O5
C\J CVl ^~ CO LO •*""
if r\"i ^ ^^
T
"~
•* odd
•3
3
D
3
3
3
3
3
3
3
co
o un co oj T-
05 CV J
h °° T
XT 000
)
3
3
3
3
t-
5
3
CD
«* T~ .w
C O - • i ~ ~
— \ C
— CO C c~
i- O
CO
UJ
j . 07 iZ
x"C>
x'Z.
O DC
&,
o
0)
£
u?
o
o
j=
c
0
3
.Q
in
• —
>,
"c o
o
0
a.
E
CO
u
Q
.
C
• - .9
J C " c o
i_
F
3 X
m
Q.
« §
~O
i-
5i 2
" a> ®
' C D Q >
11
=J cr
M ?
—
c u
- D E
-2 -g
a> 2
5
5 UJ
^OT
J ^
"~^
x"
LO QJ
CD "D
CO
•*-"
«
D ^
(D O
*-
O
( U I I
S
Lt
So
2 x
0) CD
N
-0
C D .E
-° .—
0 ~
Q
^ E
.> <
0
?- c
^
Z -
" to ^ J
o S
S o.
E a
C O C O
D?g
Q j I I
~O CO
S d
SOME
INTERPRETATION
PRINCIPLES
139
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 137/185
canonical co r r e l a t io n ana lys i s ( c f . Cohen
& Cohen,
1983,
p.
4 5 6 ; L e v i n e ,
1977 , p . 2 0 ; Mere d i th , 196 4 ; Thompson, Z O O O a ) , w h e n w e i g h ts an d s t r u c t u r e
c oe f f i c i e n t s
a r e
u n e q u a l , s e n s i b l e i n t e r p r e t a t i o n u s u a l ly r e q u i r e s e x a m i n a t i o n
o f bo th se t s o f r e su l t s . S t ruc tu re c o e f f i c i en t s a r e equa l ly im p o r t an t i n C FA
(cf .
G r a h a m
et
a l . , 2 003 ;
Thompson,
1997) .
Thompson, Cook, and
H e a t h
( 2 0 0 3 ) p r e s e n t e d
i l l u s t r a t i ve
s t r a t eg i e s fo r r e p o r t i n g r e s u l t s fo r co r r e l a t ed
CFA f ac to r s .
P u t d i f f e r e n t l y , a s B e n t l e r a n d Y u a n ( 2 0 0 0 ) e m p h a s i z e d , i n C F A w h e n
fac tors a re
corre la ted , "even if
a
f ac to r do es
n o t
i n f l u e n c e
a
v a r i a b l e , that
is , its p a t t e r n c o e f f i c i e n t o r w e i g h t is zero , th e c o r r e s p o n d i n g s t r u c t u r e
c oe f f i c i e n t , r ep resen t ing the co r r e l a t io n or c o v a r i a n c e of the v a r i a b l e w i t h
that
f ac to r , gene ra l ly w i l l
not be
zero"
(p.
3 2 7 ) . There
are
th ree po ss i b l e
s i t ua t io ns th a t a r i s e w i th co r r e l a t ed f ac to r s , no t c o un t ing co m bina t io ns o f
t h e s e s i t u a t i o n s .
nivoc l
Models
O n e
po ss i b i l i t y
is
t h a t
each
m e a s u r e d v a r i a b l e
re f lec ts th e
i n f l u e n c e s
of only a s ing le one of the co r r e l a t e d f ac to r s . Mo d e l 4 p resen ted in Table
11 .3 invo lves such a case.
F o r such "un ivo ca l " m o de l s
in
w h i c h each m e a s u r e d v a r i a b l e s p e a k s
o n
beh a l f
o f a
si n g le l a t e n t v a r i a b l e ,
th e
freed s t a n d a r d i z e d p a r a m e t e r s
an d
th e f ac to r s t ruc tu re co e f f i c i en t s a lw a ys exac t ly match. F o r e x a m p l e , P E R I
had a pa t ter n c oe ff ic ie nt o f Facto r 1 in Mod el 4 o f .950; the corresp ond ing
s t r u c t u r e
c oe f f i c i e n t also w a s exa c t ly .950 .
H o w e v e r , note that none of the s t r u c t u r e c oe f f i c i e n t s a s so c i a t ed w i th
p a t t e r n c o e f f i c ie n t s fixed to equa l zeroes were themselves zeroes . For e x a m p l e ,
P E R I
h ad a pa t t e rn co e f f i c i en t o n Factor I I i n Mo de l 4 f i xed to equa l ze ro ,
b u t the
c o r re s p o n d i n g s t r u c t u r e c o e f f i c ie n t
was
.398.
In other w o rds , w h en fa c to r s a r e co r r e l a t ed , a l l m easured v a r i ab l e s a r e
co r re l a t ed w i th a l l f ac to r s . This dyn am ic m us t h e co ns ide red a s pa r t o f r e su l t
i n t e r p r e t a t i o n . The h i g h e r the f a c t o r c o r r e l a t i o n s are in a b s o l u t e v a l u e s ,
t h e m o r e a l l v a r i a b l e s
w i l l
be co r r e l a t ed w i th a l l t h e f ac to r s .
Correlated Error Variances
A unique f e a t u r e of CFA v e r s u s EFA is the p o s s i b i l i t y of estimating
covariances
or
correlations
of m e a s u r e m e n t e r ro r variances. If the model is
correctly
spec i f ied , then
the
e r r o r
variances of the
m e a s u r e d v a r i a b l e s r e f l e c t
score u n r e l i a b i l i t y ( s e e Thompson, 2 0 0 3 ) .
For
e x a m p l e ,
the var iance of the
m e a s u r e d v a r i a b l e P E R I
was 1.847.
In Model 4 the estimated error variance for this v a r i a b l e was .178. This
means that,
if the
model
is correctly spec i f ied , the u n r e l i a b i l i t y of t h i s
140
EXPLORATORY
A N D
C O N F I R M A T O R Y
FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 138/185
~ "
0
C O
LL
O
O
T
~
||
C;
C O
t_
0
C O
'-*-
"O
0
jo
2 ?
o
O
o>
•
• —
'co
O
CL
~ TT7
""" T3
UJ O
C O
< c ®
-C
o
C O
0
5 5
'-4-J
rri
U_
T3
CO
C D
03
E
co
01
03
E
jO
C O
CL
CD
C D
T3
o
CD
|-r
O
C O
—
c
C D
JJ
CO
Q _
=
—
in
*
"o
T3
o
C D
-3
u
C O
^—
c
J D
D -
=
—
^
C D
-o
o
^^
2
r
CO
—
—
c
t=
CL
—
—
— .y
CO
.^0
'^_ •» —
C O
5
>C/3
r ^ L O h ~ T t - c o o c o o >
cococ\ i coa>a>a i ^
oooooooo
roc\joocDc\jcMC\ia>
c
c
•> a
3 C£
3
a
D C O C O C *
5 ^1
3OOOOOOC
i-
>
^- oo co i-
C O
O CO T-
O3
CD O) t
dddd
CT> C \ J OO CD
^~
co co h i i
c n c o C D o o I 1
dddd
CD fljOOCOOOh-QOCOCM
w c fc o^^ ^C TJcno
o 'CM o d o
OJI^OCDCOOCOCT)
CO
C O C O C O O) O ) O) t*^
dddddddd
OOOCD-r-OOT-^ -
COO31-COC500CDOO
c
n a
D r
- 0
C O C O C -T
0
^
dddddddd
CM h- ^
J- aD
1
C O O C O O )
I o C D cn
dddd
CO
O CD 1-
CO
CD T— 00 i I
CD
OO h
1
CO
| |
d d o o
h * o ) T —
c o o c o r ^ - T —
N T —
a > i — mcDcor--
C M
^3- y-j ^_: a> a) o
i '
cvi
o o d
CDCDOOCOCOOCOO3
c o c o c M c o c n c n c D r ^
oooooooo
c
D C
n
c
3 r-
~ O O r
- •*;
•
Uj i^- *jj i^- wj \j±i
i^ij
vj
c n c o c D o o c o c o c o r o
oooooooo
CM (^ ^
C O O 0
CD C3 > C
J-
00
o
a
o
h
D
d
ddd
o
a>
o h-
m r - cn r- i i
O) CO CD 00 | |
c
D C
3 C
3 C
V
3
•» -
1 1~ » tr\ iv
f jocooinr- ' -o jT-
^ T - C O - T - C O L O G O O O
C "^" " O5 O5 O
' co odd
CD
Q)
v
— ^[
a: cn cc ct a: cc ir cr > ^ <£
U J L U L U U J U J U J U J L I J Ocoa j
•^<\r~
|J
-
lJ
-2
0 - C L C L Q - Q - C L Q _ C L OLL ?-i~C3 X~Z- OCC
'
c a
0)
£
8
"
^
II
<
LU
< n
2
CL
X
0
•Q
£Z
iH
(D
>
c 5
CO
Q.
o
_
o
X
O
-o
._ c
^
S
o
£=
LL
Z .
0)
n
0]
CJ
"a.
8 -
0
c:
II
cc
c
r^
J
CO
(D
J
w
ro
•a
fa
-o
- ±
0
a_
ff
CO
LO
<u
N
S
Q
T3
0
X
a >
5
11
X
-4= O
S
a.
W
CL
Q)
"3
a5
*o
£ -
CO
O
s ®
Q J
QJ co
Q
=
g
CT
SOM INTERPRET TION PRIN IPLES
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 139/185
measured var iab le could
he
e s t ima ted
as
.178
/
1.847
=
.096
(o r
9 .6%).
The
corresponding re l i a b i l i ty of this measured var iab le could be es t imated as
(1.847 - .178) / 1.847 = 1.669 / 1.847 = .904 (or
90 .4% ;
or 100% - 9 .6%) .
Measurement error variances in classical s tatis t ics are usua l ly taken as
be ing random
and
therefore independent . However ,
in
some cases elements
of measurement error may be sys temat ic and overlapping. For example , in
a s tudy involv ing pe rcep t ions of gender roles , if two i t ems in a given s tu d y
invoked s t rong hom ophob ic emo tional reac t ions , the error var ianc es for
these
tw o
i t ems migh t
b e
corre la ted even t h o u g h t h i s d y n a m i c
is
ex t raneous
to the p r im ary f ea tu re s of the model .
Model
5
allows
the two
error var iances
for the
Factor
I
measured
var iab les ,
"Giving users in d iv id u a l a t tent ion" and "Employees who deal
with users in a caring fashion," to c o v a r y . T h i s add i t iona l s ing le pa rame te r
es t imate cos ts one addi t ional degree of f reedom, as reported in
Ta b l e
11.3.
B ut
only a l lowing error covariances to be nonzero does no t
change
the
s t ruc tu re coeff icients associated w ith freed pat tern coef f ic ients, and the
freed
pat tern coef f ic ients
and
re la ted s t ruc ture coef f ic ients
wi l l
remain equal .
Multivocal Models
Model
6 is
presented graph ica l ly
in
Figure 11 .2 .
In
this model there
is o ne
measured va r iab le (PER8) th rough which b o t h factors speak. Note
t h a t
fo r such var iab les the freed pat tern coef f ic ients no longer equal the
associated structure coeff icients (i .e . , .206
^
.489,
an d
.711
^
.793) .
C O M P A R I S O N S
OF MODEL FITS
In any CFA (o r SEM) m odel tes t ing, mu l t ip le models ma y f it the same
da ta . T h u s , i t is not best practice to test only a single, preferred model
(Th o mps o n , 2 0 0 0 c ) .
Ins tead, i t i s des i rab le to tes t m ul t ip le plau s ib le r iv a l
models .
Then,
if the p referr ed mod el is the only m odel w ith reasona ble f i t ,
the case for t h a t mo del i s correspondingly s t ronger .
In a fac tor analyt ic context ,
r iva l
models would of ten inc lude a model
speci fying independence
of the
measured v ar iab les ( i .e .,
no
fac tors ) ,
a
mode l
posit ing a s ingle factor, and m odels posit in g
b o t h
unco rrelated an d correlated
factors . T a b l e 11.4 presents
fit
s tatis t ics
for
these models
fo r
these data.
The s ta t is t ica l s ignif icance tes t for the b ase l ine independence model
pos i t ing no factors results in re jec t ion of the nu l l hypothes is t ha t the mode l
fits.
N o t e
t h a t th e ^ C A L C U L A T E D v a l u e for this test is reported in scientif ic
nota tion as "1.45E-134." This means t h a t af ter the decimal , the p is 133
zeroes fol lowed by
"145."
142
EX PL O RA T O RY
A N D
CO N FJ RMA T O RY FA CT O R ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 140/185
e1 ) e2
) (
e3 j e4
/ V y V y
1
1
1
51-1
Per2 Per3
Per4
ServAff
V
LibPlacd
1
~ n
Per5
i
r
P e r G
I
Per?
i I
^
A ^
1 1 1
Per t
t
1
t
r
i
^-v.
e5 i e6 e7 e8
Figure
11.2.
C on f irmatory fac to r ana lys is M odel #6 (Fac to rs I an d II mu l t ivoca l throug h
P E R 8 ) . ServAf f
=
"Service Affect" ; LibPlace
=
"Library
as
P lace. "
SOM E
I N T E R P R E T A T I O N
P R I N C I P L E S 1 4 3
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 141/185
^ — ~
o
CC
LL
0
O
T
—
II
C
co
CD
o
C D
.2
Q
C D
73
0
-*•
°
i- £
T C O
m
UJ (T
C D
"g
< CC
C O
CD
73
O
13
O
LL
L_
4
—
C O
o
to
-2
C O
~
1 1
73
~
O
CD
C D
U)
^
U
C O
D C
~
O
1 1
^
s
73
% _ - »
• * * • •
P
L
U
L
T
E
D
:+.-
3
cy
^
73
O
i-
CD
O CD
m co
0CD
O CD
O
i-
OCD
0C
8%
OCD
00
C O C O
C O C M
in
CM T-
CO C3) O
T TT
LU LU LU
in in K
^~ r o
^ CD en
00 O CO
C vJ C M
in
i- N. f
m C D o o
•* m co
(M 00 CO
r -~ C M -< t
CD
o
I
n
O
n
a
o
D
i
e
e
o >
0
05
in
o
CO
CO
C 3>
i—
•*
c\ i
O5
CO
o
0
o
C)
o
O J
f>x_
CM
CO
^_^^
-n C O
U
n
e
a
e
f
a
o
s
C
o
e
a
e
T—
CO
o
CM
CO
O
in
en
CD
in
CD
T— .
in i-
co o
C O O
CO O
o
o
C D 0
O >
i-
CD T-
CO
O5
• CD
COT
_
^
f
a
o
s
D
i
e
e
|j
LL
^
1
"o
if)
CD
O
Q
M
CO
OO
E
-Q
~ 5
o
OJ
.c
"(Q
1
03
ra
c
cO
0)
E
"5
O ,_ :
^
=
•—
c
gX
|
£ "o
S0
1 |
CO W
^ cz
-
1
o j "5
"O o
CD *"~
_C II
OD
<
> LJJ
0 C/)
« CT
c . „
0) <U
Cj-
:• —
CO ^
^ T
^
LLJ
~
2
- a
"- E
.18
75
ii
"5 JT
C 7^
u
u
rt x"
A
o
e
I
n
s
e
n
m
e
n
144
E X P LO R A TO R Y A N D CO NF I R M A TO R Y F A CTO R A NA LYS IS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 142/185
The
p
is not
e q u a l
to
zero, regard less
o f
w h a t
the
s ta t i s t i ca l packa ge
m ay
p r in t to r e la t ive ly f ew dec im al p laces . A common e r ro r in pub l i shed
l i t e ra tu re is er roneously repor t ing that p = .000, wh ich means th a t the
researcher impossibly ob ta ined an impossible
resul t
(wh ich is imposs ib le ) .
Such
sm all values can b e acc ura tely repor ted as p < .001, o r exac t va lues
can be
derived using Excel spreadsheet functions such
as:
= c h i d i s t ( 7 2 4 . 5 1 , 2 8 ) ,
which here y ie ld s p = "1 .45E-134 ."
For
these tes ts , Model 4 provides reasonable fit. Th is is tr u e n o tw i th -
s tand ing
th e p
value ( .03685) associated with
th e
tes t .
A s
noted
in
chap te r
10, the p values fo r m o d e l fit tests are inflated by the large sam ple sizes
usually used
in
CF A ,
as are
s ta t is t ical s ignif icance tes ts
in o ther
contexts
( T homp son ,
1996) .
The s ignif icance tes t is more
useful
in look ing at comparative fit of
nested models . In th i s example , Model 4 posi ted correla ted factors, e s t im a te d
17 paramete r s (8 pat tern coeff ic ien ts , 8 er ror var iances , and 1 covar iance) ,
and reta ined 19 (36 - 17) degrees o f fr eedom. Model 3 posi ted uncorrela ted
factors , es t imated 1 6 paramete r s (8 pat tern coeff ic ien ts , 8 e r ro r var iances ) ,
and
reta ined 20 (36 - 16) degrees of f re e d o m .
T he
models
are
nes ted because Model
3 is
Model
4
w i th
one
p a r a m e te r
( a covar iance) r emoved . Th i s m e a n s t h a t
th e
differences
in the
m o d e l
fi ts
c a n b e q u a n t i f i e d b y s u b t r a c ti n g the %
:
va lu es (48 .27 - 31.36 = 16.91) an d
the
corresponding degrees
o f
f reedom
( 2 0 - 19 = 1 ) . Th e
i m p r o v e m e n t
in
Model 4 over Model 3 is
s ta t i s t ical ly
s ignif icant (p =
.00004) .
H I G H E R- ORD E R M OD E L S
A s
exp la ined
in
chap te r
6, in EFA
h ig her -o rder fac to r s
can b e
ex t r ac ted
where lower -order factors
a re
correla ted . G i v e n
the
G L M ,
the
same pr inci -
ples
a p p ly
i n
CFA.
One di f ference in the
analyses , however ,
is th a t
h igher -
order factors
in CFA a re
e x t r a c t e d w i th
th e
factor covar iances /correla t ions
f ixed to b e
zeroes.
In CFA
h igher -order analys is ,
t he
factor covariances
are
posi ted
to he
accounted
for by the
h igher -order factors .
Three First-Order Factors
Figure 11.3 presents a h igher -order m odel invo lv ing 12 measured var i -
ables and 3 f i r s t -order factors for the data provided by the 100 faculty. In
C FA
h igher -o rder ana lyses , usual ly there
are
th ree
or
more
first-orde r
factors
so tha t
th e
h igher -o rder paramete r s wi l l become ident if ied (see Byrne, 1998,
p p . 1 7 0 - 1 7 3 ) .
SOME
INTERPRETATION PRINCIPLES
145
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 143/185
N /
e2 ) ( e3 ( e4
Per5
Per6
Per?
_J
i
]1
J L
.
Per8
ti
i
^_L.^
ti
e5 ( e6 e7 e8 )
Figure 11.3.
Second-order model wi th three f i rs t-order factors. ServAff
=
"Service
A ffec t " ;
LibPlace = "Library as Place";
InfoAcc
= "Information
A cces s . "
The m odel g raph ica l ly po r t r aye d in F igu re 11 .3 has 78 deg rees o f
f reedom
( [12 x (12 + 1) ] / 2) . The model expends 27 of the 78 degrees of
f reedom to es t im ate 12 er ror var ianc es of the me asured var iab les , 9 f i rs t -
o rder f ac to r pa t te rn coefficients , 3 var i anc es of the f i r st -order factors, and
3
second -o rder f ac to r pa t te rn coef f ic ien ts
(12 + 9 + 3 + 3 =
2 7 ) . Three
first-order pattern coefficients were fixed to equal Is , and the second-order
fac tor ' s v a r i a n c e was f ixed to equa l 1 .
146
E X P L O R A T O R Y A ND CO NF I R M A TO R Y F A C T O R A N A L Y S I S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 144/185
Table 11.5 presents
th e
p a r a m e t e t e s t i m a t e s
and f i t
s ta t i s t ics
fo r
th is
analys i s . As
b e f o r e,
note
that pa t t e rn coef f i c i en t s
fixed to
equa l ze ro
do not
i m p l y
that co r respond ing s t ruc tu re
o r
o ther co r re l a t ions
a re
also zero .
Note that these r esu l t s sugges t
a
r easonab le model
fit.
These three
latent
v a r i a b l e s
d o
m e a s u r e
a
second -o rder f ac to r .
But the
" In fo rmat ion
Access"
first-order
fac to r
has the
h ighes t co r re l a t ion wi th
the
second -o rder
f ac t o r
(r = .990). The f i rs t -order factor , "Library as Place," has the lowes t
of
th e three
co r re l a t ions wi th
th e
second-order factor
(r =
. 6 0 2 ) .
This
a n a ly s i s reflects
the
r ichness
o f
s e c o n d - o r d e r f a c t o r a n a l y s i s
and i t s
capac i ty
t o m o d e l c om p l e x h i e r a r c h i c a l d y n a m i c s .
Two First Order Factors
A
second-order
CFA
c o r r e s p o n d i n g
to
M o d e l
#4 can be
execu ted ,
if
addi t ional const rain ts
are
imposed
so
that
the
model becomes
iden t i f ied ,
because there
are
on ly
two first-order
factors .
The
constraint i l lus t ra ted
in
Figure 11.4 is impos ing a res t r ic t ion that the two second -o rder p a t t e r n
coef f ic ien ts for the f i rs t -order fac to r s a re to be es t imated , b u t that these
es t imates m us t be equa l . In SEM/CFA s o f t w a r e , s u c h r e s t r i c t i o n s a r e c o m m u -
n ica t ed b y u s ing a s ing le l e t t e r fo r two o r more p aram ete r s cons t r a ined to
b e e q u a l .
Table
11.6 presents
the
paramete r es t imates
and the fit
s ta t i s t ics associ-
a t ed wi th th i s mode l . No te that the f i t s ta t i s t ics repor ted in Table 11.6 fo r
t h i s second -o rder m ode l
are
iden t i ca l
to the fit
s ta t i s t ics repor ted
in Table
11.3 for the correlated f i rs t -order fac to r mode l , Model #4 .
In
CFA, h igher -o rd er f ac to r ana ly s i s r eexp resses
the
correlat ions amo ng
the f i r s t - o rde r
fac to r s
b y
m o d e l i n g
these
d y n a m i c s
in an
a l t e rna t ive model .
H o w e v e r , the fits of these h igh er -o rder mode l s and the co r respond ing first-
order m odels are the s a m e , w h e n t h e r e a re f ew er than four f i r s t - o rde r fac to r s
fo r
a g iven second -o rder f ac to r . N ever the l ess , t he h igh er -o rd er ana lys i s does
p rov ide add i t iona l i n s igh t , because on ly in th i s mode l a re the co r re l a t ions
of
m e a s u r e d v a r i a b l e s
and f i r s t -order
f a c t o r s w i t h h i g h e r - o r d e r c o n s t r u c ts
q u a n t i f i e d .
M A J O R
CON CEPTS
The decis ion of how to
iden t i fy
a fac to r mode l i s a rb i t r a ry and does
no t impa c t f i t s t a t i s ti cs . Ho w eve r , s ty l i s t i ca l ly , mode l
id e n t i f i c a t ion
v i a f ix ing
f a c to r var i ances
to 1 has
s o m e a d v a n t a g e s , s u c h
a s
m a k i ng f ac to r c o v a r i a n c e s
a l so equ a l f ac to r co r re l a t ions
an d
a l l o w i n g c o m p a r i s o n s
o f
e s t i m a t e d p a t t e r n
coef f ic ien ts
f o r a l l t h e m e a s u r e d v a r i a b l e s o n a g i v e n f a c t o r .
SOME
IN T E R P R E T A T IO N PRINCIPLES 147
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 145/185
C O
CD
.a
C O
-
i_
C O
"0
C D
C O
C O
C D
>
CM
J
C O
c o
^*»
C O
^
<
0
C O
LL
C D
E
95
LO r ^
•r
OO
i- OCO
CD LL
03
o°
<M
1
~
1— CO II
0c-
'•*—
viz*
C O
'
C O
c / 5
II
T3
C
C O
£ 2
"c
CD
]o
C D
o
O
o
^
2
JD
co
a.
0
0
C O
C O ^
C D
O "o *"•
III
o CD c j
^ E
CO
i_
C D
O
1
j
i
U_
(ii
ii
^ _ «
o
CO
~
—
—
E
Q J
^
O
a.
—
^
—
~ C D
S-S
(0 .«
ca
^o
C D
C O
i~
^ ~ T~ " C D i" ~ C ^ T— O ) C
LO
i
—
CO i —
CD
^~ tOCO
f**-
C3 L
c o C D ^~ c o L O LO L O ^~ L O r*-- c
0000O
O h
LO
C
^ r^ oo CD
3
r - ~ - o LO
D C
f
D -5
0O0000C
i- LO LO CD CD CO C
*- i— C D a
00000
LO C
CO Cl
15 O O3 CO
D O ) C D C O
C O C O C M C O C 5
00O0O
- CM <M
CvJ
0
J CO O) CO O)
O 5 C O C D 0 0 C O
OOOOO
o o o o o o c
CO
C M 00 C
O CO CO T
T CD U
f
CM C
T
U
D
0
CO
0) h- CO t CO C
OO0000C
O O •* K
CO
CD T
oo
a) co en o> LO c
1
co co co co ^- ^i- a
o o o o o o c
C l
I
h
U
D C D C
- i- «
• > r - ~ «
5 C
3 a
D ^
d d d c
•* 0
CO
CO
0)
d
J C M C M
•< t
co a> co
O ) C O C D C O
C O C M C O
o co en
0) o- r--
ddd
I
I
O
C3
OO
i- CM CO ^J- LO
D C D C D C D C D C
LU U
J LU LU LU
C L D _ Q . C L C L
O i- O
D C DC DC DC £ E 5
LU
LU LU U
J LU UJ U
Q . D _ C L C L C L D _ Q
T3
O
O
CD
CO
CD CM O
O5 OCT>
CD CD CD
odd
CD CM O
0 ) 0 0 )
CD CD O)
^ i- LO CM 00 O
O LO CO CM h~ CD
rn O5 O> O
CD
000
LU
«
LL
2
O D C
LU
CO
05
E o
2'15
148
EXPL ORAT ORY AND CONFIRMAT ORY F A C T O R
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 146/185
© 1 ) © 2 © 3 04
ServAffHH v1
Pe
— i
> r 5
t
1
Per6
Pe
i
\\1
1
i
|
Per8
e5 e6
)
e7 e8
gure
n.4. Second-order model (Sec-O rder) with two first-order
factors. S e r v A f f :
Service Affect ; LibPlace
=
Library
a s
Place.
S O M E INTERPRETATION P R I N C I P L E S 149
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 147/185
12
TESTING MODEL INVA RIAN CE
The
sine qua non
of
research
is f inding effects
that repl icate under
stated conditions. When factor analytic results replicate across samples or
analytic decisions, the results are said to be invariant. Confirmatory factor
analysis (CFA) is ideally suited for testing invariance, because CF A allows
constraining different types of result equalit ies (e.g., invariance of pattern
coefficients,
invariance
of
factor correlations)
in
different models
and
quant i -
fies the
degrees
of
mo del invarian ce across these variation s.
Here three sets o f invariance analyses a re reported using t he Appendix
A da ta . The examples involve all 12 measured variables and the graduate
s tudent
(HI =
100)
and
faculty
(n^ =
100) data. These analyses illustrate
testing factor invariance across different respondent groups. Al ternat ives
are equally plausible, such as investigating invariance over random ly created
subgroups,
across data
collected
from
the
same people over t ime,
or
across
the same role group but different geographic origins (e.g., s tudents from
dif feren t
regions of the country) .
SAME MODEL, DIFFERENT
ESTIMATES
The
least restrictive test
of
factor invariance presumes only that
the
same model
fits in
different groups
b ut
that
the
parameters estimated
in the
different groups are independently estimated in each group. Figure 12.1
specifies
an input model for such an analysis.
153
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 148/185
e1
e2 e3 e4
e9 e10
6 1 2
1
r
Perl
1
Per2
V
\
Pe
t
1
r
1
I
> r3
1
r
Per4
Pe
r
;r9
,—
--.
1
t i
1
r
i
Perl 6
Peril
] 1 _ _
^er1S
(LibPlacd
1
x\
e5
e6 e7 e8
F/gure
12.1.
Confirmatory factor analysis factor model with
1 2
measured variables
and 3
correlated factors with factor variances constrained
to
equal 1.0. ServAff
=
"Service
Affect";
InfoAcc = "Information Access"; and LibPlace = "Library as Place."
In
this analysis
in both
groups
there are 78
degrees
of
freedom,
because
there
are 12
measured variables ([1 2
x (12 + 1 )] / 2). The specified
model
posits
that
each of three correlated factors, with
factor
variances con-
s t rained
to equal Is, are estim ated . Parameters freed to be est imated in both
groups are 12
factor pattern coefficients,
1 2
error variance s
for the
measured
variables,
and 3
factor covariances/correlations
(12 + 12 + 3 =
2 7 ) .
154
EXPLORATORY
AN D
CONFIRM ATOR Y FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 149/185
Thus,
there
are in both groups 51 (78 - 27) unspent degrees of f r e e d om ,
for
a
t o t a l
of 102 (2 X 51)
avai lable degrees
of
freedom after paramete r
es t imat ion .
Table
12 .1 presen ts
the
s tandard ized fac tor pa t te rn
coefficients , the
factor s t ruc tu re
coefficients ,
the factor covariances (here also correlations,
because
the
fac tor var iances
are all
fixed
to
equal
1), and fi t
statistics
for
the analysis. The fit s ta t is t ics suggest that the m od e l is a plaus ib le fit to the
data provided by both groups.
There are some differences in the paramete rs es t imated ind ependen t ly
across the two groups. For example , the i tem PER3 ("Employees who d e a l
with users in a caring fashion") is more correlated with the "Service Affect"
factor for
graduate s tudents (r
s
2
=
.835
2
=
69.7%)
than fo r
faculty (r
s
2
=
.692 = 47.9% ). The
d i f ference
in the squared s tructure
coefficients ,
wh i ch
can be
directly compared because squaring
has
been invoked,
is
no tewor thy .
S tu d e n t s def ine
affect of
service
in
l ibraries more tha n faculty
in
t e rms
of
being deal t wi th
in a
caring fashion
by library staff.
Con versely , the i tem PER6 ("A m edi t a t ive place") is more correla ted
wi th the "Library as Place" factor for faculty (r
s
2
= .908
Z
= 82 .4%)
than
for
s tuden ts (r
2
=
.753
2
=
56.7%).
In
this case, faculty percept ions
of the
library
are more correla ted wi th a quie t and contemplat ive a tmosphere.
The factor correla t ions are reasonably similar. The biggest differences
occur on the correla tion of "Service Af fect" with "Library as Place" for
gradua te
s tuden ts
(r |
x
u =
.569
=
32 .4%) versus facul ty
(r j
x
\\ =
.418
2
=
1 7 . 5 % ) .
Nevertheless ,
the
single model
fit
reasonably well
in both
groups,
and
the parameter es t imates were reasonably s imilar across groups. Furthe rmo re,
the
detected differences seem reasonably
to
reflect n a tu ra l
differences in
needs and
interests
of the two
g roups
o f
l ibrary users.
SAME
MODEL,
EQUAL ESTIMATES
Figure 1 2.2 presents a more con strained test of factor inv arianc e across
the two groups. In this case the three un s tan dard ized factor pat tern
coeffi-
cien ts
are
con strained
to
equal
Is. The
remain ing nine pa t te rn
coefficients
are
freed
to be
es t imated ,
but
w i t h
the
constrain t that
the set of
nine under-
standardized estimates mus t be equal across the two groups, as reflected by the
use of
le t ters wi t hin
the
d rawing
of the
mode l .
Also
es t imated independen t ly
in the two g roups are 12 error variances for the measured variables , the
variances
of the 3
factors,
and the 3
factor covariances .
Thus, from
the
avai lable
total of 156
degrees
of
freedom
(78 X 2), a
total of 45 (9 + 2 [12 + 3 + 3] = 9 + 2
[18]
= 9 + 36)
degrees
of
f reedom
are spent on parameter e s t imation , leaving 1 11 degrees of freedom un spen t .
TESTING MODEL INVARIANCE 1 5 5
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 150/185
TABLE 12.1
Factor
Pattern/Structure
Coefficients and Fit
Statistics
for
First-Order
Factor
Analysis With 12
Measured Variables
(n, = 100 Graduate Students; n
2
= 100 Faculty)
Group/
variable/
statistic
Graduate
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
PER9
PER10
PER11
PER12
O x I I
r
\ x III
1 XIII
Faculty
PER1
PER2
PER3
PER4
PER5
PER6
PER?
PER8
PER9
PER10
PER11
PER12
Ixll
XIII
1 XIII
f
df
f/df
NFI
CFI
RMSEA
Pattern
II
students
0.945 —
0.886 —
0.835 —
0.822 —
— 0.907
— 0.753
— 0.890
— 0.834
— —
— —
— —
— —
0.569
0.642
0.567
0.944 —
0.882 —
0.692 —
0.882 —
— 0.933
— 0.908
— 0.932
— 0.798
— —
— —
— —
—
—
0.418
0.689
0.596
135.03
102
1.32
0.924
0.980
0.040
III I
— 0.945
— 0.886
— 0.835
— 0.822
—
0.516
— 0.429
— 0.507
— 0.475
0.827 0.531
0.739 0.474
0.813 0.522
0.613 0.393
— 0.944
— 0.882
— 0.692
— 0.882
—
0.390
— 0.380
— 0.390
— 0.334
0.576 0.397
0.716
0.493
0.663 0.456
0.480 0.331
Structure
II
0.538
0.504
0.475
0.468
0.907
0.753
0.890
0.834
0.469
0.419
0.461
0.347
0.395
0.369
0.290
0.369
0.933
0.908
0.932
0.798
0.343
0.426
0.395
0.286
III
0.607
0.569
0.536
0.528
0.514
0.427
0.505
0.473
0.827
0.739
0.813
0.613
0.650
0.607
0.477
0.608
0.556
0.541
0.555
0.475
0.576
0.716
0.663
0.480
Note. Parameter estimates "fixed"
to be
zeroes
are
reported
as dashes
("—").
All
critical ratios
> |2.0|.
NFI
= normed fit index; CFI = comparative fit index; RMSEA = root-mean-square error of approximation.
1 5 6
EXPLORATORY AND CONFIRMATORY
FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 151/185
e-\J ( e2 ) ( e3 ) ( e4 )
1
Perl
V ~ : y
1 n
e9)
(e1o)
ellj
1612,,
Per2 Per3
1
J_
1 1
Per4
Per9| Perl6
Peril
^er1
i I I
Per6
Per? Per8
Figure 12 2 Confirmatory factor analysis factor model with 12 measured variables
and 3
correlated factors with equality constraints.
ServAff
=
"Service
Affect";
InfoAcc
=
"Information Access"; and LibPlace = "Library as Place."
In
other
words the equality constraint means
that
the Figure 12.2 model
is
more
parsimonious than the Figure 12.1 model by a factor of 9 degrees
of
freedom
(111-102).
Table 12.2 presents the standardized factor pattern coefficients the
factor structure coefficients
the
factor
correlations, and fit
statistics
for the
TESTING MODEL
1NVARIANCE
7
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 152/185
TABLE
12.2
Factor Pattern/Structure Coefficients and Fit Statistics for First-Order
Factor Analysis With Equality Constraints
on
Unstandardized Factor
Pattern Coefficients n - i = 100 Graduate Students; n
2
= 100
Faculty)
Group/
variable/
statistic
Pattern
II
III
I
Structure
II
III
Graduate students
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
PER9
PER10
PER11
PER12
Ixll
X
III
r \\
x H I
Faculty
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
PER9
PER10
PER11
PER12
O x I I
r
\ x
III
X
III
f
df
I
z
df
NFI
CFI
R MS EA
(0.944)
—
0.893 —
0.812
—
0.834 —
— (0.907)
—
0.789
— 0.884
— 0.816
— —
— —
— —
— —
0.568
0.641
0.565
(0.947) —
0.870 —
0.735 —
0.874
—
—
(0.932)
—
0.898
— 0.935
—
0.814
— —
— —
— —
— —
0.422
0.699
0.603
144.47
111
1.30
0.919
0.980
0.039
—
0.944
— 0.893
—
0.812
— 0.834
— 0.515
—
0.448
—
0.502
— 0.463
0.825 0.529
(0.763)
0.489
0.798 0.511
0.607 0.389
— 0.947
— 0.870
— 0.735
—
0.874
—
0.393
— 0.379
— 0.395
— 0.344
0.589 0.412
(0.658) 0.460
0.690 0.482
0.488
0.341
0.536
0.507
0.461
0.474
0.907
0.789
0.884
0.816
0.466
0.431
0.451
0.343
0.400
0.367
0.310
0.369
0.932
0.898
0.935
0.814
0.355
0.397
0.416
0.294
0.605
0.572
0.520
0.535
0.513
0.446
0.500
0.461
0.825
0.763
0.798
0.607
0.662
0.608
0.514
0.611
0.562
0.542
0.564
0.491
0.589
0.658
0.690
0.488
Note.
Parameter estimates
"fixed"
to be zeroes are reported as dashes ( — ). Parameter estimates
"fixed"
to equal
1.0 for the
unstandardized model
are
reported
in
parentheses.
All
critical ratios
>
|2.0|.
NFI =
normed
fit index; CFI = comparative it index; RMSEA = root-mean-square error of approximation.
158
E X P L O R A T O R Y AND CONFIRMATORY
FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 153/185
analysis .
The fi t s tatistics suggest
that
the mod el is again a plau sible fi t to
the
da ta p rov ided
by
both groups.
The fit is not as
good
as the fit of the
less con strained m odel ,
but
never the less remains plaus ib le . Further res tr ic t ions on
equal i t ies
might be
imposed (e.g. , constraining
factor
covariances
to be
equal ) ,
bu t
such addi-
t ional con strain ts w ould necessi tate some reasons
as to why each set of
add i t iona l paramete rs might be expected to be i n v a r i an t .
Al te rna t ive ly ,
a
sequence
of
equal i ty cons t ra in t s m ight
be
imposed
to
detect
w here, i f a t a l l , invariance breaks down across groups (Byrn e, 2001;
Joreskog
&.
Sorbom, 1993).
The
sequence
of
invariance tes ts wi th equal i ty
cons t ra in t s might
proceed as follows: (a )
e q u a l i t y
of
uns tand ard ized pa t te rn
coefficients; (b)
equal i ty
of
u n s t an d a rd i ze d p a t t e rn
coefficients an d
factor
covariances; (c) equal i ty of uns ta nd ardized pat ter n
coefficients,
factor covari-
ances,
and
factor variances;
and (d)
equal i ty
of
u n s t an d a rd i z e d p a t t e rn
coefficients,
factor covariances , factor varianc es ,
an d
e r ror var iances
of
mea-
sured
variables.
SAME MODEL, TEST INVARIANCE
OF
PARAMETERS
The ul t ima te tes t of mod el in varian ce is us ing some or all of the specific
parameter es t imates
from on e
sample
in a
model tes t wi th
an
i n d e p e n d e n t
sample. Here
the
uns tand ard ized pa t te rn
coeff icients for the
model es t imated
wi th g rad ua te s tuden ts were
f i t to the
da ta p rov ided
by the
facul ty .
Figure
12.3
presen ts
the
inpu t mode l .
Table 12 .3 presen ts the s tandardized factor pat tern coefficients, the
factor
s t ruc tu re
coefficients,
the factor correlations, and fi t s tatistics for the
analysis .
The f i t s ta t is t ics
suggest
tha t the model is near the bou nd ary of
plausible
f i t to the data provided by the
facul ty .
Ne ver th e less , the deg ree
of fit is
s t r ik ing g iven
that
selected
specific
parameters from
the
g radua te
s tudents were used
in the
tes ted model .
In
pract ice , researchers usually hope t ha t
the
same cons t ruc ts
are
measured
across group s but usually do not expect factor pat tern
coeff icients
to be
completely invarian t . Nevertheless , such tes ts i l lus tra te
the
power
of
CFA to find the
boundar ies
of
resul t invariance.
MAJOR CONCEPTS
Science i s abou t i so la t ing and und ers tan d ing la ten t cons t ruc ts. We
seek con structs that repl icate unders ta ted condi t ions and that are useful for
stated purposes.
In
other words ,
w e
seek factors that
are
invar ian t over
samples and analyses .
TESTING
MODEL INVARIANCE
159
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 154/185
e 1 ) e 2 ) 6 3 e 4
1
e9j
(elQ) ( e11J e12
Per9 RerlCJ eril
v
(LibPlacq
1
Per5
Per6 Per?
65 ) 66 67 68
Figure 12.3. Confirmatory factor analysis factor model
with 12
measured variables
and 3 correlated factors
with
1 2 unstandardized pattern coefficients from graduate
students constrained as fixed. ServAff = Service Affect";
InfoAcc
= Information
Access";
and
LibPlace
=
"Library
as
Place.
For exam ple, do w e obtain the same or simila r factors across ind epen -
dent samples of people? Do we obtain similar factors when we analyze two
similar but different
sets
of
measured
variables? Do we
obtain similar factors
when we invoke different estimation theories?
CFA is ideally suite d to tes tin g invarian ce propositions, because of its
flexibility
and
capacity
to quantify
degrees
of
model fi t .
When
rival models
for
a given da ta set are nested, we can quant i fy the degree to which more
160 EXPLORATORY A ND CONFIRMAT ORY
FACTOR
ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 155/185
TABLE 12.3
Factor
Pattern/Structure Coefficients and Fit Statistics for First-Order
Factor Analysis for Faculty Data n = 100) Fitting Graduate Students'
Unstandardized Factor Pattern Coefficients
Variable/
statistic
PER1
PER2
PER3
PER4
PER5
PER6
PER7
PER8
PER9
PER10
PER11
PER12
1.x II
XIII
1l x III
x
2
df
f/df
NFI
CFI
RMSEA
Pattern
II
[0.966] —
[0.902] —
[0.816]
—
[0.901] —
— [0.917]
— [0.837]
— [0.929]
— [0.801]
— —
— —
— —
— —
0.460
0.783
0.575
106.71
63
1.69
0.879
0.946
0.084
III
I
— 0.966
— 0.902
—
0.816
— 0.901
— 0.422
—
0.385
— 0.428
— 0.369
[0.708] 0.554
[0.753] 0.590
[0.808] 0.632
[0.613] 0.480
Structure
II
0.445
0.415
0.375
0.415
0.917
0.837
0.929
0.801
0.407
0.433
0.464
0.352
III
0.756
0.706
0.638
0.705
0.527
0.481
0.534
0.461
0.708
0.753
0.808
0.613
Note. Parameter estimates fixed to be zeroes are reported as dashes ( — ). Parameter estimates fixed
to
equal
unstandardized pattern coefficients for the graduate students are reported in brackets. A ll critical
ratios > |2.0|. NFI = normed fit index; CFI = comparative fit index; RMSEA = root-mean-square error of
approximation. 78-(12
+ 13)
freed estimates
= 36.
parsimonious models estimating
fewer
parameters
fit
less wel l ,
as noted in
chapter
11. Remember that less parsimonious models (i.e., models estimating
more parameters) fit as well as or better than more parsimonious models,
until at the extreme w e estimate all possible parameters (i.e., df = 0),
at
which
point
every
j u s t - i den t i f i ed
model
with
zero degrees
of
freedom
fits perfectly.
Because
% s for nested models are additive, we can subtract the %
2
for
th e
less parsimonious model (e.g.,
%
2
=
31.36, df
= 19) from the %
2
for the
more parsimonious model (e.g., Jj = 48.27,
df
= 20) to quantify h ow
much
better
the
less parsimonious model
fits
(e.g.,
2
48.27
-
31.36
=
16.91).
Because degrees of freedom
from
nested models are additive, we can also
subtract the degrees of freedom (e.g., 20 - 19 = 1). This allows calculation
of
th e p
value (e.g.,
p =
.00004
fo r f =
16.91 with
df = 1) for the
null
hypothesis that
the
nested models
fit
equally
wel l .
TESTING MODEL
INVARIANCE
1 6 1
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 156/185
In CFA, we can fi t s imi la r but d i f ferent models to d i f feren t types o f
samples or to the same sample over time. We can fit the same models to
different samples but free the parameters to be est imated independent ly in
each
data set . Or we can fit the same model and also use some or all of the
est imates
in one
sample with
the
da ta
from
an
independent sample (see
Thompson
et
a l . , 2003) . Analyses
of the
last kind
are the
u l t ima te
in
tests
of invar iance ,
and the
u l t im a t e
in
parsimony, because
at the
ex t reme
we
are es t ima t ing no pa rameter s wi th new da ta when w e fi t only the parameters
from
a
prior study.
] 62
EXPLORATORY
AND
CONFIRMAT ORY FACTOR ANALYSIS
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 157/185
A P P E N D I X
A
LibQUAL+™ Da t a
0 0 1 2 8 7 5 5 3 2 3 3 6 5 5 7
002 2 5 7 5 5 4 5 5 6 3 4 3 6
0 0 3
2 6 5 5 6 5 3 5 3 3 5 4 7
0 0 4 2 5 5 4 6 4 4 4 4 2 1 5 4
0 0 5
2 5 5 5 5 5 4 6 2 4 4 5 7
006 2 7 7 7 8 7 8 7 6 6 8 7 5
007
2 8 8 7 7 6 7 6 7 6 5 6 5
008 2 7 7 7 7 5 3 3 6 5 4 5 8
9 2 1 3 1 1 1 1 1 1 1 1 1 1
0 1 0
2 9 9 9 8 5 7 5 2 7 6 7 7
1 1 2 9 9 9 7 7 9 9 7 4 6 5 9
1 2 2 4 3 3 5 6 5 4 5 2 5 1 4
013
2 6 6 5 7 7 7 7 7 6 5 6 7
0 1 4
2 7 7 2 7 2 3 2 3 6 6 5 9
0 1 5 2 9 9 9 9 7 5 4 5 4 1 1 4
016
2 9 9 9 9 5 5 5 7 6 7 6 6
017
28899766868
9 9
0 1 8
2 8 5 8 5 4 8
5
8 4 3 4 7
019 2 9 9 9 9 6 7 6 6 7 9 9 9
0 2 0 2 8 7 7 9 5 6 6 4 5 6 2 9
021 2 9 9 8 9 9 9 9 8 8 6 6 7
022
2 7 7 7 8 7 6 7 6 7 8 8 7
023
2 7 8 7 6 5 6 5 6 6 5 7 7
0 2 4 2 8 9 7 8 5 3 4 5 9 9 7 7
025 2 8 8 8 8 6 6 6 7 9 6 9 9
026
2 8 9 9 9 6 6 6 6 8 8 9 8
027
2 8 8 8 8 8 7 8 8 8 9 8 7
028
2 9 9 9 7 6 7 7 6 9 9 6 9
0 2 9
2 6 5 6 5 5 6 6 5 4 5 4 5
0 30 2 8 8 7 8 7 8 8 7 8 7 6 7
031 2 7 7 7 8 8 7 8 7 7 6 8 8
3 2 2 1 1 2 1 1 1 1 1 6 6 5 9
0 33 2 8 8 8 8 8 7 7 8 6 1 6 7
0 34
2 8 8 6 7 5 6 5 6 8 9 8 7
035
2 8 8 8 5 7 4 8 5 6 6 6 7
0 36 2 8 8 8 8 5 5 6 5 7 7 7 8
0 3 7 2 6 4 5 6 7 7 7 8 5 5 4 4
1 6 3
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 158/185
0 3 8
2 5 4 5 3 3 4 3 4 6 7 3 3
0 3 9
2 5 7 6 7 7 5 6 7 6 6 5 5
0 40 7 7 7 8 6 7 7 7 7 6 8 8
0 41 2 6 8 7 7 8 7 7 7 7 9 7 7
0 42
2 7 8 8 5 6 6 6 6 6 5 5 6
0 43 2 8 8 7 7 5 6 5 6 6 7 6 7
0 4 4 2 7 7 5
7 4 5 4 4 8 7 6 8
045 2 9 9 9 9 8 9 8 8 9 7 7 9
046 2 7 7 7 7 7 7 7 7 7 7 7 1
047 2 7 5 7 6 4 7 5 4 7 9 6 7
0 4 8
2 1 3 3 1 4 5 4 4 9 6 7 7
0 49 2 7 8 6 7 8 8 8 8 7 6 7 7
0 5 0 2 8 8 7 8 8 7 7 8 7 8 7 7
051
2988864686767
052
2 9 9 8 9 8 7 8 7 8 7 9 9
053 2 6 6 5 6 8 4 6 7 7 7 6 7
0 5 4
2 8 8 8 4 5 4 5 8 7 9 6 7
0 55
2 8 8 7 9 2 2 2 1 5 2 6 8
0 5 6
2 8 8 6 8 3 3 5 6 6 5 4 7
057
2 9 9 9 9 2 2 2 2 8 9 9 9
058 2 8 8 8 8 7 7 8 8 8 8 8 8
059
2 6 5 6 7 6 7 6 7 5 6 6 7
0 6 0
2 8 9 7 7 6 5 6 6 7 7 6 7
0 6 1
2 7 7 7 7 7 7 7 8 7 7 7 7
0 6 2 2 6 7 7 6 4 4 5 7 6 3 7 3
063
2 8 8 7 8 8 6 7 9 7 7 7 8
064
2 8 7 8 8 4 4 4 3 8 2 6 4
0 65 2 8 9 8 7 7 7 7 7 6 6 7 7
066 2 8 8 9 8 8 8 8 8 7 9 6 7
0 6 7 2 7 7 7 7 7 7 6 8 7 6 7 8
0 6 8 2 6 3 7 9 4 3 3 3 9 4 5 5
069
2 9 8 8 9 8 8 7 8 6 7 6 9
0 7 0 2 6 6 6 5 4 4 4 4 5 256
0 71
2 8 8 9 8 7 7 7 7 8 6 7 8
0 7 2 2 8 8 8 8 7 6 8 7 7 7 7 8
0 73 2 9 9 7 8 4 5 6 7 6 7 6 7
0 7 4 2 8 8 8 8 6
6
6 6 8 8 8 8
075
2 8 8 8 7 8 7 8 7 6 8 8 9
076 2 8 8 7 8 4 5 6 8 8 5 6 7
0 77
2 6 6 6 6 7 7 5 6 6 5 3 7
0 7 8
2 7 5 5 4 3 5 5 6 2 2 6 8
079
2 4 4 5 8 7 8 8 8 6 4 6 9
0 8 0
2 9 9 8 7 8 1
5
99699
64 APPENDIX A
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 159/185
0 8 1
2 8 5 8 7 3 1 6 1 7 7 6 9
0 82 2 9 7 7 9 9 8 7 9 8 8 8 7
0 83 2 8 8 8 8 6 4 7 7 7 9 6 6
0 8 4
2 7 7 7 7 6 6 6 7 6 7 6 8
8 5 2 4 2 3 4 1 4 1 1 1 1 1 1
0 86
2 6 7 5 6 5 6 6 6 5 6 6 7
0 87
2 6 6 7 7 8 5 6 9 6 7 6 7
0 8 8
2 8 8 7 8 6 5 7 7 8 5 6 9
089 2 9 9 9 9 9 6 9 9 8 9 9 7
090 2 5 5 5 5 3 3 44.7 5 5 7
091 2 8 8 8 8 7 7 7 7 6 9 7 9
092 2 5 5 4 5 7 7 7 7 7 6 7 5
093
2 8 8 8 8 6 6 6 6 4 4 6 7
094
2 8 8 9 8 8 6 8 8 9 5 7 9
0 95
2 8 7 7 7 8 6 7 6 7 7 7 8
0 96 2 8 8 8 8 7 7 7 6 7 5 7 8
0 97 2 7 8 7 8 7 5 8 5 7 8 7 7
0 9 8 2 4 5 2 1 5 1 2 2 1 2 6 2
0 99
2 8 8 1 8 3 5 8 3 4 2 1 7
100
2 9 8 8 9 7 4 9 8 9 8 9 9
1 1 3 7 7 7 7 6 7 6 6 7 6 6 8
1 2 3
8 8 6 5 4 4 4 7 6 7 5 8
1 3
3 5 5 5 5 5 7 5 6 4 5 5 5
104 3 9 9 9 9 6 6 6 7 5 6 9 7
1 5
3 9 9 8 8 5 3 5 6 8 6 6 8
106 3 6 7 5 7 6 5 5 6 6 7 8 4
1 07 3 8 6 9 7 6 8 8 7 5 5 6 7
108 3 8 7 8 8 6 4 6 7 6 5 6 7
1 0 9 3 6 6 6 5 5 5 5 4 6 7 5 8
1 1 0 3 8 7 7 8 6 8 4 6 6 8 7 7
3 9 8 9 9 9 9 9 9 8 5 8 9
1 1 2
3 7 7 8 7 6 8 7 5 6 6 7 5
1 1 3 3 7 7 7 6 6 6 6 6 4 6 7 6
1 1 4
3 7 7 7 8 1 2 2 1 7 7 8 3
1 1 5 3 7 7 8 7 7 7 7 7 7 5 64
1 1 6 3 8 9 8 8 6 5 6 9 4 3 5 7
1 1 7 3 5 4 5 5 7 7 5 5 1 3 5 3
1 1 8
3 9 9 8 9 7 7 7 9 7 6 8 9
1 1 9 3 8 8 8 8 7 7 7 8 6 7 8 8
1 2
3 7 6 8 8 7 6 7 6 7 8 7 5
1 2 1 3 8 8 8 9 8 8 8 8 4 7 6 7
1 2 2 3 8 7 8 8 5 4 6 5 6 7 6 7
1 2 3 3 5 3 4 5 2 2 2 2 2 2 2 7
APPE NDIX A 165
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 160/185
1 2 4 3 9 9 9 9 8 8 8 8 6 7 6 9
1 2 5 3 88 7 93 3 3 5 4 5 68
1263 999998896 7 8 8
1 2 7
3 9 9 8 9 8 8 8 8 8 7 8 9
1 2 8
3 9 9 7 9 7 7 7 6 8 7 8 7
1 2 9 3 3 3 7 3 7 5 6 3 7 3 6 6
1 30
3 9 9 8 9 6 7 7 7 7 7 8 8
1 3 1 3 8 8 6 7 3 5 4 5 8 6 8 8
1 3 2
5 5 5 5 2 3 3 2 6 6 5 5
1 3 3
3 7 6 8 8 5 5 5 6 5 4 5 8
1 3 4 3 8 8 9 8 6 6 6 7 8 7 8 8
1 3 5 3 8 8 9 8 8 9 8 8 7 3 7 7
1 3 6 3 7 6 8 8 4 4 4 6 8 6 7 8
1 3 7 3 8 8 8 8 2 2 2 2 1
1 6 4
1 3 8
3 8 7 8 9 6 7 5 7 6 8 6 6
139 3 7 8 8 7 6 7 6 7 4 8 8 9
140 3 9 9 7 8 7 8 7 8 5 6 8 8
1 4 1
3 8 7 7 8 6 7 8 7 7 6 6 7
1 4 2
3 8 7 6 7 6 6 6 6 7 2 5 8
1 4 3
3 7 7 7 5 3 2 5 3 9 3 7 4
144 3 8 8 8 8 9 8 8 8 8 8 8 7
1 4 5 3 8 7 8 7 3 5 5 5 6 8 5 6
1 46 3 7 7 6 7 6 5 6 7 7 5 7 6
1 4 7
3 8 8 7 8 5 6 5 6 6 3 5 3
1 48
3 5 7 5 7 6 6 6 6 5 5 6 7
149 3 9 9 9 9 7 7 7 7 9 9 9 7
1 5 3 8 8 7 8 5 5 5 7 1 5 6 8
1 5 1 3 8 5 8 8 8 8 8 8 7 2 6 6
1 5 2 3 9 9 9 9 5 6 5 9 9 8 8 5
1 5 3 3 9 9 8 9 8 8 8 9 8 9 8 8
1 5 4 8 8 8 8 7 7 7 8 8 8 7 7
1 5 5 3 6 5 5 5 3 2 3 2 1 2 5 3
1 5 6
3 5 5 3 5 5 5 5 7 5 3 7 7
1 5 7
3 7 6 7
7 3 3 4 4 3 6
7 4
1 5 8 3 7 6 6 7 5 5 7 6 7 4 8 6
1 5 9 3 9 8 9 9 6 6 6 7 7 9 8 9
160
3 7 7 7 7 6 6 6 2 8 8 6 7
1 6 1 3 8 8 3 7 5 5 5 4 7 5 6 7
162
3888888897 3 72
1 6 3 3 9 9 9 9 1 1 1 1 6 5 6 9
164
3 9 9 9 9 6 6 7 6 6 3 8 7
1 6 5 3 7 7 6 7 6 7 6 6 6 7 5 6
1 6 6 3 9 5 9 8 9 9 9 3 9 8 9 7
66 APPENDIX A
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 161/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 162/185
A P P E N D I X
B
SPSS Syntax
to
Execute
a
Best-Fit Factor Rotation
SET PRINTBACK=LISTING .
COMMENT Type in the pattern matrix that is the target as
matrix A ; Then type in the pattern matrix to be
rotated to
'best-fit'
position with the target as
matrix B ; Type in a matrix N_A consisting of all
zeroes with the same number or rows as A and B ;
Type in a square matrix of all zeroes with the number
of
rows and columns equal to the number of orthogonal
factors
in A and B .
MATRIX
.
COMMENT 100 Grad Students,
COMPUTE A =
{ .92134, .18690,
.21945,
.28211,
.28360,
.84662,
.85348,
.82875,
.80378,
.13580,
Varimax Pattern Matrix
.84248,
.75606,
.80837,
.22292,
.17110,
.29969,
.22454,
.30781,
.15795,
.27222,
.18737
.25625
.35335
.19958
.26935
.03794
.22423
.26075
.81976
.23077,
.80478
.20840, .80207 } .
COMMENT 100 Faculty, Varimax Pattern Matrix
COMPUTE B =
.18675
;
.21537
.16198
.24099
.21722
.21639
.19870
.14477
.66769
.79192
.71967
}
{ .20284
.10828,
.23051,
.16361,
.90807,
.89654,
.90832,
.79337,
.33664,
.18185,
.11897,
COMPUTE
N_A =
{ -0 ;
. 0 ;
.0
;
.0 ;
.0 ;
.89659,
.87403,
.73176,
.87285,
.12646,
.13470,
.17788,
.32388,
.10530,
.22943,
.33942,
169
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 163/185
.0
;
-0
.0 ;
0 ;
0 ;
.0
} .
COMPUTE DIAG_M
=
{ . 0 , . 0 , . 0 ;
0 , . 0 , . 0 ;
.0,
.0, .0 } .
COMPUTE N_B=N_A .
PRINT A /
FORMAT='F8.2 ' /
TITLE=
1
First Pattern Matrix (Target)' /
SPACE=4
/
RLABELS=Perl, Per2, Per3, Per4, PerS, Per6,
Per7, Per8, Per9, PerlO, Peril
/
CLABELS=Fact_I, Fact_II, Fact-Ill
/ .
COMPUTE A_N=A
.
-LOOP #1=1
TO
NROW(A)
.
+
LOOP #J = 1 TO NCOL(A) .
COMPUTE A_N(#I,#J)=A(#I,#J) ** 2 .
+
END
LOOP
.
-END LOOP .
A_N /
FORMAT='F8.4' /
TITLE='First Pattern Matrix (Target) Squared' /
SPACE=4 /
RLABELS=Perl, Per2, Per3, Per4, Per5, Per6,
PerV, Per8, Per9, PerlO, Peril
/
CLABELS=Fact_I, Fact_II, Fact_III
/ .
-LOOP #J=1 TO NCOL(A) .
+
LOOP #1=1 TO NROW(A) .
COMPUTE N_A(#I)=A_N(#I,#J) + N_A(#I) .
+
END LOOP .
-END LOOP
.
N_A /
FORMAT='F8.3' /
TITLE=
>
Row Sum of Squares for First Pattern Matrix' /
SPACE=4
/
RLABELS=Perl, Per2, Per3, Per4, PerS, Per6,
PerV,
PerS, Per9, PerlO, Peril / .
LOOP
#1=1
TO
NROW(A)
.
COMPUTE N_A(#I) = 1.0 / (N_A(#I) ** .5) .
END LOOP .
N_A /
FORMAT='F8.3'
/
TITLE='Normalization Factor
for
Rows'
/
SPACE=4
/
RLABELS=Perl, Per2, PerS, Per4, PerS, Per6,
170
APPENDIX
B
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 164/185
Per7, Per8, Per9, PerlO, Peril / .
-LOOP #J=1 TO NCOL(A) .
+ LOOP #1=1
TO
NROW(A)
.
COMPUTE A_N(#I,#J)=A(#I,#J)
*
N_A(#I)
.
+ END LOOP .
-END LOOP
.
PRINT A_N /
FORMAT='F8.4' /
TITLE='First Pattern Matrix (Target) Normalized' /
SPACE=4 /
RLABELS=Perl, Per2, Per3, Per4, Per5, Per6,
PerV, Per8, Per9,
PerlO,
Peril
/
CLABELS=Fact_I, Fact_II,
Fact__III
/ .
PRINT B /
FORMAT='F8.2'
/
TITLE='Second Pattern Matrix'
/
SPACE=4
/
RLABELS=Perl, Per2, Per3, Per4, PerS, Per6,
PerV,
PerS, Per9,
PerlO,
Peril /
CLABELS=Fact_I, Fact_II, Fact_III
/ .
COMPUTE B_N=B .
-LOOP #1=1 TO NROW(B) .
+ LOOP #J=1
TO
NCOL(B)
.
COMPUTE B_N(#I,#J)=B(#I,#J)
** 2 .
+ END LOOP .
-END LOOP .
B_N /
FORMAT='F8.4' /
TITLE='Second
Pattern Matrix Squared' /
SPACE=4 /
RLABELS=Perl, Per2, Per3, Per4,
Per5,
Per6,
PerV, PerS, Per9, PerlO, Peril /
CLABELS=Fact_I, Fact_II, Fact_III / .
-LOOP #J=1 TO NCOL(B) .
+
LOOP #1=1 TO NROW(B) .
COMPUTE N_B(#I)=B_N(#I,#J)
+
N_B(#I)
.
+ END LOOP .
-END LOOP
.
PRINT N_B /
FORMAT='F8.3'
/
TITLE='Row Sum of Squares for Second Pattern Matrix' /
SPACE=4 /
RLABELS=Perl,
Per2,
Per3,
Per4, Per5, Per6,
PerV, PerS, Per9, PerlO, Peril / .
LOOP #1=1 TO NROW(B) .
COMPUTE N_B(#I)
= 1.0 /
(N_B(#I)
** .5) .
END
LOOP .
N_B /
FORMAT='F8.3' /
TITLE='Normalization Factor for Rows' /
A P P E N D I X
B 171
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 165/185
SPACE=4
/
RLABELS=Perl, Per2, Per3, Per4, Per5, Per6,
Per7, PerS, Per9, PerlO, Peril / .
-LOOP #J=1 TO NCOL(B) .
+
LOOP #1=1 TO NROW(B) .
COMPUTE B_N(#I,#J)=B(#I,#J)
*
N_B(#I)
.
+ END LOOP .
-END LOOP .
PRINT B_N /
FORMAT='F8.4'
/
TITLE='Second Pattern Matrix Normalized'
/
SPACE=4
/
RLABELS=Perl, Per2,
Per3,
Per4, PerS, Per6,
Per?, PerS, Per9,
PerlO,
Peril /
CLABELS=Fact_I, Fact_II, Fact-Ill / .
COMPUTE
A_T=TRANSPOS
(A_N)
.
A_T /
FORMAT=
'F8.2' /
TITLE='A_N Transpose'
/
SPACE=4
/
RLABELS=Fact_I, Fact_II, Fact-Ill /
CLABELS=Perl, Per2, Per3, Per4, PerS, Per6,
Per7, PerS, Per9, PerlO, Peril
/ .
COMPUTE B_T=TRANSPOS(B_N) .
PRINT B_T /
FORMAT='F8.2'
/
TITLE='B_N
Transpose' /
SPACE=4
/
RLABELS=Fact_I,
Fact_II, Fact-Ill /
CLABELS=Perl, Per2, Per3, Per4, PerS, Per6,
PerV,
PerS, Per9, PerlO, Peril / .
COMPUTE RI=A_T * B_N .
RI /
FORMAT= > F8.3 ' /
TITLE='A_N Transpose times B_N' /
SPACE=4 /
RLABELS=Fact_I, Fact_II, Fact-Ill /
CLABELS=Fact_I,
Fact_II, Fact-Ill / .
COMPUTE RI_T=TRANSPOS(RI) .
RI_T /
FORMAT='F8.3' /
TITLE='Transpose of (A_N Transpose times B_N)' /
SPACE=4 /
RLABELS=Fact_I, Fact_II, Fact-Ill /
CLABELS=Fact_I, Fact_II, Fact-Ill
/ .
COMPUTE QUAD=RI * RI_T .
PRINT QUAD
/
FORMAT='F8.3' /
TITLE='A_N Trans * B_N * Trans of (A_N Trans * B_N)' /
SPACE=4 /
] 72
APPENDIX
B
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 166/185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 167/185
TITLE=
>
D= trans (trans A times B) times Eigenvectors' /
SPACE=4 /
RLABELS=Fact_I,
Fact_II, Fact_III /
CLABELS=Fact_I, Fact_II, Fact_III / .
-LOOP J=l TO NCOL(A) .
COMPUTE EE=EIG(J) .
+ LOOP
1=1 TO
NCOL(A)
.
COMPUTE D(I,J)=D(I,J) * EE .
+ END LOOP .
-END LOOP .
PRINT D /
FORMAT=
>
F9.3' /
TITLE='D
= D
times Eigenvalues
**
-1.5'
/
SPACE=4 /
RLABELS=Fact_I,
Fact_II, Fact_III /
CLABELS=Fact_I,
Fact_II, Fact_III / .
COMPUTE D_T=TRANSPOS(D) .
PRINT D_T /
FORMAT='F9.3' /
TITLE='D transposed' /
SPACE=4
/
RLABELS=Fact_I,
Fact_II, Fact_III
/
CLABELS=Fact_I,
Fact_II, Fact_III / .
COMPUTE C=EIGVEC
* D_T .
C I
FORMAT='F9.3' /
TITLE='Factor Correlations (Cosines)'
/
SPACE=4
/
RLABELS=Fact_Ia, Fact_IIa, Fact_IIIa /
CLABELS=Fact_Ib,
Fact_IIb, Fact_IIIb
/ .
COMPUTE C=D * VEC_T .
COMPUTE B_ROT=B * C .
B-ROT /
FORMAT='F8.3'
/
TITLE='B rotated to Best-Fit with A' /
SPACE=4
/
RLABELS=Perl,
Per2, Per3, Per4, Per5, Per6,
PerV, Per8, Per9, PerlO, Peril /
CLABELS=Fact_I, Fact_II, Fact_III / .
COMPUTE BROT_N=B_ROT
.
-LOOP
#1=1 TO NROW(A) .
+ LOOP #J=1 TO NCOL(A) .
COMPUTE BROT_N(#I,#J)=B_ROT(#I,#J)
** 2 .
+
END
LOOP
.
COMPUTE N_A(#I)= .0 .
-END LOOP .
PRINT BROT_N /
FORMAT='F8.4' /
TITLE=
>
Best
Fit Pattern
Matrix
(Target)
Squared'
/
SPACE=4 /
174
APPENDIX
B
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 168/185
RLABELS=Perl, Per2, Per3, Per4, Per5, Per6,
PerV, Per8, Per9, PerlO, Peril
/
CLABELS=Fact_I, Fact_II,
Fact__III
/ .
-LOOP
#J=1 TO NCOL(A) .
+
LOOP
#1=1 TO
NROW(A)
.
COMPUTE N_A(#I)=BROT_N(#I,#J) + N_A(#I) .
+
END
LOOP
.
-END LOOP .
PRINT N_A /
FORMAT='F8.3' /
TITLE=
1
Row
Sum of Squares for
Best
Fit
Matrix'
/
SPACE=4 /
RLABELS=Perl, Per2, Per3, Per4, Per5, Per6,
PerV,
Per8, Per9, PerlO, Peril / .
LOOP #1=1 TO NROW(A) .
COMPUTE N_A(#I)
= 1.0 /
(N_A(#I)
** .5 } .
END
LOOP
.
PRINT N_A /
FORMAT='F8.3'
/
TITLE='Normalization Factor
for
Rows'
/
SPACE=4
/
RLABELS=Perl, Per2, Per3, Per4, Per5, Per6,
PerV, Per8, Per9,
PerlO,
Peril
/ .
-LOOP
#J=1
TO
NCOL(A)
.
+
LOOP #1=1 TO NROW(A) .
COMPUTE BROT_N(#I,#J)=B_ROT(#I,#J)
*
N_A(#I)
.
+ END
LOOP
.
-END LOOP .
PRINT BROT_N /
FORMAT='F8.4' /
TITLE=
>
Best Fit Pattern Matrix (Target) Normalized' /
SPACE=4 /
RLABELS=Perl, Per2, Per3, Per4, Per5, Per6,
PerV, Per8, Per9,
PerlO,
Peril
/
CLABELS=Fact_I, Fact_II, Fact_III
/ .
COMPUTE BROTN_T=TRANSPOS(BROT_N)
.
COMPUTE T_M=A_N * BROTN_T .
COMPUTE TEST=DIAG(TJVI) .
TEST /
FORMAT='F8.3'
/
TITLE='Test Vector Cosines for Variables' /
SPACE=4 /
RLABELS=Perl, Per2, Per3, Per4, Per5, Per6,
PerV, Per8 , Per9, PerlO, Peril / .
END
MATRIX
.
APPENDIX B 175
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 169/185
NOTATION
C v x v
T he
symm etr ic va r iance/cova r iance
(or
simply covar iance) ma tr ix wi th
var iances on the diagonal and covar iances of measured va r iables wi th each
other
off the
diagonal.
D
2
,
The Ma ha lano bi s distan ce of the i th person's set of scores from the set of
means
on the
measured va r i ab les .
F
N x V
The scores on the latent factors.
I P X F
The
ident i ty ma tr ix which, when postmult ipl ied times
an
original conform-
able matr ix , y ie lds
as an
answer
the
or iginal matr ix ( l ike
the
n u mb e r
1
in algebra) .
P V X F The f irs t-order
factor pat tern coeff ic ient matr ix .
P F X H T he
second-order fa ctor pa ttern coefficient m a tri x.
P V X H The
product
(first-order b y
second-order) factor pat tern coefficie nt m atrix .
R F X F The
symmetr ic ma t r ix
of
bi v ar i a t e correla t ions (e .g. , Pearson
r's)
among
the f factors.
R V X V The
symm etr ic ma t r ix
of
biv ar ia te correla t ions (e .g. , Pearson
r's)
among
the v
measured var iables .
R V X V The
reproduced
correlation
matrix (i .e.,
the
portion
of the
intervar iable
correla t ion matr ix that the factors [i.e., the pattern coefficients]
can
re-
produce).
R V X V The
residual correla t ion ma tr ix ( i .e . ,
the
por t ion
of the
i n t e r v a r i ab le
correla t ion matr ix
that
the factors [i.e., the pattern coefficients]
cannot
re-
produce) .
R V x v ' The
inver ted intervar iable
correlation
matrix, which satisfies
th e
equat ion:
RyxV
RvxV
=
IvxV'
S \ x F
The f i r s t -order
factor s t ructure
coefficient
m a t r i x .
W
V
x F
The factor score coefficient m at rix.
Z
N
x
v The measured va r iables in z score form, such
that
the columns have means
of
zero
and
s tandard devia t ions
(and
v ar iances)
of one.
177
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 170/185
GLOSSARY
Communality coefficient h
2
). A statistic in a squared metric indicating how much
of the variance in a measured variable the factors as a set can reproduce, or
conversely, how much of the variance of a given measured variable w as useful
in delineating
the
factors
as a
set.
Component.
A set of pattern
coefficients
for the measured variables that can be
used
to estimate factor scores, derived without iterative estimation of communality
coefficients. (Some researchers consider factor a synonymous term, but others
feel that factor is a term that should be reserved fo r weights derived when
iteration is invoked.)
Critical ratio or
or Wald statistic).
A parameter estimate divided by its standard
error,
which is expected to be greater
than
roughly |2.0| for parameters
that
one
expects
to he
nonzero.
Eigenvalues
A . ) .
Area-world, variance-accounted-for statistics that characterize the
amount of information present in a given factor or function. (Eigenvalues a re
also sometimes synonymously called characteristic roots. )
Factor rotation. Graphic v isua l
or
mathematical movement
of the
axes measuring
the factor space used in exploratory factor
analysis
so that the factors can be
more readily interpreted.
Factor
scores. The latent (synthetic or composite) variables computed on each
factor (like the Y scores in multiple regression, or the discriminant function
scores
in
descriptive discriminant
analys is)
that
ma y
then
b e
used
in
fur ther
statistical analyses (e.g., t tests, analyses
of
variance)
in place of the
mea-
sured variables.
First-order factors. The pattern or pattern/structure coefficients for the factored
entities (measured variables in the commonly used R-technique analysis).
General linear model GLM).
The
notion
that all statistical analyses are correla-
tional and can yield
effect
sizes analogous to r
2
, and
apply
weights to measured
var i ab les to obtain scores on the composite variables that are actually the
focus of all these analyses.
Identification.
A model is said to be identified when, for a given research problem
and
data set,
sufficient
constraints
a re
imposed such
that there
is a
single
s et
of parameter estimates yielded
by the
analysis.
Inadmissible. A statistical estimate that is mathematically implausible (e.g., a
negative variance, or an r
that
is outside th e range of
— 1
to +1) or a solution
that contains
one or
more estimates that
are
implausible.
Invariance. The
degree
to
which
a
factor
or a
solution
can be
reproduced across
samples, across
different
analytic methods, or
both.
Inversion. The
matrix algebra process
of
solving
to find a
matrix that when post-
mult ipl ied by the original matrix yields the identity matrix; inverted matrices
are
used
in
matrix algebra
to
accomplish division.
17 9
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 171/185
Iteration. A
sta t i s t ica l
process of est i ma t in g a set of
results,
and then successively
tweaking them u n t i l two consecutive sets of es t imates are sufficiently close
that the i terat ion is deemed to have converged.
Loadings.
A n amb iguous and con fusing slang term used by some au thors for pat tern
coefficients,
and b y some au thors for structure coefficients, and by some auth ors
for both even w i th in the same ar t icle even when the coefficients are not equal .
Measurement error. The
var iance
in
scores that
is
deemed unrelia ble; this area-
world
proport ion can be est ima ted a s 1.0 minus a rel ia bi l i t y coefficient.
Modification
index.
The est imated decrease ( improvement) in the model fit chi-
square resulting from freeing
a
given prev ious ly
fixed
parameter instead
to
be est imated.
Oblique. An angle of more or less than 90° between some rotated factors, which
means that the factors (or components) and factor scores are correlated.
Orthogonal. The 90° angle
of
al l unrotated and some rotated factors , which s tat is t i -
cally
means the factors (or components) are perfectly uncorrelated.
Outliers.
Persons with aberrant
or
anomalous scores
on one or
more var iables with
regard
to one or
more statistics.
Pattern coefficients. The weights appl ied to the measured var iables to obtain scores
on the
factor analysis latent va r iables (cal led
factor
scores). These weights
a re
analogous
to the (3
weights
in
mu lt ip le regression,
the
s tandardized d iscriminant
func t ion coefficients
in
descr ipt ive discr iminant analys is ,
and the
s tandardized
canonical funct ion coefficients
in
canonical correlat ion ana lysis .
Pattern/structure coefficients.
The
term
for the
entries
in the
single exploratory
factor
analysis
coefficient
ma tr ix produced
when
factors are first extracted or
after rotat ion when or thogonal rotat ion
has been
invoked,
and in CFA
when
factors are
constrained
to be
uncorrelated.
Procrustean rotation. A method that rotates a pattern (or pat te rn/s t ruc ture) coeffi-
cient mat r ix
to
best-f i t posi t ion wi th
a
theoretical
or
actual pat tern coeffi-
cient ma t r ix .
Reflection. The process of chan ging a ll the signs of pat tern and s tru ctu re coefficients
on a
given factor
so
tha t
all or
most
of the
largest values
on the
factor
are
pos i t i ve in sign.
Sampling error. The uni que , id iosyncra t ic var iance in a given sample that does
not exist (a) in the popula t io n , which has no idiosyncrat ic var iance due to a
sampling process, or (b) in other samples, which do have sam pling error, but
sampling error var iances that are at least somewhat different.
Second-order factors.
The pat tern or pat tern/s tructure coefficients for the second-
order factors defined
by the first-order
factors.
Simple structure. A theoret ical pat tern of the dispersion of near-zero and large
nonzero pattern coefficients tha t opt imizes interp reta bi l i t y of the factors by
there being (a) enough v ar iab les ref lect ing a construct and (b) most v ar iab les
180 GLOSSARY
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 172/185
reflect ing
only
a
single factor such
that the
factors
can be both
labeled
and
dis t inguished from each other.
Structure
coefficient.
The
Pearson
product— moment
cor re la t ion
of a
factored ent i ty
(measured variables
in the commonly
used
R-technique
analysis) with
a la-
tent factor.
GLOSSARY 181
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 173/185
REFERENCES
Anderson,
T. W., &
R u b i n ,
H.
(1956). Statistical inference
in
factor analysis.
Proceedings
of
the
Third Berkeley Symposium
on
Mathem atical Statistics
and
Proba-
bility,
5,
111-150.
Arbuckle , J. L, &
Wothke,
W. (1999).
AMOS
4.0 user s
guide.
Chicago: Small-
waters.
Armstrong,
J. S.
(1967) . Der iva t ion
of
theory
by
means
of
factor analysis
or
Tom Swift and his electric factor analysis machine. American Statistician,
2 1 ( 5 ) ,
17-21.
Bagozzi, R. P . , Fornell, C. , & Larcker, D. F. (1981). Canonical correlation analysis
as a
special case
of a
structural relations model. Multivariate Behavioral Research,
16,
437-454.
Bandalos, D. , & Finney, J. S. (2001). I tem parceling issues in structural equat ion
modeling. In G. A. Ma rcoulides & R. Schumacker
(Eds.) , New
developments
and
techniques
in structural equation modeling (pp.
269-295).
Mahwa h, NJ: Er lbaum.
Bart let t ,
M. S.
(1937).
The
stat is t ical conception
of
mental factors. British Journal
o f Psychology, 28 , 97-104.
Bart let t ,
M. S.
(1950).
Tests of
significance
in
factor analysis. British Journal of
Psychology, 3(2 ) , 77-85.
Bentler, P. M. (1990). Comparative fit indexes in structural models. Psychological
Bulletin, 107, 238-246.
Bentler, P. M. (1994). On the qual i ty of test statistics in covariance structure
analysis:
Caveat
emptor. In C . R . Reynolds (Ed.) , Cognitive assessment:
A
multidisciplinary perspective
(pp. 237— 260). N ew Y ork: Plenum Press.
Bentler,
P. M.
(1995).
E Q S
structural equations prog ram manual. Encino,
C A :
M ul t i-
var iate Software.
Bentler, P. M., & Bonnett, D. G. (1980). Significance tests and goodness of fit in
the ana lysis of covariance structures. Psychological
Bulletin,
88, 588— 606.
Bentler, P. M., &. Yuan, K. -H. (2000). On adding a
mean
structure to a covariance
structure model. Educational an d Psychological Measurement, 60, 326-339.
Bollen, K. A . (1 987). Outliers a nd improper solutions: A co nfirma tory factor ana lysis
example. Sociological Methods and Research, 15,
375— 384-
Bollen, K. A. (1989). Structural equations
with
latent variables. N ew York: W iley.
Borrello, G. M., &
Thompson,
B. (1990). An hierarchical analysis of the Hendrick-
Hendrick measure of Lee's typology of love. Journal
of Social
Behavior
and
Personality, 3,
327-342.
Byrne, B. M. (1994). Structural
equation
modeling with EQS and
E QS/Windows: Basic
concepts,
applications,
and
programming.
Thousand
Oaks,
CA:
Sage.
Byrne, B. M.
(1998).
Structural equation modeling with
LISREL, PRE LIS, and
SIMPL IS: Basic concepts, applications, and programm ing. Ma hwa h, N J: Erlb aum .
183
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 174/185
Byrne, B . M. (2001). Structural
equation
modeling
with AMOS:
Basic
concepts, applica-
tions,
and
programming.
Mahwah , N J: Erlbaum.
Carr , S. (1992) . A primer on the use of Q-technique factor analysis. Measurement
and
Evaluation in Counseling and Development, 25, 133-138.
Cattell, R . B .
(1953).
A
quanti ta t ive analysis
of the
changes
in the
culture pat tern
of
Great Brita in ,
1837-1937, by
P-technique. Acta
Psychologia, 9 ,
99-121.
Cattel l , R . B . (1966a). Th e data box: Its ordering of total resources in terms of
possible
relational systems. In R. B.
Cattell
(Ed.) , Handbook of m ultivariate
experimental
psychology
(pp. 67-128).
Chicago:
Rand McNal ly .
Cattel l ,
R . B .
(1966b).
The
scree test
for the
number
of
factors. M ultivariate Behav-
ioral Research, 1, 245-276.
Cattel l ,
R. B.
(1978). T he scientific use o f factor analysis in behavioral and
l i f e
sciences.
N ew
York: Plenum.
Cliff ,
N .
(1987). Analyzing multivariate
data..
Sa n
Diego,
C A :
Ha rcourt Brace Jova-
novich.
Cohen,
J.
(1968). Mu ltip le regression
as a
general da ta -an aly tic system. Psychological
Bulletin, 70, 426-443.
Cohen, J., &
Cohen,
P. (1983).
Applied multiple regression/correlation analysis
for the
behavioral
sciences. Hillsdale, N J: Er lbaum.
Cohen, J., Cohen, P ., West, S. G., & A iken, L. S. (2003). Applied multiple regression/
correlation analysis for the behavioral sciences (3rd ed.). Mahwah, N J: Er lbaum.
Cook, C ., Heath, F., &. Thompson, B . (2001). Users' hierarchical perspectives on
library
service quality:
A
L i b Q U A L + ™ s t u d y .
College
an d Research Libraries,
62, 147-153.
Cook, C. , &. Thompson, B.
(2000). Higher-order factor analytic perspectives
on
users'
perceptions of lib rary service quality. Library Information Science
Research,
22,
393-404.
Cook, C., &.
Thompson,
B.
(2001). Psychometric properties
of
scores from
the
W eb-based L ibQU AL +™ s tudy of perceptions of library service qua lity. Library
Trends,
49,
585-604.
Cooley, W . W ., &Lohnes, P . R . (1971). M ultivariate data analysis. N ew York: W iley.
Courville, T., &.
Thompson,
B. (2001). Use of structure coefficients in published
multiple regression articles:
(3 is not
enough. Educational
and
Psychological
Measurement, 61,
229-248.
Cronkhite, G., & Liska, J. R . (1980). The judgment of communicant acceptabil i ty .
In M. R. Roloff & G. R. Miller (Eds.) , P ersuasion: N ew
directions
in theory an d
research (pp. 101-139). Beverly Hills, C A : Sage.
Grossman, L. L. (1996). Cross-validation ana lysis for the ca nonical case. In
B.
Thompson (Ed.) , Advances in social
science methodology
(Vol. 4, pp.
95-106).
Greenwich,
C T: JAI
Press.
Cudeck , R .
(1989).
The analysis of correlation mat rices using covariance s truc ture
models. Psychological
Bulletin, 105,
317-327.
184 R E FE R E NCE S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 175/185
Gumming,
G., & Finch, S. (2001). A primer on the understanding, use and calcula-
tion of confidence intervals
that
are based on central and noncentral d i s t r ibu-
tions. Educational
and
P sychological
Measurement, 61,
532-575.
Davis,
W. R.
(1993).
The FC1
rule
of
identification
for
confirmatory
factor
analysis:
A general sufficient condition.
Sociological Methods
and Research, 21,
403-437.
Diaconis,
P., &.
Efron,
B.
(1983). Computer-intensive methods
in
statistics. Scientific
American, 248(5),
116-130.
Duncan, O. D. (1975). Introduction
to
structural
equation
models. New York: Aca-
demic Press.
Dunlap,
W. P., &
Landis,
R. S.
(1998). Interpretations
of
multiple regression
borrowed
from factor
analysis
and canonical correlation. Journal
of
General
Psychology, 125, 397-407.
Fabr igar , L. R.,
Wegener,
D. T.,
MacCallum,
R. C., & Strahan, E. J.
(1999).
Evaluating
the use of exploratory
factor
analysis in psychological research.
Psychological Metho ds,
4,
272-299.
Fan,
X.,
Thompson,
B., &
Wang,
L.
(1999).
The
effects
of sample
size, estimation
methods,
and
model specification
on SEM fit
indices.
Structural Equation
Model-
ing, 6,
56-83.
Fan, X .,
Wang,
L., & Thompson, B . (1997, March). Effects
of data
nonnormality on
fi t indices and
parameter estimates
for true and
misspecified
SE M
models. Paper
presented at the annual meeting of the American Educational Research Associ-
ation, Chicago. (ERIC Document Reproduction Service
No. ED 408
299)
Fouladi,
R. T.
(2000). Performance
of modified
test statistics
in
covariance
and
correlation structure analysis under conditions of multivariate nonnormality.
Structural Equation
Modeling,
7,
356-410.
Frankiewicz,
R. G., &
Thompson,
B. (1979, April). A comparison of
strategies
for
measuring teacher
brinkmanship behavior. Paper presented
at the
annual meeting
of
the American Educational Research Association, San Francisco. (ERIC
Document Reproduction Service
No. ED 171
753)
Gorsuch,
R. L.
(1983). Factor analysis (2nd ed.). Hillsdale,
NJ:
Erlbaum.
Gorsuch,
R. L.
(2003). Factor analysis.
In ]. A.
Schinka
& W. F.
Velicer (Vol. Eds.),
Handbook
of
psychology: Vol. 2. Research methods in
psychology
(pp.
143-164).
Hoboken, NJ: Wiley.
Graham,
J. M.,
Guthrie,
A. C., 6k
Thompson,
B.
(2003). Consequences
of not
interpreting structure
coefficients
in published CFA research: A reminder.
Structural Equation Modeling, 10, 142-153.
Guil ford, J . P . (1946). N ew standards for test evaluation.
Educational and
Psychological
Measurement, 6, 427-439.
Guil ford, J . P .
(1967).
T h e
nature of human intelligence.
N ew
York: McGraw-Hill.
Guadagnoli, E., & Velicer, W. (1988). Relation of sample
size
to the stability of
component patterns.
Psychological
Bulletin, 103,
2 6 5 — 2 7 5.
Guttman,
L.
(1954). Some necessary conditions
for
common-factor analysis. Psycho-
metrika, 19, 149-161.
R E F E R E N C E S 185
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 176/185
Ha yduk ,
L . A. (1987). Structural equation modeling with
LISREL: Essent ia ls an d
advances.
Balt imore:
Johns Hopkins
U niv ersi ty 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.
Henson,
R . K .
(1999).
Mul t i va r i a t e normal i ty :
What
is it and how is it assessed?
In B. Thompson (Ed. ) ,
Advances
in social science methodology (Vol. 5, pp.
193-212). Stamford, CT: JAI Press.
Heywood,
H. B. (1931). On
fini te
sequences of
real numbers. Proceedings
of the
Royal
Statistical Society of London, 134,
486-501.
Holzinger, K. J., &. Swineford, F.
(1939).
A study in factor analysis: The
stability
of
a bi-factor solution
(No. 48).
Chicago:
U n i ve r s it y
of
Chicago Press.
Horn, J. L. (1965). A ra t ionale and test for the number of factors in factor analysis.
Psychometrika, 30,
179-185.
Horst , P . (1966). Psychological
measurement
an d
prediction,
Belmont, C A :
W adsworth .
Hu, L, &. Bender, P. M.
(1999).
C utoff cr i teria for
fit
indexes in
covariance
st ructure
analysis:
Conventional
cri ter ia versus
new
al terna t ives.
Structural Equation
Modeling, 6, 1-55.
Huher ty , C . (1994). Applied
discrim inant analysis.
N ew York: W i ley.
Huber ty , C. J., & Morris, J. D. (1988). A single
contrast test
procedure. Educational
and Psychological Measurement, 48, 567-578.
Jackson, D.
(2001). Sample size
and number of parameter
estimates
in ma x i mum
likelihood confirmatory factor analysis: A
Monte Carlo
invest iga t ion.
Structural
Equation Modeling,
8,
205-223.
Jones, H. L., Thompson, B., & M iller, A. H. (1980). How
teachers
perceive s imi lar i t -
ies and differences among v arious teaching models.
Journal
of Research in
Science
Teaching,
17, 321-326.
Joreskog, K. G. (1969). A general
approach
to confi rmatory maximum l ikel ihood
factor analysis.
Psychometrika,
34, 183-202.
Joreskog,
K. G., & Sorbom, D .
(1984).
LISREL VI user s
guide (3rd ed.). Mooresville,
IN :
Scientific Software.
Joreskog,
K. G., &
Sorbom,
D .
(1993).
LISREL 8:
User s
reference gu ide.
Chicago:
Scientific Software
International.
Kaiser, H. F.
(1958).
The
va r i ma x criterion
for
analyt ic rotation
in
factor analysis.
Psychometrika, 23, 187-200.
Kerlinger,
F. N.
(1979).
Behavioral research:
A
conceptual
approach.
N ew
York:
Holt,
Rinehart
&.
Winston.
Kerlinger, F. N. (1984).
Liberalism
an d
conservatism:
T he
nature
an d
structure
of social
attitudes.
Hillsdale,
N J:
Erlbaum.
Kerlinger, F. N.
(1986). F oundations
of behavioral
research
(3rd ed.). New York: Holt,
R i neh a r t & Winston.
186
REFERENCES
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 177/185
Knapp , T. R . (1978). Canonical correlation a nalysis: A general param etric signifi-
cance testing system.
Psychological
Bulletin, 85, 410-416.
Lautenschlager, G. F., Lance, C. E., & Flaherty , V. L. (1989). Parallel analysis
cri ter ia: R ev ised equa t ions fo r es t ima t ing the latent roots of random da ta
correla t ion matrices.
E ducational
an d
Psychological
Measurement, 49,
339-345.
Lev ine, M. S. (1977) . Canonical analysis an d factor comparison. Beverly Hills,
CA:
Sage.
Lunneborg,
C . E .
(2000). Data analysis
by
resampling:
Concepts and
applications.
Pacific Grove, CA: D uxbury.
MacCa l lum, R. C., Browne, M. W., & Sugawara , H. M. (1996). Power analysis and
determina t ion
of
sample
size for
covariance structural modeling.
Psychological
Methods, I, 130-149.
MacCa l lum, R. C., W i da m a n, K. F., Zhang, S., & Hong, S. (1999). Sample size
in factor analysis. Psychological Methods, 4, 84-99.
Marcoulides, G. A., &.
Schumacker,
R . E .
(2001) .
N ew
developments
an d
techniques
in structural equation modeling. Mahwah , N J: Er lbaum.
Mered i th , W . (1964)- Canonical correlations with fallible data .
Psychometrika,
29,
55-65.
Montanell i , R. G., Jr., & Humphreys , L. G. (1976 ). La tent roots of random data
correlation matrices with squared multiple correlations on the diagonal: A
Monte Carlo study. Psychometrika, 41,
341-348.
Mula ik ,
S. A . (Ed.) . (1992). Theme issue on prin cipa l components analysis. Multwar-
iate
Behavioral
Research, 27(3).
Mulaik ,
S. A. , Ja mes, L. R . , Van A lst ine, J . , Bennett , N. , Lind, S. , &. Sti l lwell ,
C . D. (1989). A n ev alua tion of goodness of fit indic es for stru ctur al equa tion
models. Psychological Bulletin, 105, 430-445.
Nasser , F., Benson, J. , & W isenb aker, J. (200 2 ). The performance of regression-
based variations of the v i sua l scree for determining the number of common
factors. Educational and Psychological
Measurement,
62,
397-419.
Nasser, F., & W isenbaker, J. (2003) . A M onte C arlo study invest igat ing the impact
of i tem parcel ing on measures of fit in confirmatory factor analysis.
E ducational
and Psychological
Measurement, 63, 729-757.
Nunna l l y ,
J . C .
(1978).
Psychometric
theory (2nd ed.).
N ew
York: McG raw-Hi l l.
O'Connor,
B. P.
(2000) .
SPSS
and SAS
programs
for
determining
the
number
of
components using parallel analysis a nd Velicer's M A P test. Behavior Research
Methods, Instruments, and Com puters, 32, 396-402.
Ogasawara, H. (2000). Some relationships between factors and components. Psycho-
metrika, 65, 167-185.
Pedhazur , E. J. (1982) . Multiple regression
in
behavioral research:
Explanation and
prediction
(2nd ed.).
N ew
York: H o l t, R ineha r t
&.
W ins ton .
R ussell, D . W .
(2002) .
In
search
of
underlying dimensions:
The use
(and abuse)
of
factor analysis in Personality and Social Psychology
BuUetin.
Personality and
Social
Psychology
Bulletin,
28,
1629-1646.
R E F E R E N C E S
187
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 178/185
Satorra , A., & Bender , P. M. (1994) .
Corrections
to test stat ist ics and s tandard
errors
in
covariance structural analysis.
In A. von Eye & C. C. Clogg
(Eds.) ,
Latent
variable
analysis: Applications fo r
developmental
research (pp.
399-419).
Thousand
Oaks,
CA:
Sage.
Schmid ,
J., &
Le i man ,
J.
( 1 9 5 7 ) .
The
development
of
hierarchica l factor solutions.
Psychometrika,
22,
53-61.
Schumacker , R . E. , & M arcoulides , G. A .
(1998).
Interaction
and
nonlinear
e f f e t s
in s tructural equation modeling.
M a h w a h ,
N J:
E r l b a u m .
Snook,
S. C., & Gorsuch, R. L. (1989). Component analysis versus common factor
analysis: A Monte Carlo study.
Psychological
Bulletin, 106, 148-154.
Spearman,
C.
(1904) .
General
intelligence, ob ject iv ely determined
and
mea sured .
American journal of Psychology, 15, 201— 293.
Steiger,
J. H., &.
Lind ,
J. C.
(1980, June). Statistically
based tests for the
number
o f
common
factors.
Paper presented
at the
annual meet ing
of the
Psychometr ic
Society,
Iowa
C i t y , 1A .
Stephenson,
W . (1953) .
T he study of behavior:
Q-technique
and its
m ethodology .
Chicago:
U n i ve r si t y
of Chicago
Press.
Ta t suoka ,
M. M.
( 19 7 1 ) . M ultwariate a nalysis: Techniques fo r
educational
an d
psycho-
logical research. New York : W i ley .
Thompson, B.
(1980a). Comparison
of two
strategies
for
collecting
Q-sort
da ta .
Psychological
Reports,
47, 547-551.
Thompson,
B. (1980b) . V ali di t y of an ev alu ato r typology. Ed ucational Ev aluation
and
Policy
Analysis , 2, 59-65.
Thompson,
B . (1984). C anonical
correlation
analysis:
Uses
an d interpretation.
N e w b u ry
Park ,
CA: Sage.
Thompson, B. (1988).
Program
FACSTRAP: A
program
that
computes boots t rap
estimates
of
factor s t ructure . Educational
an d
Psychological Measurement,
48,
681-686.
Thompson,
B. (1989). M eta-an alysis of
factor
stru cture studies: A case study exam ple
wi th Bern's androgyny mea sure. Journal
of Experimental
Education, 57,
187-197.
Thompson, B. (1990a). MULTINOR: A FORTRAN program
that
assists in e va lu a t -
in g
mul t i v a r i a te normal i ty . Educa tiona l
and Psychological
M easurement,
50,845-
848.
Thompson, B. (1990b).
SECONDOR:
A program that computes a second-order
pr incipal
components analys is and v ar iou s in terpreta t ion a ids . Educational and
Psychological
Measurement,
50,
575-580.
Thompson, B.
(1991).
A
pr imer
on the
logic
and use of
canonica l correlation
analysis. M easurement and Ev aluation in Counseling and D evelopment, 24, 80-95.
Thompson, B.
( 19 9 2a ) .
DISCSTRA: A
computer program that computes boots t rap
resampling es timates of descr ipt ive discr imina nt ana lys is funct ion and s t ructure
coefficients and
group centroids. Educational
an d Psychological
Measurement,
52,
905-911.
188 R E F E R E N C E S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 179/185
Thompson, B. (1992h). A partial test distribution for cosines among factors across
samples.
In B.
Thompson (Ed.),
Advances
in social science
methodology (Vol.
2 , .
pp. 81-98).
Greenwich,
CT: JAI
Press.
Thompson,
B.
(1993a). Calculation
of
standardized,
noncentered
factor scores:
An
alternative to conventional
fac tor
scores.
Perceptual and
Motor
S k i f f s ,
77, 1128-
1130.
Thompson,
B.
(1993b).
The use of
statistical significance tests
in
research: Bootstrap
and other alternatives,
journal
of Experimental
Education,
61, 361-377.
Thompson, B. (1994). The pivotal role of replication in psychological research:
Empi r i ca l ly
evaluating
th e
replicability
of
sample results. Journal of Personality,
62 ,
157-176.
Thompson, B. (1995). Exploring the replicability of a study's results: Bootstrap
statistics
for the multivariate case.
Educational and
Psychological
Measurement,
55, 84-94-
Thompson,
B.
(1996). AERA editorial policies regarding statistical significance
testing: Three suggested reforms.
Educational
Researcher, 25(2),
26-30.
Thompson,
B.
(1997).
The
importance
of
structure coefficients
in
structural equa-
tion modeling confirmatory factor analysis.
Educational and
Psychological
Mea-
surement,
57, 5-19.
Thompson, B.
(2000a). Canonical correlation analysis.
In L.
Grimm
& P.
Yarnold
(Eds . ) , Reading and understanding
more multivariate
statistics (pp. 285-316).
Washington,
DC:
American Psychological Association.
Thompson, B. (2000b). Q-technique factor analysis: One variation on the two-
mode
fac tor
analysis of variables. In L. Grimm & P. Yarnold (Eds.),
Reading
and understanding
more multivariate
statistics (pp. 2 0 7— 2 2 6 ) .
Washington, D C :
American Psychological Association.
Thompson,
B.
(2000c).
Ten
commandments
of
structural equation modeling.
In
L. Grimm & P. Yarnold (Eds.), Reading and understanding more
multivariate
statistics
(pp. 261-284). Washington,
DC:
American Psychological Associ-
ation.
Thompson, B. (2002). Statistical, practical, and clinical : How many kinds of
significance
do counselors
need
to consider? Journal
of Counseling Develop-
ment, 80,
64-71.
Thompson,
B .
(Ed.). (2003).
Score
reliability:
Contemporary
thinking on
reliability
issues.
Newbury Park,
CA:
Sage.
Thompson,
B., & Borrello, G. M. (1985). The importance of structure coefficients
in regression research.
Educational
and Psychological
Measurement,
45, 203-209.
Thompson, B., & Borrello, G. M. (1986). Second-order factor structure of the MBTI:
A construct
v a l i d i t y
assessment.
Measurement
and
Evaluation
in Counseling and
Development, 18, 148-153.
Thompson, B., &
Borrello,
G. M.
(1987, J an u a r y ) . Comparisons of
factors extracted
from the
correlation
matrix
versus
th e
covariance
matrix: A n
example using
th e Love
Relationships
Scale.
Paper presented at the annual meeting of the Southwest
REFERENCES
189
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 180/185
Educational Research Association, Dallas. (ERIC Document Reproduction
Service
No. ED 280
862)
Thompson, B.,
Cook,
C., &
Heath,
F.
(2001).
How
many dimensions does
it
take
to
measure users' perceptions
of
libraries?:
A
LibQUAL+™
study,
portal:
Libraries
and the
Academy,
1 ,
129-138.
Thompson, B., Cook, C., &
Heath,
F.
(2003). Structure
of
perceptions
of
service
qual i ty
in
libraries:
A
LibQUAL+™ study. Structural
Equation
Modeling,
10,
456-464.
Thompson, B., Cook, C., & Thompson, R. L.
(2002). Reliability
and
structure
of
LibQUAL+™ scores: Measuring perceived library service quality, portal: Lib rar-
ies and the Academy, 2 ,
3-12.
Thompson,
B., &. Daniel, L. G. (1996). Factor analytic evidence for the construct
va l id i ty of
scores:
A n
historical overview
an d
some guidelines. E ducational
a nd
Psychological Measurement,
56, 197-208.
Thompson, B., &.
Pi t t s ,
M. C .
(1982).
The use of
factor adequacy coefficients.
Journal
of
Experimental
Education, 50,
101— 104.
Thurstone, L. L. (1935). The vectors
of the
mind.
Chicago:
University of
Chicago
Press.
Thurstone, L. L.
(1947).
Multiple
factor analysis.
Chicago:
University
of
Chicago
Press.
Tucker,
L. R.
(1966).
Some
mathematical notes
on
three-mode factor analysis.
Psychomemka, 31, 279-311.
West, S. G., Finch, J. F., & Curran, P. J. (1995). Structural equation models with
nonnormal data. In R. H. Hoyle
(Ed.) ,
Structural
equation
modeling
(pp.
56-75).
Thousand Oaks,
CA: Sage.
Wilcox, R. R. (1998). How many discoveries have been lost by ignoring modern
s ta t is t ical methods? American Psychologist, 53, 300-314.
Wilkinson, L., & Task Force on Statistical Inference, APA Board of Scientific
Affairs. (1999). Statistical methods in psychology journals: Guidelines and
explanations. American
Psychologist,
54,
594-604.
[Reprint available
from
http://www.apa.org/journals/amp/amp548594.html]
Vacha-Haase, T. (1998). Reliability generalization: Exploring variance in measure-
ment error
affecting
score
rel iab i l i ty
across studies.
E ducational
and
Psychological
Measurement,
58,
6-20.
Vacha-Haase, T., Kogan, L. R., & Thompson, B. (2000). Sample compositions and
variabi l i t ies
in
published studies versus those
in
test manuals: Validity
of
score
re l iab i l i ty inductions. Educational
an d
Psychological
Measurement, 60 , 5 0 9— 5 2 2 .
Veldman, D. J. (1967). Fortran programming
for the
behavioral
sciences.
N ew York:
Holt, Rinehart
&
Winston.
Zwick, W . R ., &.
Velicer,
W . F . (1986).
Factors influencing
five rules for
determining
the number of components to retain. Psychological B ulletin, 99, 432-442.
190 R E F E R E N C E S
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 181/185
INDEX
ADF estimation theory. See A symptot i -
cally
distribution-free est ima tion
theory
Alpha factor extract ion, 38
AMOS,
111, 120, 131
Analysis of
var i ance (ANOVA) ,
12
Ana ly t i c purpose, 70
A nderson-R ubin est imation method,
4 4
Area world statistics, 11, 12
Asymptotically distribution-free (ADF)
estimation theory, 127, 13 8
Bartlett method,
58
Bell curves, 121-122
Beta (]3) weights, 16
Binet , Alfred, 3
Bivar ia te
normali ty,
122
Bootstrap
methods,
102-108
in CFA/SEM, 129-130
in EFA, 128-129
Canonical factor analysis, 38
CFA. See C onfirmatory factor analysis
C F1 (comparat ive fit index), 130
Characteristic values. Se e Eigenvalues
Communality coefficients, 19-21
in i t ia l estimates, 38, 51
as rel iabil i t ies , 99
as
R
2
v alues, 61-64
Comparat ive fi t index (CFI), 130
Composite variables (latent variables,
synthetic va riables), 9, 10
Confirmatory factor analysis (CFA),
5-7, 30,
133-151
alternative estimation theories in ,
138, 139
and comparisons of model fits, 142 ,
144-145
decision
sequence in, 109-132
EFA
vs., 110-114
and
higher-order models, 1 4 5 — 1 5 0
matrix of
association coefficients
in ,
120-121
maximum likelihood theory
in, 138
model identifica t ion in,
118-120,
134-137
mult ivar ia te
normali ty in,
121-123
outl ier screening in , 123—126
and parameter est imation theory,
126-127
pattern and
structure
coefficients in ,
138,
140-143
postanalysis decisions in , 127—131
preanalysis
decisions
in ,
114—127
Confirmatory
rotat ion,
93-95
Conformabili ty, 13-14, 25
Constructs,
5
Convergence, 37
C orrelated error v ari anc es, 140,
14 2
Correlat ion coefficient (r), 10-11
Correlat ions
Interpret by squar ing, 19
Covariance matrix,
30—31
C ovariance structure analysis.
Se e
Struc-
tural
equation
modeling (SEM)
Crit ical rat io (t test), 105, 123
Cross-validat ion, EFA, 101-103
Data
intervally
scaled,
29, 120
Data box (C at te l l) , 83-86
Delta
values,
44
Diagonal
of a
matr ix ,
14
Disconfirmability, 115-118
Division of matrices, 15, 25
EFA. See Exploratory factor analysis
Eigenvalue
greater than 1.0 rule (Kl
rule), 32
Eigenvalues
(character is tic v alues),
21-22
Elbow (scree plots ),
33
191
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 182/185
Equality constraints, 113, 14 7
Error(s)
correlated error variances, 140, 14 2
measurement, 50-52, 99
model specification, 99-100
root-mean-square error of approxima-
t ion ( R M S E A ) , 130
sampling,
50-52,
1 00
Est imates
admissible, 135
Estimation theory(-ies)
asymptotically distribution-free, 127,
138
maximum likelihood,
126—127
ordinary least squares (OLS), 138
Exploratory
factor analysis (E FA ),
5-7,9
bootstrap, EFA, 102-108
C FA vs., 110-114
cross-validat ion, EFA, 101-103
factor
extraction method in , 36—38
factor rotat ion in ,
38-44
factor scores
in , 44—47
linear sequence of decisions in,
27-28,
47-48
matrix of
association coefficients
in ,
28-31
number
of
factors
in , 31—36
orthogonal vs. oblique solutions in,
71
sample
size in , 2 4 — 2 5
Ex terna l replicab ility analyses, 100
Factor adequacy, 94
Factor analysis
development
of, 3
and
nature
of
constructs,
5
an d
parsimony,
5
purposes
of, 4-5
two major classes of, 5— 7
and val id i ty , 4 -5
Factor extraction
alpha,
38
image,
38
maximum likelihood, 38
Factor interpretation
naming of factors,
97—98
salience,
97
Factor rotation, 38-44. Se e
also
Oblique
rotations
Factors
naming of ,
97-98
number of (in EFA), 31-36
possible number
of, 17—18
reflection
of , 96-97
Factor scores,
16 ,
44-47
Anderson— R ubin , 4 4
Bart let t ,
58
in different extraction methods,
57-61
First-order
factors, 72
Fi t statistics, 127-130
General
linear
model (GLM), 7, 109
interpretat ions in , 7 1 — 7 2
weights in , 15—16
G
theory
of
intelligence,
3
Guilford, Joy, 4 -5
Heywood cases,
20
Higher-order factors
CFA, 145-150
EFA, 72-81
Identif ication,
model,
118-120,
134-137
Ident i ty matr ix , 14-15
111 conditioned
matrices,
15
Image factor extraction, 38
Inadmissabil i ty, 20
Independence model,
11 8
Intell igence,
3
Internal replicabili ty
methods,
99-108
bootstrap, EFA, 102-108
cross-validation, EFA, 101-103
external vs.,
100
and sampling error, 99-100
Invariance, test ing fo r model,
153-162
with same model, different esti-
mates, 153-156
with same model, equal estimates,
155,
157-159
with
same model, test invar iance
of
parameters,
159—161
Inverse,
matr ix ,
15
IQ tests, 3
Iterat ion,
37
192 I N D E X
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 183/185
Kl rule (eigenv alue greater than 1.0
rule) ,
32
Laten t var i ables . S ee
Composite v a r i a b l e s
L i b Q U A L + ™ s t u d y ,
28, 39
LISREL, 111
Loadings, 19
Mahalanobis dis tance, 124
M a t r i x
algebra, 9,
12-15,
25
Matrix(-ces) , 13-15
of association coeffi cients ,
28—31
comformable ,
13-14
covar iance ,
30—31
div i s ion
of, 15, 25
ident i ty , 14-15
ill
conditioned,
15
inverse of, 15
mul t ip l ica t ion
of ,
13—14,
25
symmetr ic , 14
transpose of, 13
M a x i m u m
likelihood factor extract ion,
38
Maximum likelihood theory, 126-127
M ean ,
11
Measured (ob served) var i ables , 9
Measurement error
defined,
99
sampl ing
error vs.,
5 0 — 5 2
M edian , 11
Mediated ranking, 88
Model
proof
of, 115
Model fit statistics,
127-130
Model specification error, 99-100
Model specif ica t ion searches, 131
Models
nested r iva l , 12 9
plausible, 115
Modes, 84
Modif ica t ion indices ,
131
Monte Car lo s imula t ions , 24, 128
Mult ip l ica t ion of matrices , 13-14, 25
Multivariate
Behav ioral Research, 53
M ul t i v a r i a t e
no rma l i t y ,
121-123
Mult ivoca l models , 142
N a m i n g of factors , 97-98
N F1
(normed fit index), 129
N orma l i t y
b i v a d a t e , 1 2 2
m ul t i v a r i a t e , 1 2 1 — 1 2 3
univar i a te ,
121-122
Normed fi t index (NF1), 129
O bl imin ,
44
Oblique rotat ions , 42-44, 67-81
and higher-orde r factors, 72-81
a nd m ax im iz in g sample f i t vs.
replicabi l i ty,
69-72
Observed variables . S ee Measured
va r i ab les
Occam's razor, 70-71
OLS.
Se e
Ordinary leas t squares
One-factor model,
118
Ordina ry leas t squares (OLS), 126, 138
Orthogonal
ro ta t ion , 42, 67
O-technique
factor analysis , 85
Outlier screening, 123-126
Para l le l
analysis ,
34-36
Parameter es t imat ion theory,
126—127
Parcels
I t em, 123
Parsimony, 5, 70-71
Pars imony-adjusted s ta t is t ics ,
130
Pa t tern coefficients
in
CFA, 138, 140-143
in GSM , 16-17
Pa t tern /s t ruc ture
coefficients, 19
PD A (predic t ive d i sc iminant ana lys i s ) , 71
Pearson correlation coefficients, 10, 61
Pearson K
factors affecting,
30, 124
Pearson product-moment biv aria te
corre la t ion mat r ix , 29
Pearson
r, 2 9 — 3 1
Pencil test, 33
Pos tmul t ip l ica t ion , 15
Predic t ive d i sc iminant a na lys i s (PD A ) ,
71
Principal-axes factor analysis ,
37
i terat ion process for ,
51
principal-components analysis vs . ,
53-55
Pr inc ipa l components ana lys i s , 36-38
pr inc ipa l axes analysis vs . , 5 3 — 5 5
Procrustean rotat ion,
43-44
I N D E X 193
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 184/185
Promax rotat ion, 43-44
P-technique factor analysis,
Q
sorts,
87-88
Q-technique
factor analysis, 84 ,
data collection
in,
87-88
uses of, 86
6-91
R .
See
Correlation coefficient
Rank/ranking,
13, 88
R eflection
of
factors,
96-97
Regression method,
44
Rel iab i l i ty
coefficients,
12, 99
Rel iab i l i ty generalizat ion (RG) analysis,
38
Repl icab i l i ty
analysis
external,
100
internal,
99-108
R eproduced correlat ion ma trix,
16-17,
22-23
Resamples,
10 3
Residua l
correlation matrix,
23-24
inspection
of , 33-34
R G
analysis ,
38
R oot-mean-square error
of
approximation
( R M S E A ) , 130
R otat ion, factor.
See
Factor
rotation
R -techruque factor analysis,
84
Salience
of variables, 97
Sampling error, 100
measurement
error
vs., 50-52
Schmid-Leiman solution, 74
Scores,
factor.
See
Factor scores
Score
world statistics,
11
Scree
test,
32-33
SD (standard deviat ion), 11
Second-order factors,
72
SEM.
Se e
Structural equation modeling
Sequent ia l
factor extraction,
22-24
Simple structure, 41
Spearman's rho, 29, 30
Specific factors, 3
SPSS,
22, 36, 44, 45, 51, 58, 61-63, 73 ,
74-81, 85
Standard deviat ion
(SD), 11
Standard errors.
Se e
Error(s)
Standardized regression coefficients, 1 6
Statist ical significance tests, 31-32
S-technique factor ana lysis, 85
Structural equat ion modeling (SEM),
109-110
Structural
models, 110
Structure coefficients
in
CFA, 138, 140-143
in EFA, 61
in GSM, 18-19
Superscripts, 11
Suppressor effect,
72
Symmetric matrix,
14
Synthetic variables.
Se e
C omposi te
variables
Third-order factors,
72
Three-mode analysis, 84
Transapose(s), 13
T-technique factor analysis, 85
t test. See
C ri t ical rat io
Two-mode analysis,
83-91
O-technique, 85
P-technique,
84
Q-technique, 84, 86-91
R -technique ,
84
S-technique,
85
T-technique, 85
Uncorrelated factors model,
118
Uniqueness,
74
Unit weight ing, 70
Univar ia te normal i ty , 1 2 1 — 1 2 2
Univocal models, 140, 14 1
Val idi ty ,
4-5
Variables
composite (latent, synthetic), 9, 10
measured
(observed),
9
number
of, 10, 49-50
salience
of, 97
Variance, 11
Visual
scree test,
33
W ald s ta t is t i c ,
105,
123
W eights/weight ing,
9, 15-16
b e t a ( W , 16
uni t ,
70
194
7NDEX
8/19/2019 Bruce Thompson-Exploratory and Confirmatory Factor Analysis_ Understanding Concepts and Applications-America…
http://slidepdf.com/reader/full/bruce-thompson-exploratory-and-confirmatory-factor-analysis-understanding 185/185
ABOUT
THE
AUTHOR
Bruce
Thompson
is a professor and distinguished research scholar at the
Department of Educational Psychology at Texas A&M University, ad junc t