See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/209835984

Evaluating the use of Exploratory Factor Analysis in psychological research

Article in Psychological Methods · January 1999

DOI: 10.1037//1082-989X.4.3.272

CITATIONS READS
2,756 7,372

4 authors, including:

Leandre R Fabrigar Duane T. Wegener

Queen's University The Ohio State University
100 PUBLICATIONS 11,560 CITATIONS 167 PUBLICATIONS 24,653 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Leandre R Fabrigar on 15 May 2014.

The user has requested enhancement of the downloaded file.

Psychological Methods Copyright 1999 by (he American Psychological Association, Inc.
1999, Vol.4. No. 3.272-299 1082-989X/99/S3.00

Evaluating the Use of Exploratory Factor Analysis in

Psychological Research
Leandre R. Fabrigar Duane T. Wegener
Queen's University Purdue University

Robert C. MacCallum Erin J. Strahan

Ohio State University Queen's University

Despite the widespread use of exploratory factor analysis in psychological research,

researchers often make questionable decisions when conducting these analyses.
This article reviews the major design and analytical decisions that must be made
when conducting a factor analysis and notes that each of these decisions has
important consequences for the obtained results. Recommendations that have been
made in the methodological literature are discussed. Analyses of 3 existing em-
pirical data sets are used to illustrate how questionable decisions in conducting
factor analyses can yield problematic results. The article presents a survey of 2
prominent journals that suggests that researchers routinely conduct analyses using
such questionable methods. The implications of these practices for psychological
research are discussed, and the reasons for current practices are reviewed.

Since its initial development nearly a century ago of factor analysis or Tom Swift and his electric factor
(Spearman, 1904, 1927), exploratory factor analysis analysis machine," argued that factor analysis had
(EFA) has been one of the most widely used statistical limited utility for aiding in the development of theory,
procedures in psychological research. Despite this because it could not be relied on to provide meaning-
long history and wide application, the use of factor ful insights into data.1 He attempted to demonstrate
analysis in psychological research has often been this point by creating artificial data with a known
criticized. Some critics have raised concerns about structure and then ostensibly showing that EFA failed
fundamental limitations of factor analysis for contrib- to accurately represent the structure. 2 Other critics
uting to theory development (e.g., Gould, 1981; Hills, have not challenged the fundamental utility of EFA
1977; Overall, 1964). For instance, Armstrong (1967), but have instead criticized the manner in which it is
in an article entitled "Derivation of theory by means

Leandre R. Fabrigar and Erin J. Strahan, Department of

1
Psychology, Queen's University, Kingston, Ontario, The name Tom Swift refers to a character in a popular
Canada; Duane T. Wegener, Department of Psychology, series of juvenile science fiction novels published in the
Purdue University; Robert C. MacCallum, Department of 1960s. In each novel, Tom Swift makes use of futuristic
Psychology, Ohio State University. devices with near-miraculous powers. Armstrong's refer-
Erin J. Strahan is now at Department of Psychology, Uni- ence to this character in the title of his article highlighted
versity of Waterloo, Waterloo, Ontario, Canada. what he regarded as the naive belief by researchers in fun-
This research was supported by a grant from the Social damental utility of EFA.
2
Sciences and Humanities Research Council of Canada and It is important to note that Armstrong (1967) was not
Grant PO1-MH/DA56826 from the National Institutes of the first person to examine the effectiveness of EFA proce-
Health. We thank Stephen G. West for his comments on an dures using data sets with a known underlying structure
earlier version of this article. (e.g., see Thurstone, 1947; Cattell & Dickman, 1962; Cattell
Correspondence concerning the article should be ad- & Sullivan, 1962; Cattell & Jaspers, 1967). Interestingly, in
dressed to Leandre R. Fabrigar, Department of Psychology, these other cases, the authors concluded that appropriate
Queen's University, Kingston, Ontario, K7L 3N6, Canada. EFA procedures were reasonably effective in revealing the
Electronic mail may be sent to [email protected]. known underlying structure of the data.

272
 USE OF FACTOR ANALYSIS 273

sometimes applied (e.g., Comrey, 1978; Ford, Mac- lum, Widaman, Zhang, & Hong, 1999; Velicer &
Callum, & Tail, 1986; Gorsuch, 1983; McNemar, Fava, 1998; Weiss, 1976). To the extent that a re-
1951; Skinner, 1980). In this article, we primarily searcher makes poor decisions, the analysis is more
address the latter issue. That is, we explore the man- likely to provide misleading results. However, re-
ner in which factor analysis is applied in psychologi- searchers often appear to be unaware of the issues
cal research and evaluate the soundness of current involved in these decisions.
practices. However, we contend that these two issues
are intertwined. The utility of factor analysis for Decisions in Conducting an EFA
theory development is dependent on the manner in Study design. As with any statistical procedure,
which it is implemented (see Cattell, 1978; Comrey, the utility of the results obtained in EFA is in large
1978). Furthermore, we suggest that some critics who part determined by the soundness of the design of the
have questioned the fundamental value of factor study from which the data are collected. Within the
analysis have not been sufficiently sensitive to this context of EFA, one design issue that is especially
relationship. important is what measured variables to include in the
We begin our discussion by reviewing some of the study (Cattell, 1978). If a researcher inadequately
major methodological decisions that researchers must samples measured variables from the domain of in-
make when conducting a factor analysis. Next, we terest, he or she may fail to uncover important com-
illustrate with published data sets how poor choices mon factors. Conversely, if measured variables irrel-
when making these decisions can substantially distort evant to the domain of interest are included, then
the results. We then turn our attention to the extent to spurious common factors might emerge or true com-
which current use of factor analysis reflects sound mon factors might be obscured. Therefore, research-
practice-. We conclude with discussions of the impli- ers should carefully define their domain of interest
cations of current factor analytic practices for psycho- and specify sound guidelines for the selection of mea-
logical theory and reasons for the prevalence of cer- sured variables.
tain practices. Research suggests that EFA procedures provide
more accurate results when each common factor is
Methodological Issues in the Implementation of represented by multiple measured variables in the
Factor Analysis analysis (i.e., when common factors are "overdeter-
mined"; MacCallum et al., 1999; see also Velicer &
Perhaps more than any other commonly used sta- Fava, 1998). Methodologists have recommended that
tistical method, EFA requires a researcher to make a at least three to five measured variables representing
number of important decisions with respect to how the each common factor be included in a study (MacCal-
analysis is performed (see Finch & West. 1997). Spe- lum et al., 1999; Velicer & Fava, 1998). Thus, when
cifically, there are at least five major methodological designing studies for which EFA is likely to be used,
issues that a researcher should consider when con- a researcher should consider the nature and number of
ducting a factor analysis. First, he or she must decide common factors he or she expects might emerge. The
what variables to include in the study and the size and total number of measured variables included should
nature of the sample on which the study will be based. be at least 3 to 5 times the number of expected com-
Second, a researcher must determine if EFA is the mon factors, and the selected variables should include
most appropriate form of analysis given the goals of multiple variables likely to be influenced by each of
the research project. Third, assuming that EFA is ap- the common factors. Alternatively, in cases in which
propriate, a specific procedure to fit the model to the there is little or no basis to anticipate the number and
data must be selected. Fourth, the researcher must nature of common factors, a researcher should attempt
decide how many factors should be included in the to delineate as comprehensively as possible the popu-
model. Finally, it is usually necessary for a researcher lation of measured variables for the domain of inter-
to select a method for rotating the initial factor ana- est. He or she should then include in the study a
lytic solution to a final solution that can be more sample of these measured variables that is as large as
readily interpreted. Each of these decisions can have feasible (see Cattell, 1978).
important consequences for the results obtained (see Sound selection of measured variables also requires
Armstrong & Soelberg, 1968; Cattell, 1978; Comrey, consideration of psychometric properties of measures.
1978; Ford et al., 1986; MacCallum, 1983; MacCal- When EFA is conducted on measured variables with
274 FABRIGAR, WEGENER, MAcCALLUM, AND STRAHAN

low communalities (i.e., variables for which the com- al., 1999). However, under more moderate conditions
mon factors explain little variance), substantial distor- a sample size of at least 200 might be needed; when
tion in results can occur (MacCallum et al., 1999; these conditions are poor it is possible that samples as
Velicer & Fava, 1998). There are a number of reasons large as 400 to 800 might not be sufficient.
why communalities for measured variables might be It is worth noting that obtaining parameter esti-
low. One obvious reason is low reliability. As ex- mates that closely approximate population values is
plained later, variance due to random error cannot, by only one criterion a researcher might consider when
definition, be explained by common factors. Because determining sample size. In some situations, addi-
of this, variables with low reliability will have low tional concerns might also play a role. Most notably,
communalities and thus should be avoided. A second when EFA involves the testing of formal hypotheses
reason why a variable might have a low communality regarding model fit or parameter estimates (as is
is that the variable is unrelated to the domain of in- sometimes done in maximum likelihood [ML] EFA),
terest and thus shares little in common with other statistical power might also be considered. A re-
measured variables in that domain. Therefore, to the searcher could specify a hypothesis of interest, a de-
extent such information is available, a researcher sired level of power, and an assumed population value
should consider the validity (e.g., face validity, con- for model fit. The sample size necessary to achieve
vergent validity) of measured variables when select- these objectives can then be calculated (see MacCal-
ing items to include in the analysis. lum, Browne, & Sugawara, 1996).
A second important design decision is the selection Researchers should also consider the nature of the
of the sample. A researcher must determine how large sample on which the study is based. Psychologists
the sample should be and how that sample will be often select samples based on convenience. In many
selected from the population of interest. Methodolo- cases, this practice will not pose a problem. However,
gists have proposed a host of rough guidelines for if the sample is considerably more homogeneous than
estimating an adequate sample size for an EFA. Most the population on the common factors, this can lead to
of these guidelines involve determining sample size restriction of range in the measures, thereby attenuat-
based on the number of measured variables included ing correlations among variables. This attenuation can
in the analysis—with more measured variables requir- result in falsely low estimates of factor loadings and
ing larger sample sizes. Sometimes such guidelines correlations among factors (see Comrey & Lee, 1992;
also specify a minimum necessary sample size regard- Gorsuch, 1983; Tucker & MacCallum, 1997). Addi-
less of the number of measured variables. tionally, selection biases related to a single measured
Unfortunately, there are serious drawbacks to such variable in the analysis can also distort results (Tucker
guidelines. One problem is that these recommenda- & MacCallum, 1997). For these reasons, researchers
tions vary dramatically. For instance, Gorsuch (1983) should consider the nature of the measured variables
suggested a ratio of 5 participants per measured vari- they are investigating and the manner in which their
able and that the sample size never be less than 100. sample is selected. Overly homogeneous samples and
In contrast, Nunnally (1978) and Everitt (1975) samples whose selection is related to measured vari-
proposed ratios of 10 to 1. More important, recent ables in the analysis should be avoided. Thus, when
research has suggested that such guidelines are not there is a substantial basis to expect that convenience
sufficiently sensitive to a variety of important char- samples will not be appropriate, a researcher should
acteristics of the data (MacCallum et al., 1999; consider obtaining a sample representative of the
Velicer & Fava, 1998). The primary limitation of such population of interest. Alternatively, a researcher
guidelines is that adequate sample size is not a func- might wish to select a sample to maximize variance
tion of the number of measured variables per se but is on measured variables relevant to the constructs of
instead influenced by the extent to which factors are interest and minimize variance on measured variables
overdetermined and the level of the communalities of irrelevant to the constructs of interest (see Cattell,
the measured variables. When each common factor is 1978).
overdetermined (i.e., at least three or four measured Determining whether EFA is appropriate. The
variables represent each common factor) and the com- primary purpose of EFA is to arrive at a more parsi-
munalities are high (i.e., an average of .70 or higher), monious conceptual understanding of a set of mea-
accurate estimates of population parameters can be sured variables by determining the number and nature
obtained with samples as small as 100 (MacCallum et of common factors needed to account for the pattern
 USE OF FACTOR ANALYSIS 275

of correlations among the measured variables. That is, measured variables by estimating the pattern of rela-
EFA is used when a researcher wishes to identify a set tions between the common factor(s) and each of the
of latent constructs underlying a battery of measured measured variables (i.e., as indexed by factor load-
variables. Before using EFA. a researcher should ings).
carefuily consider if this is a goal of the research In contrast, PCA does not differentiate between
projec. common and unique variance. Rather, this approach
In reaching this decision, it is important to recog- defines each measured variable as a linear function of
nize that the goal of identifying latent constructs (i.e., principal components, with no separate representation
understanding the structure of correlations among of unique variance. Mathematically, these principal
measured variables) is different from that of data re- components can be defined as linear composites of the
duction. Data reduction involves taking scores on a original measured variables and thus contain both
large set of measured variables and reducing them to common and unique variance. Therefore, principal
scores on a smaller set of composite variables that components are not latent variables, and because of
retain as much information from the original variables this it is not conceptually correct to equate them with
as possible. Data reduction does not attempt to model common factors (although in practice researchers of-
the structure of correlations among the original vari- ten do so). Furthermore, whereas the goal of common
ables. This distinction is important, because different factor analysis is to explain correlations among mea-
methods have been designed to achieve these two sured variables, the goal of PCA is to account for
objectives. If the goal is to arrive at a parsimonious variance in the measured variables. That is, the ob-
representation of the associations among measured jective of PCA is to determine the linear combinations
variables, EFA can be an appropriate form of analysis. of the measured variables that retain as much infor-
If the goal is data reduction, principal components mation from the original measured variables as pos-
analysis (PCA) is more appropriate. Many researchers sible. Thus, although PCA is often referred to and
mistakenly believe that PCA is a type of EFA when in used as a method of factor analysis, it is not factor
fact these procedures are different statistical methods analysis at all.
designed to achieve different objectives (for a discus- Nevertheless, some methodologists have argued
sion of the distinction, see Bentler & Kano, 1990; that PCA is a reasonable substitute for analyses of
Bookstein, 1990; Gorsuch, 1990: Loehlin, 1990; common factors and might even be superior (e.g., see
McArdie, 1990; Mulaik, 1990; Rozeboom. 1990; Velicer & Jackson, 1990a, 1990b). They noted that
Schonemann, 1990; Steiger, 1990; Velicer & Jackson, PCA is computationally simpler than common factor
1990a, 1990b; Widaman, 1990). analysis and therefore requires less computer memory
EFA is based on the common factor model (Thur- and processing time (e.g., Velicer & Jackson, 1990a).
stone, 1947). This model postulates that each mea- They also argued that the two approaches generally
sured variable in a battery of measured variables is a produce very similar results (e.g.. Arrindell & van der
linear function of one or more common factors and Ende, 1985; Velicer, 1977; Velicer & Jackson, 1990a;
one unique factor. Common factors are unobservable Velicer, Peacock, & Jackson, 1982). Also, they noted
latent variables that influence more than one mea- that common factor analysis procedures sometimes
sured variable in a battery and are presumed to ac- produce "Heywood cases" (e.g., Velicer & Jackson,
count for the correlations (covariances) among the 1990a)—situations in which a communality for a
measured variables (i.e., two measured variables are measured variable (i.e., the proportion of variance in
assumed to be correlated, because they are influenced the measured variable accounted for by the common
by one ;>r more of the same common factors). Unique factors) is estimated to be at 1 or greater than 1. Be-
factors are latent variables that influence only one cause it is impossible to account for more than 100%
measured variable in a battery and do not account for of the variance in a variable, such an estimate is po-
correlations among measured variables. Unique fac- tentially problematic. Finally, advocates of PCA note
tors are assumed to have two components: a specific that it is determinate, whereas the common factor
factor component (i.e., systematic latent factors that model is not (e.g., Steiger, 1979, 1990; Velicer &
influence only one measured variable) and an error of Jackson, 1990a). That is, it is possible to compute an
measurement component (i.e., unreliability in a mea- individual person's score on a principal component,
sured variable). The goal of the common factor model whereas it is not possible to do so for a common
is to understand the structure of correlations among factor.
276 FABRIGAR. WEGENER, MAcCALLUM, AND STRAHAN

There are, however, reasons to question these ar- lates, predictors, or consequences of other variables
guments. Advances in the speed and memory capa- and obtain estimates of the relevant parameters with-
bilities of computers have made the advantage of the out ever estimating factor scores. Additionally, recent
computational simplicity of PCA trivial. Also, al- developments have provided useful nonstructural
though PC A and EFA do often produce similar re- equation modeling procedures for estimating the cor-
sults, there are some contexts in which this is not the relations between factors and other variables without
case (e.g., see Bentler & Kano, 1990; Borgatta, the need to compute factor scores (Gorsuch, 1997).
Kercher, & Stull, 1986; Gorsuch, 1988, 1990; Hub- Finally, advocates of common factor analysis note
bard & Allen, 1987; McArdle, 1990; Snook & Gor- that it has certain advantages over PCA when the goal
such, 1989; Tucker, Koopman, & Linn, 1969; Wida- is to identify latent constructs. They argue that most
man, 1990, 1993). Differences in results are most measures used in psychological research contain some
likely when communalities are low (e.g., .40) and random error. Because EFA procedures reflect a rec-
there are a modest number of measured variables ognition of this fact, whereas PCA does not, the com-
[e.g., three) per factor (Widaman, 1993). Furthermore, mon factor model is a more realistic model of the
when the data correspond to assumptions of the com- structure or correlations (e.g., see Bentler & Kano,
mon factor model, EFA produces more accurate re- 1990; Gorsuch, 1973; Loehlin, 1990). Additionally,
sults than PCA (e.g., McArdle, 1990; Snook & Gor- the common factor model is testable, whereas the
such, 1989; Tucker et al., 1969; Widaman, 1990, PCA model is not (e.g., see Bentler & Kano, 1990;
1993). In contrast, when the data are relatively con- McArdle, 1990). The common factor model specifies
sistent with the assumptions of PCA (e.g., little certain hypotheses about the data. Thus, it can be fit to
unique variance present in measured variables), ex- data and the model rejected if the fit is poor. In con-
traction of common factors does as well as extraction trast, because PCA does not involve a specific hy-
of principal components (e.g., see Gorsuch, 1990; pothesis to be tested, it does not provide information
McArdle, 1990; Velicer et a!., 1982). on which one could base a decision to reject the
The occasional occurrence of Heywood cases "model."
should also not be regarded as a flaw of common Therefore, there are clear conceptual distinctions
factor analysis. Heywood cases often indicate that a between PCA and EFA. Although these approaches
misspecified model has been fit to the data or that the often produce similar results, this is not true in certain
data violate assumptions of the common factor model contexts. When the goal of the analysis is to identify
(van Driel, 1978). Thus, Heywood cases can have latent constructs underlying measured variables, it is
diagnostic value (Velicer & Jackson, 1990a). In con- more sensible to use EFA than PCA (see Cattell,
trast, because such cases do not occur in PCA, such 1978; Gorsuch, 1983; McDonald, 1985; Mulaik,
problems are not solved but simply go unnoticed 1972).
; McArdle, 1990).3 Assuming that identifying latent variables that ac-
Indeterminacy of individual factor scores in the count for the correlations among measured variables
common factor model need not be a problem. In most is the goal of the research project, a researcher must
applications of EFA, this issue is irrelevant in that the then decide if an exploratory or confirmatory ap-
objective is to identify common factors that account proach will be used. Both EFA and confirmatory fac-
ror the structure of the correlations among the mea- tor analysis (CFA) are based on the common factor
sured variables. This goal does not require the com- model, and both seek to represent the structure of
putation of factor scores but rather only factor load- correlations among measured variables using a rela-
ings and factor intercorrelations (McArdle, 1990). In
those situations where factor scores might be of in-
terest, researchers usually obtain scores for the pur- 1
Furthermore, there might be some cases in which a
pose of assessing correlations of factors with other
Heywood case would not represent a misspecified model
variables, or using factors as independent variables or and might not be problematic. For example, circumstances
dependent variables in regression models. However, in which population parameters are close to logical bound-
the development of structural equation modeling aries (e.g., if true communalities are close to 1.0) could
makes it unnecessary to estimate factor scores to ob- produce Heywood cases simply because of expected vari-
tain such information. One can use structural equation ability in parameter estimates resulting from sampling error
modeling to specify a model with factors as corre- (van Driel, 1978).
 USE OF FACTOR ANALYSIS 277

lively small set of latent variables. However, EFA is mon factor model to the data. A number of model-
primarily a data-driven approach. No a priori number fitting methods (i.e., factor-extraction procedures) are
of common factors is specified and few restrictions available. Most widely used among these are ML,
are placed on the patterns of relations between the principal factors with prior estimation of communali-
common factors and the measured variables (i.e., the ties, and iterative principal factors. These are different
factor loadings). EFA provides procedures for deter- methods for estimating the parameters (factor load-
mining an appropriate number of factors and the pat- ings and unique variances) of the same model, the
tern of factor loadings primarily from the data. In common factor model.4
contrast, CFA requires a researcher to specify a spe- Although these procedures fit the same model, each
cific number of factors as well as to specify the pat- method does have certain advantages and disadvan-
tern of zero and nonzero loadings of the measured tages. The primary advantage of ML is that it allows
variables on the common factors. Alternatively, a re- for the computation of a wide range of indexes of the
searcher might specify a priori a small set of compet- goodness of fit of the model. ML also permits statis-
ing models postulating differing numbers of factors, tical significance testing of factor loadings and corre-
different patterns of factor loadings, or both. lations among factors and the computation of confi-
In situations in which a researcher has relatively dence intervals for these parameters (Cudeck &
little theoretical or empirical basis to make strong O'Dell, 1994). The primary limitation of ML estima-
assumptions about how many common factors exist or tion is its assumption of multivariate normality. When
what specific measured variables these common fac- this assumption is severely violated, this procedure
tors are likely to influence, EFA is probably a more can produce distorted results (Curran, West, & Finch,
sensible approach than CFA. EFA is likely to be more 1996; Hu, Bentler, & Kano, 1992). On the other hand,
desirable in these situations, because the number of principal factors methods (both iterated and noniter-
plausible alternative models might be so large that it ated) have the advantage of entailing no distributional
would be impractical to specify and test each one in assumptions. Principal factors are also less likely than
CFA. Additionally, when a strong basis does not exist ML to produce improper solutions (i.e., a solution
for identifying a single model or a few specific com- with a Heywood case or a solution that fails to con-
peting models, it is quite possible that a researcher verge on a final set of parameter estimates; Finch &
might fail to identify a number of plausible models. West, 1997). The major drawback of the principal
Therefore, in these contexts, the data-driven approach factor methods is that they provide a much more lim-
of EFA seems advisable. ited range of goodness-of-fit indexes and generally do
However, when there is sufficient theoretical and not allow for computation of confidence intervals and
empirical basis for a researcher to specify the model significance tests. Regardless of these differences,
or small subset of models that is the most plausible, when the common factor model holds reasonably well
CFA is likely to be a better approach. This is because in the population and severe violations of distribu-
CFA allows for focused testing of specific hypotheses tional assumptions are not present, solutions provided
about the data (e.g., see Finch & West, 1997; Wege- by these methods are usually very similar.
ner & Fabrigar, in press). Also, the a priori nature of Selecting the number of factors. Determining how
CFA makes it less likely that a researcher will capi- many factors to include in the model requires the
talize on chance characteristics in the data. researcher to balance the need for parsimony (i.e., a
It is also often useful to use EFA and CFA in con- model with relatively few common factors) against
junction with one another. An EFA can be conducted the need for plausibility (i.e., a model with a sufficient
in an initial study to provide a basis for specifying a number of common factors to adequately account for
CFA model in a subsequent study. Alternatively, if
the sample size in a single study is sufficiently large,
the sample could be randomly split in half. An EFA
could then be conducted on one half of the data pro- 4
Although the common factor model makes a further
viding the basis for specifying a CFA model that can conceptual distinction by decomposing unique variance into
be fit to the other half of the data. specific variance and error variance, methods of fitting the
Choice of model-fitting procedure. If EFA is the common factor model allow for estimates of the amount of
most appropriate form of analysis, it is then necessary unique variance only and do not provide separate estimates
to decide what procedure will be used to fit the com- of specific variance and error variance.
278 FABRIGAR, WEGENER, MAcCALLUM, AND STRAHAN

the correlations among measured variables). In other of factors. A number of procedures for answering this
words, the goal of the researcher is to determine the question have been proposed. Perhaps the best known
number of "major" factors underlying a battery of of these procedures is the Kaiser criterion of comput-
measures. Importantly, errors in selection of the num- ing the eigenvalues for the correlation matrix to de-
ber of factors in a model can have a substantial effect termine how many of these eigenvalues are greater
on the results obtained (e.g., Comrey. 1978; Fava & than 1 (for discussion of the nature of eigenvalues, see
Velicer, 1992; Levonian & Comrey, 1966; Wood, Ta- Gorsuch, 1983). This number is then used as the num-
taryn, & Gorsuch, 1996). ber of factors. Although this procedure is appealing
Traditionally, methodologists have regarded speci- for its simplicity and objectivity, the approach has
fying too few factors in a model (i.e., underfactoring) significant problems. First, it is often misapplied by
as a much more severe error than specifying too many referring to the eigenvalues of the correlation matrix
factors (i.e., overfactoring; see Cattell, 1978; Rum- with communality estimates in the diagonal (i.e., the
mel, 1970; Thurstone, 1947). Empirical research has reduced correlation matrix) rather than eigenvalues of
generally supported this notion. When too few factors the correlation matrix with unities in the diagonal
are included in a model, substantial error is likely (i.e., the unreduced correlation matrix; see Guttman.
(Fava & Velicer, 1992; Wood et al., 1996). Measured 1954; Kaiser, 1960). Application of this rule to the
variables that load on factors not included in the eigenvalues of the reduced correlation matrix is an
model can falsely load on factors included in the erroneous procedure (Gorsuch, 1980; Horn, 1969).
model, and poor estimates of the factor loadings can Second, as with any mechanical rule, the procedure
he obtained for measured variables that do actually can to some extent be arbitrary. For instance, it is not
load on the factors included in the model (Wood et al., really meaningful to claim that a common factor with
1996). Such distortions can result in rotated solutions an eigenvalue of 1.01 is a "major" factor whereas a
in which two common factors are combined into a common factor with an eigenvalue of 0.99 is not.
single common factor (thereby obscuring the true fac- Finally, in numerous studies involving both principal
i.or structure) and in solutions with complex patterns components and common factors, this procedure has
of factor loadings that are difficult to interpret (see been demonstrated to lead to substantial overfactoring
Comrey. 1978). and occasionally to underfactoring (Cattell & Jaspers,
Empirical research suggests that overfactoring in- 1967; Cattell & Vogelmann, 1977; Hakstian, Rogers.
troduces much less error to factor loading estimates & Cattell, 1982; Linn, 1968; Tucker et al., 1969;
ihan underfactoring (Fava & Velicer. 1992; Wood et Zwick & Velicer, 1982, 1986). In fact, we know of no
..;!., 1996). Such models often result in rotated solu- study of this rule that shows it to work well.
i ions in which the major factors are accurately repre- Another widely known approach for determining
sented and the additional factors have no measured the number of factors is the "scree test" (Cattell, 1966;
variables that load substantially on them or have only Cattell & Jaspers, 1967). In this procedure, the eig-
a single measured variable that loads substantially on envalues of the correlation matrix (or the reduced cor-
each additional factor. Nonetheless, overfactoring relation matrix) are computed and then plotted in or-
should be avoided (see Comrey, 1978; Comrey & der of descending values. This graph is then examined
Lee, 1992). Solutions with too many factors might to identify the last substantial drop in the magnitude
prompt a researcher to postulate the existence of con- of the eigenvalues. A model with the same number of
s tracts with little theoretical value and thereby de- common factors as the number of eigenvalues prior to
velop unnecessarily complex theories. Additionally, this last substantial drop is then fit to the data. This
overfactoring can accentuate poor decisions made at procedure is often used with eigenvalues from the
other steps in a factor analysis. For example, PCA unreduced correlation matrix. However, when the
tends to produce loadings that are larger than factor goal is to identify common factors, it is more concep-
loadings (see Snook & Gorsuch, 1989; Widaman. tually sensible to examine the plot of eigenvalues
1993). Thus, solutions using this approach can some- from the reduced correlation matrix, because this ma-
times make minor components appear to be major trix more directly corresponds to the common factor
components (Wood et al., 1996). model. Regardless of which matrix is used, this pro-
Given these consequences, it is not surprising that cedure has been criticized (e.g.. Kaiser. 1970) because
an extensive methodological literature has developed of its subjectivity (i.e., there is no clear objective defi-
exploring the issue of determining the optimal number nition of what constitutes a "substantial" drop in mag-
 USE OF FACTOR ANALYSIS 279

nitude). Additionally, sometimes the obtained pattern a researcher to adopt a somewhat different approach
of eigenvalues is ambiguous in that no clear substan- to determining the optimal number of factors.
tial drop may be present. However, when strong com- One can conceptualize the number-of-factors issue
mon faciors are present in data, studies indicate that as choosing the most appropriate model from a series
this procedure functions reasonably well (Cattell & of alternative factor analysis models that differ in their
Vogelmann. 1977; Hakstian et al., 1982: Tucker et al., complexity (i.e., the number of factors). The objective
1969). is to select a model that explains the data substantially
A third factor-number procedure that has been in- better than simpler alternative models (i.e., models
vestigated in the methodological literature is parallel with fewer factors) but does as well or nearly as well
analysis (Horn, 1965; Humphreys & Ilgen, 1969; as more complex alternative models (i.e., models with
Humphreys & Montanelli, 1975; Montanelli & Hum- more factors). Because factor analysis is a special
phreys, 1976). This approach is based on a compari- case of structural equation modeling, many of the pro-
son of eigenvalues obtained from sample data to eig- cedures used in structural equation modeling for
envalues one would expect to obtain from completely model selection can be applied to EFA. That is, the
random data (i.e., the predicted means of eigenvalues use of ML estimation for EFA introduces a vast array
producec by repeated sets of random data). Suppose a of goodness-of-fit information that can be used to
set of measured variables observed in a given sample determine the appropriate number of factors (see
depends on m major common factors. Parallel analy- Browne & Cudeck, 1992; Hu & Bentler, 1998; Marsh,
sis is based on the notion that the m largest sample Balla, & McDonald, 1988; Mulaik et al., 1989). These
eigenvalues of the reduced correlation matrix should fit measures, which are intended to assess the degree
be larger than the m largest expected values of eigen- to which a given model provides an approximation of
values obtained from repeated corresponding sets of observed correlations or covariances, can be applied
random data (based on the same sample size and num- to the number-of-factors problem. The general proce-
ber of variables). The eigenvalues that would be ex- dure is to fit models with a range of numbers of fac-
pected from random data are then compared with the tors, beginning with zero, and increasing through
eigenvalues actually produced by the data, and a some maximally interesting number. Fit measures for
model is specified with the same number of common each model can then be evaluated, and a decision can
factors as real eigenvalues that are greater than the be made as to the appropriate number of factors to
eigenvalues expected from random data. Like other retain. The desired model is that which constitutes a
objective mechanical rules, this procedure can some- substantial improvement in fit over a model with one
times be arbitrary in that a factor just meeting the fewer factor but for which a model with one more
criterion is retained, whereas a factor falling just be- factor provides little if any improvement in fit.5
low the criterion is ignored. This procedure is also not One common statistic for assessing fit in ML factor
available in major statistical programs. Nonetheless, analysis solutions is the likelihood ratio statistic
simulation research suggests parallel analysis func- (Lawley, 1940). If N is sufficiently large and the dis-
tions fairly well (Humphreys & Montanelli, 1975). tributional assumptions underlying ML estimation are
Parallel analysis methods have also been developed adequately satisfied, the likelihood ratio statistic ap-
for use in components analysis and found to function proximately follows a chi-square distribution if the
well in ;his context (Allen & Hubbard. 1986: Lauten- specified number of factors is correct in the popula-
schlager, 1989; Longman, Cota, Holden, & Fekken,
1989; Zwick & Velicer, 1986).
The methods described to this point for determining
s
the number of factors to retain involve analysis of Ideally, the preferred model should not just fit the data
eigenvalues obtained in principal factors or PCA. Al- substantially better than simpler models and as well as more
complex models. The preferred model should also fit the
though >uch methods of factoring are widely used, the
data reasonably well in an absolute sense. When this is not
ML method of factor extraction is becoming increas-
the case, a researcher should exercise some caution in in-
ingly popular. The ML method has a more formal terpreting the results. A preferred model that fits the data
statistical foundation than the principal factors meth- poorly might do so, because the data do not correspond to
ods and thus provides more capabilities for statistical assumptions of the common factor model. Alternatively, it
inference, such as significance testing and determina- might suggest the existence of numerous minor common
tion of confidence intervals. These capabilities allow factors.
280 FABRIGAR, WEGENER, MAcCALLUM, AND STRAHAN

tion. Thus, this statistic can be used to test the null obtained from one sample will generalize to other
hypothesis that the model holds exactly in the popu- samples. Although no guidelines exist for interpreting
lation with a given number of common factors. This ECVI in absolute terms, it is useful for relative com-
test can be applied to a series of numbers of factors, parisons among alternative models. The smaller the
beginning with zero and continuing until a nonsignif- value of ECVI, the better the expected cross-
icant test statistic is obtained, indicating that the com- validation of the model. An important property of
mon factor model with the corresponding number of ECVI discussed by Cudeck and Henly (1991) and
factors is not rejected as a correct model in the popu- illustrated by Browne and Cudeck (1992) is that it is
lation. Although this procedure is intuitively appeal- sensitive to sample size. ECVI will tend to support the
ing in that it is based on a clear statistical rationale, it retention of simpler models (fewer factors) when N is
is susceptible to serious problems. Most importantly, small and more complex models (more factors) when
it is highly influenced by sample size. When N is N is large. One can think of ECVI as addressing the
large, even trivial discrepancies between the model following question; Based on the available data, for
and the data are likely to give rise to rejection of the how many factors can we obtain accurate and gener-
model with any reasonable number of factors (Haks- alizable parameter estimates?
tian et al., 1982; M. L. Harris & Harris, 1971; Hum- A major advantage of both these indexes is the
phreys & Montanelli, 1975; MacCallum, 1990). In availability of confidence intervals. Thus, a researcher
situations where N is small, even large discrepancies can compare the point estimates and corresponding
between the model and the data may not be statisti- confidence intervals of these fit indexes for a series of
cally significant, thereby leading to underfactoring models with varying numbers of factors. A model can
(Humphreys & Montanelli, 1975). In addition, the use be selected that shows fit that is substantially better
of the likelihood ratio test is problematic, because the than simpler models and fit that is comparable to that
null hypothesis of perfect fit is an unrealistic standard of more complex models. Of course, the logic of the
in that all models are approximations of reality approach illustrated by Browne and Cudeck (1992) of
(Browne & Cudeck, 1992; Cudeck & Henly, 1991; using RMSEA and ECVI to determine the number of
MacCallum, 1990). The realistic goal in factor analy- factors in EFA can be extended to other descriptive fit
sis is to obtain a parsinomious solution that provides a indexes. Thus, a researcher might choose to examine
good approximation to the real world; thus, the hypoth- additional indexes of model fit when determining the
esis of perfect fit is not generally of empirical interest. number of factors (for a recent evaluation of fit in-
Because of these problems, methodologists have dexes, see Hu & Bentler, 1998).
developed an array of alternative "descriptive" mea- Although the use of RMSEA, ECVI, and other de-
sures of fit. These developments began with the work scriptive measures of model fit has yet to be exten-
of Tucker and Lewis (1973), who proposed a "reli- sively tested within the context of determining the
ability coefficient" for ML factor analysis solutions, number of factors in EFA, there is a compelling logic
and have continued through the present day. Our view to this approach. Furthermore, many of these indexes
is that developments reviewed and illustrated by have been tested within the context of more general
lirowne and Cudeck (1992) provide a promising ap- covariance structure models. Thus, users of ML factor
nroach for assessing fit of these models. Browne and analysis would do well to consider the use of these fit
Cudeck focused on the Root Mean Square Error of indexes when determining the number of factors to
Approximation (RMSEA) fit index (Steiger & Lind, retain.
1980) and the Expected Cross-Validation Index In summary, a number of procedures exist to de-
i ECVI; Browne & Cudeck, 1989) and illustrated their termine the appropriate number of common factors.7
use for determining the number of factors in EFA.
RMSEA is an estimate of the discrepancy between the
model and the data per degree of freedom for the 6
Although these guidelines for RMSEA are generally
model. It has been suggested that values less than 0.05 accepted, it is of course possible that subsequent research
constitute good fit, values in the 0.05 to 0.08 range might suggest modifications.
acceptable fit, values in the 0.08 to 0.10 range mar- 7
One method that has been found to perform well in
ginal fit, and values greater than 0.10 poor fit (see determining the number of principal components to retain is
Browne & Cudeck, 1992; Steiger, 1989).6 ECVI, on the minimum average partial procedure (Velicer, 1976;
the other hand, is an estimate of how well the solution Zwick & Velicer, 1982, 1986). We do not discuss this pro-
 USE OF FACTOR ANALYSIS 281

Some of these procedures are highly problematic (i.e., structure (Thurstone, 1947). Thurstone proposed that
eigenvalues greater than 1, the likelihood ratio statis- for any given set of mathematically equivalent solu-
tic), whereas others are likely to perform reasonably tions, the solution with the best "simple structure"
well s'. least under a range of conditions (e.g., scree would generally be the most easily interpretable,
test, parallel analysis, descriptive indexes of model fit psychologically meaningful, and replicable. Thurst-
such as RMSEA). However, even the best procedures one used the term simple structure to refer to solu-
are not infallible. Furthermore, it is important to re- tions in which each factor was defined by a subset
member that the decision of how many factors to of measured variables that had large loadings rela-
include in a model is a substantive issue as well as a tive to the other measured variables (i.e., high with-
statistical issue. A model that fails to produce a ro- in-factor variability in loadings) and in which each
tated -olution that is interpretable and theoretically measured variable loaded highly on only a subset
sensible has little value. Therefore, a researcher of the common factors (i.e., low factorial complex-
should always consider relevant theory and previous ity in defining variables). Therefore, Thurstone sug-
research when determining the appropriate number of gested that factors be rotated in multidimensional
factors to retain. It should also be noted that because space to arrive at the solution with the best simple
any individual data set is likely to have its own idio- structure.
syncrasies, decisions regarding the appropriate num- A number of analytic rotation methods have been
ber of factors can be further improved by examining developed to seek simple structure, and numerous ar-
the replicability of the decision over multiple data sets ticles have been published comparing the utility of
(Catteil, 1978). When a given number of factors for a these various rotation procedures (e.g., Crawford &
battery of measured variables is shown to be appro- Ferguson, 1970; Dielman, Catteil, & Wagner, 1972;
priate in more than one data set (e.g., sensible rules Gorsuch, 1970; Hakstian, 1971; Hakstian & Boyd,
for determining the number of factors give similar 1972; Hofmann, 1978). Although these rotation
results, rotated solutions produce similar interpretable methods differ in a number of respects, perhaps the
patterns of factor loadings), a researcher can be more most fundamental distinction that can be made is be-
confident that the optimal number of factors has been tween orthogonal and oblique rotations. Orthogonal
extracled. rotations constrain factors to be uncorrelated. Though
Factor rotation. For any given solution with two a number have been developed, varimax (Kaiser,
or more factors (or principal components), there exists 1958) has generally been regarded as the best or-
an infinite number of alternative orientations of the thogonal rotation and is overwhelmingly the most
factors in multidimensional space that will explain the widely used orthogonal rotation in psychological re-
data equally well. This means that EFA models with search.
more than one factor do not have a unique solution. In contrast to orthogonal rotations, oblique rota-
Therefore, a researcher must select a single solution tions permit correlations among factors. One common
from among the infinite number of equally fitting so- misconception among researchers is that oblique ro-
lutions. tations require factors to be correlated. This is not the
The criterion most commonly used for selecting case. If the solution with the best simple structure
among solutions in EFA is the property of simple involves orthogonal factors, a successful oblique ro-
tation will provide estimates of the correlations
among factors that are close to zero and produce a
cedure, because it is a method for use with principal com- solution that is quite similar to that produced by a
ponents and has not been developed for use with common successful orthogonal rotation (Harman, 1976). How-
factors. Another approach to determining the number of ever, in situations in which the best simple structure is
components is to examine the comparability of component a solution with correlated factors, successful oblique
scores across split halves of a sample (Everett, 1983). Un-
rotations will produce solutions with correlated fac-
fortunately, formal approaches for determining the number
tors. Unlike orthogonal rotation, there is no single
of common factors based on the stability of common factors
have nor been extensively developed and tested. Nonethe- method of oblique rotation that is clearly dominant in
less, approaches based on this logic seem like a promising psychological research. Several oblique rotation pro-
direction for future research. Catteil (1978) suggested that cedures are commonly used and have been found to
true factors should replicate across samples whereas spuri- generally produce satisfactory solutions. These in-
ous factors should be unstable over samples. clude direct quartimin rotation (Jennrich & Sampson,
282 FABRIGAR, WEGENER, MAcCALLUM, AND STRAHAN

1966), promax rotation (Hendrickson & White, 1964), some options are clearly more optimal than others.
and Harris-Kaiser orthoblique rotation (Harris & Kai- Thus, the utility of an EFA is likely to be a function
ser, 1964).8 of the decisions made in the design of the study and
Some researchers have indicated a preference for the implementation of the analysis. Indeed, it is inter-
orthogonal rotation because of its simplicity and con- esting that some notable examples of the supposed
ceptual clarity (e.g., Nunnally, 1978). However, there failure of EFA to provide valid insights into data in-
are a number of reasons to question the wisdom of this volved poor factor analytic methodology. For in-
view. First, for many constructs examined in psychol- stance, Armstrong's (1967) classic demonstration of
ogy (e.g., mental abilities, personality traits, atti- the alleged limitations of EFA used a PCA, the eig-
tudes), there is substantial theoretical and empirical envalues-greater-than-1 rule, and varimax rotation. In
basis for expecting these constructs (or dimensions of light of the questionable nature these choices, it is not
these constructs) to be correlated with one another. surprising the analysis failed to uncover the true struc-
Therefore, oblique rotations provide a more accurate ture of the data.9 Given these facts, some clear rec-
and realistic representation of how constructs are ommendations can be made regarding how EFA
likely to be related to one another. Thus, from a sub- should be conducted.
stantive perspective, the restriction of uncorrelated /. Study design. Because EFA results are likely
factors imposed by varimax and other orthogonal ro- to be more accurate if sensible decisions are made in
tations is often unwarranted and can yield misleading selecting measured variables and samples, it is essen-
results. Second, because orthogonal rotations require tial that researchers carefully consider both of these
factors to be oriented at 90° angles from one another issues. Obviously, because EFA is used in situations
in multidimensional space (i.e., uncorrelated factors) where there is relatively little prior theory and empiri-
whereas oblique rotations allow for orientations of cal evidence, variable selection can be difficult. None-
less than 90° (i.e., correlated factors), orthogonal ro- theless, to the extent it is possible, researchers should
tations are likely to produce solutions with poorer try to anticipate the number and nature of the factors
simple structure when clusters of variables are less they expect to obtain and use this as a guide for se-
than 90° from one another in multidimensional space lecting variables. We suggest researchers include at
(i.e., when the true underlying factor structure is least four measured variables for each common factor
based on correlated factors). Finally, oblique solutions they expect to emerge and perhaps as many as six
provide more information than orthogonal rotations. given that there is usually considerable uncertainty
Oblique rotations produce estimates of the correla- about the nature of the common factors and their re-
tions among common factors. Knowing the extent to lations to the measured variables. Furthermore, al-
vvhich factors are correlated with one another can of- though EFA should be based on comprehensive sam-
ren be useful in interpreting the conceptual nature of
i he common factors. Indeed, the existence of substan-
8
lial correlations among factors suggests that higher Unfortunately, the selection of oblique rotations offered
order factors may exist. Correlation matrices of fac- in major statistical packages is quite limited. SPSS offers
tors can in turn be analyzed to gain insight into the promax and Harris-Kaiser orthoblique rotations. Direct obli-
min rotation is a family of rotations defined by different
number and nature of these higher order factors and
values of the delta parameter, which governs the oblique-
thereby further refine a researcher's understanding of
ness of the solution. The default value for this parameter in
the data (see Gorsuch, 1983). Because orthogonal ro- SPSS is 0. which corresponds to a direct quartimin rotation.
tations do not provide correlations among factors, it is A new EFA program. CEFA (comprehensive exploratory
impossible to determine if one or more higher order factor analysis; Browne. Cudeck, Tateneni. & Mels, 1998)
factors are present in the data. offers a much wider range of rotations.
9
Additionally, because the data created by Armstrong
Summary and Recommendations have a number of properties that violate assumptions of the
common factor model (e.g., nonlinear relations between fac-
The implementation of EFA requires a researcher tors and measured variables), his data are not really appro-
to address five major types of methodological deci- priate for EFA. Nonetheless, we found that if one uses more
sions. There are a number of options available for optimal EFA procedures, the analysis produces a solution
each decision, and the issues involved in selecting that provides a good representation of the underlying struc-
among these options are sometimes complex. None- ture of the data, thus invalidating Armstrong's arguments
theless, these decisions are not arbitrary in that against the use of factor analysis.
 USE OF FACTOR ANALYSIS 283

pling of measured variables in the domain of interest, 4. Determining the number of factors. We sug-
we recommend that researchers carefully consider the gest that researchers rely on multiple criteria when
psychometric properties of variables. Variables with deciding on the appropriate number of factors to in-
low reliabilities (e.g., below .70) should be avoided, clude in a model. In the use of principal factors meth-
and when validity information is available, it should ods, we recommend the scree test and parallel analy-
be used as a basis for selecting items. With respect to sis using eigenvalues from the reduced correlation
determining adequate sample size, the properties of matrix. In ML factor analysis, we encourage the use
the measured variables should be taken into account. of descriptive fit indexes such as RMSEA and ECVI
Under good conditions (communalities of .70 or as discussed by Browne and Cudeck (1992) along
higher, four to five variables for each factor), a sample with more traditional approaches such as the scree test
size of 100 might well be adequate (although it is and parallel analysis. A sensible strategy in most
always best to have larger sample sizes if possible). cases would be to use multiple methods to make this
Under conditions of moderate communalities (e.g., decision and then carefully examine the rotated solu-
.40 to .70) and moderate overdetermination of factors, tion for the suggested model to confirm that it is in-
a sample of 200 or more seems advisable. Finally, terpretable and theoretically plausible. In situations in
under poor conditions, no sample size may be suffi- which the sample size is sufficiently large, a re-
cient to produce accurate estimates of the population searcher might also wish to randomly split the data
parameters. However, it seems likely that a sample and examine the stability of the solution across the
less than 400 will lead to distorted results. two halves (for a discussion of assessing factor com-
2. Appropriateness of EFA. We urge researchers parability across samples, see Gorsuch, 1983). In con-
to carefully consider if EFA is the most appropriate texts in which procedures suggest different numbers
form of analysis to meet their research objectives. of factors or in which the procedures produce some-
EFA should be used when the primary goal is to iden- what ambiguous results, the researcher should exam-
tify latent constructs and there is insufficient basis to ine the subset of models that these procedures suggest
specify an a priori model (or small subset of models). are most plausible. The rotated solutions for these
CFA should be used when the goal is to identify latent models can then be examined to see which model
constructs and a substantial basis exists to specify an produces the most readily interpretable and theoreti-
a priori model or small subset of models. PCA should cally sensible pattern of results (Comrey, 1978; Ford
not be used as a substitute for EFA. et al., 1986; Hakstian & Muller, 1973; Hakstian et al.,
3. M,idel-fitting procedures. With respect to se- 1982; C. W. Harris, 1967) and when possible which
lecting one of the major methods of fitting the com- of these solutions is most stable over different data
mon factor model in EFA (i.e., principal factors, iter- sets or split halves of data sets.
ated principal factors, maximum likelihood), all three 5. Rotation. Given the advantages of oblique ro-
are reasonable approaches with certain advantages tation over orthogonal rotation, we see little justifica-
and disadvantages. Nonetheless, the wide range of fit tion for using orthogonal rotation as a general ap-
indexes available for ML EFA provides some basis proach to achieving solutions with simple structure
for preferring this approach. However, researchers (Gorsuch, 1983). Instead, it is most sensible to first
should rscognize that this procedure can produce mis- examine the solutions produced by one or more of the
leading results when assumptions of multivariate nor- common methods of oblique rotation (e.g., promax,
mality are severely violated (see Curran et al., 1996; orthoblique, or direct quartimin). If this solution in-
Hu et a!., 1992). Therefore, we recommend that the
distributions of measured variables be examined prior 10
to conducting ML EFA. If nonnormality is severe Another approach to addressing violations of normal-
ity is to create "item parcels" (i.e., composites of several
(e.g., skew > 2; kurtosis >7; West, Finch, & Curran,
measured variables that tap on the same common factor).
1995), one of several remedies might be employed Creation of such item parcels often results in measured vari-
(see West et al., 1995). Measured variables could be ables with more normal distributions than the individual
transformed to normalize their distributions. Correc- items (West et al., 1995). However, such an approach will
tions to fit indexes and standard errors could be per- often not be feasible in EFA, because little is known re-
formed (Bentler & Dudgeon, 1996; Browne, 1984; garding which measured variables are influenced by the
Satorra & Bentler, 1994). Alternatively, one might same common factors, and thus it will be difficult to form
wish to use a principal factors procedure.10 item parcels.
284 FABRIGAR, WEGENER, MAcCALLUM, AND STRAHAN

dicates that the factors are uncorrelated, then it is (perhaps) behavior factors when examining relevant
reasonable for the researcher to conduct a varimax data sets.
rotation and use this as the basis for interpretation We addressed this question by reanalyzing three
(though this solution will likely be very similar to the data sets from this literature. First, we analyzed two
oblique solution). On the other hand, if at least some 9-item matrices from Breckler (1984). In these matri-
of the factors are found to be correlated with one ces, three items each had been designed to assess
another, it is most defensible to use the oblique solu- affective, cognitive, and behavioral aspects of atti-
tion as the basis for interpretation. tudes, respectively (see Breckler, 1984, for detail on
the measures). We took the third data set from Crites
Illustration of Consequences of Choices et al. (1994), in which general measures were devel-
in Analysis oped to assess the affective and cognitive bases of
attitudes (three types of measures for each). In order
As noted earlier, some critics of EFA have been to take advantage of the existence of an additional
insufficiently sensitive to the extent to which particu- measure of affect and an additional measure of cog-
lar decisions about how the analysis is conducted can nition specific to one of the attitude objects (snakes;
influence the results. In order to illustrate some of the Breckler, 1984), we analyzed 8 items from Study 1 of
effects that these decisions can have, we reanalyzed the Crites et al. (1994) article, in which people re-
three data sets from a well-developed psychological sponded to the attitude object "snakes" (see Breckler,
literature on the tripartite model of attitude structure. 1984; Crites et al., 1994, for additional detail on the
According to the tripartite model (Rosenberg & measures).
Hovland, 1960; Smith, 1947), attitudes have affective, We wished to use these data sets to illustrate how
cognitive, and behavioral components. That is, feel- different procedures in conducting EFA analyses
ings experienced when in the presence of an attitude could lead one to either accurately or inaccurately
object, beliefs about the attributes of the object, and identify the number and nature of the common factors
past or present behaviors relevant to the object make underlying these data sets. We submitted the matrices
up the structure underlying overall summary evalua- to a PCA, examining both a varimax (orthogonal) and
tions (see Eagly & Chaiken, 1998; Petty, Priester, & a direct quartimin (oblique) rotation. We also submit-
Wegener, 1994; Wegener & Gregg, in press, for more ted the same matrices to a ML factor analysis, again
recent discussions). Numerous studies have investi- examining both the varimax and direct quartimin ro-
gated the hypothesized tripartite nature of attitudes tations. In order to determine the number of factors,
(e.g., Breckler, 1984; Kothandapani, 1971; Ostrom, we examined the eigenvalues from the unreduced cor-
•969), and in particular, the affective and cognitive relation matrices applying the eigenvalues-greater-
components have received a great deal of attention as than-1 rule. For the ML factor analysis models, we
-.eparable aspects of attitude structure. For example, also examined the RMSEA and ECVI indexes of
assessments of the affect and cognition associated model fit, as well as the scree test and parallel analysis
with particular attitude objects have independently (Montanelli & Humphreys, 1976) for the eigenvalues
predicted attitudes in a variety of settings (e.g., Abel- from the reduced correlation matrix.
•on. Kinder, Peters, & Fiske, 1982; Crites, Fabrigar, The sample sizes were 138, 105, and 164, for
& Petty, 1994), and confirmatory analyses have sup- Breckler (1984) Studies 1 and 2 and Crites et al.
ported multifactor (tripartite) rather than single-factor (1994) Study 1, respectively. Though the sample sizes
models of these attitude structures (Breckler, 1984). are a bit smaller than one might want, the sampling
Moreover, attitudes primarily based on affect versus error that might occur in small samples would only
cognition predict different classes of behaviors (e.g., work against the observation of consistent results
Millar & Tesser, 1986, 1989) and are differentially across data sets. Therefore, consistency across the
influenced by affective versus cognitive persuasive samples would increase confidence in the obtained
appeals (e.g., Edwards, 1990; Fabrigar & Petty, 1999; results, despite any concerns about small samples.
Millar & Millar, 1990). Therefore, there are compel- The initial communality estimates (squared multiple
ling theoretical and empirical rationales for separating correlations) are somewhat variable in each study (see
affect and cognition in attitude structure. Because of Table 1). For Breckler, Study I , the estimates ranged
this, there is a strong basis for expecting that EFA from .61 to .25 with M = .42. For Breckler, Study 2.
should ideally recover separable affect, cognition, and the estimates ranged from .81 to .37 with M = .62.
 USE OF FACTOR ANALYSIS 285

Table 1 in applied research (see later discussion). Of course,

Squared Multiple Correlations (SMCs) for Items in most variables in real data will have substantial
Breckler (1984; Studies I and 2) and Crites, Fabrigar, unique variance, and such conditions should be espe-
and Petn (1994; Studv I) cially likely to lead to differences between EFA
SMCs (which separately represents common and unique
Item Study 1 Study variance in the model) and PC A (which does not dif-
ferentiate between common and unique variance).
Breckler
Unfortunately, information about the distributional
Affect
Thurstone affect 0.43 0.54 properties of the items in the Breckler (1984) data sets
Mood Checklist (+) 0.25 0.37 was unavailable. For the Crites et al. (1994) data,
Mood Checklist (-) 0.33 0.51 however, the skew (-0.77 to 0.73, lAfl = 0.41) and
Behavior kurtosis (-1.02 to 3.28. IA/1 = 0.79) of the items were
Action sequence 0.53 0.73 far smaller than the recommended thresholds for
Distance 0.37 0.81 questioning the adequacy of ML estimation methods
Thurstone behavior 0.56 0.77 (skew > 2, kurtosis > 7; West et al., 1995). Therefore,
Cognition across the data sets we analyzed, consistency in solu-
Thurstone cognition 0.38 0.50 tions could provide a compelling illustration of how
Semantic differential 0.61 0.76
the various decisions can influence the results for real
Listed thoughts 0.35 0.58
data.
Crites et al.
Affect Number of Factors
Multiresponse Checklist 0.82
Dichctomous Checklist 0.78 Table 2 provides the eigenvalues for the reduced
Semantic differential 0.78 and unreduced correlation matrices for each data set.
Thurstone affect 0.34 As can be seen from the eigenvalues for the unre-
Cognition duced correlation matrix, the eigenvalues-greater-
Multiresponse Checklist 0.85
than-1 rule suggested two-factor and one-factor mod-
Dichotomous Checklist 0.79
Semantic differential 0.85
els for the Breckler (1984) data sets (Studies 1 and 2,
Thurstone cognition 0.56 respectively). For the Crites et al. (1994) data, the rule
suggested a one-factor model (see Table 1). Scree
plots of these values, though a bit ambiguous, would
For Crites et al., Study 1, the estimated ranged from perhaps suggest one-factor models for all three stud-
.85 to 34 with M = .72. However, these average ies. However, given that the goal is to identify com-
communalities, especially for Study 2 of Breckler mon factors, it is more sensible to examine the scree
(1984) and Study 1 of Crites et al. (1994), are actually plots of the eigenvalues from the reduced correlation
higher man the initial communalities that often occur matrices. These plots would once again perhaps sug-

Table 2
Eigenvalues for Breckler (1984; Studies 1 and 2) and Crites, Fabrigar, and Petty (1994;
Studv I), From the Unreduced and Reduced Correlation Matrices
Factor number
Correlation matrices
Breckler Study 1
Unreduced 3.84 1.20 0..98 0.85 0.65 0.47 0.41 0.35 0.25
Reduced 3.30 0.60 0,.39 0.20 0.02 -0.10 -0.12 -0.21 -0.27
Breckler Study 2
Unreduced 5 .59 0.91 0,.70 0.50 0.42 0.34 0.23 0.19 0.13
Reduced 5 .24 0.40 0,.34 0.05 -0.04 -0.07 -0.09 -0.13 -0.14
Crites et al. Study 1
Unreduced 5 .72 0.88 0,.46 0.38 0.19 0.17 0.12 -0.09
Reduced 5.48 0.50 0.07 0.01 -0.05 -0.06 -0.08 -0.09
286 FABRIGAR, WEGENER, MAcCALLUM, AND STRAHAN

gest a one-factor model for the three data sets. How- the fourth observed eigenvalue falling far below the
ever, on the basis of these scree plots, one might also fourth value for random data. For Crites et al. (1994)
argue for a four-factor model in the Breckler Study 1 Study 1, two of the four eigenvalues expected for
data set, a three-factor model in the Breckler Study 2 random data (0.42, 0.27, 0.19. 0.11) fall below the
data set, and two-factor model in the Crites et al. data observed eigenvalues (5.48. 0.50, 0.07, 0.01).'' Thus,
set. the parallel analyses question the one-factor decision
When one examines the parallel analysis performed that might come from the eigenvalues-greater-than-1
on the reduced correlation matrix (Montanelli & rule. They suggest a three-factor solution in one of the
Humphreys, 1976), a one-factor solution appears less Breckler data sets and a two-factor solution in the
reasonable. For Breckler (1984) Study 1, the first four Crites et al. data set (both as expected given past data
eigenvalues expected for random data (0.51, 0.34, and theory), but they suggest a four-factor solution for
0.25, 0.17) fall below the observed eigenvalues from Study 1 of the Breckler (1984) data.
the reduced matrix (3.30, 0.60, 0.39, 0.20). For Breck- The goodness-of-fit indexes, however, strongly
ler (1984) Study 2, three of the first four eigenvalues suggest the theoretically expected factor structures
expected for random data (0.60, 0.40, 0.29, 0.20) fall (see Table 3). For the Breckler (1984) data sets.
at or below the observed eigenvalues from the re- RMSEA shows poor fit with the one-factor model and
duced correlation matrix (5.24. 0.40, 0.34, 0.05) with presents substantial improvement with a three-factor
model. Moreover, overall fit becomes quite good (see
Table 3 Browne & Cudeck, 1992) when the third factor is
Selected Fit Indexes for Analyses of Breckler (1984; added. There is also very little overlap in the 90%
Studies I and 2) and Crites, Fabrigar, and Petty (1994; confidence intervals (CI) for the two-factor (RMSEA
Study 1): RMSEAs and ECVls From ML and PACE = 0.12, CI = 0.09-0.16 for Study 1; RMSEA =
Factor Analyses 0.09, CI = 0.04-0.14 for Study 2) and three-factor
Number of factors models (RMSEA = 0.03, CI = 0.00-0.10 for Study
Model-fitting procedure 1 2 3 4 1; RMSEA = 0.00, CI = 0.00-0.00 for Study 2) but
complete overlap between the three- and four-factor
Breckler Study 1 (n = 138)
models (RMSEA = 0.00, CI = 0.00-0.07 for Study
RMSEA
0.12" 0.00"
1: RMSEA = 0.00, CI = 0.00-0.00 for Study 2). The
ML 0.15 0.03"
PACE 0.13 0.14 0.08 0.25 ECVI has it lowest value (i.e., the smallest point es-
ECVI timate for a predicted discrepancy in a new sample)
ML 1.04 0.79 0.58 0.59 for the three-factor model. This pattern clearly sug-
PACE 0.93 0.87 0.65 0.97 gests the three-factor model as superior to the one- or
Breckler Study 2 (n = 105) two-factor models. For the Crites et al. (1994) data,
RMSEA the expected two-factor structure is strongly sug-
ML 0.14 0.09 0.00 0.00 gested. The model has quite poor fit with one factor
PACE 0.14 0.09 0.00 0.00
ECVI
ML 1.16 0.85 0.68 0.76
1
PACE 1.11 0.85 0.69 0.76 ' As noted earlier, there have also been parallel analyses
Crites et al. Study 1 (n = 164) developed for determining the number of components to
RMSEA retain (e.g., Longman et al., 1989). Given our interest of
ML 0.23 0.06 0.02 determining the number of factors underlying the data, how-
PACE 0.23 0.06 0.04 ever, it seems most reasonable to use the parallel analysis
ECVI method designed for determining the number of common
ML 1.41 0.42 0.40 factors (i.e., parallel analysis with squared multiple corre-
PACE 1.38 0.41 0.41 lations in the diagonal of the correlation matrix). Using
h'ote. Noniterated principal factoring and PACE factoring pro- parallel analyses for the unreduced matrix suggests extrac-
cjce similar solutions with no boundary estimates. (See Footnote tion of a single principle component (though such a conclu-
12 fora description of this analysis.) RMSEA = root mean square sion is questioned by the parallel analysis on the reduced
eiror of approximation; ECVI = expected cross-validation index; matrix and by the indexes of EFA model fit). Moreover, one
ML = maximum likelihood; PACE = partitioned covariance es-
timator.
would want to be cautious about accepting a single factor
"The solutions noted included one or two boundary estimates (0) for model in this context given that underfactoring is often
unique variances. thought to be a more severe error than overfactoring.
 USE OF FACTOR ANALYSIS 287

and acceptable fit with two factors. Moreover, there is tions reported in Tables 4, 5, and 6, there are 28 times
no overlap in the 90% CIs for the one-factor (RMSEA that an item is found to load on more than one factor
= 0.23, CI = 0.20-0.26) and two-factor models (or component). However, the six direct quartimin
(RMSEA = 0.06, CI = 0.00-0.11) but nearly com- rotations reported in Tables 4, 5, and 6 produce only
plete overlap between the two- and three-factor mod- eight cross-factor loadings. Examining the loadings
els (RMSEA = 0.02, CI = 0.00-0.10). Again, ECVI below 0.30, the varimax rotation produced only 13%
fails to improve beyond the level obtained with the with magnitudes (ignoring sign) of less than 0.10;
theoretically expected number of factors. 42.6% had values between 0.10 and 0.20, and 44.4%
When taken together, these analyses provide a use- had values greater than 0.20. Using the direct quarti-
ful illustration of how nonoptimal procedures can lead min rotation, 60.8% of the loadings below 0.30 had
a researcher to misleading conclusions regarding the magnitudes equal or less than 0.10; 25.3% had values
appropriate number of factors. The eigenvalues- between 0.10 and 0.20, and only 13.9% had values
greater-i ban-1 rule consistently suggested fewer fac- greater than 0.20 (despite including the 25 loadings
tors than past research and theory would indicate was that fell below 0.30 using direct quartimin but were
most plausible for each data set. The scree test pro- above 0.30 using varimax). Therefore, it seems quite
vided somewhat ambiguous information. In contrast, clear that direct quartimin (the oblique rotation) pro-
parallel analysis suggested the most plausible number vided superior simple structure.
of factors for two of the three data sets, and model fit Besides the "cleaner" solutions provided by the ob-
suggested the most plausible number of factors in all lique rotation, a researcher relying on an orthogonal
three data sets. Thus, consistent with the methodologi- rotation would also forfeit any knowledge of the ex-
cal literature, a researcher clearly would have been isting correlations among factors. Although some of
better served relying on parallel analysis and model fit the varimax solutions for the current data (e.g., the
than the eigenvalues-greater-than-1 rule when deter- ML factor analysis solution for Breckler, 1984, Study
mining the number of factors for these data. 1) might have led the researcher to think in terms of
"affect, behavior, and cognition," he or she might
Orthogonal Versus Oblique Rotations have thought that there was little reason to think of the
Researchers are often tempted to seek "conceptu- three factors as correlated (given the orthogonal rota-
ally distinct" factors by conducting varimax (orthogo- tion). Yet, when the direct quartimin rotation was
nal) rotations in factor analyses. It is sometimes used on the same data, it not only produced better
thought that this retention of statistically independent simple structure clearly breaking along affective, be-
factors "cleans up" and clarifies solutions, making havioral, and cognitive lines, but it also revealed that
them easier to interpret. Unfortunately, this intuition the factors were correlated in the .4-.6 range (see
is exact.ly the opposite of what the methodological Tables 4, 5, and 6).
literature suggests, and the present results bear out the
methodology-based conclusions. Tables 4, 5, and 6 Factor Analysis Versus Principal Components
present PCA and ML factor analysis results (the left
and right columns) using varimax and direct quarti- Comparing the PCA and EFA columns also pro-
min rotations (the top and bottom portions of the vides some useful information. As noted earlier,
tables). For ease of examination, loadings above 0.30 Widaman (1993) used simulation data to show that
are presented in bold face. We have ordered factors PCA inflates loadings when compared with factor
similarly for each analysis (for the Breckler data sets, analysis. Widaman found that "salient" loadings (i.e.,
with the most "affective" factors first, most "behav- the highest loadings on a given factor) are higher in
ioral" factors second, and most "cognitive" factors PCA than in factor analysis and that such inflation is
third; for the Crites et al. data, with the most "affec- greater when the salient loadings are more moderate
tive" factor first, and most "cognitive" factor second). in value (e.g., 0.40 in the population) rather than high
For each of the studies, it would have been unfor- (e.g., 0.80 in the population). Because factor and com-
tunate for a researcher to rely solely on a varimax ponent analyses lead to some differences in the pat-
rotation. Examining only those loadings above 0.30, terns of loadings across factors in the current data,
there are many more cross-factor loadings (i.e., items direct comparisons of loadings might not always be
that load above 0.30 for more than one factor) with meaningful. Yet, an overall pattern consistent with
the orthogonal rotation. Across the six varimax rota- Widaman's (1993) analyses seems to be present.
288 FABRIGAR, WEGENER, IVUcCALLUM, AND STRAHAN

Table 4
Loadings for Principal Components and Common Factors Using Varimax and Direct
Quartimin Rotations: Breckler (1984: Study I)
Principal components Maximum likelihood
analysis (components) factor analysis (factors)
Rotation
Item 1 2 3 1 2 3
Varimax
Affect
Thurstone affect 0.83 0.29 0.09 0.99 0.10 0.11
Mood Checklist (+) 0.47 0.57 -0.19 0.43 0.15 0.13
Mood Checklist (-) 0.77 -0.05 0.34 0.46 0.23 0.16
Behavior
Action sequence 0.25 0.66 0.37 0.24 0.86 0.20
Distance -0.05 0.80 0.25 0.15 0.49 0.34
Thurstone behavior 0.19 0.66 0.46 0.28 0.60 0.40
Cognition
Thurstone cognition 0.03 0.11 0.78 0.10 0.17 0.58
Semantic differential 0.12 0.39 0.77 0.17 0.26 0.95
Listed thoughts 0.26 0.17 0.68 0.23 0.27 0.47
Direct quartimin
Affect
Thurstone affect 0.82 0.18 -0.06 1.08 -0.14 -0.05
Mood Checklist (+) 0.40 0.57 -0.35 0.43 0.04 0.03
Mood Checklist (-) 0.81 -0.20 0.26 0.45 0.13 0.03
Behavior
Action sequence 0.14 0.64 0.23 0.03 0.97 -0.12
Distance -0.20 0.84 0.12 -0.01 0.47 0.20
Thurstone behavior 0.08 0.64 0.33 0.10 0.56 0.21
Cognition
Thurstone cognition 0.00 0.04 0.78 -0.01 -0.01 0.62
Semantic differential 0.04 0.34 0.71 -0.01 -0.04 1.02
Listed thoughts 0.23 0.08 0.63 0.12 0.13 0.43
Correlations among factors or components
j _
2 .32 .51
3 .18 .28 — .42 .60 —
Note. Loadings in bold are values above 0.30.

If one takes all the loadings that are above 0.30 for ined possible inflation of salient loadings, some of the
at least one of the analyses (PCA or ML factor analy- inflations in the current data are "nonsalient" loadings
sis) and compares the magnitude of those loadings from factor analyses that become large enough for
across methods, the PCA loadings tend to be higher. some researchers to consider them salient in defining
For example, the salient loadings for Component 2 of a component. Such inflations would pose a potentially
the varimax rotation in Study 1 (in bold on Table 4) major problem, because they could change the pre-
have values of 0.57, 0.66, 0.80, 0.66, and 0.39 (M = sumed nature and interpretation of the constructs un-
0.62). The corresponding loadings for the ML factor der investigation. One possible instance of such infla-
analysis (Factor 2) are 0.15, 0.86, 0.49, 0.60, and 0.26 tion is the creation of a sizeable loading of the Affect
(M = 0.47). There are 16 such comparisons across Dichotomous Checklist (0.57) in the PCA solution for
the analyses presented, and 12 of the comparisons Crites et al. (1994) Study 1 in which the direct quar-
show higher salient loadings for components than for timin rotation was used (see Table 4). The same load-
factors. Although existing work has generally exam- ing (of an "affect" item on a primarily "cognitive"
 USE OF FACTOR ANALYSIS 289

Table 5
Loadings for Principal Components and Common Factors Using Varimax and Direct
Quartimin Rotations: Breckler (1984; Study 2)
Principal components Maximum likelihood
analysis (components) factor analysis (factors)
Rotation
Item 1 2 3 1 2 3
Varimax
Affect
Thurstone affect 0.72 0.43 0.18 0.75 0.32 0.20
Mood Checklist (+) 0.89 0.18 0.07 0.63 0.24 0.12
Mood Checklist (-) 0.31 0.69 0.19 0.40 0.53 0.28
Behavior
Action sequence 0.27 0.81 0.27 0.39 0.69 0.34
Distance 0.25 0.89 0.19 0.32 0.90 0.24
Thurstone behavior 0.24 0.84 0.29 0.36 0.73 0.36
Cognition
Thurstone cognition 0.08 0.22 0.94 0.13 0.22 0.76
Semantic differential 0.36 0.55 0.65 0.43 0.44 0.70
Listed thoughts 0.53 0.49 0.39 0.57 0.38 0.39
Direct Quartimin
Affect
Thurstone affect 0.68 0.24 0.03 0.81 0.02 0.01
Mood Checklist (+) 0.98 -0.11 -0.04 0.71 -0.01 -0.05
Mood Checklist (-) 0.11 0.72 -0.01 0.22 0.49 0.08
Behavior
Action sequence 0.01 0.87 0.04 0.10 0.72 0.10
Distance -0.05 1.01 -0.07 -0.09 1.10 -0.10
Thurstone behavior -0.04 0.90 0.07 0.04 0.79 0.11
Cognition
Thurstone cognition -0.05 -0.07 1.02 -0.06 -0.01 0.84
Semantic differential 0.18 0.36 0.56 0.24 0.19 0.63
Listed thoughts 0.42 0.32 0.26 0.51 0.14 0.25
Correlations among factors or components

.59 .73
.35 .56 .52 .64
Note. Loadings in bold are values above 0.30.

component) was 0.09 in the factor analysis. It is clear error from the factors, the relations among factors are
to see that the direct quartimin ML solutions are sub- more likely to approach the population values. Such
stantially "cleaner" than the same rotation of the PCA differences are especially evident in analyses of
solutions. Breckler (1984) Study 1. The correlations among
Another notable difference between the PCA and components (i.e., .32, .18, and .28) were roughly half
ML solutions concerns the identified correlations the magnitude of the correlations among factors (i.e.,
among factors. Consistent with Widaman's (1993) .51, .42, and .60). If a researcher had used a PCA, he
simulations, the correlations identified by oblique ro- or she would have been tempted to conclude that the
tation of the PCA solution were substantially lower components were largely independent. However, us-
than the correlations identified by the same rotation of ing an EFA, the same researcher would have realized
the ML factor analysis. It makes sense that PCAs that the smallest of the interfactor correlations was
should generally underestimate relations among the actually greater than 0.4. These would certainly be
constructs, because random error is included in the rather different conclusions. Similar deflations of cor-
components. Because factor analyses remove random relations occur for the PCAs of Breckler (1984) Study
290 FABRIGAR, WEGENER, MAcCALLUM, AND STRAHAN

Table 6
Loadings for Principal Components and Common Factors Using Varimax and Direct
Quartimin Rotations: Criles, Fabrigar, and Petty (1994; Study 1)
Principal components Maximum likelihood
D t t- audi^M* ^uiiipuiicHUV idt-iui analysis uacioisj
Item 1 i 1 2
Varimax
Affect
Multiresponse Checklist 0.63 0.66 0.81 0.48
Dichotomous Checklist 0.67 0.61 0.80 0.44
Semantic differential 0.57 0.70 0.74 0.52
Thurstone affect 0.91 0.09 0.56 0.20
Cognition
Multiresponse Checklist 0.27 0.89 0.38 0.87
Dichotomous Checklist 0.31 0.86 0.40 0.81
Semantic differential 0.28 0.90 0.41 0.85
Thurstone cognition 0.15 0.83 0.38 0.64
Direct Quartimin
Affect
Multiresponse Checklist 0.42 0.64 0.82 0.15
Dichotomous Checklist 0.49 0.57 0.84 0.09
Semantic differential 0.35 0.69 0.70 0.24
Thurstone affect 0.94 -0.04 0.66 -0.09
Cognition
Multiresponse Checklist -0.04 0.95 -0.04 0.98
Dichotomous Checklist 0.01 0.90 0.02 0.89
Semantic differential -0.04 0.95 0.01 0.94
Thurstone cognition -0.15 0.90 0.11 0.66
Correlations among factors or components
1 — —
2 .46 — .77 —
Note. Loadings in bold are values above 0.30.

2 (see Table 5) and Crites et al. (1994) Study 1 (see (which could not be tested for the Breckler data sets), we
Table 6). l 2 conducted noniterated principal factor analyses for these
The analyses of the Breckler (1984) and Crites et data. These analyses produced results very similar to those
al. (1994) data provide examples of how questionable of the ML analyses. Additionally, the noniterated principal
factor solutions for the four-factor models confirmed that a
four-factor model was in fact inappropriate for these data.
12
It is worth noting that some parameter estimation prob- The four-factor models (as was the case for the ML analy-
k m s were encountered in the ML analyses of the Breckler ses) failed to produce readily interpretable or plausible so-
(1984) Study 1 data set. Heywood cases were encountered lutions. No estimation problems were encountered in any of
f.ir the two-, three-, and four-factor models. It is perhaps not the analyses of the Crites et al. (1994) data set, and not
surprising that difficulties were encountered for one of the surprisingly the noniterated principal factor analyses of
f lur-factor models given that this model is almost certainly these data produced solutions very similar to the ML analy-
inappropriate for the present data (i.e., the model is over- ses. Finally, to further test the stability of the ML solutions
factored). One common reason for encountering such esti- across all three data sets, we conducted analyses using Par-
mation problems is when a misspecified model is fit to the titioned Covariance Matrix Estimator Factor Analysis
data. In the case of the three-factor model for Study 1, the (PACE; Cudeck. 1991). This procedure is available in the
theoretical plausibility of the solution and the comparability program CEFA (Browne et al.. 1998). It is a noniterated
of this solution with the three-factor model solution for common factor model-fitting procedure that permits the
Study 2 and the two-factor solution from the Crites et al. computation of the same fit indexes available in ML. The
( ! 994) data suggest that the existence of Heywood cases PACE analyses produced patterns of fit indexes across the
d.ies not constitute a serious problem. Nonetheless, to ex- models that were comparable to those of the ML analyses.
amine if the results of the ML analyses had produced un- The solutions obtained from the PACE analyses were also
u-ual results as a function of severe nonnormality in the data quite similar to those of the ML analyses.
 USE OF FACTOR ANALYSIS 291

procedures can produce results that are rather confus- psychology) in which EFA has been widely used. We
ing and misleading. For all three studies, an EFA with also chose these particular journals, because both are
an oblique rotation provides much better simple struc- among the most prestigious journals in their respec-
ture, more interpretable results, and more theoretically tive areas, so the articles found in these journals
plausible representations of the data than a PCA with should presumably reflect methodologically rigorous
an orthogonal rotation. Moreover, the oblique rotation work.
(especially when paired with EFA rather than PCA) In our review, we examined every article published
provides important information concerning the rela- in these journals during the specified time period to
tions among extracted factors. Across the example determine if any of the statistical analyses reported in
data sets, we see that different procedures for deter- the article addressed an exploratory factor analytic
mining the appropriate number of factors, different question. If one or more analyses of this type was
analytic methods (PCA vs. EFA), and different rota- conducted, we then examined the description of each
tion procedures all have an effect on results and in- analysis. Studies were coded for the ratio of measured
terpretations. variables to factors, average reliability (or communal-
ity) of the measured variables, sample size, examina-
tion of common factors versus principal components,
The Use of Factor Analysis in Current
Psychological Research model-fitting procedure, method for determining the
number of factors, and rotation procedure. The results
of this review are presented in Table 7. The first two
One obvious question that arises from our review
columns of the table indicate the number of articles
and illustration is the issue of the extent to which
and corresponding percentage of total number of ar-
researchers actually use the procedures we have
ticles falling into each category for JPSP. The third
shown to produce misleading results. An answer to
and fourth columns present the same information for
this question is not readily available. Although meth-
JAP.
odologists have often commented on the use of factor
There are several aspects of these results that merit
analysis in psychology (e.g., Comrey, 1978; Ford et
commentary. First, our review indicates that EFA
al., 1986; McNemar, 1951; Skinner, 1980; see also
continues to be an extremely popular statistical pro-
Gorsuch. 1983), most of these criticisms have been
cedure in psychological research. A total of 159 of the
based on subjective impressions rather than system-
883 articles published in 60 issues of JPSP over a
atic reviews of published applications. One exception
5-year period reported EFAs. A total of 58 of 455
is a review of EFA practices reported by Ford et al.
articles published in 30 issues of JAP over the same
(1986). They systematically examined applications of
5-year period reported EFAs. Thus, the typical issue
EFA published between 1975 and 1984 in the area of
of these journals contained two or three articles using
industrial-organizational psychology. They con-
EFA.
cluded that factor analytic practices within this area
The first section of Table 7 shows the distribution
were generally inconsistent with the methodological
of variable to factor ratios. These results indicate that
literature. However, their review did not fully address
the great majority of analyses were based on ratios of
study design issues, was confined to a single area of
at least 4; 1. Thus, in most analyses, factors were prob-
psychology, and covered applications published 14 or
ably adequately overdetermined. However, there was
more y;:ars ago.
a nontrivial number of articles (i.e., about one in five)
with ratios below 4:1. This finding is noteworthy,
Assessing Current Factor Analytic Practices because one condition in which PCAs produce sub-
stantially inflated estimates of factor loadings is when
To explore the question of how factor analysis is the ratio is 3:1 or less (Widaman, 1993).
currently used in psychological research, we con- The next section of Table 7 shows the distribution
ducted a systematic review of articles published from of the average reliability of measured variables in
1991 through 1995 in the Journal of Personality and analyses. These results show that when researchers
Social Psychology (JPSP) and the Journal of Applied reported the reliability of their measured variables, the
Psychology (JAP). We selected these journals, be- values were generally high (i.e.. .70 or greater). One
cause they represent two areas of psychology (person- might conclude from this finding that the psychomet-
ality-social psychology and industrial-organizational ric properties of variables analyzed in EFA are typi-
292 FABRIGAR, WEGENER, MAcCALLUM, AND STRAHAN

Table 7 cally quite good. However, in both journals about

Summary Information of Current Practices in the Use of 40% of analyses did not include reports of the reli-
Factor Analysis ability of the measured variables. A large number of
Journal of the analyses in which reliabilities were not reported
Personality Journal of involved single-item measures, whereas those for
and Social Applied which reliabilities were reported involved multi-item
Psychology Psychology measures. Because single-item variables are likely to
Variable N % N % be considerably less reliable than multi-item vari-
Ratio of variable to factors ables, it seems probable that many of the analyses
Less than 3:1 1 0.6 1 1.7 reported in these journals were based on variables
3:1 28 17.6 9 15.5 with less than optimal psychometric properties. Also
4:1 26 16.4 10 17.2 of interest is the fact that we found virtually no cases
5:1 14 8.8 10 17.2 in which authors reported the communalities of their
6:1 13 8.2 6 10.3 measured variables. This practice is unfortunate given
More than 6: 1 74 46.5 18 31.0
2 that, in some ways, the communalities are more in-
Unknown 1.3 4 6.9
Average reliability of variables formative than the reliabilities regarding the sound-
Less than .60 3 1.9 2 3.4 ness of the EFA results. However, for 18 data sets we
.60-.69 6 3.8 5 8.6 were able to obtain the correlation matrices on which
.70-79 33 20.8 9 15.5 the authors' factor analyses were based. This allowed
.80-.89 33 20.8 11 19.0 us to index the communalities by examining the
.90-1.00 14 8.8 9 15.5 squared multiple correlations associated with each
Unknown 70 44.0 22 37.9 data set. The average communality associated with
Sample size each data set ranged from .12 to .65 with the median
100 or less 30 18.9 8 13.8 of these averages being .425.
101-200 44 27.7 14 24.1
The third section of Table 7 presents the distribu-
201-300 25 15.7 9 15.5
301^00 8.2 2 3.4
tion of sample sizes across articles. More than a third
13
More than 400 47 29.6 25 43.1 of the articles across the two journals conducted EFA
Type of analysis based on modest-to-small sample sizes (i.e., samples
Principal components 84 52.8 28 48.3 of 200 or less). A somewhat smaller number of ar-
Common factors 31 19.5 13 22.4 ticles used moderate sample sizes (i.e., 201 to 400),
Multiple methods 8 5.0 2 3.4 and about a third or more used large samples (i.e.,
Other 1 0.6 0 0.0 greater than 400). Thus, there was a substantial num-
Unknown 35 22.0 15 25.9 ber of articles based on sample sizes sufficiently small
Factor-component number procedure that results could have been distorted if the measured
Eigenvalue > 1 .0 25 15.7 11 19.0 variables included in the analysis were less than op-
Scree Test 24 15.1 9 15.5
timal or the factors were underdetermined.
Parallel analysis 1 0.6 0 0.0
0
The next section of Table 7 indicates the greater
Model fit 0 0.0 0.0
Theory 2 1.3 4 6.9 popularity of PCA relative to EFA. Approximately
Interpretability 4 2.5 0 0.0 half of the published applications reported using the
Multiple methods 35 22.0 12 20.7 PCA method. This method was used despite the fact
Other 2 1.3 0 0.0 that in the vast majority of these articles, the primary
Unknown 66 41.5 22 37.9 goal was to identify latent constructs underlying mea-
Factor-component rotation sured variables rather than data reduction per se. In
Varimax 87 54.7 28 48.3 contrast, only about 20% of analyses used some form
Harris-Kaiser 1 0.6 1 1.7 of EFA (with some type of principal factors analysis
Promax 2 1.3 2 3.4 accounting for three out of every four articles using
Direct quartimin 21 13.2 9 15.5
common factor analysis). In approximately a fourth of
No rotation 23 14.5 4 6.9
3 2 3.4
the articles in both journals, it was impossible to de-
Multiple methods 1.9
Other 1 0.6 0 0.0 termine what method was used.
Unknown 21 13.2 12 20.7 The next section of Table 7 reports the distribution
of articles across different procedures for determining
 USE OF FACTOR ANALYSIS 293

the number of factors. The eigenvalue-greater-than-1 Implications of Current Practices for

rule was the most popular single procedure in both Psychological Research
journals, followed closely by the scree test. Use of
other methods of determining the number of factors Given the prevalence of poor decision making by
was relatively rare. A fairly sizable number of articles researchers in the use of EFA, many readers might
(i.e., about 20%) reported using multiple methods. In wonder what the implications of such practices are for
the overwhelming majority of these cases, multiple the interpretations and conclusions reached in these
articles. A precise answer to this question is not pos-
methods involved some combination of the eigenval-
sible. Although the methodological literature suggests
ues-greaier-than-1 rule, scree test, or factor interpret-
(and the examples presented in this article illustrate)
ability-a priori theory. However, it was extremely
that different choices in designing studies and select-
common (i.e., about 40% of the time) for authors to
ing factor analytic procedures can produce substantial
fail to clearly indicate how they arrived at their deci-
differences in results, there are clearly many cases in
sions as to the number of factors to include.
which this does not occur. The methodological litera-
The final section of Table 7 provides the distribu-
ture has delineated some of the conditions under
tion of articles across different methods of rotation.
which differences in procedures will produce substan-
Varimax rotation was clearly the most commonly tial differences in results. However, understanding of
used rotation with approximately half of all factor these issues is far from complete. Furthermore, in
analyses in both journals using this procedure. The many cases, it is impossible to know how often these
second most popular method of rotation was direct conditions exist in actual data. One method for an-
quartimin although this rotation was used in only swering this question would be to conduct compre-
about a third to a fourth as many articles as varimax hensive reanalyses of published data sets. Unfortu-
rotation. A nontrivial number of articles (i.e., 13% to nately, such an approach is not practical. Beyond the
21%) failed to indicate the rotation used. fact that such a study would require the reanalysis of
When viewed in their entirety, the results reported hundreds of data sets, access to data sets is very dif-
in Table 7 are discouraging. A fair number of analyses ficult. It is relatively rare for researchers to report the
were based on studies in which at least one design correlation matrix on which their factor analyses are
feature was marginal (i.e., low measured variable to based, and requesting access to data is often unsuc-
factor ratios, modest sample sizes, or both). A sizable cessful. 13
number of analyses provided no information regard- Nonetheless, we were able to address this question
ing the psychometric properties of the measured vari- in a somewhat informal manner by reanalyzing those
ables. Trie most popular types of analysis and rotation data sets for which we had access to the correlation
(i.e., PC A and varimax rotation) were not optimal or matrices (a total of 18 data sets). These analyses re-
even ne;;.r-optimal choices. In the case of determining vealed a number of cases in which extracting factors
the number of factors, the largest proportion of pub- versus components, using different procedures to de-
lished articles indicated that multiple methods were termine the number of factors, or using different ro-
used. Such an approach seems sensible in light of the tation methods produced changes in results suffi-
methodological literature. However, exclusive use of ciently substantial to have altered the conclusions a
the largely discredited eigenvalues-greater-than-1 rule
was the second most popular choice. Thus, it appears
that the same poor choices found to be problematic in
11
the methodological literature and demonstrated to Interestingly, when we contacted authors to request
produce the misleading results in our examples are their data, we found that a relatively small proportion of
popular among researchers. Equally disconcerting, the them agreed to our request. In the majority of cases, authors
results in Table 7 indicated that researchers often did not answer our request even after reminder letters were
sent. In other cases, authors indicated that the data were no
failed to provide adequate information about the pro-
longer available. Finally, in a few cases, authors indicated
cedures used. In 15.7% of JPSP factor analyses and that they would send the data but did not or directly refused
22.4% of JAP factor analyses it was impossible to to send the data at all. This last situation is particularly
determine what procedures were used for at least two interesting given that current APA ethical guidelines specify
of the three major analytic decisions. In more than that authors publishing in APA journals must make their
half of all factor analyses reviewed, information was data available to third parties for a period of 5 years fol-
not provided concerning at least one of the decisions. lowing publication.
294 FABRIGAR, WEGENER. MAcCALLUM, AND STRAHAN

researcher might have reached concerning the data. tionally, our survey indicated that many researchers
For example, we compared the solutions produced by made more than one poor decision. The combination
ML factor analysis and PCA for models with the same of several poor decisions is likely to produce even
number of factors retained by the authors and using a more serious distortions. For example, it was rela-
rotation comparable to that used by the authors. We tively common for researchers to conduct a PCA, re-
found that there were differences between PCA and tain as many factors as eigenvalues greater than 1, and
ML sufficiently substantial that a researcher could conduct a varimax rotation. This particular "package"
have reached a different interpretation of the factors in of decisions is especially likely to result in poor re-
5 of the 18 data sets. That is, for 5 of these data sets, covery of the underlying factors. IS
one or more measured variables changed the factors
on which they had their primary loadings across the General Discussion
two types of analysis. Interestingly, for 4 of the data
sets, it was impossible to compare the PCA and ML We began our article by noting that researchers
solutions. The solutions could not be compared, be- must consider five major methodological issues when
cause the authors in their original analyses retained so conducting a factor analysis. We argued that the
many components relative to the number of measured methodological literature suggests that not all options
variables that the corresponding ML factor analysis available to researchers for each of these decisions are
models would have had negative degrees of freedom. equally sound and that poor choices in designing the
Using the same models (i.e., common factor or study and conducting the factor analysis can produce
principal component) and fitting procedures used by poor results. To further illustrate this point, we reana-
:he authors and the same number of factors retained lyzed previously published data sets and showed that
by the authors, we also compared varimax rotation the (mis)use of EFA can produce misleading results.
and direct quartimin rotation. We found that direct We demonstrated that the misleading results obtained
quartimin rotations often produced slightly better in these analyses did not really reflect a fundamental
simple structure than varimax rotation when the fac- weakness in EFA but instead demonstrated the con-
lors were correlated, but the pattern of loadings was sequences of using questionable analytic procedures.
substantially different (i.e., there were changes in the We then went on to establish that the same question-
factors on which one or more of the measured vari- able procedures used in our examples are in fact quite
ables had their primary loadings) in only 1 of the data prevalent in current empirical research using factor
sets. analysis.
Finally, using the eigenvalues-greater-than-1 rule, We believe that our review highlights two impor-
-cree test, and model fit (as indexed by RMSEA), we tant points that should be brought to the attention of
also examined the issue of the appropriate number of researchers using EFA and to readers of articles in
factors. In 7 of the 18 data sets, the scree test, model which EFAs are reported. First, contrary to what
T
it, or both contradicted the eigenvalues-greater- many researchers probably believe, the decisions in
i.han-1 rule. Interestingly, when we examined the re- the design of studies and in selecting factor analytic
sults of the scree test and model fit (two of the better procedures are not arbitrary and inconsequential.
procedures for determining the number of factors), we There is reason to consider some design features and
found that the results often contradicted the decisions
made by the authors in the original articles. In 3 of the
data sets, both model fit and the scree test clearly
14
suggested a different number of factors than that re- In those cases in which model fit clearly contradicted
tained by the authors. In 5 of the data sets, model fit either the eigenvalues-greater-than-1 rule or the original de-
clearly suggested a different number of factors. In 1 cisions made by the authors but the scree test did not, it was
almost never the case that the scree test supported the eig-
data set, the scree test clearly suggested an alternative
envalues-greater-than-1 rule or the original decision. In-
number of factors. 14
stead, in these cases, the scree test was sufficiently ambigu-
Although this small sample of data sets cannot be ous that it was difficult to argue that it clearly indicated an
regarded as fully representative of the entire applied appropriate number of factors.
EFA literature, it does suggest that differences in re- 1?
It is perhaps not surprising that this particular choice of
sults due to changes in EFA procedures might be options is so popular. This set of decisions are the default
more common than many researchers realize. Addi- options for the factor analysis procedure in SPSS.
 USE OF FACTOR ANALYSIS 295

EFA procedures to be markedly better than others. they believe that these options would not be the de-
Second, this article suggests that the quality of EFAs faults unless they were the most acceptable methods
reported in psychological research is routinely quite currently available. Additionally, if a particular pro-
poor. Researchers sometimes base their analyses on cedure is not offered in these programs, it is unlikely
studies with less than optimal features, commonly that a researcher will or can even be expected to use
make questionable choices when selecting analytic it. Unfortunately, the factor analytic procedures of-
procedures, and do not provide sufficient information fered in the major statistical programs are far from
for readers to make informed judgements of the ideal (e.g., see MacCallum, 1983; Wood et al., 1996).
soundness of the EFA being reported. Given the inadequacies of these programs, it is not
One obvious question that arises from these obser- surprising that factor analyses conducted in psycho-
vations is why factor analysis is so poorly used. There logical research are often far from optimal.
are several possible reasons. One is that researchers Although it is relatively easy to understand why
are ill-informed regarding the use of EFA. Although EFA is often misused, it is more difficult to formulate
there is a substantial methodological literature on how such practices might be changed. It seems un-
EFA, much of this literature is relatively complex and likely that psychology departments will be able or
published in quantitatively oriented journals that most willing to invest the resources necessary to substan-
psychologists are unlikely to read on a regular basis. tially upgrade their statistics courses. However, meth-
With a few exceptions (e.g., Finch & West, 1997), odologists could accept a greater responsibility for
nontechnical, concise, and up-to-date reviews of this educating the research community regarding the use
literature have generally not been available to re- of EFA and other statistical procedures. This would
searchers. Furthermore, most researchers probably re- require them not only to write highly technical papers
ceive little formal training in EFA. Although it is targeted at the quantitative methodology community
common for graduate programs to require their stu- (which are undeniably important) but also to write
dents to take courses on the use of analysis of variance less technical papers clearly explaining the practical
(ANOVA), it is much rarer for this to be the case for implications of this methodological research. Further-
EFA (-,ee Aiken, West, Sechrest, & Reno, 1990). more, editors must be willing to publish these articles
Many graduate programs do not offer courses that in nonquantitative journals that are likely to be read
cover factor analysis or devote only 1 week or 2 to by researchers. The use of EFA might also be im-
this topic as part of a broad survey course covering proved by editors of journals adopting higher stan-
different multivariate statistical procedures. There- dards for the manner in which factor analyses are
fore, it is not surprising that many researchers are conducted and reported. At the very least, researchers
relatively uninformed regarding the implications of should be required to report what procedures they
study design features and choosing different factor used when conducting an EFA. Researchers should
analytic procedures. also be expected to offer a brief rationale for their
A second major reason why factor analyses are design decisions and choices of EFA procedures. Fi-
poorly conducted is simple tradition. There is a strong nally, developers and users of EFA should more ac-
tendency for researchers to conduct analyses in a tively pressure the manufacturers of the major statis-
manner that is similar to what has been done before. tical programs to improve their products. Taken
Researchers do so because (a) they wish for their re- together, these initiatives would be a good start to-
sults to be directly comparable to past studies, (b) they ward improving the use of EFA in psychological re-
naively believe that procedures must be reasonable if search.
so many people have used them in the past, or (c) they References
feel (perhaps correctly) that the surest way to avoid
difficulties in the peer review process is to do what Abelson. R. P., Kinder, D. R., Peters, M. D., & Fiske. S. T.
has been done before. (1982). Affective and semantic components in political
Finally, another reason for the poor use of EFA has person perception. Journal of Personality of Social Psy-
to do w ith the statistical software currently popular in chology. 42, 619-630.
psychological research. Such programs are likely to Aiken, L. S., West, S. G.. Sechrest, L., & Reno, R. R.
exert a tremendous influence on the way analyses are (1990). Graduate training in statistics, methodology, and
conducted. For example, many researchers probably measurement in psychology: A survey of PhD programs
follow the default options of their programs, because in North America. American Psychologist. 45, 721-734.
296 FABRIGAR, WEGENER, MAcCALLUM, AND STRAHAN

Allen, S. J., & Hubbard, R. (1986). Regression equations for Cattell, R. B., & Jaspers, J. (1967). A general plasmode (No.
the latent roots of random data correlation matrices with 30-10-5-2) for factor analytic exercises and research.
unities in the diagonal. Multivariate Behavioral Re- Multivariate Behavioral Research Monographs, No. 67-3.
search, 21, 393-398. Cattell, R. B., & Sullivan, W. (1962). The scientific nature
Armstrong, J. S. (1967). Derivation of theory by means of of factors: A demonstration by cups of coffee. Behavioral
factor analysis or Tom Swift and his electric factor analy- Science, 7, 184-193.
sis machine. The American Statistician, 21, 17-21. Cattell, R. B., & Vogelmann, S. (1977). A comprehensive
Armstrong, J. S., & Soelberg, P. (1968). On the interpreta- trial of the scree and KG criteria for determining the
tion of factor analysis. Psychological Bulletin, 70, 361- number of factors. Multivariate Behavioral Research, 12,
364. 289-325.
Arrindell, W. A., & van der Ende, J. (1985). An empirical Comrey, A. L. (1978). Common methodological problems
test of the utility of the observations-to-variables ratio in in factor analytic studies. Journal of Consulting and
factor and components analysis. Applied Psychological Clinical Psychology, 46, 648-659.
Measurement, 9, 165-178. Comrey, A. L., & Lee, H. B. (1992). A first course in factor
Bentler, P. M., & Dudgeon, P. (1996). Covariance structure analysis (2nd ed.). Hillsdale, NJ: Erlbaum.
analysis: Statistical practice, theory, and directions. An- Crawford, C. B., & Ferguson, G. A. (1970). A general ro-
nual Review of Psychology, 47, 563-592. tation criterion and its use in orthogonal rotation. Ps\-
Bentler, P. M., & Kano, Y. (1990). On the equivalence of chometrika, 35, 321-332.
factors and components. Multivariate Behavioral Re- Crites, S. L., Jr., Fabrigar, L. R., & Petty, R. E. (1994).
search, 25, 67-74. Measuring the affective and cognitive properties of atti-
Bookstein, F. L. (1990). Least squares and latent variables. tudes: Conceptual and methodological issues. Personality
Multivariate Behavioral Research, 25, 75-80. and Social Psychology Bulletin, 20, 619-634.
Borgatta, E. F., Kercher, K., & Stull, D. E. (1986). A cau- Cudeck, R. (1991). Noniterative factor analysis estimators
tionary note on the use of principal components analysis. with algorithms for subset and instrumental variable se-
Sociological Methods and Research, 15, 160-168. lection. Journal of Educational Statistics, 16, 35-52.
Breckler, S. J. (1984). Empirical validation of affect, behav- Cudeck, R., & Henly, S. J. (1991). Model selection in co-
ior, and cognition as distinct components of attitude. variance structures and the "problem" of sample size: A
Journal of Personality and Social Psychology, 47, 1191- clarification. Psychological Bulletin, 109, 512-519.
1205. Cudeck, R., & O'Dell, L. L. (1994). Applications of stan-
Browne, M. W. (1984). Asymptotically distribution-free dard error estimates in unrestricted factor analysis: Sig-
methods for the analysis of covariance structures. British nificance tests for factor loadings and correlations. Psy-
Journal of Mathematical and Statistical Psychology, 37, chological Bulletin, 115, 475-487.
62-83. Curran, P. J., West, S. G., & Finch, J. F. (1996). The ro-
Browne, M. W., & Cudeck, R. (1989). Single sample cross- bustness of test statistics to nonnormality and specifica-
validation indices for covariance structures. Multivariate tion error in confirmatory factor analysis. Psychological
Behavioral Research, 24, 445-455. Methods, 1, 16-29.
Browne, M. W., & Cudeck, R. (1992). Alternative ways of Dielman, T. E., Cattell, R. B., & Wagner, A. (1972). Evi-
assessing model fit. Sociological Methods and Research, dence on the simple structure and factor invariance
21, 230-258. achieved by five rotational methods on four types of data.
Browne, M. W., Cudeck, R., Tateneni, K., & Mels, G. Multivariate Behavioral Research, 7, 223-242.
(1998). CEFA: Comprehensive exploratory factor analy- Eagly, A., & Chaiken, S. (1998). Attitude structure. In D.
sis. Authors. [Computer program]. To obtain this pro- Gilbert, S. Fiske, & G. Lindsay (Eds.), Handbook of so-
gram, see URL http://auantrm2.psy.ohio-state.edu/ cial psychology (Vol. 1, 4th ed., pp. 269-322). New
browne/. York: McGraw-Hill.
Cattell, R. B. (1966). The scree test for the number of fac- Edwards, K. (1990). The interplay of affect and cognition in
tors. Multivariate Behavioral Research, I, 245-276. attitude formation and change. Journal of Personality and
Cattell, R. B. (1978). The scientific use of factor analysis in Social Psychology, 59, 202-216.
behavioral and life sciences. New York: Plenum. Everett, J. E. (1983). Factor comparability as a means of
Cattell, R. B., & Dickman, K. (1962). A dynamic model of determining the number of factors and their rotation. Mul-
physical influence demonstrating the necessity of oblique tivariate Behavioral Research, 18, 197-218.
simple structure. Psychological Bulletin, 59, 389-400. Everitt, B. S. (1975). Multivariate analysis: The need for
 USE OF FACTOR ANALYSIS 297

data, and other problems. British Journal of Psychiatry, Hakstian, A. R., Rogers, W. T., & Cattell, R. B. (1982). The
126, 237-240. behavior of nutnber-of-factors rules with simulated data.
Fabrigar, L. R., & Petty, R. E. (1999). The role of the af- Multivariate Behavioral Research, 17, 193-219.
fective and cognitive bases of attitudes in susceptibility to Harman, H. H. (1976). Modern factor analysis (3rd ed.).
affectively and cognitively based persuasion. Personality Chicago: The University of Chicago Press.
and Social Psychology Bulletin, 25, 91-109. Harris, C. W. (1967). On factors and factor scores. Psy-
Fava, J. L., & Velicer, W. F. (1992). The effects of overex- chometrika, 32, 363-379.
traction on factor and component analysis. Multivariate Harris, C. W., & Kaiser, H. F. (1964). Oblique factor ana-
Behavioral Research, 27, 387^15. lytic solutions by orthogonal transformations. Psy-
Finch, .1. F., & West, S. G. (1997). The investigation of chometrika, 29, 347-362.
personality structure: Statistical models. Journal of Re- Harris, M. L., & Harris, C. W. (1971). A factor analytic
search in Personality, 31, 439^1-85. interpretation strategy. Educational and Psychological
Ford, J K., MacCallum, R. C., & Tail, M. (1986). The ap- Measurement, 31, 589-606.
plications of exploratory factor analysis in applied psy- Hendrickson, A. E., & White, P.O. (1964). Promax: A
chology: A critical review and analysis. Personnel Psy- quick method for rotation to oblique simple structure.
chology, 39, 291-314. British Journal of Statistical Psychology, 17, 65-70.
Gorsucn, R. L. (1970). A comparison of biquartimin, max-
Hills, M. (1977). A review of "An introduction to multi-
plane, promax, and varimax. Educational and Psycho-
variate techniques for social and behavioral sciences."
logical Measurement, 30, 861-872.
Applied Statistics, 26, 339-340.
Gorsuch, R. L. (1973). Using Bartlett's significance test to
Hofmann, R. J. (1978). Complexity and simplicity as ob-
determine the number of factors to extract. Educational
jective indices descriptive of factor solutions. Multivari-
and Psychological Measurement, 33, 361-364.
ate Behavioral Research, 13, 247-250.
Gorsuch, R. L. (1980). Factor score reliabilities and domain
Horn, J. L. (1965). A rationale and technique for estimating
validities. Educational and Psychological Measurement,
the number of factors in factor analysis. Psychometrika,
40, 895-897.
30, 179-185.
Gorsuch, R. L. (1983). Factor analysis (2nd ed.). Hillsdale,
Horn, J. L. (1969). On the internal consistency reliability of
NJ: Frlbaum.
factors. Multivariate Behavioral Research, 4, 115-125.
Gorsuch, R. L. (1988). Exploratory factor analysis. In J.
Nessdroade & R. B. Cattell (Eds.), Handbook of multi- Hu, L., & Bentler, P. M. (1998). Fit indices in covariance
variaie experimental psychology (pp. 231-258). New structure modeling: Sensitivity to underparameterized
York: Plenum. model misspecification. Psychological Methods, 3, 424-
453.
Gorsuch, R. L. (1990). Common factor analysis versus prin-
cipal component analysis: Some well and little known Hu, L., Bentler, P. M., & Kano, Y. (1992). Can test statistics
facts. Multivariate Behavioral Research, 25, 33-39. in covariance structure analysis be trusted? Psychological
Gorsuch. R. L. (1997). New procedures for extension analy- Bulletin, 112, 351-362.
sis in exploratory factor analysis. Educational and Psy- Hubbard, R., & Allen, S. J. (1987). A cautionary note on the
chological Measurement, 57, 725-740. use of principal components analysis: Supportive empiri-
Gould, S. J. (1981). The mismeasure of man. New York: cal evidence. Sociological Methods and Research, 16,
Norton. 301-308.
Guttman, L. (1954). Some necessary conditions for com- Humphreys, L. G., & Ilgen, D. R. (1969). Note on a crite-
mon cactor analysis. Psvchometrika, 19, 149-161. rion for the number of common factors. Educational and
Hakstian, A. R. (1971). A comparative evaluation of several Psychological Measurements, 29, 571-578.
prominent methods of oblique factor transformation. Ps\- Humphreys, L. G., & Montanelli, R. G., Jr. (1975). An in-
chometrika, 36, 175-193. vestigation of the parallel analysis criterion for determin-
Hakstian, A. R., & Boyd, W. M. (1972). An empirical in- ing the number of common factors. Multivariate Behav-
vestigation of some special cases of the general "ortho- ioral Research, 10, 193-205.
max" criterion for orthogonal factor transformation. Ed- Jennrich, R. I., & Sampson, P.P. (1966). Rotation for
ucational and Psychological Measurements, 32, 3—22. simple loadings. Psychometrika, 31, 313-323.
Hakstian. A. R., & Muller, V. J. (1973). Some notes on the Kaiser, H. F. (1958). The varimax criterion for analytic ro-
number of factors problem. Multivariate Behavioral Re- tation in factor analysis. Psychometrika, 23, 187-200.
search. 8, 461^75. Kaiser, H. F. (1960). The application of electronic comput-
298 FABRIGAR, WEGENER, MUcCALLUM, AND STRAHAN

ers to factor analysis. Educational and Psychological McNemar, Q. (1951). The factors in factoring behavior.
Measurement, 20, 141-151. Psychometrika, 16, 353-359.
Kaiser, H. F. (1970). A second generation Little-Jiffy. Psy- Millar, M. G., & Millar, K. U. (1990). Attitude change as a
chomelrika, 35, 401-415. function of attitude type and argument type. Journal of
Kothandapani, V. (1971). Validation of feeling, belief, and Personality and Social Psychology, 59, 217-228.
intention to act as three components of attitude and their Millar, M. G., & Tesser, A. (1986). Effects of affective and
contribution to prediction of contraceptive behavior. cognitive focus on the attitude-behavior relation. Journal
Journal of Personality and Social Psychology, 19, 321- of Personality and Social Psychology, 51, 270-276.
333. Millar, M. G., & Tesser, A. (1989). The effects of affective-
Lautenschlager, G. J. (1989). A comparison of alternatives cognitive consistency and thought on the attitude-
to conducting monte carlo analyses for determining par- behavior relation. Journal of Experimental Social Psy-
allel analysis criteria. Mu/tivariate Behavioral Research, chology, 25, 189-202.
24, 365-395. Montanelli, R. G., Jr., & Humphreys, L. G. (1976). Latent
Lawley, D. N. (1940). The estimation of factor loadings by roots of random data correlation matrices with squared
the method of maximum likelihood. Proceedings of the multiple correlations on the diagonal: A Monte Carlo
Royal Society of Edinborough, 60A, 64-82. study. Psychometrika, 41, 341-348.
Levonian, E., & Comrey, A. L. (1966). Factorial stability as Mulaik, S. A. (1972). The foundation of factor analysis.
a function of the number of orthogonally-rotated factors. New York: McGraw-Hill.
Behavioral Science, 11, 400-404.
Mulaik, S. A. (1990). Blurring the distinction between com-
Linn, R. L. (1968). A Monte Carlo approach to the number
ponent analysis and common factor analysis. Mu/tivari-
of factors problem. Psychometrika, 33, 37-72.
ate Behavioral Research, 25, 53-59.
Loehlin, J. C. (1990). Component analysis versus common
Mulaik, S. A., James, L. R., Van Alstine, J., Bennett, N.,
factor analysis: A case of disputed authorship. Mu/tivari-
Lind, S., & Stilwell, C. D. (1989). Evaluation of good-
ate Behavioral Research, 25, 29-31.
ness-of-fit indices for structural equation models. Psycho-
Longman, R. S., Cota, A. A., Holden, R. R., & Fekken,
logical Bulletin, 105, 430^45.
G. C. (1989). A regression equation for the parallel analy-
Nunnally, J. C. (1978). Psychometric theory (2nd ed.). New
sis criterion in principal components analysis: Mean and
York: McGraw-Hill.
95th percentile eigenvalues. Multivariate Behavioral Re-
search, 24, 59-69. Ostrom, T. M. (1969). The relationship between the affec-
MacCallum, R. C. (1983). A comparison of factor analysis tive, behavioral, and cognitive components of attitude.
Journal of Experimental Social Psvcholog\, 5, 12—30.
programs in SPSS, BMDP, and SAS. Psychometrika, 48,
223-231. Overall, J. E. (1964). Note on the scientific status of factors.
MacCallum, R. C. (1990). The need for alternative mea- Psychological Bulletin, 61, 270-276.
sures of fit in covariance structure modeling. Multivariate Petty, R. E., Priester, J. R., & Wegener, D. T. (1994). Cog-
Behavioral Research, 25, 157-162. nitive processes in attitude change. In R. S. Wyer & T. K.
MacCallum, R. C., Browne, M. W., & Sugawara, H. M. Srull (Eds.), Handbook of social cognition (Vol. 2, 2nd
(1996). Power analysis and determination of sample size ed., pp. 69-142). Hillsdale, NJ: Erlbaum.
for covariance structure modeling. Psychological Meth- Rosenberg, M. J., & Hovland, C. L (1960). Cognitive, af-
ods, 1, 130-149. fective, and behavioral components of attitudes. In C. I.
MacCallum, R. C., Widaman, K. F., Zhang, S., & Hong, S. Hovland & M. J. Rosenberg (Eds.), Attitude organization
(1999). Sample size in factor analysis. Psychological and change: An analysis of consistency among attitude
Methods, 4, 84-89. components (pp. 1-14). New Haven, CT: Yale University
Marsh, H. W., Balla, J. R., & McDonald, R. P. (1988). Press.
Goodness-of-fit indexes in confirmatory factor analysis: Rozeboom, W. W. (1990). Whatever happened to broad
The effect of sample size. Psychological Bulletin, 103, perspective? Multivariate Behavioral Research, 25,
391-410. 1-28.
McArdle, J. J. (1990). Principles versus principals of struc- Rummel, R. J. (1970). Applied factor analysis. Evanston,
tural factor analyses. Multivariate Behavioral Research, IL: Northwestern University Press.
25, 81-87. Satorra, A., & Bentler, P.M. (1994). Corrections to test
McDonald, R. P. (1985). Factor analysis and related meth- statistics and standard errors in covariance structure
ods. Hillsdale, NJ: Erlbaum. analysis. In A. Von Eye & C. C. Clogg (Eds.), Analysis of
 USE OF FACTOR ANALYSIS 299

latent variables in developmental research (pp. 399- Velicer, W. F., & Fava, J. L. (1998). Effects of variable and
419). Newbury Park, CA: Sage. subject sampling on factor pattern recovery. Psychologi-
Schonerrann, P. H. (1990). Facts, fictions, and common cal Methods, 3, 231-251.
sense ;ibout factors and components. Multivariate Behav- Velicer, W. F., & Jackson, D. N. (1990a). Component
ioral Research, 25, 47-51. analysis versus common factor analysis: Some issues in
Skinner, H. A. (1980). Factor analysis and studies of alco- selecting an appropriate procedure. Multivariate Behav-
hol. Journal of Studies on Alcohol, 41, 1091-1101. ioral Research, 25, 1-28.
Smith, M. B. (1947). The personal setting of public opin- Velicer, W. F., & Jackson, D. N. (1990b). Component
ions: A study of attitudes toward Russia. Public Opinion analysis versus common factor analysis: Some further
Quarterly, 11, 507-523. observations. Multivariate Behavioral Research, 25, 97-
Snook, S C., & Gorsuch, R. L. (1989). Component analysis 114.
versus common factor analysis: A Monte Carle study. Velicer, W. F., Peacock, A. C., & Jackson, D. N. (1982). A
Psychological Bulletin, 106, 148-154. comparison of component and factor patterns: A Monte
Spearman, C. (1904). General intelligence, objectively de- Carlo approach. Multivariate Behavioral Research, 17,
termired and measured. American Journal of Psychol- 371-388.
ogy, 15, 201-293. Wegener, D. T., & Fabrigar, L. R. (in press). Analysis and
Spearman, C. (1927). The abilities of man. New York: Mac- design for nonexperimental data: Addressing causal and
mil Ian noncausal hypotheses. In H. T. Reis & C. M. Judd (Eds.),
Steiger, .1. H. (1979). Factor indeterminacy in the 1930's Handbook of research methods in social psychology.
and the 1970's: Some interesting parallels. Psychomet- New York: Cambridge L'niversity Press.
rika. <14, 157-167. Wegener, D. T., & Gregg, A. P. (in press). Attitude struc-
Steiger, .1. H. (1989). EzPATH: A supplementary module for ture. In A. Kazdin (Ed.), Encyclopedia of Psychology.
SYSTAT and SYCRAPH. Evanston, 1L: SYSTAT. New York: American Psychological Association.
Steiger, I. H. (1990). Some additional thoughts on compo- Weiss, D. (1976). Multivariate procedures. In M. D. Dun-
nents, factors, and factor indeterminacy. Multivariate Be- nete (Ed.), Handbook of industrial/organizational psy-
havioral Research, 25, 41-^-5. chology. Chicago, IL: Rand McNally.
Steiger, .1. H., & Lind, J. (1980, May). Statistically based West, S. G., Finch, J. F., & Curran, P. J. (1995). Structural
tests for the number of common factors. Paper presented equation models with nonnormal variables: Problems and
at the annual meeting of the Psychometric Society, Iowa remedies. In R. H. Hoyle (Ed.), Structural equation mod-
City. eling: Concepts, issues and applications (pp. 56-75).
Thurstone, L. L. (1947). Multiple factor analysis. Chicago: Newbury Park, CA: Sage.
University of Chicago Press. Widaman, K. F. (1990). Bias in pattern loadings represented
Tucker, L. R., Koopman, R. F., & Linn, R. L. (1969). by common factor analysis and component analysis. Mul-
Evaluation of factor analytic research procedures by tivariate Behavioral Research, 25, 89-95.
means of simulated correlation matrices. Psychometrika, Widaman, K. F. (1993). Common factor analysis versus
34, 421^159. principal component analysis: Differential bias in repre-
Tucker, L. R., & Lewis, C. (1973). A reliability coefficient senting model parameters? Multivariate Behavioral Re-
for maximum likelihood factor analysis. Psychometrika, search, 28, 263-311.
38, 1-10. Wood, J. M., Tataryn, D. J., & Gorsuch, R. L. (1996). Ef-
Tucker, L. R., & MacCallum, R. C. (1997). Exploratory fects of under- and overextraction on principal axis factor
factor analysis. Unpublished manuscript, Ohio State Uni- analysis with varimax rotation. Psychological Methods,
versity, Columbus. I, 354-365.
Zwick, W. R., & Velicer, W. F. (1982). Factors influencing
van Driel, O. P. (1978). On various causes of improper so-
four rules for determining the number of components to
lutions in maximum likelihood factor analysis. Psy-
retain. Multivariate Behavioral Research, 17, 253-269.
chometrika, 43, 225-243.
Zwick, W. R., & Velicer, W. F. (1986). Comparison of five
Velicer, W. F. (1976). Determining the number of compo-
rules for determining the number of components to retain.
nents from the matrix of partial correlations. Psy-
Psychological Bulletin, 17, 253-269.
chometrika, 41, 321-327.
Velicer, W. F. (1977). An empirical comparison of the simi- Received September 24, 1997
larity of principal component, image, and factor patterns. Revision received February 3, 1999
Multivariate Behavioral Research, 12, 3—22. Accepted February 10, 1999 •

View publication stats