54 USING CFA Bryant et al.

• CONFIRMATORY FACTOR ANALYSIS

Statistical Methodology:
VIII. Using Confirmatory Factor Analysis (CFA)
in Emergency Medicine Research
FRED B. BRYANT, PHD, PAUL R. YARNOLD, PHD,
EDWARD A. MICHELSON, MD

Abstract. How many underlying characteristics (or using a patient satisfaction survey. They then ex-
factors) does a set of survey questions measure? plain, within this research context, how CFA is used
When subjects answer a set of self-report questions, to evaluate the explanatory power of a factor model
is it more appropriate to analyze the questions indi- and to decide which model or models best represent
vidually, to pool responses to all of the questions to the data. The information that must be specified in
form one global score, or to combine subsets of related the analysis to estimate a CFA model is highlighted,
questions to define multiple underlying factors? Fac- and the statistical assumptions and limitations of
tor analysis is the statistical method of choice for an- this analysis are noted. Analyzing the responses of
swering such questions. When researchers have no 1,614 emergency medical patients to a commonly-
idea beforehand about what factors may underlie a used ‘‘patient satisfaction’’ questionnaire, the authors
set of questions, they use exploratory factor analysis demonstrate how to: 1) compare competing factor-
to infer the best explanatory model from observed models to find the best-fitting model; 2) modify mod-
data ‘‘after the fact.’’ If, on the other hand, research- els to improve their goodness-of-fit; 3) test hypotheses
ers have a hypothesis beforehand about the under- about relationships among the underlying factors; 4)
lying factors, then they can use confirmatory factor examine mean differences in ‘‘factor scores’’; and 5)
analysis (CFA) to evaluate how well this model ex- refine an existing instrument into a more streamlined
plains the observed data and to compare the model’s form that has fewer questions and better conceptual
goodness-of-fit with that of other competing models. and statistical precision than the original instrument.
This article describes the basic rules and building Finally, the role of CFA in developing new instruments
blocks of CFA: what it is, how it works, and how re- is discussed. Key words: CFA; surveys; research meth-
searchers can use it. The authors begin by placing odology; statistical methods; confirmatory factor analy-
CFA in the context of a common research applica- sis; emergency medicine. ACADEMIC EMERGENCY
tion — namely, assessing quality of medical outcome MEDICINE 1999; 6:54–66

Q UESTIONNAIRES are commonly used in

emergency medicine (EM) as a research tool
to compare various treatments, to survey practi-
lenging. When subjects answer a set of self-report
questions, is it more appropriate to analyze the
questions individually, to pool responses to form
tioners to define state of the art, and to assess pa- one global score, or to group subsets of related
tient and practitioner satisfaction with certain questions to define the underlying characteristics
types of care and clinical experience.1,2 Although (or factors) being measured? Can the results of
writing a survey instrument at first seems fairly such an analysis be used to guide improvements
straightforward, analysis of the survey responses in the survey instrument, or to shorten or stream-
for statistically valid conclusions becomes chal- line the questions without sacrificing measure-
ment outcome reliability?
Factor analysis is the statistical method of
From the Department of Psychology, Loyola University Chi-
cago (FBB), Chicago, IL; the Divisions of General Internal choice for understanding how people respond to
Medicine (PRY) and Emergency Medicine (EAM), Northwest- self-report surveys. The basic focus of factor anal-
ern University Medical School, Chicago, IL; and the Depart- ysis is on the ‘‘structure’’ underlying a set of mea-
ment of Psychology, University of Illinois at Chicago (PRY), sures collected from a group of subjects. By struc-
Chicago, IL.
Series editor: Roger J. Lewis, MD, PhD, Department of Emer- ture, we mean the ways in which responses to the
gency Medicine, Harbor – UCLA Medical Center, Torrance, CA. individual measures interrelate (if they do) to de-
Received April 3, 1996; revision received October 17, 1996; ac- fine one or more underlying characteristics, or fac-
cepted September 14, 1998. tors. When researchers have a hypothesis before-
Address for correspondence and reprints: Fred B. Bryant, PhD,
Department of Psychology, Loyola University Chicago, 6525
hand about the factors that underlie survey
North Sheridan Road, Chicago, IL 60626. Fax: 773-508-8713; responses, they can use confirmatory factor anal-
e-mail: [email protected] ysis (CFA) to evaluate how well the survey results
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ACADEMIC EMERGENCY MEDICINE • January 1999, Volume 6, Number 1 55

match the hypothesized measurement model. An Factor analysis was originally developed for use
explicit structural framework for explaining re- in situations in which investigators had no prior
sponses to a set of questions is known as a mea- knowledge about the structure underlying the an-
surement model. In addition, researchers can com- swers to survey questions, and so-called ‘‘explora-
pare how well competing models fit the collected tory factor analysis’’ was used as an empirical
data, compared with the hypothesized model. On method of analysis. CFA has several important ad-
the other hand, the factors underlying a set of sur- vantages over exploratory factor analysis. First,
vey questions may not be obvious to the investi- CFA enables investigators to systematically test
gator prior to data analysis. In this instance, ex- specific prior hypotheses about the structure un-
ploratory factor analysis may be used to infer the derlying survey results and to compare alternative
best measurement model to ‘‘explain’’ observed measurement models with respect to explanatory
data ‘‘after the fact.’’ This article describes the ba- power. Researchers can also use CFA to refine a
sic rules and building blocks of CFA: what it is, suboptimal model into a simpler form that is both
how it works, and how researchers can use it. parsimonious and reliable, thus improving concep-
Besides determining the appropriate number of tual and statistical precision. This makes CFA a
factors or dimensions underlying responses to a set valuable tool for both theory testing and theory
of survey questions, factor analysis also enables re- building.
searchers to interpret the meaning of each factor, Confirmatory factor analysis provides a means
to label each factor in theory-relevant terms. Fac- of systematically testing hypotheses about factor
tor analysis quantifies how strongly each question structure. CFA enables researchers to test and
characterizes each underlying factor, thereby pin- compare alternative models of the factors under-
pointing the specific subsets of questions that de- lying a set of questions, to verify or confirm the
fine the factors. Questions that strongly reflect a most appropriate measurement model to explain
particular factor are said to have strong ‘‘loadings’’ responses to the questions. Researchers often be-
on that factor or to ‘‘load’’ highly on that factor (i.e., gin by testing a ‘‘null’’ model that assumes there
each question’s factor loadings reflect how strongly are no underlying factors (i.e., that the survey
each underlying factor influences responses to that questions share nothing in common). Researchers
question). usually wish to reject this null model in favor of a
Think of the set of survey questions as a bank more conceptually meaningful structure. Thus, the
of telescopes, each providing a separate field of null model serves as a baseline against which to
vision with more or less clarity of focus. Factor contrast the fit of more complex structural models.
analysis integrates the separate images from the After testing the null model, one then evaluates
bank of telescopes (analyzes the interrelationships increasingly complex models, starting with a one-
among the questions) to determine the number of factor (unidimensional) model that assumes the
separate objects (underlying factors, if any) being set of questions reflects a single, global underlying
viewed, the clarity with which each telescope fo- factor. Next, one tests multifactor (multidimen-
cuses on its target object (factor loadings), and the sional) models, the simplest of which includes two
identity of the objects (factors) under scrutiny. underlying factors that could be either uncorre-
When more than one factor underlies a set of lated (orthogonal model) or correlated (oblique
questions (i.e., a multidimensional model), factor model).
analysis can also be used to determine how Table 1 shows two areas assessed by a fre-
strongly the multiple factors relate to one another. quently used ED patient satisfaction survey
For example, researchers can test the hypothesis (Press, Ganey Associates, Inc., South Bend, IN).
that the underlying factors are unrelated to one The survey developer grouped questions under two
another (i.e., an orthogonal model), or that the un- general subheadings: Nurses and Doctors. Some
derlying factors are correlated with one another questions are virtually parallel (e.g., ‘‘. . . nurses
(i.e., an oblique model). In the latter case, factor were courteous . . .’’ and ‘‘. . . doctors were courte-
analysis computes the correlations among the fac- ous . . .’’), whereas other questions are unique for
tors. For example, if it were determined that factor physicians (e.g., ‘‘The doctors were concerned
1 (e.g., ratings of satisfaction with care received about my comfort’’) and for nurses (e.g., ‘‘The
from physicians) correlated at r = 0.5 with factor 2 nurses were technically skilled’’). We use factor
(e.g., ratings of satisfaction with care received from analysis to see how well the intended Nurses and
nurses), then these factors would share 0.52 ⫻ Doctors groupings explain patients’ responses to
100%, or 25% of their variance in common. Thus, the survey questions.
although the two factors (i.e., ratings of physicians Both exploratory and confirmatory factor anal-
and of nurses) are correlated, they each measure yses operate on matrices of correlations among re-
primarily unique aspects of care (i.e., the remain- sponses to the survey questions. (Although CFA
ing 75% of their variance is unique). can also analyze covariances, i.e., the correlation
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
56 USING CFA Bryant et al. • CONFIRMATORY FACTOR ANALYSIS

TABLE 1. Areas Examined by Use of the Questionnaire*

ED Staff LISREL Code Name Patient Perception Item

Nurses
ncourtesy 1. The nurses were courteous to me.
ntookser 2. The nurses took my problem seriously.
nattentn 3. The nurses paid enough attention to me.
ninform 4. The nurses informed me about my treatment.
nprivacy 5. The nurses maintained my privacy.
nskill 6. The nurses were technically skilled.

Emergency physicians
waittime 7. Before I was seen by a doctor, my waiting time in the
treatment area was satisfactory.
dcourtesy 8. The doctors were courteous to me.
dtookser 9. The doctors took my problem seriously.
dcomfort 10. The doctors were concerned about my comfort.
dexptest 11. The doctors explained my tests and treatment.
dexpprob 12. The doctors explained my problem.
dresident 13. The residents (wearing blue coats) gave me good care.
dselfcar 14. I received good advice about self-care at home, or about
follow-up care.

*The questionnaire used was The Emergency Department Survey, from Press, Ganey Associates, Inc., South Bend, IN. All items
were scored on a Likert-type scale from 1 = very poor to 5 = excellent. The LISREL code names identify measured variables in
the LISREL analysis (see Fig. 1 and Tables 3 and 6). The complete questionnaire consists of separate, explicitly titled sections
evaluating perceptions of registration, nurses, emergency physicians, tests and other staff, and final ratings. For present pur-
poses, only the sections on nurses and emergency physicians were included in the analyses. We selected these sections because
they had mostly complete data, whereas the other sections had more missing data (subjects with any missing data are auto-
matically dropped from confirmatory factor analysis, i.e., casewise deletion). In addition, the use of only two factors greatly
streamlined the presentation. The above items have been abridged from the original questionnaire.

between two measures multiplied by both stan- ware is available for use with either mainframe or
dard deviations, the use of correlations standard- personal computers. Some CFA programs, such as
izes results, making them easier to interpret.) Fac- LISREL, EQS, and AMOS,9 are available in Win-
tor analysis requires that the data are measured dows-version and even provide publication-quality
on at least a continuous, interval scale (e.g., age, diagrams of the model being evaluated (e.g., Fig.
blood pressure, temperature). When data are not 1).
interval-scale (e.g., yes/no, true/false), researchers In conducting CFA, the computer program:
often use procedures to estimate what the corre- 1. uses the raw data to compute the actual, ob-
lations would have been had the variables been served correlations among the survey questions;
measured on an interval scale.4,5 2. uses the CFA model hypothesized by the user
Although most available multivariate statisti- to predict what the observed correlations among
cal software packages (e.g., SPSS, BMDP, SAS) of- the survey questions should have been, assuming
fer exploratory factor analysis capability, not all that the hypothesized model is accurate;
packages are capable of conducting CFA to impose 3. determines the differences between the corre-
specific models on the data, to systematically com- lations predicted by the user’s model and the cor-
pare these models with alternative models for relations that were actually observed; and
goodness-of-fit, and to refine models to make them 4. computes a maximum-likelihood chi-square
more parsimonious and reliable. The three com- value estimating the probability (p-value) that the
puter programs most commonly used to perform differences between the predicted and actual ob-
CFA are: LISREL,5 which stands for LI near Struc- served correlations would occur by chance, assum-
tural RELationships (supported by Scientific Soft- ing the hypothesized model is correct.
ware International, Chicago, IL); EQS,6 pro- Contrary to other inferential statistical tests for
nounced ‘‘X’’ (supported by Multivariate Software, which significant p-values represent a positive re-
Inc., Encino, CA); and CALIS,7 which stands for sult, with CFA a statistically significant chi-square
Covariance A nalysis of LInear Structural equa- denotes a model that fails to predict the observed
tions (supported by SAS, Cary, NC).8 LISREL has data accurately (i.e., the intercorrelations pre-
been available the longest and is currently the dicted by the CFA model are different from the ac-
most popular, a larger literature having accrued tual, observed intercorrelations). CFA users thus
concerning its use and its accuracy under condi- seek models that have nonsignificant chi-square
tions violating statistical assumptions. CFA soft- values. If a CFA model fits the data well, then the
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ACADEMIC EMERGENCY MEDICINE • January 1999, Volume 6, Number 1 57

Figure 1. This graphical representation shows that

we believe answers to the 14 questions on the pa-
tient satisfaction questionnaire are influenced by
two fundamental factors, one due to nurses, one due
to doctors (Table 1), and that these two underlying
factors are correlated. In diagramming a confirma-
tory factor analysis (CFA) model, each observed in-
dicator is enclosed in a square and designated by a
Roman letter x. In the hypothesized model, there
are 14 observed indicators, labeled x1 – x14. The effect
of measurement error on each observed indicator is
marked by a small straight line to the indicator, and
each unique error term is designated by a Greek
letter delta (␦). The correlation matrix of these
unique-error terms is referred to as Theta delta (⌰␦).
In the hypothesized model, the unique errors for the
14 observed indicators (survey questions) are des-
ignated ␦1 – ␦14, all assumed to be independent of one
another (i.e., no correlated-error terms are in-
cluded). Each underlying factor is enclosed in a cir-
cle and designated by a Greek letter xi (␰). The hy-
pothesized model has two underlying factors —
Nurses and Doctors — labeled ␰1 and ␰2, respectively.
The effect of a factor on an observed indicator (i.e.,
a factor loading) is marked by a straight line from
the factor to that indicator and is designated by a
Greek letter lambda (␭). The matrix of factor load-
ings is referred to as Lambda x (⌳x). In the hypoth-
esized model, there are 14 estimated factor loadings
in ⌳x; the first six indicators (x1 – x6) have loadings
(␭1,1 – ␭6,1) on the first factor (␰1); and the last eight
indicators (x7 – x14) have loadings (␭7,2 – ␭14,2) on the
second factor (␰2). All other factor loadings (i.e., for
x1 – x6 on ␰2; and for x7 – x14 on ␰1) have been fixed at
zero. (This specifies that the first six indicators ap-
ply only to Nurses, and that the following eight in-
dicators apply only to Doctors.) The correlations among the underlying factors are marked with curved paths in
CFA diagrams. In LISREL notation, each factor intercorrelation is designated by a Greek letter phi (␾), and the
matrix of factor variances and factor intercorrelations is referred to as Phi (⌽).

correlations that it predicts are not statistically ues (⌬␹2) across models may be more informative
different from the actual, observed correlations — than the chi-square values themselves.
resulting in a nonsignificant p-value. Two models are said to be ‘‘nested’’ when they
To ensure sufficient statistical power for hy- are identical except that one model includes more
pothesis-testing purposes, researchers typically ‘‘restrictions’’ than the other (i.e., more factor load-
aim for a large total sample (e.g., 500 – 1,000 sub- ings or factor intercorrelations are fixed at a spe-
jects). A rough guideline for the minimum ratio of cific value, such as zero, instead of being esti-
subjects to questions is 5:1. Models that allow each mated). When competing models are nested, then
survey question to tap more than one underlying their chi-squares can be directly contrasted to test
factor require a larger sample size than models the hypothesis that one model fits the data better
that force each question to tap only one factor. than the other. In the present example, an orthog-
Because the maximum-likelihood chi-square onal (uncorrelated factors) version of the two-fac-
statistic obtained from CFA is sensitive to sample tor model is obtained when the correlation between
size, researchers usually rely on other criteria to the Nurses and Doctors factors is fixed at zero.
gauge how well a given model fits the data. With This orthogonal model has all of the same param-
large samples, even reasonable models are likely eters as the less restrictive, oblique (correlated
to produce statistically significant chi-square val- factors) two-factor model, except that the factor
ues. For this reason, differences in chi-square val- correlation has been omitted, in contrast to the
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
58 USING CFA Bryant et al. • CONFIRMATORY FACTOR ANALYSIS

oblique model in which the factor correlation is ries its own assumptions, is the engine that drives
included. Thus, the orthogonal model is nested CFA. Maximum-likelihood estimates assume that
within the oblique model, and the chi-square for a multivariate normal distribution underlies the
the latter model can be subtracted from that for data and that observations are sampled randomly
the former to obtain a difference in chi-squares and independently. Multivariate normality is often
(⌬␹2), to test the hypothesis that the Doctors and difficult to obtain in applied research. To meet this
Nurses factors are correlated. criterion, not only must patient responses on every
Besides contrasting the chi-square values of question analyzed be normally distributed, but
nested models to test ‘‘incremental fit,’’ researchers also the distribution of scores obtained by any pos-
also report a variety of different relative-fit indices, sible linear function of the questions— e.g., [0.5 ⫻
including the goodness-of-fit index (GFI) and the (question 1)] ⫺ [0.75 ⫻ (question 7)] ⫹ [0.33 ⫻
adjusted (for the number of parameters estimated question 18)])— must also be normally distributed.
in the model) goodness-of-fit index (AGFI) (Table Because violating this criterion invalidates the
2). Despite variations in their formulas, most com- type I error estimates obtained using CFA, re-
parative fit indices basically reflect how much bet- searchers usually ignore the model’s p-value as an
ter the given factor model fits the data, relative to absolute measure of fit, focusing instead on mea-
the baseline ‘‘null’’ model, which assumes there are sures of relative fit (which have fewer distribu-
no underlying factors. Thus, fit indices reflect the tional assumptions). When interval-scale data are
proportion of shared variance that the model ex- not multivariate normal, researchers sometimes
plains among the set of survey questions. All fit transform them into a matrix of asymptotic covar-
indices range between 0 and 1, with higher values iances and use the weighted least-squares method
indicating better fit, and 0.90 generally considered of estimation, to obtain more accurate results.4,5
to be a minimum acceptable level of fit. In the present example, the patient satisfaction
The maximum-likelihood procedure, which car- data diverge from multivariate normality, re-

TABLE 2. Summary Table for Confirmatory Factor Analysis (CFA)

Alternate names and Also known as measurement modeling, structural equation modeling with latent factors, covariance
related methods structure modeling, and analysis of latent covariance-structures. Related exploratory techniques include
principal-components analysis and common-factor analysis.

Data type A matrix of correlations or covariances among a set of interval- or ratio-scale variables.

Assumptions Multivariate normality, sufficient sample size (see Limitations, below).

Principal results 1. Maximum-likelihood goodness-of-fit chi-square (␹2) and its associated degrees of freedom, for testing
the hypothesis that the data matrix predicted by the particular factor model differs from the observed
data matrix. Smaller chi-squares reflect better goodness-of-fit.
2. Indices of relative fit. These measures of relative fit include the goodness-of-fit and adjusted goodness-
of-fit index. Each fit index reflects how much better the particular model fits the data, relative to a
null model that assumes there is no common variance (i.e., that sampling error alone explains ob-
served correlations among survey questions). These indices range from 0 (the fit of the absolute,
worst null model) to 1 (perfect fit), with 0.90 as a rough guideline in judging minimally acceptable
fit. Also included are modification indices, useful in improving model fit.
3. Standardized estimates of factor loadings relating survey questions to hypothesized underlying fac-
tors, factor intercorrelations controlling for measurement error, and the proportion of common vari-
ance that the particular model explains in each survey question.

Strengths CFA allows systematic comparison of alternative structural models, to test hypotheses about relative fit
and to determine the most appropriate measurement model(s) for a set of measured variables. CFA can
be used to refine a model to increase its parsimony and statistical precision.

Limitations 1. Low sample size reduces CFA’s power to detect meaningful differences between the predicted and
the observed data matrices and thus biases models to have better apparent goodness-of-fit. A rough
guideline for the minimum ratio of respondents to measured variables is 5:1.
2. Skewness and range restriction in measured variables attenuate correlations and factor loadings,
worsening goodness-of-fit and making multivariate normality difficult to obtain. To correct this prob-
lem, data are sometimes transformed before analysis.
3. Missing data are omitted using casewise deletion, which can shrink sample size considerably when
missing responses are scattered throughout the data set. To avoid this problem, researchers some-
times construct and analyze data matrices using pairwise deletion. Although this maximizes average
sample size, it sometimes produces matrices that are ill-conditioned for factor analysis, when there
are numerous missing responses within subsets of respondents.
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ACADEMIC EMERGENCY MEDICINE • January 1999, Volume 6, Number 1 59

TABLE 3. Correlations, Means, and Standard Deviations (SDs) for the 14 Patient Satisfaction Questions*
Nurses Doctors

Question 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Nurses
1. ncourtsy
2. ntookser 86
3. nattentn 80 85
4. ninform 75 78 86
5. nprivacy 73 74 79 81
6. nskill 78 77 75 75 77

Doctors
7. waittime 47 51 57 56 51 47
8. dcourtsy 54 57 56 55 53 56 53
9. dtookser 55 62 58 57 54 56 52 83
10. dcomfort 55 61 59 58 57 57 51 80 81
11. dexptest 54 58 56 60 54 55 50 74 78 78
12. dexpprob 52 57 56 59 52 54 49 70 75 74 86
13. dresidnt 57 61 58 58 54 56 50 74 75 74 75 76
14. dselfcar 52 55 54 58 52 54 47 67 69 67 76 78 69

Mean 4.4 4.3 4.1 4.0 4.1 4.3 3.8 4.4 4.4 4.3 4.3 4.2 4.3 4.2
SD 0.9 0.9 1.0 1.1 1.0 0.9 1.2 0.8 0.9 0.9 1.0 1.0 0.9 1.0

*N = 1,614. Decimals have been omitted from correlations. For the meaning of the questions, see Table 1.

sponses to all survey questions being negatively Having discussed the motivation, data require-
skewed. Because the data violate the multivariate ments, assumptions, and logic underlying CFA, we
normality assumption, and because sample size is are ready to discuss the components of the CFA
relatively large, we have chosen to ignore CFA model. Aspects of the CFA model that the investi-
models’ p-values as indicators of absolute fit, and gator must specify in order to conduct the analysis
to use relative fit indices instead to gauge how well include:
models explain the data. (Analyzing the asymp- 1. the number of factors underlying the set of
totic covariance matrix for the present data using questions;
least-squares estimation revealed that the final 2. which questions measure which factor(s);
model identified using maximum likelihood pro- 3. the nature of the relationship(s) between fac-
vides an excellent fit to the data, ␹2 = (19, n = tors, if the model has more than one factor — i.e.,
1,614) = 64.1; GFI = 0.99; AGFI = 0.99. are the factors correlated (oblique) or independent
In addition to relative fit indices, the output of (orthogonal) of one another?; and
CFA includes other information useful in gauging 4. the concept of an error term (i.e., unexplained
model fit. For example, the root mean square re- variance) for each survey question, and the inter-
sidual (RMSR) is a measure of the average abso- relationships (if any) among these error terms.
lute difference between the interrelations pre- Finally, to improve models with inadequate fit,
dicted by the user’s model and the actual, observed researchers often examine the modification index
interrelations. As RMSR decreases and approaches (MI) for each parameter (i.e., each factor loading,
zero, the fit of the given model improves. The CFA factor interrelation, or correlated-error term) that
output also includes information about the set of has a fixed value of zero in the hypothesized model,
survey questions that is useful in assessing model in search of ways to improve model fit. The MI es-
fit. For example, the ‘‘squared multiple correlation’’ timates the reduction in chi-square that would re-
for each survey question estimates the proportion sult if a particular fixed parameter were free to be
of variance that the model explains in subjects’ re- estimated in the model, as well as the expected
sponses to that question. The ‘‘total coefficient of change in the value of the particular parameter,
determination’’ reflects how well the survey ques- given its inclusion in the model. This exploratory
tions jointly serve as measures of the underlying probing is known as a ‘‘specification search.’’ Based
factors, and approaches 1.0 as the model’s fit im- on specification searches, researchers often free
proves. Table 2 provides a summary of critical in- additional factor loadings or allow for correlated
formation about CFA; and Figure 1 is a schematic measurement errors in their CFA models, partic-
diagram (with LISREL notation) of the CFA model ularly when these modifications are plausible and
originally hypothesized for the 14 patient satisfac- improve the fit of models that are well-grounded
tion questions in Table 1. in preexisting theory.
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
60 USING CFA Bryant et al. • CONFIRMATORY FACTOR ANALYSIS

TABLE 4. Goodness-of-fit Statistics for Various Factor Models of the Patient Satisfaction Survey*
Measures of Fit

Factor Model ␹2
df GFI AGFI RMSR

One factor (total score) 6,200.2 77 0.48 0.29 0.092

Two correlated factors: Nurses and Doctors 1,556.9 76 0.87 0.81 0.042
Modified two-factor model (dual ␭s for waiting time) 1,426.3 75 0.87 0.82 0.026
Modified two-factor model with two correlated ␦s 948.9 73 0.92 0.88 0.025
Modified two-factor model with four correlated ␦s 756.3 71 0.94 0.91 0.020

*N = 1,614. GFI = the goodness-of-fit index. AGFI = the adjusted goodness-of-fit index. Each of these indices reflects how much
better the particular model fits the data, relative to a null model that assumes there is no common variance (i.e., that sampling
error alone explains observed correlations among the survey questions). GFI and AGFI approach 1 as the fit of the given model
improves, with 0.90 considered a minimally acceptable level of fit. RMSR = root mean square residual (i.e., a measure of the
average difference between the correlations actually observed among the survey questions and the correlations predicted by a
particular model), which approaches 0 as the fit of the given model improves.

SAMPLE DATA lated. In estimating these models, we standardized

the underlying factors (i.e., fixed their means at
Data were obtained for an urban, 800 bed univer- zero and their variances at one) to define the mea-
sity-based Level 1 trauma center. Staffed by at- surement scale for each factor. This also allowed
tending physicians 24 hours a day, the ED has an us to interpret the standardized factor interrela-
annual census of 48,000 patients. The hospital tionships as correlation coefficients. We used two
serves as a primary site for an EM residency, and goodness-of-fit indices (GFI and AGFI) and the
housestaff from other specialties rotate through RMSR to gauge model fit. We compared the chi-
the ED for one-month periods. The majority of pa- squares of nested models to assess incremental fit.
tients are cared for by both an attending and one Table 4 presents the goodness-of-fit statistics
or more resident physicians. for these three measurement models. Contrary to
For the purpose of this study, data collection in- the view of patient satisfaction as unidimensional,
volved retrieval of archival data: information about the one-factor model explained less than half of the
patient identity was not recorded, and the study variance that the satisfaction questions have in
did not require institutional review board ap- common (GFI = 0.48). As hypothesized, the in-
proval. tended two-factor model in oblique (correlated fac-
Each year the ED surveys all discharged pa- tors) form fit the data better than the one-factor
tients who were not admitted, mailing them a two- model, ⌬␹2(1, n = 1,614) = 4,643.3; p < 0.00001.
page questionnaire that is widely used to assess Confirming the appropriateness of correlated fac-
satisfaction with various aspects and components tors, the orthogonal version of the two-factor model
of care (Press, Ganey Associates, Inc., South Bend, (with the factor intercorrelation fixed at zero) pro-
IN), one week after their visit to the ED. For pres- duced a set of predicted correlations among the
ent purposes, we used questions from two sections survey questions that was mathematically impos-
of the questionnaire — Nurses and Doctors — in a sible to analyze (i.e., the matrix of predicted cor-
CFA (Table 1). We analyzed the responses of 1,614 relations was ‘‘not positive definite’’). LISREL’s in-
patients who returned the survey between April ability to estimate the orthogonal model indicates
and September 1995, and who completed every that the orthogonal model is untenable in the face
question in the Nurses and Doctors sections. of the data (i.e., a case of so-called ‘‘model mis-
specification’’). However, an oblique version of the
ANALYSIS intended two-factor model fell just short (0.87) of
the minimum acceptable GFI value (0.90).
How well does the intended two-factor (Nurses and In search of an appropriate measurement
Doctors) model explain the pattern of correlations model, we closely inspected the LISREL solution
among the 14 satisfaction questions (Table 3)? To for the oblique two-factor model, with an eye for
answer this question, we used LISREL5 to impose ways to improve model fit. Examining modification
three different measurement models on the data: indices for factor loadings, we found that allowing
1) a one-factor model that assumes patient satis- Doctors question 1 (i.e., satisfaction with time in
faction is unidimensional; 2) the hypothesized, the treatment area before being seen by a doctor)
oblique two-factor model (correlated Nurses and to load also on the Nurses factor would signifi-
Doctors factors); and 3) an orthogonal version of cantly improve the two-factor model’s fit, MI =
the intended two-factor model in which ratings of 127.2, p < 0.00001. The modification analysis fur-
Doctors and Nurses are assumed to be uncorre- ther revealed that this ‘‘waiting time’’ question had
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ACADEMIC EMERGENCY MEDICINE • January 1999, Volume 6, Number 1 61

an estimated factor loading of 0.36 on the Nurses to pairs of questions (e.g., situational variation in
factor (the remaining 13 questions, in contrast, all patient load or time demands on staff ) or 2) shared
had estimated factor loadings below 0.08 on their sources of error variance (e.g., patients’ subjective
alternate factor). In retrospect, it seems plausible impressions of the personalities of nurses or phy-
that patients may attribute responsibility for their sicians). We decided to rerun the modified two-fac-
waiting time in the treatment area to both nurses tor model, first freeing the two largest correlated
and physicians. We therefore decided to modify the errors.
oblique two-factor model to allow the ‘‘waiting As seen in Table 4, the modified two-factor
time’’ question to load on both factors. model (with dual loadings for the ‘‘waiting time’’
Using this procedure to modify an initial model question) fit the data significantly better when it
represents a blend of exploratory and confirmatory included the two correlated-error terms with the
approaches, and it is common practice when mod- largest MIs, ⌬␹2(1, n = 1,614) = 477.4; p < 0.0001.
els that are grounded in strong theory achieve less Although the model’s GFI was now 0.91, its AGFI
than adequate goodness-of-fit. However, these was still below 0.90. We thus freed the other two
modified models need to be replicated with an in- correlated errors with sizable MIs and reestimated
dependent sample of respondents, before they can the model. As seen in Table 4, including all four
truly be termed ‘‘confirmed’’ measurement models correlated-error terms significantly improved the
in a technical sense. Thus, when sample size is model’s fit, compared with including only the two
large, researchers sometimes randomly split their largest correlated-error terms, ⌬␹2(1, n = 1,614) =
data set in half, using one (training) sample to de- 192.6; p < 0.0001. The two-factor model with four
velop a final measurement model and the other correlated errors also had acceptable goodness-of-
(hold-out) sample to confirm the model’s cross-sam- fit when assessed by both GFI (0.94) and AGFI
ple generalizability. Although not reported for (0.91), making it an appropriate measurement
brevity, this procedure yielded identical models for model for the 14-item instrument. In the standard-
the present data. When the sample size is too ized LISREL solution for this final model (Chart
small to split the data set into separate training 1), the Nurses and Doctors factors correlate 0.73
and hold-out samples, an independent sample is (p < 0.00001), indicating that about half (i.e., 0.732
needed to confirm the model’s replicability. = 0.53) of the variance in patient satisfaction with
Confirmatory factor analysis revealed that this emergency physicians is related to patient satis-
modified model (i.e., allowing the ‘‘waiting time’’ faction with emergency nurses, and vice versa.
question to reflect both Doctors and Nurses fac- Chart 2 shows the LISREL 7 computer code used
tors) fit the data significantly better than the orig- to conduct this final analysis.
inal two-factor model, ⌬␹2(1, n = 1,614) = 130.6; p Confirmatory factor analysis also spotlights
< 0.0001. However, the modified model still had survey questions that are poorly focused. For ex-
goodness-of-fit indices below 0.90 (Table 4). Accord- ample, the LISREL solution (Chart 1) reveals that
ingly, we made another specification search, this 13 of the 14 patient satisfaction questions have rel-
time inspecting the modification indices of the atively small unique errors (␦s ⱕ 0.36) and sizable
unique-error terms in the LISREL solution for the squared multiple correlations (i.e., ⱖ0.64); thus,
modified model. Two correlated-error terms had the two-factor model explains nearly twice as
MIs greater than 200 (p < 0.00001)— one for much common variance as it leaves unexplained
Nurses questions (between ‘‘courtesy of the nurses’’ for all but one of the questions. The single discrep-
and ‘‘degree to which the nurses took your problem ant question — wait time (‘‘waiting time in the
seriously’’) and one for Doctors questions (between treatment area, before you were seen by a
‘‘doctors’ concern to explain your test and treat- doctor’’)— had a relatively high proportion of
ment’’ and ‘‘advice you were given about caring for unique error (0.602) and a relatively low squared
yourself at home, or obtaining follow-up medical multiple correlation (0.398); thus, the model ex-
care’’). Two other correlated-error terms had MIs plains only two-thirds as much common variance
between 100 and 200 (p < 0.00001), both for Doc- as it leaves unexplained for this question. These
tors questions: 1) between ‘‘overall explanation of findings suggest that patients attribute respon-
your illness/injury’’ and ‘‘advice you were given sibility for the time they spend waiting in the
about caring for yourself at home, or obtaining fol- treatment area largely to factors other than emer-
low-up medical care’’; and 2) between ‘‘courtesy of gency medical staff. For example, perhaps patients
the doctor’’ and ‘‘degree to which the doctor took blame delays in treatment on bureaucratic hospi-
your problem seriously’’ (paralleling the measure- tal procedures (e.g., the need to check records) or
ment error shared between comparable questions on circumstances beyond the control of emergency
for Nurses). These reliable correlated-error terms medical staff (e.g., the number of other sick pa-
reflect either: 1) unmeasured sources of influence tients being treated).
other than Nurses or Doctors that affect responses Having determined an appropriate measure-
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
62 USING CFA Bryant et al. • CONFIRMATORY FACTOR ANALYSIS

CHART 1. LISREL Output for the Final Two-factor Model*

continued

ment model for the patient satisfaction question- correlated factor scores— one score for satisfaction
naire, one can then use the factors — rather than with nurses, and another score for satisfaction
the original variables — as more reliable and more with doctors.
parsimonious measures of patient satisfaction, The most common method for obtaining factor
which in turn may be used as outcome measures scores is known as unit weighting. In the present
in the study of potential predictors of patient sat- case, because all 14 questions were measured us-
isfaction. In the present context, this would re- ing the same five-point scale, unit weighting sim-
quire that, for each patient, the responses to the ply involves summation of the responses to the dif-
14 questions (Table 1) be summarized using two ferent questions (if different measurement scales
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ACADEMIC EMERGENCY MEDICINE • January 1999, Volume 6, Number 1 63

CHART 1 (cont).

*The above is taken directly from the LISREL 7 output for the final two-factor solution. For the meaning of the survey questions,
see Table 1.

were used, then the responses would have to be identical factor loadings in the final measurement
transformed into a common metric before unit model (Chart 1), and therefore theoretically should
weighting could meaningfully be conducted). Using not be weighted equivalently. Unit weighting also
the full 14-item measurement model, for example, ignores the influence of the four correlated error
each patient’s score on the Nurses factor would be terms (Table 4, Chart 1). However, simulation re-
computed by summing the patient’s responses to search has shown that unit-weighted factors are
the six questions in the Nurses section, plus the often nearly perfectly correlated with ‘‘true’’ fac-
first question (waiting time) in the Doctors section: tor scores (e.g., computed by LISREL, EQS, or
scores on this Nurses factor could range between CALIS), which include all aspects of the measure-
7 and 35. Each patient’s score on the Doctors factor ment model. Indeed, using the present data, unit-
would be computed by summing the patient’s re- weighted and true factor scores correlated 0.99 for
sponses to the eight questions in the Doctors sec- both factors. Nevertheless, as user-friendly soft-
tion: scores on the Doctors factor could range be- ware becomes more widely available, increasing
tween 8 and 40. Note that if one’s hypothesis numbers of researchers are reporting results based
involves comparing mean scores on the Nurses and on true factor scores, rather than unit-weighted
Doctors factors within subjects, then it is essential factor scores.
that the metric underlying the factor scores is com- Besides indicating how to obtain appropriate
mon. To contrast mean scores on the two factors Nurses and Doctors factor scores based on the cur-
within subjects, we could divide scores on the rent 14-item patient satisfaction questionnaire,
Nurses factor by 7 (the number of questions) and CFA may also be useful in creating a more stream-
divide scores on the Doctors factor by 8. This would lined survey having fewer questions than the orig-
allow us to meaningfully compare 1) each respon- inal questionnaire, but equivalent goodness-of-fit.
dent’s level of satisfaction with nurses with 2) the For example, the two-factor model for the 14-item
same respondent’s level of satisfaction with doc- data set indicated that it was more appropriate to
tors, using the original ‘‘metric’’ of the five-point let the ‘‘waiting time’’ question load on both Nurses
rating scale for each factor score. and Doctors factors, rather than on the Doctors
Clearly, in the present context, unit weighting factor only, as was originally postulated. However,
ignores the fact that the variables did not have the factor loading coefficient for the waiting time
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
64 USING CFA Bryant et al. • CONFIRMATORY FACTOR ANALYSIS

CHART 2. LISREL 7 Computer Code for the Final Two-factor Model

question was relatively low for both Nurses (0.36) ‘‘advice you were given about caring for yourself at
and Doctors (0.31), suggesting that this question home, or obtaining follow-up medical care,’’ with
did not tap either factor well in an absolute sense loadings of 0.85 and 0.80, respectively, on the Doc-
(all other questions loaded at least 0.80). Thus, tors factor; 3) ‘‘doctors’ concern to explain your test
eliminating the waiting time question would im- and treatment’’ and ‘‘advice you were given about
prove the conceptual clarity and reliability of the caring for yourself at home, or obtaining follow-up
Nurses and Doctors factors. Although this reduces medical care,’’ with loadings of 0.89 and 0.80, re-
the number of questions only by one in the present spectively, on the Doctors factor; and 4) ‘‘degree to
case, it is not unusual to distill a smaller ‘‘core’’ of which the doctor took your problem seriously’’ and
survey questions from instruments that were orig- ‘‘courtesy of the doctor,’’ with loadings of 0.89 and
inally larger. 0.86, respectively, on the Doctors factor. For each
In addition, the final two-factor model con- pair of questions with correlated errors, we omit-
tained four correlated error terms. Correlated er- ted the question with the lower factor loading and
rors are theoretically problematic, since a basic as- then reestimated the final two-factor model for this
sumption of classic measurement theory involves new subset of ten questions.
the absence of shared measurement error between Results revealed large modification indices for
variables.10 We therefore inspected the factor load- two more correlated-error terms in the model: ‘‘doc-
ings for each of the four pairs of questions that tors’ concern to explain your test and treatment’’
shared measurement error in the final two-factor and ‘‘overall explanation of your illness/injury’’ (MI
model, in order to discard the question with the = 232.8); and ‘‘degree to which the doctor took your
lower factor loading. Questions with shared mea- problem seriously’’ and ‘‘doctors’ concern for your
surement error and their factor loadings were: 1) comfort while treating you’’ (MI = 106.4). In both
‘‘degree to which the nurses took your problem se- cases, the latter question had a lower loading on
riously’’ and ‘‘courtesy of the nurses,’’ with loadings the Doctors factor. Omitting the question with the
of 0.89 and 0.86, respectively, on the Nurses factor; lower loading for each pair yielded a final refined
2) ‘‘overall explanation of your illness/injury’’ and survey containing five Nurses questions and three
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ACADEMIC EMERGENCY MEDICINE • January 1999, Volume 6, Number 1 65

Doctors questions, for a total of eight questions (as taining copies of a wide variety of self-report ques-
opposed to 14 questions on the original survey). tionnaires, survey forms, projective measures, and
CFA revealed that two correlated factors — Nurses observational checklists in the health and social
and Doctors — provided an acceptable measure- sciences. HaPI enables developers of new instru-
ment model for this subset of eight questions, ments to build on the prior work of others without
␹2(19, n = 1,614) = 350.5; GFI = 0.95; AGFI = 0.90; ‘‘reinventing the wheel,’’ and it provides a means
RMSR = 0.021. Also, the correlation between the of systematically selecting criterion measures for
Nurses and Doctors factors was 0.73 — identical to use in assessing the convergent and discriminant
that found analyzing all 14 questions. validity of new measures. By examining methods
The refined measurement model achieves a that prior researchers have used to construct ques-
goodness-of-fit comparable to that of the full 14- tions and response scales, and by studying findings
item model, but contains no correlated errors and of factor analyses of these prior instruments, re-
uses only about half as many questions (i.e., is searchers not only may gain insights into how they
more parsimonious). As mentioned earlier, how- wish to pursue their specific measurement needs,
ever, before relying on the refined questionnaire in but may also avoid pitfalls encountered in prior re-
actual practice, it is best to collect data using it search.
with an independent random sample, and then to
impose the measurement model from the original CONCLUSION
analysis on these new data to evaluate the model’s
goodness-of-fit. Several benefits can be gained by applying CFA in
The logic of the CFA approach to measurement EM research. First, CFA improves conceptual pre-
modeling can also help researchers in generating cision (i.e., specificity) by enhancing our ability to
new questionnaires. That is, by knowing in ad- label the variables under investigation in theory-
vance how CFA can be used to identify the struc- relevant terms. CFA thus contributes to the con-
ture underlying responses to an instrument, one struct validity of empirical research by refining our
can design a new questionnaire from the outset to understanding of the underlying factors being
have good measurement properties. For example, measured. Theoretical concepts that are otherwise
an important issue in questionnaire construction ‘‘fuzzy’’ can be explicated by dissecting them into
their constituent parts, to improve conceptual clar-
concerns the number of questions that should be
ity.
included to measure each underlying factor. Using
A second benefit of CFA is improved statistical
the same number of questions for each factor pro-
precision (i.e., reliability) in assessing dependent
vides direct comparability of mean factor scores
variables. CFA improves the researcher’s ability to
within subjects, and direct comparability of esti-
detect true relationships, by statistically control-
mates of the internal consistency and stability of
ling for unreliability in the underlying factors that
factor scores without requiring correction for at- is due to measurement error. By helping research-
tenuation due to differences in the number of com- ers develop parsimonious, reliable measurement
posite questions. If possible, all questions should models for their data, CFA improves the validity of
share the same response scale. In addition, re- the statistical conclusions drawn in empirical re-
sponse scales should be designed so as to avoid ex- search.
treme values of skewness, because skewed data at- Confirmatory factor analysis also enables re-
tenuate correlations and constrain relative-fit searchers to take fuller advantage of the multivar-
indices to maximum values less than one.10 Finally, iate nature of their data. Rather than analyzing
the complexity of the language used to construct dependent measures piecemeal or in arbitrary
questions should be comparable both within and sets, researchers can use CFA to determine the
across factors, and questions with highly similar measurement model that best explains subjects’ re-
wording should be avoided to prevent correlated sponses to the entire questionnaire, to obtain the
errors. most reliable measure of the research variables in
Another important resource for researchers at- question. CFA can thus improve conceptual and
tempting to construct new questionnaires is the statistical precision in EM research.
Health and Psychosocial Instruments (HaPI) da- The reader interested in learning more about
tabase of measures in the health and social sci- CFA may consult several sources that provide a
ences.11 Produced by Behavioral Measurement Da- thorough and comprehensive overview of the logic,
tabase Services (Pittsburgh, PA) and available on application, and interpretation of CFA.3,5 – 7,9,12 – 15
line and on CD-ROM from Ovid Technologies (New
York, NY), HaPI provides information about more References
than 63,000 records, including summary descrip-
1. Bjorvell H, Stieg J. Patients’ perception of the health care
tions, information about reliability, validity, and received in an emergency department. Ann Emerg Med. 1991;
appropriate populations, and instructions for ob- 20:734 – 8.
 15532712, 1999, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1553-2712.1999.tb00096.x by National Institutes Of Health Malaysia, Wiley Online Library on [20/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
66 USING CFA Bryant et al. • CONFIRMATORY FACTOR ANALYSIS

2. Carey RG, Seibert JH. A patient survey system to measure tor analysis programs: EQS, EzPATH, LINCS, LISCOMP,
quality improvement: questionnaire reliability and validity. LISREL7, SIMPLIS, and CALIS. Appl Psychol Meas. 1993;
Med Care. 1993; 31:234 – 45. 17:73 – 100.
3. Bryant FB, Yarnold PR. Principal-components analysis and 9. Arbuckle JL. AMOS 3.5 User’s Guide. Chicago, IL:
exploratory and confirmatory factor analysis. In: Grimm LG, SmallWaters Corporation, 1994.
Yarnold PR (eds). Reading and Understanding Multivariate 10. Magnusson D. Test Theory. Reading, MA: Addison-Wesley,
Statistics. Washington, DC: American Psychological Associa- 1967.
tion Books, 1995, pp 99 – 136. 11. Health and Psychosocial Instruments (HaPI)-CD Version
4. Joreskog KG, Sorbom DG. PRELIS: A Program for Multi- 7.06. New York: Ovid Technologies, 1998.
variate Data Screening and Data Summarization. Chicago, IL: 12. Bollen KA. Structural Equations with Latent Variables.
Scientific Software, 1989. New York: Wiley, 1989.
5. Joreskog KG, Sorbom DG. LISREL 7 User’s Reference 13. Byrne BM. A Primer for LISREL: Basic Applications and
Guide. Chicago, IL: Scientific Software, 1989. Programming for Confirmatory Factor Analytic Models. New
6. Bentler PM. EQS Structural Equations Program Manual. York: Springer-Verlag, 1989.
Los Angeles, CA: BMDP Statistical Software, 1989. 14. Hayduk LA. Structural Equation Modeling with LISREL.
7. SAS. User’s Guide. Version 6. Vol 1. Cary, NC: SAS Insti- Baltimore, MD: Johns Hopkins University Press, 1987.
tute, 1990. 15. Long JS. Confirmatory Factor Analysis. Beverly Hills, CA:
8. Waller NG. Software review — seven confirmatory fac- Sage, 1983.

Instructions for Contributors to Commentaries

On occasion, the Academic Emergency Medicine Editorial same issue of AEM. Commentary authors may wish to point
Board asks individuals with special expertise to develop com- out strengths and weaknesses of the study that is under
mentaries (editorials) regarding an article published in AEM. consideration, but should remember that significant peer re-
AEM manuscript submission requirements correspond view has already occurred. Instead, the commentary authors
with the ‘‘uniform requirements for manuscripts submitted might wish to expand on the concept or message that is be-
to biomedical journals’’ (JAMA. 1997; 277:927 – 34). Solicited ing demonstrated in the work they are reviewing. Commen-
commentaries should be sent to the SAEM Editorial Office tary authors are also selected because of their noted exper-
(address below) and should be accompanied by a copyright tise in the area under consideration. It is therefore
and author disclosure form that has been reviewed and appropriate for authors to bring their expertise into play as
signed and an electronic copy on a computer disk. Any word they discuss the article.
processing program is acceptable for this submission; the Commentaries do not require any specific formatting.
type of processing program should be indicated on the label Written permission must be obtained if the author decides
of the disk. to reproduce figures or tables from any other publications.
Send three hard copies and the electronic version of the The sources of reproduced material must be acknowledged
commentary to: Academic Emergency Medicine, Society for in the manuscript.
Academic Emergency Medicine, 901 North Washington Av- Commentaries that include material under considera-
enue, Lansing, MI 48906. Phone: (517) 485-5484; fax: (517) tion by another publication and/or duplicate materials pub-
485-0801; e-mail: [email protected]. lished or submitted elsewhere by the authors cannot be con-
Commentaries should be no longer than ten double- sidered. Copies of similar manuscripts under review or
spaced, single-sided, typewritten pages. Appropriate refer- published elsewhere should be provided when submitting
ences should be incorporated, but should not exceed ten ref- the manuscript for consideration as a commentary.
erences. Tables and figures are also acceptable for Any specific questions regarding these guidelines offered
commentaries, although they are rarely submitted. for authors of commentaries can be directed to the Editor-
The commentary is generally solicited to respond to or in-Chief of Academic Emergency Medicine, Michelle Biros,
reflect upon a specific article that will be published in the MS, MD, at (612) 347-5683.