Content uploaded by Karin Schermelleh-Engel
Author content
All content in this area was uploaded by Karin Schermelleh-Engel
Content may be subject to copyright.
Methods of Psychological Research Online 2003, Vol.8, No.2, pp. 23-74 Department of Psychology
Internet: http://www.mpr-online.de © 2003 University of Koblenz-Landau
Evaluating the Fit of Structural Equation Models:
Tests of Significance and
Descriptive Goodness-of-Fit Measures
Karin Schermelleh-Engel1 and Helfried Moosbrugger
Goethe University, Frankfurt
Hans Müller
University of Erfurt
For structural equation models, a huge variety of fit indices has been developed. These
indices, however, can point to conflicting conclusions about the extent to which a
model actually matches the observed data. The present article provides some guide-
lines that should help applied researchers to evaluate the adequacy of a given struc-
tural equation model. First, as goodness-of-fit measures depend on the method used
for parameter estimation, maximum likelihood (ML) and weighted least squares
(WLS) methods are introduced in the context of structural equation modeling. Then,
the most common goodness-of-fit indices are discussed and some recommendations for
practitioners given. Finally, we generated an artificial data set according to a "true"
model and analyzed two misspecified and two correctly specified models as examples
of poor model fit, adequate fit, and good fit.
Keywords: Structural equation modeling, model fit, goodness-of-fit indices, standardized
residuals, model parsimony
In structural equation modeling (SEM), a model is said to fit the observed data to
the extent that the model-implied covariance matrix is equivalent to the empirical co-
variance matrix. Once a model has been specified and the empirical covariance matrix is
given, a method has to be selected for parameter estimation. Different estimation meth-
ods have different distributional assumptions and have different discrepancy functions
to be minimized. When the estimation procedure has converged to a reasonable solu-
1 Correspondence concerning this article should be addressed to Dr. Karin Schermelleh-Engel, Goethe
University, Institute of Psychology, Mertonstrasse 17, 60054 Frankfurt am Main, Germany. E-mail:
schermelleh-engel@psych.uni-frankfurt.de.
24 MPR-Online 2003, Vol. 8, No. 2
tion, the fit of the model should be evaluated. Model fit determines the degree to which
the structural equation model fits the sample data. Although there are no well-
established guidelines for what minimal conditions constitute an adequate fit, a general
approach is to establish that the model is identified, that the iterative estimation proce-
dure converges, that all parameter estimates are within the range of permissible values,
and that the standard errors of the parameter estimates have reasonable size (Marsh &
Grayson, 1995). Furthermore, the standardized residuals should be checked for patterns
in the residual matrix as a sign of ill fit. As Hayduk (1996) states, it is the difference
between the empirical and the model-implied covariance matrix "that drives all tests of
overall fit, and systematic differences here, even if small, warrant caution" (p. 198).
Applied researchers often have difficulty determining the adequacy of structural
equation models because various measures of model fit point to conflicting conclusions
about the extent to which the model actually matches the observed data. Software pro-
grams such as LISREL (Jöreskog & Sörbom, 1996), EQS (Bentler, 1995), Mplus
(Muthén & Muthén, 1998), AMOS (Arbuckle & Wothke, 1999), SEPATH (Steiger,
1995), or RAMONA (Browne & Mels, 1992), among others, provide a variety of fit indi-
ces for model evaluation. As there does not exist a consensus about what constitutes a
"good fit" (Tanaka, 1993), the fit indices should be considered simultaneously.
In the following, we will first give a short overview over two methods frequently used
for parameter estimation in structural equation modeling, i.e., maximum likelihood
(ML) and weighted least squares (WLS)2. Second, we will discuss some common good-
ness-of-fit measures provided by the LISREL program, as most of these indices are pro-
vided by other software programs, too. Third, we will give some recommendations for
evaluating the fit of structural equation models. Finally, using an artificial data set we
will evaluate the fit of four alternative models as examples of poor model fit, adequate
fit, and good fit.3
2 Besides these, several other methods exist, but they will not be discussed here in detail. For further
information on estimation methods we recommend the respective book chapters in Bollen (1989), Bollen
and Long (1993), and Kaplan (2000).
3 To facilitate orientation, a detailed table of contents is provided as a separate file next to this article at
http://www.mpr-online.de
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 25
Methods for Parameter Estimation
Maximum Likelihood (ML)
Maximum Likelihood (ML) is the most widely used fitting function for structural
equation models. Nearly all of the major software programs use ML as the default esti-
mator. This method leads to estimates for the parameters θ which maximize the likeli-
hood L that the empirical covariance matrix S is drawn from a population for which the
model-implied covariance matrix )(
θ
Σ is valid. The log-likelihood function log L to be
maximized is (Bollen, 1989, p. 135)
cθθNL )(trlog)1(
2
1
log }{ ][ 1++−−= −
SΣ)Σ( (1)
where
log is the natural logarithm,
L is the likelihood function,
N is the sample size,
θ is the parameter vector,
)(θΣ is the model-implied covariance matrix and )(θΣ its determinant,
tr is the trace of a matrix, and
c is a constant that contains terms of the Wishart distribution that do not change
once the sample is given (Bollen, 1989, p. 135).
Maximizing log L is equivalent to minimizing the function
pθθF−+−= −])([trlog)(log 1
ML SΣSΣ (2)
where
F
ML is the value of the fitting function evaluated at the final estimates (cf. Hayduk,
1989, p. 137), and
p is the number of observed variables.
The ML estimator assumes that the variables in the model are multivariate normal,
i.e., the joint distribution of the variables is a multivariate normal distribution. Fur-
26 MPR-Online 2003, Vol. 8, No. 2
thermore it is assumed that )(θΣ and S are positive definite, which implies that these
matrices must be nonsingular. ML estimators have several important properties (cf. Bol-
len, 1989). If the observed data stem from a multivariate normal distribution, if the
model is specified correctly, and if the sample size is sufficiently large, ML provides pa-
rameter estimates and standard errors that are asymptotically unbiased, consistent, and
efficient. Furthermore, with increasing sample size the distribution of the estimator ap-
proximates a normal distribution. Thus, the ratio of each estimated parameter to its
standard error is approximately z-distributed in large samples.
An important advantage of ML is that it allows for a formal statistical test of overall
model fit for overidentified models. The asymptotic distribution of (N − 1)FML is a χ2
distribution with df = s – t degrees of freedom, where s is the number of nonredundant
elements in S and t is the number of free parameters. Another advantage of ML is that
its estimates are in general scale invariant and scale free (Bollen, 1989, p. 109). As a
consequence, the values of the fit function do not depend on whether correlation or co-
variance matrices are analyzed, and whether original or transformed data are used.
A limitation of ML estimation is the strong assumption of multivariate normality, as
violations of distributional assumptions are common and often unavoidable in practice
and can potentially lead to seriously misleading results. Nevertheless, ML seems to be
quite robust against the violation of the normality assumption (cf. Boomsma &
Hoogland, 2001; Chou & Bentler, 1995; Curran, West & Finch, 1996; Muthén &
Muthén, 2002; West, Finch, & Curran, 1995). Simulation studies suggest that under
conditions of severe nonnormality, ML parameter estimates are still consistent but not
necessarily efficient. Using the χ2 as a measure of model fit (see below) will lead to an
inflated Type I error rate for model rejection (Curran, West, & Finch, 1996; West,
Finch, & Curran, 1995). For ML estimation with small samples, bootstrapping (Efron &
Tibshirani, 1993) may be an alternative (Shipley, 2000).
Corrections have been developed to adjust ML estimators to account for nonnormal-
ity (for an overview of robust, corrective statistics, see Satorra & Bentler, 1994). The
Satorra-Bentler scaled χ2 is computed on the basis of the model, estimation method, and
sample fourth-order moments and holds regardless of the distribution of the observed
variables (Hu & Bentler, 1995, p. 79). As simulation studies demonstrate, robust maxi-
mum likelihood estimators based on the Satorra-Bentler scaled χ2 statistic have rela-
tively good statistical properties compared to least squares estimators (Boomsma &
Hoogland, 2001; Hoogland, 1999).
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 27
In robustness studies, the scaled χ2 statistic outperformed the standard ML estimator
(Curran, West, & Finch, 1996; Chou, Bentler, & Satorra, 1991), and robust standard
errors yielded the least biased standard errors, especially when the distributions of the
observed variables were extremely nonnormal (Chou & Bentler, 1995). But as robust
maximum likelihood estimation needs relatively large sample sizes of at least N ≥ 400
(Boomsma & Hoogland, 2001) or even N ≥ 2,000 (Yang-Wallentin & Jöreskog, 2001),
further studies are necessary to determine the advantages and disadvantages of these
methods in small to moderate sample sizes.
Weighted Least Squares (WLS)
If the data are continuous but nonnormal, the estimation method most often recom-
mended is the asymptotically distribution free (ADF) method (Browne, 1984), although
simulation studies suggest that ML estimation with or without a correction for non-
normality seems to perform better than ADF and should be preferred (Boomsma &
Hoogland, 2001; Hu, Bentler, & Kano, 1992; Olsson, Foss, Troye, & Howell, 2000). The
ADF method is available in LISREL under the name "weighted least squares (WLS)"
and in EQS under "arbitrary distribution generalized least squares (AGLS)". In contrast
to ML, raw data are needed for data analysis. This method may also be used if some of
the observed variables are ordinal and others continuous, if the distributions of the con-
tinuous variables deviate considerably from normality, or if models include dichotomous
variables.
WLS minimizes the fit function
[][]
)(
'
)( 1
WLS θθFσσ −−= −sWs (3)
where
s is the vector of nonredundant elements in the empirical covariance matrix,
)(θσ is the vector of nonredundant elements in the model-implied covariance ma-
trix,
θ is the (t × 1) vector of parameters,
W
–1 is a (k × k) positive definite weight matrix with k = p( p + 1)/2 and p = num-
ber of observed variables.
28 MPR-Online 2003, Vol. 8, No. 2
WLS requires that the matrix W is a consistent estimate of the asymptotic covari-
ance matrix of the sample variances and covariances (or correlations) being analyzed
(Browne, 1984; see also Kaplan, 2000, p. 81f.). The elements of W –1 are usually ob-
tained by inverting a matrix W which is chosen in such a way that a typical element
wijgh is proportional to a consistent estimate of the asymptotic covariance of sij with sgh:
ijgh
jgihjhigghij
ijgh N
N
ssacovNw κ
1
σσσσ),()1( −
++=−= (4)
where
κijgh is the fourth-order cumulant, a component of the distribution related to the
multivariate kurtosis,
sij , sgh are sample covariances, and
σig, σjh, σih, and σjg are population covariances of Xi with Xg, Xj with Xh, Xi with Xh,
and Xj with Xg, respectively.
The WLS method has several advantages, yet also some disadvantages (Bollen, 1989,
p. 432). One main advantage is that it requires only minimal assumptions about the
distribution of the observed variables. Simulation research with nonnormal data shows
that the WLS test statistic is relatively unaffected by distributional characteristics
(Hoogland & Boomsma, 1998; West, Finch, & Curran, 1995). Another advantage is that
WLS may also be used as a means of analyzing correlation matrices, if the correspond-
ing matrix W contains the covariances of the correlations rij and rgh. It should be noted
that this matrix differs from the one containing the covariances sij and sgh. In general,
WLS produces an accurate χ2 test statistic and accurate standard errors if sample size is
sufficiently large.
A limitation of the WLS method can be seen in the fact that the weight matrix
grows rapidly with increasing numbers of indicator variables. As the asymptotic covari-
ance matrix is of order (k × k), where k = p( p + 1)/2 and p is the number of observed
variables, the weight matrix of a model containing 10 variables would be of order (55 ×
55) with 1540 nonredundant elements. Thus, the WLS method compared to ML re-
quires large samples in order to obtain consistent and efficient estimates. If the distribu-
tion of the observed variables does not deviate from the normal distribution by a con-
siderable amount, one may also apply ML. Consistent with previous findings (cf. Chou,
Bentler, & Satorra, 1991; Muthén & Kaplan, 1985, 1992), Chou and Bentler (1995) do
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 29
not recommend WLS for practical applications when models are complex and when the
sample size is small.
Recent developments of WLS-based estimators under nonnormality suggest that
WLS for categorical outcomes as implemented in Mplus (Muthén & Muthén, 1998) does
not require the same large sample sizes as WLS for continuous non-normal data (cf.
Kaplan, 2000, p. 85ff.). But further studies are needed to determine the advantages and
disadvantages of this method compared to other robust estimators.
Special cases of WLS estimation are the Generalized Least Squares (GLS) and the
Unweighted Least Squares (ULS) estimators.
Under the assumption of multivariate normality, the WLS fitting function FWLS can
be rewritten as
21
WLS }{ )]([tr
2
1−
−= VΣSθF (5)
(Bollen, 1989, p. 428), where
tr is the trace of the matrix,
S is the empirical covariance matrix,
)(θΣis the model-implied covariance matrix,
θ is the (t × 1) vector of parameters, and
V–1 is a p × p weight matrix (a weight matrix of lower dimensions).
Equation 5 is the general form of the Generalized Least Squares (GLS) estimator in
which the k × k weight matrix W –1 of Equation 3 is replaced by the p × p weight matrix
V–1.
Inserting S for V (the most common choice which is found both in LISREL and in
EQS) leads to the Generalized Least Squares (GLS) estimator, a special case of WLS,
with the fitting function
21
GLS }{ )]([tr
2
1−
−= SΣSθF, (6)
where
tr is the trace of the matrix,
30 MPR-Online 2003, Vol. 8, No. 2
S is the empirical covariance matrix and S–1 its inverse,
)(θΣis the model-implied covariance matrix, and
θ is the (t × 1) vector of parameters.
Generalized Least Squares (GLS) is a frequently used estimation method that is as-
ymptotically equivalent to FML. As GLS is based on the same assumptions as ML, this
estimation method is used under the same conditions. But as it performs less well in
small samples, maximum likelihood should generally be preferred with small sample
sizes.
Finally, with the identity matrix I chosen as the weight matrix V –1, GLS reduces to
Unweighted Least Squares (ULS) estimator, another special case of WLS. For parame-
ter estimation, the following fit function is minimized:
2
ULS })]([tr
2
1{θFΣS−= , (7)
where
tr is the trace of the matrix,
S is the empirical covariance matrix,
)(θΣis the model-implied covariance matrix, and
θ is the (t × 1) vector of parameters.
The fitting function FULS minimizes the sum of squares of each element in the residual
matrix ))(( θΣS−. ULS offers the advantage that it leads to a consistent estimator of θ
comparable to ML and WLS, yet in contrast to ML, distributional assumptions are not
necessary. Disadvantages are that ULS does not provide the most efficient estimates for
θ and that these estimates are neither scale invariant nor scale free (Bollen, 1989). Fur-
thermore, some software programs do not provide χ2 statistics and standard errors if
ULS is applied. Others estimate standard errors and χ2 for ULS only under the assump-
tion of multivariate normality. Therefore, ULS results should be interpreted with cau-
tion.
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 31
Evaluation of Model Fit
In structural equation modeling, evaluation of model fit is not as straightforward as
it is in statistical approaches based on variables measured without error. Because there
is no single statistical significance test that identifies a correct model given the sample
data, it is necessary to take multiple criteria into consideration and to evaluate model
fit on the basis of various measures simultaneously. For each estimation procedure, a
large number of goodness-of-fit indices is provided to judge whether the model is consis-
tent with the empirical data. The choice of the estimation procedure depends on the
type of data included in the model.
Generally, the fit criteria of a structural equation model indicate to what extent the
specified model fits the empirical data. Only one goodness-of-fit measure, i.e., the χ2 test
statistic, has an associated significance test, while all other measures are descriptive.
Thus, following successful parameter estimation, model evaluation can be assessed infer-
entially by the χ2 test or descriptively by applying other criteria.
For inferential statistical evaluation, only the χ2 test is available, whereas for descrip-
tive evaluation, three main classes of criteria exist, i.e., measures of overall model fit,
measures based on model comparisons, and measures of model parsimony (cf.
Schumacker & Lomax, 1996, p. 119 ff.). Most of the descriptive fit criteria are based on
the χ2 statistic given by the product of the sample size (N − 1) and the optimized fitting
function.
In the following, we will present a selection of fit indices that are provided by the
LISREL program (Jöreskog & Sörbom, 1993). Most of these indices are reported by
other software programs as well so that this presentation should be also beneficial for
users of EQS, Mplus, and other software programs for SEM.
Part I: Tests of Significance
χ
2 Test Statistic
The χ2 test statistic is used for hypothesis testing to evaluate the appropriateness of a
structural equation model. If the distributional assumptions are fulfilled, the χ2 test
evaluates whether the population covariance matrix Σ is equal to the model-implied
covariance matrix )(θΣ, i.e., it tests the null hypothesis that the differences between
32 MPR-Online 2003, Vol. 8, No. 2
the elements of Σ and )(θΣ are all zero: 0ΣΣ =− )(θ. Being population parameters,
these matrices are unknown, so researchers examine their sample counterparts, the em-
pirical covariance matrix S and the model-implied covariance matrix )
ˆ
(θΣ, where θ
ˆ is
the (t × 1) vector of estimated parameters. If the null hypothesis is correct, the mini-
mum fit function value times N − 1 converges to a χ2 variate
)]
ˆ
(,[)1()(χ2θFNdf ΣS−= (8)
with df = s − t degrees of freedom, where
s is the number of nonredundant elements in S,
t is the total number of parameters to be estimated,
N is the sample size,
S is the empirical covariance matrix, and
)
ˆ
(θΣ is the model-implied covariance matrix.
LISREL provides different χ2 test statistics for ML, WLS, GLS, or ULS, so that the
obtained χ2 value depends on the estimation method. In general, high χ2 values in rela-
tion to the number of degrees of freedom indicate that the population covariance matrix
Σ and the model-implied covariance matrix )(θΣ differ significantly from each other.
As the residuals, namely, the elements of )
ˆ
(θΣS−, should be close to zero for a good
model fit, the researcher is interested in obtaining a nonsignificant χ2 value with associ-
ated degrees of freedom. If the p-value associated with the χ2 value is larger than .05,
the null hypothesis is accepted and the model is regarded as compatible with the popu-
lation covariance matrix Σ. In this case the test states that the model fits the data.
But still an uncertainty exists that other models may fit the data equally well (cf. Lee
& Hershberger, 1990; Stelzl, 1986).
There are several shortcomings associated with the χ2 test statistic.
- Violation of assumptions. The χ2 test is based on the assumptions that the observed
variables are multivariate normal and that the sample size is sufficiently large. How-
ever, these assumptions are not met in many practical applications.
- Model complexity. One disadvantage of the χ2 value is that it decreases when pa-
rameters are added to the model. Thus, the χ2 value of a more complex, highly pa-
rameterized model tends to be smaller than for simpler models because of the reduc-
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 33
tion in degrees of freedom. For example, the saturated model with as many free pa-
rameters as there are variances and covariances in S yields a χ2 of zero, whereas the
independence model, a very restrictive model, usually has a very large χ2 value (cf.
Mueller, 1996, p. 89). Thus, a good model fit may result either from a correctly
specified model or from a highly overparameterized model.
- Dependence on sample size. With increasing sample size and a constant number of
degrees of freedom, the χ2 value increases. This leads to the problem that plausible
models might be rejected based on a significant χ2 statistic even though the discrep-
ancy between the sample and the model-implied covariance matrix is actually irrele-
vant. On the other hand, as sample size decreases, the χ2 value decreases as well and
the model test may indicate nonsignificant probability levels even though the dis-
crepancy between the sample and the model-implied covariance matrix is consider-
able. Therefore not too much emphasis should be placed on the significance of the χ2
statistic. Jöreskog and Sörbom (1993) even suggest to use χ2 not as a formal test
statistic but rather as a descriptive goodness-of-fit index. They propose to compare
the magnitude of χ2 with the expected value of the sample distribution, i.e., the
number of degrees of freedom, as E(χ2) = df. For a good model fit, the ratio χ2/df
should be as small as possible. As there exist no absolute standards, a ratio between
2 and 3 is indicative of a "good" or "acceptable" data-model fit, respectively. How-
ever, the problem of sample size dependency cannot be eliminated by this procedure
(Bollen, 1989, p. 278).
χ
2 Difference Test
In applications of covariance structure analysis, researchers often face the problem of
choosing among two or more alternative models. The choice of which measure to use for
selecting one of several competing models depends on whether or not the models are
nested.
A specific model (Model A) is said to be nested within a less restricted model (Model
B) with more parameters and less degrees of freedom than Model A, if Model A can be
derived from Model B by fixing at least one free parameter in Model B or by introduc-
ing other restrictions, e.g., by constraining a free parameter to equal one or more other
parameters. For example, in multi-sample comparisons of factorial invariance, any
model with some parameters constrained to be invariant over the multiple groups is
nested under the corresponding model in which the respective parameters are uncon-
strained, and the model in which all parameters are invariant is nested under both these
34 MPR-Online 2003, Vol. 8, No. 2
models. Any two models are nested when the free parameters in the more restrictive
model are a subset of the free parameters in the less restrictive model.
As the test statistic of each of the nested models follows a χ2 distribution, the differ-
ence in χ2 values between two nested models is also χ2 distributed (Steiger, Shapiro, &
Browne, 1985), and the number of degrees of freedom for the difference is equal to the
difference in degrees of freedom for the two models. Under appropriate assumptions, the
difference in model fit can be tested using the χ2 difference test
)()()( B
2
BA
2
Adiff
2
diff dfdfdf χ−χ=χ (9)
(Bentler, 1990; Bollen, 1989; Jöreskog, 1993), where
2
A
χ denotes the χ2 value of Model A, a model that is a restricted version of Model B,
i.e., Model A has less free parameters and more degrees of freedom (dfA) and is thus
nested within Model B,
2
B
χ denotes the χ2 value of Model B, a model that is less restricted and therefore has
more free parameters and less degrees of freedom (dfB) than Model A, and
dfdiff = dfA − dfB.
If the χ2 difference is significant, the null hypothesis of equal fit for both models is re-
jected and Model B should be retained. But if the χ2 difference is nonsignificant, which
means that the fit of the restricted model (Model A) is not significantly worse than the
fit of the unrestricted model (Model B), the null hypothesis of equal fit for both models
cannot be rejected and the restricted model (Model A) should be favored.
The χ2 difference test applied to nested models has essentially the same strengths and
weaknesses as the χ2 test applied to any single model, namely, the test is directly af-
fected by sample size, and for large samples trivial differences may become significant.
For the χ2 difference test to be valid, at least the least restrictive model of a sequence of
models (in our example Model B) should fit the data.
It should be noted that the Satorra-Bentler scaled χ2 values resulting from robust es-
timation methods cannot be used for χ2 difference testing because the difference between
two scaled χ2 values for nested models is not distributed as a χ2 (Satorra, 2000). Re-
cently, Satorra and Bentler (2001) developed a scaled difference χ2 test statistic for
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 35
moment structure analysis. They could show that simple hand calculations based on
output from nested runs can give the desired χ2 difference test of nested models using
the scaled χ2. These calculations may be obtained from Mplus by asking for the MLM
estimator.
If models are not nested, they may be compared on the basis of descriptive goodness-
of-fit measures that take parsimony as well as fit into account, e.g., the Akaike Informa-
tion Criterion (Akaike, 1974, 1987), which can be used regardless of whether models for
the same data can be ordered in a nested sequence or not (see below). A more detailed
discussion of alternative methods for comparing competing models is given in Kumar
and Sharma (1999), Raykov and Penev (1998), and Rigdon (1999).
Part II: Descriptive Goodness-of-Fit Measures
Because of the drawbacks of the χ2 goodness-of-fit tests, numerous descriptive fit in-
dices have been developed that are often assessed intuitively. These indices are derived
from ML, WLS, GLS, or ULS, but in the following we will not differentiate between
these methods (for further information on ML-, WLS-, and GLS-based descriptive fit
indices, cf. Hu & Bentler, 1998). Many of these measures are intended to range between
zero (no fit) and one (perfect fit), but as Hu and Bentler (1995) note, the sampling dis-
tributions of goodness-of-fit indices are unknown with the exception of χ2 so that critical
values for fit indices are not defined. As a reasonable minimum for model acceptance,
Bentler and Bonett (1980) proposed a value of .90 for normed indices that are not par-
simony adjusted (cf. Hoyle & Panter, 1995), while .95 should be indicative of a good fit
relative to the baseline model (Kaplan, 2000). But recently, Hu and Bentler (1995, 1998,
1999) gave evidence that .90 might not be a reasonable cutoff for all fit indices under all
circumstances: "The rule of thumb to consider models acceptable if a fit index exceeds
.90 is clearly an inadequate rule" (Hu & Bentler, 1995, p. 95). They suggested to raise
the rule of thumb minimum standard for the CFI and the NNFI (see below) from .90 to
.95 to reduce the number of severely misspecified models that are considered acceptable
based on the .90 criterion (Hu & Bentler, 1998, 1999).
Descriptive Measures of Overall Model Fit
Due to the sensitivity of the χ2 statistic to sample size, alternative goodness-of-fit
measures have been developed. Measures of overall model fit indicate to which extent a
36 MPR-Online 2003, Vol. 8, No. 2
structural equation model corresponds to the empirical data. These criteria are based on
the difference between the sample covariance matrix S and the model-implied covari-
ance matrix )
ˆ
(θΣ. The following indices are descriptive measures of overall model fit:
Root Mean Square Error of Approximation (RMSEA), Root Mean Square Residual
(RMR), and Standardized Root Mean Square Residual (SRMR).
Root Mean Square Error of Approximation (RMSEA)
The usual test of the null hypothesis of exact fit is invariably false in practical situa-
tions and will almost certainly be rejected if sample size is sufficiently large. Therefore a
more sensible approach seems to be to assess whether the model fits approximately well
in the population (cf. Kaplan, 2000, p. 111). The null hypothesis of exact fit is replaced
by the null hypothesis of "close fit" (Browne & Cudeck, 1993, p. 146). Thus, the Root
Mean Square Error of Approximation (RMSEA; Steiger, 1990) is a measure of approxi-
mate fit in the population and is therefore concerned with the discrepancy due to ap-
proximation.
RMSEA is estimated by a
ˆ
ε,
the square root of the estimated discrepancy due to ap-
proximation per degree of freedom:
(())
a
ˆ
,1
ˆmax , 0
1
⎧⎫
⎛⎞
⎪⎪
θ⎟
⎪⎪
⎜⎟
ε=−
⎜
⎨⎬
⎟
⎜⎟
⎪⎪
⎟
⎜−
⎝⎠
⎪⎪
⎩⎭
F
df N
SΣ (10)
where
))
ˆ
(,( θFΣS is the minimum of the fit function,
df = s – t is the number of degrees of freedom, and
N is the sample size.
The RMSEA is bounded below by zero. Steiger (1990) as well as Browne and Cudeck
(1993) define a "close fit" as a RMSEA value less than or equal to .05. According to
Browne and Cudeck (1993), RMSEA values ≤ .05 can be considered as a good fit, values
between .05 and .08 as an adequate fit, and values between .08 and .10 as a mediocre
fit, whereas values > .10 are not acceptable. Although there is general agreement that
the value of RMSEA for a good model should be less than .05, Hu and Bentler (1999)
suggested an RMSEA of less than .06 as a cutoff criterion. In addition, a 90% confi-
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 37
dence interval (CI) around the point estimate enables an assessment of the precision of
the RMSEA estimate. On the basis of the CI, it is possible to say with a certain level of
confidence that the given interval contains the true value of the fit index for that model
in the population (MacCallum, Browne, & Sugawara, 1996). The lower boundary (left
side) of the confidence interval should contain zero for exact fit and be < .05 for close
fit. Note that when the model fits well in the population, the lower end of the confi-
dence interval is truncated at zero, which leads to an asymmetry of the confidence in-
terval. RMSEA is regarded as relatively independent of sample size, and additionally
favors parsimonious models (Browne & Cudeck, 1993; Kaplan, 2000).
For an understanding of RMSEA it is important to distinguish between two different
kinds of error. The error of approximation, which is of primary interest here, represents
the lack of fit of the model to the population covariance matrix Σ. The minimum fit
function value one would obtain if the model could be fitted to the population covari-
ance matrix is a possible measure of this error. In contrast, the error of estimation re-
flects the differences between the model fitted to the population covariance matrix Σ
(if this could be done) and the model fitted to the sample covariance matrix S (Browne
& Cudeck, 1993, p. 141). From the viewpoint of model fit in the population, the error of
estimation is of secondary interest only. Because the fit function value ))
ˆ
(,( θFΣS, which
refers to the sample covariance matrix, would be a biased estimator of the population
error of approximation, RMSEA includes a term inversely proportional to N – 1 which
serves as a correction for bias (Browne & Cudeck, 1993, p. 143).
Root Mean Square Residual (RMR) and Standardized RMR (SRMR)
It was already mentioned that the residuals are given by the elements of the matrix
)
ˆ
(θΣS−. These are sometimes called "fitted residuals" because they express the remain-
ing discrepancies between the covariance matrices S and )
ˆ
(θΣ once the parameters of
the model are estimated.
The Root Mean Square Residual index (RMR) of Jöreskog and Sörbom (1981, p. 41;
1989) is an overall badness-of-fit measure that is based on the fitted residuals. Con-
cretely, RMR is defined as the square root of the mean of the squared fitted residuals,
2/)1(
)
ˆ
(
11
2
+
∑∑ −
===
pp
s
RMR
p
i
i
j
ijij
σ
(11)
38 MPR-Online 2003, Vol. 8, No. 2
where
sij is an element of the empirical covariance matrix S,
ij
σ
ˆ is an element of the model-implied matrix covariance )
ˆ
(θΣ, and
p is the number of observed variables.
In principle, RMR values close to zero suggest a good fit. But as the elements of S
and )
ˆ
(θΣ are scale dependent, the fitted residuals are scale dependent, too, which im-
plies that RMR depends on the sizes of the variances and covariances of the observed
variables. In other words, without taking the scales of the variables into account it is
virtually impossible to say whether a given RMR value indicates good or bad fit.
To overcome this problem, the Standardized Root Mean Square Residual (SRMR)
has been introduced (Bentler, 1995, p. 271). Here, the residuals ijij
s
σ
ˆ
− are first divided
by the standard deviations iii ss = and jjj ss = of the respective manifest variables,
which leads to a standardized residual matrix with elements ˆ
()/()
ij ij i j
−σ =sss
ˆ/( )
ij ij i j
r−σ ss where rij is the observed correlation between the respective variables
(Bentler, 1995, p. 90). In contrast to LISREL (see below), EQS provides this matrix
explicitly. Calculating the root mean square of the such defined standardized residuals
in analogy to Equation 11 leads to the SRMR which is available in both LISREL and
EQS. Again, a value of zero indicates perfect fit, but it is still difficult to designate cut-
off values for good and for acceptable fit because of sample size dependency and sensi-
tivity to misspecified models (Hu & Bentler, 1998). A rule of thumb is that the SRMR
should be less than .05 for a good fit (Hu & Bentler, 1995), whereas values smaller than
.10 may be interpreted as acceptable.
The standardized residuals given above are similar – but in general not identical – to
the correlation residuals suggested by Bollen (1989, p. 258). Both share the main idea
that in an overall fit measure based on residuals, the observed variables should enter in
standardized form such that all matrix elements contributing to the fit measure are on
comparable scales. Thus the SRMR, same as the RMR, remains a purely descriptive fit
index.
Unfortunately the term "standardized residuals" also appears in a second meaning
that is independent of the SRMR. In this meaning, the residuals themselves get stan-
dardized: Besides the fitted residuals, the LISREL program provides a matrix of stan-
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 39
dardized residuals which are obtained by dividing each fitted residual by its large-sample
standard error (Jöreskog & Sörbom, 1989, p. 28). Being independent of the units of
measurements of the variables as the standardized residuals discussed before, they also
allow an easier interpretation than the fitted residuals. In the present case, however, the
standardized residuals can be interpreted approximately in an inferential sense, namely,
in a way similar to z scores. Provided that the SRMR or other fit indices signalize bad
fit, single standardized residuals whose absolute values are greater than 1.96 or 2.58 can
be useful for detecting the source of misfit. The largest absolute value indicates the ele-
ment that is most poorly fitted by the model. Because the kind of standardized residu-
als considered here refers to a standard error, the absolute values tend to increase with
increasing sample size if the magnitudes of the fitted residuals remain essentially con-
stant.
As the RMR and the SRMR are overall measures based on squared residuals, they
can give no information about the directions of discrepancies between S and )
ˆ
(θΣ. In a
residual analysis, regardless of whether unstandardized or standardized residuals are
used and which kind of standardization is preferred, it is important to take the sign of a
residual into account when looking for the cause of model misfit. Given that an empiri-
cal covariance is positive, a positive residual indicates that the model underestimates
the sample covariance. In this case, the empirical covariance is larger than the model-
implied covariance. A negative residual indicates that the model overestimates the sam-
ple covariance, that is, the empirical covariance is smaller than the model-implied co-
variance.
Descriptive Measures Based on Model Comparisons
The basic idea of comparison indices is that the fit of a model of interest is compared
to the fit of some baseline model. Even though any model nested hierarchically under
the target model (the model of interest) may serve as a comparison model, the inde-
pendence model is used most often. The independence model assumes that the observed
variables are measured without error, i.e., all error variances are fixed to zero and all
factor loadings are fixed to one, and that all variables are uncorrelated. This baseline
model is a very restrictive model in which only p parameters, namely the variances of
the variables, have to be estimated. An even more restrictive baseline model than the
independence model is the null model, a model in which all parameters are fixed to zero
(Jöreskog & Sörbom, 1993, p. 122) and hence, no parameters have to be estimated. The
40 MPR-Online 2003, Vol. 8, No. 2
fit index for a baseline model will usually indicate a bad model fit and serves as a com-
parison value. The issue is whether the target model is an improvement relative to the
baseline model.
Often used measures based on model comparisons are the Normed Fit Index (NFI ),
the Nonnormed Fit Index (NNFI ), the Comparative Fit Index (CFI ), the Goodness-of-
Fit Index (GFI ), and the Adjusted Goodness-of-Fit Index (AGFI ), which will be ex-
plained below in more detail.
Normed Fit Index (NFI) and Nonnormed Fit Index (NNFI)
The Normed Fit Index (NFI) proposed by Bentler and Bonnett (1980) is defined as
i
t
2
i
2
t
2
i
2
t
2
i1
χ
χ
1
χ
χχ
F
F
NFI −=−=
−
=, (12)
where
2
i
χ is the chi-square of the independence model (baseline model),
2
t
χ is the chi-square of the target model, and
F is the corresponding minimum fit function value.
NFI values range from 0 to 1, with higher values indicating better fit. When Ft = Fi,
NFI equals zero; when Ft = 0, NFI equals one, which suggests that the target model is
the best possible improvement over the independence model. Although the theoretical
boundary of NFI is one, NFI may not reach this upper limit even if the specified model
is correct, especially in small samples (Bentler, 1990, p. 239). This can occur because
the expected value of 2
t
χ is greater than zero: dfE =)χ(2
t. The usual rule of thumb for
this index is that .95 is indicative of good fit relative to the baseline model (Kaplan,
2000, p. 107), whereas values greater than .90 are typically interpreted as indicating an
acceptable fit (Marsh & Grayson, 1995; Schumacker & Lomax, 1996).
A disadvantage of the NFI is that it is affected by sample size (Bearden, Sharma, &
Teel, 1982). In order to take care of this problem, Bentler and Bonnett (1980) extended
the work by Tucker and Lewis (1973) and developed the Nonnormed Fit Index (NNFI ),
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 41
also known as the Tucker-Lewis Index (TLI ). The NNFI measures relative fit and is
defined as
)1/(1)/(
)/()/(
1)/χ(
)/χ()/χ(
ii
ttii
i
2
i
t
2
ti
2
i
−−
−
=
−
−
=NdfF
dfFdfF
df
dfdf
NNFI , (13)
where
2
i
χ is the chi-square of the independence model (baseline model),
2
t
χ is the chi-square of the target model,
F is the corresponding minimum fit function value, and
df is the number of degrees of freedom.
The NNFI ranges in general from zero to one, but as this index is not normed, values
can sometimes leave this range, with higher NNFI values indicating better fit. A rule of
thumb for this index is that .97 is indicative of good fit relative to the independence
model, whereas values greater than .95 may be interpreted as an acceptable fit. As the
independence model almost always has a large χ2, NNFI values are often very close to
one (Jöreskog & Sörbom, 1993, p. 125), so that a value of .97 seems to be more reason-
able as an indication of a good model fit than the often stated cutoff value of .95.
NNFI takes the degrees of freedom of the specified model as well as the degrees of
freedom of the independence model into consideration. More complex, i.e., less restric-
tive models are penalized by a downward adjustment, while more parsimonious, i.e.,
more restrictive models are rewarded by an increase in the fit index. An advantage of
the NNFI is that it is one of the fit indices less affected by sample size (Bentler, 1990;
Bollen, 1990; Hu & Bentler, 1995, 1998).
Comparative Fit Index (CFI)
The Comparative Fit Index (CFI; Bentler, 1990), an adjusted version of the Relative
Noncentrality Index (RNI ) developed by McDonald and Marsh (1990), avoids the un-
derestimation of fit often noted in small samples for Bentler and Bonett's (1980) normed
fit index (NFI ). The CFI is defined as
]0,)(χ,)χ[(max
]0,)[(χmax
1
i
2
it
2
t
t
2
t
dfdf
df
CFI
−−
−
−= (14)
42 MPR-Online 2003, Vol. 8, No. 2
where
max denotes the maximum of the values given in brackets,
2
i
χ is the chi-square of the independence model (baseline model),
2
t
χ is the chi-square of the target model, and
df is the number of degrees of freedom.
The CFI ranges from zero to one with higher values indicating better fit. A rule of
thumb for this index is that .97 is indicative of good fit relative to the independence
model, while values greater than .95 may be interpreted as an acceptable fit. Again a
value of .97 seems to be more reasonable as an indication of a good model fit than the
often stated cutoff value of .95. Comparable to the NNFI, the CFI is one of the fit indi-
ces less affected by sample size (Bentler, 1990; Bollen, 1990; Hu & Bentler, 1995, 1998,
1999).
Goodness-of-Fit-Index (GFI) and Adjusted Goodness-of-Fit-Index (AGFI)
The Goodness-of-Fit-Index (GFI; Jöreskog & Sörbom, 1989; Tanaka & Huba, 1984)
measures the relative amount of the variances and covariances in the empirical covari-
ance matrix S that is predicted by the model-implied covariance matrix )
ˆ
(θΣ. Accord-
ing to Jöreskog and Sörbom (1993, p. 123), this implies testing how much better the
model fits as compared to "no model at all" (null model), i.e., when all parameters are
fixed to zero. The GFI seems to be inspired by analogy with the concept of a coefficient
of determination (Mulaik et al., 1989, p. 435) and is defined as
2
n
2
t
n
t11
χ
χ
−=−= F
F
GFI , (15)
where
2
n
χ is the chi-square of the null model (baseline model),
2
t
χ is the chi-square of the target model, and
F is the corresponding minimum fit function value.
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 43
The GFI typically ranges between zero and one with higher values indicating better
fit, but in some cases a negative GFI may occur. The usual rule of thumb for this index
is that .95 is indicative of good fit relative to the baseline model, while values greater
than .90 are usually interpreted as indicating an acceptable fit (Marsh & Grayson, 1995;
Schumacker & Lomax, 1996).
Jöreskog and Sörbom (1989) also developed the Adjusted Goodness-of-Fit Index
AGFI to adjust for a bias resulting from model complexity. The AGFI adjusts for the
model's degrees of freedom relative to the number of observed variables and therefore
rewards less complex models with fewer parameters. The AGFI is given by
n
2
n
t
2
t
t
n
/
/
χ
χ
1)1(1
df
df
GFI
df
df
AGFI −=−−= (16)
where
2
n
χ is the chi-square of the null model (baseline model),
2
t
χ is the chi-square of the target model,
dfn = s = p( p + 1)/2 is the number of degrees of freedom for the null model, and
dft = s − t is the number of degrees of freedom for the target model.
AGFI values typically range between zero and one with larger values indicating a
better fit, but it is also possible that a large N in combination with small dft can result
in a negative AGFI. If the number of degrees of freedom for the target model ap-
proaches the number of degrees of freedom for the null model, the AGFI approaches the
GFI. A rule of thumb for this index is that .90 is indicative of good fit relative to the
baseline model, while values greater than .85 may be considered as an acceptable fit.
Simulation studies suggest that GFI and AGFI are not independent of sample size
(Hu & Bentler, 1995, 1998, 1999). Furthermore, both indices decrease with increasing
model complexity, especially for smaller sample sizes (Anderson & Gerbing, 1984).
Descriptive Measures of Model Parsimony
Parsimony is considered to be important in assessing model fit (Hu & Bentler, 1995;
Mulaik et al., 1989) and serves as a criterion for choosing between alternative models.
44 MPR-Online 2003, Vol. 8, No. 2
Several fit indices, among others the Parsimony Goodness-of-Fit Index (PGFI ), the
Parsimony Normed Fit Index (PNFI ), the Akaike Information Criterion (AIC ), the
Consistent AIC (CAIC ), and the Expected Cross-Validation Index (ECVI ) adjust for
model parsimony when assessing the fit of structural equation models.
Parsimony Goodness-of-Fit Index (PGFI) and
Parsimony Normed Fit Index (PNFI)
The Parsimony Goodness-of-Fit Index (PGFI; Mulaik et al., 1989) and the Parsi-
mony Normed Fit Index (PNFI; James, Mulaik, & Brett, 1982) are modifications of
GFI and NFI:
GFI
df
df
PGFI
n
t
= (17)
where
dft is the number of degrees of freedom of the target model,
dfn is the number of degrees of freedom of the null model, and
GFI is the Goodness-of-Fit Index,
and
NFI
df
df
PNFI
i
t
= (18)
where
dft is the number of degrees of freedom of the target model,
dfi is the number of degrees of freedom of the independence model, and
NFI is the Normed Fit Index.
PGFI and PNFI both range between zero and one, with higher values indicating a
more parsimonious fit. Both indices may be used for choosing between alternative mod-
els.
The effect of multiplying GFI and NFI by the respective parsimony ratio of the de-
grees of freedom is to reduce the original indices to a value closer to zero. Comparing
AGFI and PGFI , both indices adjust the GFI downward for more complex and there-
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 45
fore less parsimonious models, but the PGFI exerts a stronger penalty on complex mod-
els with less degrees of freedom than AGFI (Tanaka, 1993). Similarly, the PNFI adjusts
NFI downward.
Akaike Information Criterion (AIC)
The Akaike Information Criterion (AIC; Akaike, 1974, 1987) adjusts χ2 for the num-
ber of estimated parameters and can be used to compare competing models that need
not be nested. Various versions of the AIC exist (Hayduk, 1996, pp. 198-200) which are
essentially equivalent as long as the version is not changed during the comparisons. Fur-
thermore, all calculations must be based on the same covariance matrix.
The LISREL program provides
tAIC 2
2+χ= , (19)
where t is the number of estimated parameters.
Other software programs, e.g., EQS (Bentler, 1995), adopt AIC as
dfAIC 2
2−χ= , (20)
where df is the number of degrees of freedom.
Originally, AIC had been introduced by Akaike as "an information criterion" (Akaike,
1985, p. 12) in the form
tLAIC 2log2 +−= , (21)
where logL is the maximized value of the log likelihood function for the respective
model.
All versions share the feature that within a set of models for the same data, the
model with the minimum AIC value is regarded as the best fitting model. In other
words, AIC is actually a "badness of fit" index (Kaplan, 2000, p. 116).
The derivation of AIC unifies the information-theoretical concept of entropy and the
method of maximum likelihood (Akaike, 1985) and is rather demanding. On the other
hand, the above formulas show that the concrete calculation of AIC is quite simple once
χ2 or the log-likelihood is known. However, because the criterion (in whatever version) is
not normed, it is not possible to interpret an isolated AIC value. Therefore, AIC should
46 MPR-Online 2003, Vol. 8, No. 2
first be obtained for all models of interest, including the values for the independence
model and the saturated model (a model with zero degrees of freedom) which automati-
cally appear in the LISREL output. Researchers should then consider to select the
model with the lowest AIC value. To avoid misunderstandings, it should be recognized
that AIC is only a descriptive measure and not a test of significance although it may be
used to decide between competing models.
The main implications of AIC can be understood as follows. Provided a set of com-
peting models for given data, the task is to select the model which serves best as an
approximation to "reality". In approximating reality, two kinds of errors can occur that
have already been addressed in the context of RMSEA, namely, systematic discrepan-
cies (the error of approximation) introduced by the population model, and random dis-
crepancies (the error of estimation) introduced by the use of sample data. In contrast to
RMSEA which implies a focus on the population model while disregarding the estima-
tion of its parameters, AIC takes the view that only the estimated model can be used
for the prediction of further observations.
From this predictive point of view, extremely imprecise parameter estimates would
jeopardize the practical applicability of an otherwise appropriate model. Considering the
overall error, it can be better to tolerate a model that is slightly oversimplified if this is
compensated by a reduction of sampling fluctuations. Therefore, AIC reflects the search
for a compromise between the approximation and estimation errors that minimizes the
overall error. For example, note that the LISREL version of AIC increases both with χ2
(which is connected to the approximation error) and the number of free parameters
(which is connected to the estimation error). Because AIC penalizes a high number of
estimated parameters (or, equivalently, rewards a high number of degrees of freedom), it
rewards parsimony.
To examine more closely how AIC works, it is instructive to consider the special case
that a single target model is to be contrasted with the saturated model. For the sake of
simplicity, we use the EQS version of AIC. According to Equation 20, AIC for the tar-
get model is given by AICt = χ2 − 2df, whereas for the saturated model, χ2 = 0, df = 0,
and thus AICs = 0. The target model should be selected if AICt < AICs , which is
equivalent to the condition that for the target model χ2/df < 2. This sheds new light on
the recommendation that χ2/df < 2 is indicative of a good fit (see above, subsection "χ2
Test Statistic"). Note that the result does not depend on the particular AIC version as
two models were compared.
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 47
Consistent AIC (CAIC)
After the introduction of AIC, similar information indices have been proposed for
various reasons. We restrict ourselves to a single example: The LISREL program offers
the so-called Consistent AIC (CAIC; Bozdogan, 1987) in the version
tNCAIC )log1(
2++χ= , (22)
where
logN is the natural logarithm of the sample size N, and
t is again the number of estimated parameters.
Given a set of competing models, CAIC is used in the same way as AIC. In the pre-
sent context, consistency means that the correct model is selected as the sample size
tends to infinity (N → ∞).
At first sight, the only practically important difference between AIC and CAIC is
that the factor 2 in the penalty term of Equation 19 is replaced by the factor (1 +
logN ), which implies that the weight of the number of estimates now depends on the
sample size and that parsimonious models are rewarded more generously. These obvious
features may however distract from a fundamental problem inherent in CAIC. The
promise that the "true" model will be selected as N → ∞ rests on the assumption that
the true model is contained in the set of competing models. As Burnham and Anderson
(1998) pointed out, such an assumption would be rather unrealistic in the biological and
social sciences. First, if "an investigator knew that a true model existed and that it was
in the set of candidate models, would not he know which one it was?" (p. 69). Second,
even if the true model could be identified for a certain finite number of cases, this model
would hardly remain the true model as N → ∞, because in the biological and social sci-
ences, increasing the number of cases usually increases the number of relevant variables,
too.
Bozdogan himself (1987, p. 357) conceded that the concept of a true model becomes
"suspect" in view of real data, and that the virtue of consistency (which is a large-
sample property) should not be exaggerated. Therefore we generally recommend to pre-
fer the original AIC, which is intended to identify the "best fitting" model (Takane &
Bozdogan, 1987). In any case, the "magic number 2" (Akaike, 1985, p. 1) in the penalty
term of AIC is not a factor that may be arbitrarily changed.
48 MPR-Online 2003, Vol. 8, No. 2
Expected Cross Validation Index (ECVI)
Another index to be mentioned in the neighborhood of AIC is the Expected Cross
Validation Index (ECVI ) of Browne and Cudeck (1989, 1993). The ECVI is actually a
population parameter which is estimated by the statistic
1
2
))
ˆ
(,( −
+= N
t
θFc ΣS, (23)
where
))
ˆ
(,( θFΣS is the minimum value of the fit function,
t is the number of estimated parameters, and
N is the sample size.
Whereas AIC was derived from statistical information theory, ECVI is a measure of
the discrepancy between the model-implied covariance matrix in the analyzed sample
and the covariance matrix that would be expected in another sample of the same size
(Jöreskog & Sörbom, 1993, p. 120). Thus, ECVI evaluates how well a model fitted to
the calibration sample would perform in comparable validation samples (Kaplan, 2000).
When choosing between several models, the smallest ECVI estimate indicates the model
with the best fit. In addition, a 90% confidence interval allows to assess the precision of
the estimate.
Although derived independently, the ECVI estimate leads to the same rank order of
competing models as AIC, provided that the ML fitting function is used (Browne &
Cudeck, 1993, p. 148). In fact, multiplication of ECVI by N − 1 then leads to the
LISREL version of AIC given above. Therefore it suffices to take either AIC or ECVI
into account when looking for the model that minimizes the overall error, and it is not
necessary to report both indices.
Conclusions and Recommendations
Sample Size and Estimation Method
When the structural equation model is correctly specified and the observed variables
are multivariate normal, it can be analytically derived that different estimation proce-
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 49
dures, e.g., ML, WLS, and GLS, will produce estimates that converge to the same op-
timum and have similar asymptotic properties (Browne, 1984). Under ideal conditions
the choice of the estimation method is therefore quite arbitrary. But under the more
realistic assumption of models that are more or less misspecified and data that are not
multivariate normally distributed, the different procedures may not converge to the
same optimum (Olsson et al., 2000).
If all variables are measured on an interval scale, if they are normally distributed,
and if the sample size is sufficiently large, ML should be applied. Because this method is
relatively robust to violations of the normality assumption, it may also be used for mod-
els with variables that are not normally distributed, given that the deviation from nor-
mality is not too extreme. Robust maximum likelihood estimation may be an alterna-
tive but it needs relatively large sample sizes of at least N ≥ 400.
If the sample size is large and data are non-normally distributed, WLS (ADF) is of-
ten recommended. This method may also be considered if some of the observed vari-
ables are ordinal and others continuous, or if models including dichotomous variables
are being analyzed. But several simulation studies suggest that WLS may not perform
too well.
Boomsma and Hoogland (2001, p. 148) found that for a sample size of N ≤ 200, "ADF
is a disaster with more than one-third of the solutions being improper". For nonnor-
mally distributed variables, recommendations are quite divergent. Minimum sample size
for WLS estimation should be at least 1,000 (Hoogland & Boomsma, 1998), and depend-
ing on the model and data analyzed, in some cases even more than 4,000 or 5,000 cases
are required (Boomsma & Hoogland, 2001; Hoogland, 1999; Hu, Bentler & Kano, 1992).
In general, results based on small to medium sample sizes should be interpreted with
caution (Bollen, 1989, p. 432).
In a simulation study, Olsson et al. (2000) generated data for eleven conditions of
kurtosis, three conditions of misspecification, and five different sample sizes. Three es-
timation methods, ML, GLS, and WLS were compared in terms of overall fit and the
discrepancy between estimated parameter values and the true parameter values. Their
results show that WLS under no condition was preferable to ML and GLS. In fact, only
for large sample sizes of at least N = 1,000 and mildly misspecified models WLS pro-
vided estimates and fit indices close to the ones obtained for ML and GLS. In general,
ML estimation with or without a correction for non-normality (by using either robust χ2
or bootstrapping) seems to perform better than WLS and should be preferred. WLS for
50 MPR-Online 2003, Vol. 8, No. 2
categorical outcomes as implemented in Mplus may not require the same large sample
sizes as WLS for continuous, non-normal data, but further simulation studies are
needed.
For choosing the adequate estimation method, sample size is an important criterion.
Sample size requirements do not only depend on the distribution of the variables (as
data become more nonnormal, larger sample sizes are needed), but also on the size of
the model, the number of indicator variables, the amount of missing data, the reliability
of the variables, and the strength of the relationships among the variables. Therefore
one should not simply trust the rules of thumb given in the literature, not even the one
stated by Bentler (1995, p. 6) which recommends at least five times the number of free
parameters in the model (Bentler & Chou, 1987; Bentler, 1995). In reality there does
not exist a rule of thumb that applies to all situations (Muthén & Muthén, 2002).
Most sample size recommendations refer to ML estimation. As simulation research
has shown, a reasonable sample size for a correctly specified model and multivariate
normally distributed data is about N = 150 (Muthén & Muthén, 2002) or N = 200 (cf.
Hoogland & Boomsma, 1998; Boomsma & Hoogland, 2001).
For a confirmatory factor model with non-normally distributed variables and missing
data (missing completely at random), larger samples of about N = 300 may be needed
(Muthén & Muthén, 2002). With strongly kurtotic data, the minimum sample size
should be ten times the number of free parameters (Hoogland & Boomsma, 1998).
The number of indicator variables should also be considered for chosing a sufficient
large sample size. Marsh, Hau, Balla, and Grayson (1998) as well as Marsh and Hau
(1999) support Boomsma’s (1985) recommendations by stating that for confirmatory
factor analyses with 6 to 12 indicator variables per factor a sample size of N = 50 is
sufficient, whereas for 3 to 4 indicators per factor a sample size of N = 100 is necessary.
With two indicators per factor one should at least have a sample size of N ≥ 400 (cf.
Marsh & Hau, 1999; Boomsma & Hoogland, 2001). There seems to be a mutual com-
pensatory effect of sample size and number of indicators per factor: More indicators may
compensate for small sample size, and a larger sample size may compensate for few in-
dicators.
Some evidence exists that simple models could be meaningfully tested even if sample
size is quite small (cf. Hoyle, 1999; Hoyle & Kenny, 1999; Marsh & Hau, 1999). Never-
theless, models with moderate to small sample sizes should only be analyzed if a greater
sample is not available and if convergence problems or improper solutions, such as nega-
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 51
tive variance estimates or Heywood cases (cf. Chen, Bollen, Paxton, Curran, & Kirby,
2001), do not occur. As a minimum requirement for parameter estimation using the ML
method, sample size should be larger than or at least equal to the number of observed
variables (N ≥ p) (MacCallum, Browne, & Sugawara, 1996, p. 144).
Evaluation of Model Fit
There is a consensus that one should avoid to report all fit indices that have been
developed since the first days of SEM, but there is a certain disagreement on just which
fit indices to consider for model evaluation.
As the χ2 test is not only sensitive to sample size but also sensitive to the violation of
the multivariate normality assumption (Curran, West, & Finch, 1996; Hu, Bentler, &
Kano, 1992; West, Finch, & Curran, 1995), it should not serve as the sole basis for
judging model fit. Bollen and Long (1993) as well as Mueller (1996) recommend to
evaluate several indices simultaneously which represent different classes of goodness-of-
fit criteria. The following criteria form an adequate selection of indices which are fre-
quently presented in current publications: χ2 and its associated p value, χ2/df, RMSEA
and its associated confidence interval, SRMR, NNFI, and CFI. The fit indices RMSEA,
NNFI and CFI are sensitive to model misspecifications and do not depend on sample
size as strongly as χ2 (Fan, Thompson, & Wang, 1999; Hu & Bentler, 1998; Rigdon,
1996), therefore they should always be considered. Hu and Bentler (1998) recommend to
use SRMR, supplemented by NNFI, CFI, or RMSEA derived from ML and GLS estima-
tion (NNFI and RMSEA are less preferable at small sample sizes), and SRMR, NNFI,
and CFI derived from WLS estimation.
For model comparisons (nested models) it is necessary to report additionally the
value of the χ2 difference test and the AIC values of all models investigated. In most
publications, GFI and AGFI are also reported. However, these indices depend on sample
size and tend to underestimate the fit of complex models (Steiger, 1989).
Other things being equal, a model with fewer indicators per factor may have a higher
fit than a model with more indicators per factor, because more indicators per factor
provide a more powerful, precise test than a comparable model with fewer indicators
(MacCallum et al., 1996). Fit coefficients which reward parsimony, e.g., RMSEA, AIC,
PGFI, and PNFI, are one way to adjust for this tendency. For choosing between alter-
native models, parsimony indices may provide important information and should be
reported.
52 MPR-Online 2003, Vol. 8, No. 2
As we have demonstrated by discussing different goodness-of-fit indices, it is quite
difficult to decide on data-model fit or misfit, especially if various measures of model fit
point to conflicting conclusions about the extent to which the model actually matches
the observed data. Although there are no well-established guidelines for what minimal
conditions constitute an adequate fit, some rules of thumb exist. Table 1 provides an
overview over some rule of thumb criteria for goodness-of-fit indices.
Table 1
Recommendations for Model Evaluation: Some Rules of Thumb
Fit Measure Good Fit Acceptable Fit
χ2 0 ≤ χ2 ≤ 2df 2df < χ2 ≤ 3df
p value .05 < p ≤ 1.00 .01 ≤ p ≤ .05
χ2/df 0 ≤ χ2/df ≤ 2 2 < χ2/df ≤ 3
RMSEA 0 ≤ RMSEA ≤ .05 .05 < RMSEA ≤ .08
p value for test of close fit
(RMSEA < .05) .10 < p ≤ 1.00 .05 ≤ p ≤ .10
Confidence interval (CI ) close to RMSEA,
left boundary of CI = .00 close to RMSEA
SRMR 0 ≤ SRMR ≤ .05 .05 < SRMR ≤ .10
NFI .95 ≤ NFI ≤ 1.00a .90 ≤ NFI < .95
NNFI .97 ≤ NNFI ≤ 1.00b .95 ≤ NNFI < .97c
CFI .97 ≤ CFI ≤ 1.00 .95 ≤ CFI < .97c
GFI .95 ≤ GFI ≤ 1.00 .90 ≤ GFI < .95
AGFI .90 ≤ AGFI ≤ 1.00,
close to GFI .85 ≤ AGFI <.90,
close to GFI
AIC smaller than AIC for comparison model
CAIC smaller than CAIC for comparison model
ECVI smaller than ECVI for comparison model
Note. AGFI = Adjusted Goodness-of-Fit-Index, AIC = Akaike Information Criterion, CAIC = Consistent
AIC, CFI = Comparative Fit Index, ECVI = Expected Cross Validation Index, GFI = Goodness-of-Fit
Index, NFI = Normed Fit Index, NNFI = Nonnormed Fit Index, RMSEA = Root Mean Square Error of
Approximation, SRMR = Standardized Root Mean Square Residual.
aNFI may not reach 1.0 even if the specified model is correct, especially in smaller samples (Bentler,
1990). bAs NNFI is not normed, values can sometimes be outside the 0-1 range. cNNFI and CFI values of
.97 seem to be more realistic than the often reported cutoff criterion of .95 for a good model fit.
It should be clear that these rule of thumb cutoff criteria are quite arbitrary and
should not be taken too seriously. Fit indices may be affected by model misspecification,
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 53
small-sample bias, effects of violation of normality and independence, and estimation-
method effects (Hu & Bentler, 1998). Therefore it is always possible that a model may
fit the data although one or more fit measures may suggest bad fit.
Examples
Data Generation
In the following, we will evaluate the fit of four alternative models as examples of
poor model fit, adequate fit, and good fit. Using the EQS program (Bentler, 1995), we
generated data for N = 200 cases according to the "true" model depicted in Figure 1. In
this population model, two latent predictor variables and two latent criterion variables
are each measured by two indicator variables, and each predictor variable influences
both criterion variables. We generated data for the eight X- and Y-variables of the
model and computed the sample covariance matrix, which is given in Table 2.
δ1
δ2
δ3
δ4
ε1
ε2
ε3
ε4
ξ1
ξ2
η1
η2
X1
X2
X3
X4
Y1
Y2
Y3
Y4
ζ1
ζ2
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
(.38)
(.30)
(.30)
(.38)
(.36)
(.30)
(.33)
(.33)
(.64)
(.64)
(.67)
(.82)
.32 β21 =.25
γ11 =.90
γ22 =−.82
γ21 =.53
γ12 =−.40
δ1
δ2
δ3
δ1
δ2
δ3
δ4
ε1
ε2
ε3
ε4
ε1
ε2
ε1
ε2
ε3
ε4
ε3
ε4
ξ1
ξ2
ξ1
ξ2
η1
η2
η1
η2
X1
X2
X3
X4
X1
X2
X3
X4
Y1
Y2
Y3
Y4
Y1
Y2
Y3
Y4
ζ1
ζ2
ζ1
ζ2
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
(.38)
(.30)
(.30)
(.38)
(.36)
(.30)
(.33)
(.33)
(.64)
(.64)
(.67)
(.82)
.32 β21 =.25
γ11 =.90
γ22 =−.82
γ21 =.53
γ12 =−.40
Figure 1. Population model with known parameter values used for data generation.
Specifications of Models A, B, C, and D
According to Hu & Bentler (1998), there are four major problems involved in using
fit indices for evaluating goodness of fit: sensitivity of a fit index to model misspecifica-
tion, small-sample bias, estimation-method effect, and effects of violation of normality
and independence. In our present examples, we will only demonstrate the problem of
model misspecification.
54 MPR-Online 2003, Vol. 8, No. 2
Table 2
Empirical Covariance Matrix (N = 200)
Y
1 Y
2 Y
3 Y
4 X
1 X
2 X
3 X
4
Y1 1.429
Y2 1.069 1.369
Y3 0.516 0.536 1.681
Y4 0.436 0.425 1.321 1.621
X1 0.384 0.485 0.192 0.183 1.021
X2 0.494 0.424 0.181 0.191 0.640 0.940
X3 0.021 0.045 -0.350 -0.352 0.325 0.319 1.032
X4 0.035 0.013 -0.347 -0.348 0.324 0.320 0.642 0.941
δ1
δ2
δ3
δ4
ε1
ε2
ε3
ε4
ξ1
ξ2
η1
η2
X1
X2
X3
X4
Y1
Y2
Y3
Y4
ζ1
ζ2
γ22
γ12
δ1
δ2
δ3
δ1
δ2
δ3
δ4
ε1
ε2
ε3
ε4
ε1
ε2
ε1
ε2
ε3
ε4
ε3
ε4
ξ1
ξ2
ξ1
ξ2
η1
η2
η1
η2
X1
X2
X3
X4
X1
X2
X3
X4
Y1
Y2
Y3
Y4
Y1
Y2
Y3
Y4
ζ1
ζ2
ζ1
ζ2
γ22
γ12
Figure 2. Path diagram for hypothesized models. Four models were analyzed, two mis-
specified models and two correctly specified models. In the misspecified models, path
coefficients γ12 and γ22 (Model A) or γ12 (Model B) were fixed to zero (indicated by
dashed lines). In the correctly specified models, all parameters were estimated freely
(Model C) or additionally, all factor loadings pertaining to a latent construct were con-
strained to be equal (Model D).
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 55
Four models were specified and analyzed, two misspecified models and two correctly
specified models (cf. Figure 2). The first misspecified model (Model A) is severely mis-
specified with path coefficients γ12 and γ22 fixed to zero as indicated by dashed lines in
Figure 2. In the second misspecified model (Model B) only γ12 was fixed to zero. In the
first correctly specified model (Model C), all parameters were estimated freely. With the
second correctly specified model (Model D), we will demonstrate the effects of parsi-
mony. In this model, all factor loadings pertaining to a latent construct were con-
strained to be equal, i.e., all factor loadings were fixed to one.
Model Evaluation Procedure
Generally, the first gauge of model-data fit should be the inspection of the fit indices.
If there is some indication of misfit, the second gauge should be the inspection of the
residual matrix. As the fitted residuals, i.e., discrepancies between the covariance matri-
ces S and )
ˆ
(θΣ, are difficult to interpret if the manifest variables have different vari-
ances (or scale units), a more general recommendation is to inspect standardized residu-
als. Standardized residuals provided by the LISREL program (see above, section on
RMR and SRMR) that are greater than an absolute value of 1.96 ( p < .05) or 2.58
( p < .01) can be useful for detecting the cause of model misfit. The largest absolute
value indicates the element that is most poorly fit by the model. A good model should
have a high number of standardized residuals close to zero, implying high correspon-
dence between elements of the empirical and the model-implied covariance matrix.
The third gauge should be the inspection of the modification indices provided by
LISREL (or the Lagrange multiplier test provided by EQS). These indices provide an
estimate of the change in the χ2 value that results from relaxing model restrictions by
freeing parameters that were fixed in the initial specification. Each modification index
possesses a χ2 distribution with df = 1 and measures the expected decrease in the χ2
value when the parameter in question is freed and the model reestimated. Thus, the
modification index is approximately equal to the χ2 difference of two nested models in
which the respective parameter is fixed or constrained in one model and set free in the
other. The largest modification index belongs to the parameter that improves the fit
most when set free. A good model should have modification indices close to one, because
E(χ2) = df.
If the fit indices suggest a bad model-data fit, if one or more standardized residuals
are more extreme than ± 1.96 ( p < .05) or ± 2.58 ( p < .01), and if at least one modifica-
tion index is larger than 3.84 ( p < .05) or 6.63 ( p < .01), then one may consider to
56 MPR-Online 2003, Vol. 8, No. 2
Table 3
Goodness-of-Fit Indices for Misspecified and Correctly Specified Models Based on an
Artificial Data Set of N = 200
Misspecified models Correctly specified models
Model A Model B Model C Model D
Fit Index Two paths
fixed to zero
One path fixed
to zero
All parameters
estimated freely
All factor load-
ings fixed to 1
χ2(df ) 54.340 (16) 27.480 (15) 17.109 (14) 17.715 (18)
p value .000 .025 .250 .475
χ2/df 3.396 1.832 1.222 0.984
RMSEA .110 .065 .033 .000
p value for test of close fit
(RMSEA < .05) .001 .238 .668 .871
90% CI .079 ; .142 .023 ; .102 .000 ; .080 .000 ; .062
SRMR .120 .058 .016 .018
NFI .922 .963 .977 .977
NNFI .896 .966 .991 1.000
CFI .941 .982 .995 1.000
GFI .936 .967 .979 .978
AGFI .856 .920 .946 .956
PGFI .416 .403 .381 .489
PNFI .527 .516 .489 .628
Model AIC 94.340 69.480 61.109 53.715
Saturated AIC
a 72.000 72.000 72.000 72.000
Model CAIC 180.306 159.744 155.672 131.085
Saturated CAIC a 226.739 226.739 226.739 226.739
Model ECVI .474 .349 .307 .271
90% CI .380 ; .606 .294 ; .443 .291 ; .381 .271 ; .341
Saturated ECVI a .362 .362 .362 .362
Note. AGFI = Adjusted Goodness-of-Fit-Index, AIC = Akaike Information Criterion, CAIC = Consistent
AIC, CFI = Comparative Fit Index, ECVI = Expected Cross Validation Index, GFI = Goodness-of-Fit
Index, NFI = Normed Fit Index, NNFI = Nonnormed Fit Index, PGFI = Parsimony Goodness-of-Fit
Index, PNFI = Parsimony Normed Fit Index, RMSEA = Root Mean Square Error of Approximation,
SRMR = Standardized Root Mean Square Residual.
aSaturated AIC, CAIC, and ECVI serve as possible comparative values for the model AIC, CAIC, and
ECVI, respectively.
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 57
modify the model by freeing a fixed parameter. But this practice is controversial, as it
implies changing the model only in order to improve fit. It would be better to have a
substantive theory on which to base modifications, as modifications devoid of a theo-
retical basis are ill advised (Field, 2000).
In our examples, we will assume to have a theoretical basis for model modifications
according to the model depicted in Figure 1. We will modify our initial model (Model
A) three times and analyze the covariance matrix given in Table 2 repeatedly. The
goodness-of-fit indices taken from the LISREL outputs pertaining to the four models are
listed in Table 3.
Although Table 3 lists all fit measures discussed in this article, not all of these meas-
ures are necessary for evaluating the fit of the models. It is quite sufficient to assess χ2
and its associated p-value, χ2/df, RMSEA and its associated confidence interval, SRMR,
NNFI, and CFI. For model comparisons it is recommended to assess additionally the χ2
difference tests (nested models only) and the AIC values of all models investigated.
Model A (Severely Misspecified Model)
An inspection of the fit indices for Model A (Table 3) suggests to reject the model, as
χ2/df > 3, RMSEA > .08, p-value for test of close fit (RMSEA < .05) is almost zero,
lower boundary of the confidence interval does not include zero, SRMR > .10, NNFI <
.95, and CFI < .95. Only NFI > .90 and GFI > .90 are indicative of an acceptable
model fit.
An inspection of the standardized residuals in Table 4 reveals 11 significant residuals.
Because of the misspecifications in Model A, relations of Y1 and Y2 with X3 and X4 are
overestimated by the model (residuals are negative while the respective sample covari-
ances are positive, cf. Table 2), whereas relations of Y3 and Y4 with X3 and X4 are un-
derestimated (residuals are negative while the respective empirical covariances are nega-
tive).
An inspection of the modification indices (Appendix B1) reveals that the largest
modification index of 24.873 pertains to parameter γ22. This suggests that the most sub-
stantive change in the model in terms of improvement of fit would arise from relaxing
the constraint on the respective structural coefficient, which would result in an unstan-
dardized γ22 estimate of about −.715 according to the LISREL output. We followed this
suggestion and set γ22 free.
58 MPR-Online 2003, Vol. 8, No. 2
Table 4
Standardized Residuals of Model A (Severely Misspecified Model)
Y1 Y
2 Y
3 Y
4 X
1 X
2 X
3 X
4
Y1 --
Y2 -- --
Y3 -0.256 0.237 --
Y4 0.379 -0.111 -- --
X1 -1.056 2.079 1.216 1.168 --
X2 2.408 -0.587 1.268 1.425 -2.727 --
X3 -2.620 -2.390 -4.953 -4.752 0.871 0.481 --
X4 -2.540 -3.026 -5.194 -4.955 0.944 0.642 -- --
Note. Significant values are in boldface.
An inspection of the standardized residuals in Table 4 reveals 11 significant residuals.
Because of the misspecifications in Model A, relations of Y1 and Y2 with X3 and X4 are
overestimated by the model (residuals are negative while the respective sample covari-
ances are positive, cf. Table 2), whereas relations of Y3 and Y4 with X3 and X4 are un-
derestimated (residuals are negative while the respective empirical covariances are nega-
tive).
An inspection of the modification indices (Appendix B1) reveals that the largest
modification index of 24.873 pertains to parameter γ22. This suggests that the most sub-
stantive change in the model in terms of improvement of fit would arise from relaxing
the constraint on the respective structural coefficient, which would result in an unstan-
dardized γ22 estimate of about −.715 according to the LISREL output. We followed this
suggestion and set γ22 free.
Model B (Misspecified Model)
The fit indices for Model B point to conflicting conclusions about the extent to which
this model actually matches the observed data. All fit indices − with the exception of
the parsimony indices PNFI and PGFI − are improved compared to Model A, and the
model modification resulted in a significant χ2 difference test (54.34 − 27.48 = 26.86,
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 59
df = 1, p < .01). Furthermore, model AIC, model CAIC, and model ECVI are smaller
for Model B than for Model A. Indications of a good model fit are χ2/df < 2, NFI > .95,
GFI > .95, and CFI > .97, whereas other descriptive goodness-of-fit indices suggest only
an acceptable fit (RMSEA > .05, SRMR > .05, NNFI < .97). Moreover, the lower
boundary of the RMSEA confidence interval still does not include zero.
An inspection of the standardized residuals for Model B (Table 5) reveals that rela-
tions of Y1 and Y2 with X3 and X4 are still overestimated by the model (note that the
empirical covariances are positive), but that the largest standardized residuals show an
underestimation of the relation between Y3 and Y4 and an overestimation of the relation
between X1 and X2.
An inspection of the modification indices (Appendix B2) shows that two modification
indices are equally large (10.533) suggesting to either free parameter γ12 (with an ex-
pected unstandardized estimate of about –.389), or to free β12 (with an expected un-
standardized estimate of about .488). As both modification indices are of the same size,
the decision which constraint to relax can only be made on a theoretical basis. In our
example, we relaxed the constraint on parameter γ21.
Table 5
Standardized Residuals of Model B (Misspecified Model)
Y
1 Y
2 Y
3 Y
4 X
1 X
2 X
3 X
4
Y1 --
Y2 -- --
Y3 2.329 2.923 3.245
Y4 1.181 1.005 3.245 3.245
X1 -1.204 1.997 -0.069 0.033 --
X2 2.465 -0.478 -0.533 0.139 -3.246 --
X3 -2.569 -2.315 -0.920 -1.344 0.708 0.423 --
X4 -2.480 -2.927 -1.075 -1.493 0.859 0.652 -- --
Note. Significant values are in boldface.
60 MPR-Online 2003, Vol. 8, No. 2
Model C (Correctly Specified Model)
After the second model modification with parameter γ21 now set free, the fit of
Model C (correctly specified model) improved substantially with a significant drop in
the χ2 value (27.480 − 17.109 = 10.371, df = 1, p < .01) and smaller AIC, CAIC, and
ECVI values for Model C compared to Model B. The descriptive goodness-of-fit indices
point to a good model fit with χ2/df < 2, RMSEA < .05 (lower boundary of the
RMSEA confidence interval contains zero), NNFI and CFI > .97, NFI and GFI > .95,
AGFI > .90, and SRMR < .05. The only exceptions are again the parsimony fit indices
PNFI and PGFI with smaller values for Model C than for Model B. The standardized
residuals (Table 6) for Model C are now all close to zero and not significant.
Table 6
Standardized Residuals of Model C (Correctly Specified Model)
Y1 Y2 Y3 Y4 X1 X2 X3 X4
Y1 --
Y2 -- --
Y3 0.432 0.982 --
Y4 -0.667 -1.216 -- --
X1 -1.823 1.161 0.018 0.107 --
X2 1.736 -1.139 -0.297 0.291 -- --
X3 -0.134 0.381 0.224 -0.424 0.110 -0.117 --
X4 0.175 -0.400 0.356 -0.403 0.107 -0.081 -- --
An inspection of the modification indices (Appendix B3) reveals that the parameters
of four covariances between error variables should be freed to improve the model fit.
But as our theory (Figure 1) does not support this modification, such suggestions can be
ignored.
Model D (Correctly Specified Parsimonious Model)
In order to demonstrate the effects of parsimony on model fit, we fixed all factor
loadings to one. The resulting Model D (correctly specified parsimonious model) is still
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 61
in accordance with the population model depicted in Figure 1. The model fit again im-
proved with χ2/df < 1, RMSEA = .00, NNFI = 1.00, CFI = 1.00, and AGFI > .95.
Model D is more parsimonious than Model C with smaller values of AIC, CAIC, ECVI,
and RMSEA, and with larger parsimony indices PNFI, PGFI, and larger AGFI. Stan-
dardized residuals are all small (Table 7), most of them close to zero. But still modifica-
tion indices (Appendix B4) suggest to lower the constraints on four error covariances, a
suggestion that we will not follow because our theory (Figure 1) does not support such
modification.
Table 7
Standardized Residuals of Model D (Correctly Specified Parsimonious Model)
Y
1 Y
2 Y
3 Y
4 X
1 X
2 X
3 X
4
Y1 -0.108
Y2 -0.108 0.108
Y3 0.729 1.251 0.764
Y4 -0.815 -1.157 0.763 -0.764
X1 -1.326 0.811 0.107 -0.074 -0.075
X2 1.087 -0.648 -0.135 0.104 -0.075 0.075
X3 -0.139 0.379 -0.019 -0.060 0.074 -0.068 0.014
X4 0.172 -0.391 0.048 0.025 0.059 -0.050 0.014 -0.014
Discussion
Our examples demonstrate how to evaluate model fit and model misfit and how to
modify a model. If goodness-of-fit indices suggest that a model does not fit, an inspec-
tion of significant standardized residuals and the largest modification indices may be
extremely helpful in deciding at which part of the model a modification should be con-
sidered. But one should never modify a model solely on the basis of modification indices,
although the program might suggest to do so. Model modification based on modification
indices may result in models that lack validity (MacCallum, 1986) and are highly sus-
62 MPR-Online 2003, Vol. 8, No. 2
ceptible to capitalization on chance (MacCallum, Roznowski, & Necowitz, 1992). There-
fore the modifications should be defensible from a theoretical point of view.
To sum up, model fit improved substantially from Model A to Model D as expected
from the construction of our simulated example. One may ask why there is no "perfect
fit" of Models C and D although both models are correctly specified. The first reason is
that even Model D is not yet the most parsimonious model. The most parsimonious
model requires that all parameters are fixed at their true population values, but these
values are of course unknown in practice. Usually the best one can do is to fix some pa-
rameters or to impose certain equality constraints, which also leads to more parsimoni-
ous models. The second reason is that only sample covariance matrices are analyzed
instead of the population covariance matrix. Whereas the population covariance matrix
should fit a correctly specified model "perfectly", sample covariance matrices will usually
do not, as they are subject to sampling fluctuations. Of course, the χ2 test takes these
fluctuations into account and most fit indices should signalize "good fit", as we could
demonstrate for Models C and D.
As has been pointed out by Jöreskog (1993), it is important to distinguish between
three types of analyses with SEM: strictly confirmatory approach (CM), alternative
models approach (AM), and model generating approach (MG). If the analysis is not
strictly confirmatory, that is, if the first model for a given data set must be modified
several times until a good or acceptable fit is reached – either because of testing alterna-
tive models (AM) or because of developing a new model (MG) – it is possible or even
likely that the resulting model to a considerable degree only reflects capitalization on
chance. In other words, subsequent modifications based on the same data always imply
the danger that a model is only fitted to some peculiarities of the given sample. If our
example would rest on real (instead of simulated) empirical data, we would strongly
suggest to cross-validate Model D.
It should be kept in mind that even if cross-validation is successful, it is not allowed
to infer that the respective model has been "proved". As is known from the work of
Stelzl (1986), Lee and Hershberger (1990), and MacCallum, Wegener, Uchino, and
Fabrigar (1993), several different models may fit the data equally well. Different models
can imply the same covariance matrix and thus be empirically indistinguishable from
the SEM viewpoint. This even holds if the competing models are contradictory from a
causal perspective. In order to decide if the model of interest is the "best'' model one
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 63
must be able to exclude all equivalent models on logical or substantive grounds (Jöre-
skog, 1993).
Let us therefore close with the friendly reminder that it is not possible to confirm any
tested model but that it is only possible to reject improper models, as we have done
with Models A and B. At the same time, significance tests and descriptive goodness-of-
fit indices are a valuable tool in finding a structural equation model that can be ac-
cepted in a preliminary sense. If a model is accepted, it may be consistent with the
data, but still it may not be consistent with reality. "If a model is consistent with real-
ity, then the data should be consistent with the model. But, if the data are consistent
with a model, this does not imply that the model corresponds to reality" (Bollen, 1989,
p. 68). Therefore researchers should pay attention "to the important but difficult ques-
tion regarding the discrepancy between the simplified, formalized theoretical model and
the actual operating processes that govern the phenomena under study in the real em-
pirical world" (Boomsma, 2000, p. 477).
References
Akaike, H. (1974). A new look at statistical model identification. IEEE transactions on
Automatic Control, 19, 716-723.
Akaike, H. (1985). Prediction and entropy. In A. C. Atkinson & S. E. Fienberg (Eds.),
A celebration of statistics (pp. 1-24). New York: Springer.
Akaike, H. (1987). Factor analysis and AIC. Psychometrika, 52, 317-332.
Anderson, J. C., & Gerbing, D. W. (1984). The effect of sampling error on convergence,
improper solutions, and goodness-of-fit indices for maximum likelihood confirmatory
factor analysis. Psychometrika, 49, 155-173.
Arbuckle, J. L., & Wothke, W. (1999). Amos 4.0 users' guide. Chicago: SmallWaters.
Bearden, W. O., Sharma, S., & Teel, J. E. (1982). Sample size effects on chi-square and
other statistics used in evaluating causal models. Journal of Marketing Research, 19,
425−430.
Bentler, P. M. (1990). Comparative fit indexes in structural models. Psychological Bul-
letin, 107, 238−246.
Bentler, P. M. (1995). EQS structural equations program manual. Encino, CA: Multiva-
riate Software.
Bentler, P. M., & Bonett, D. G. (1980). Significance tests and goodness of fit in the
analysis of covariance structures. Psychological Bulletin, 88, 588−606.
64 MPR-Online 2003, Vol. 8, No. 2
Bentler, P. M., & Chou, C. P. (1987). Practical issues in structural modeling. Sociologi-
cal Methods & Research, 16, 78−117.
Bentler, P. M., & Dudgeon, P. (1996). Covariance structure analysis: Statistical prac-
tice, theory, and directions. Annual Review of Psychology, 47, 563–592.
Bollen, K. A. (1989). Structural equations with latent variables. New York: Wiley.
Bollen, K. A. (1990). Overall fit in covariance structure models: Two types of sample
size effects. Psychological Bulletin, 107, 256−259.
Bollen, K., & Long, J. S. (Eds.). (1993). Testing structural equation models. Newbury
Park, CA: Sage.
Boomsma, A. (1985). Nonconvergence, improper solutions, and starting values in
LISREL maximum likelihood estimation. Psychometrika, 52, 345–370.
Boomsma, A. (2000). Reporting analyses of covariance structures. Structural Equation
Modeling, 7, 461−483.
Boomsma, A., & Hoogland, J. J. (2001). The robustness of LISREL modeling revisited.
In R. Cudeck, S. du Toit, & D. Sörbom (Eds.), Structural equation models: Present
and future. A Festschrift in honor of Karl Jöreskog (pp. 139–168). Chicago: Scien-
tific Software International.
Bozdogan, H. (1987). Model selection and Akaike's information criterion (AIC): The
general theory and its analytical extensions. Psychometrika, 52, 345−370.
Browne, M. W. (1984). Asymptotic distribution free methods in the analysis of covari-
ance structures. British Journal of Mathematical and Statistical Psychology, 37,
62−83.
Browne, M. W., & Cudeck, R. (1989). Single sample cross-validation indices for covari-
ance structures. British Journal of Mathematical and Statistical Psychology, 37,
62−83.
Browne, M. W., & Cudeck, R. (1993). Alternative ways of assessing model fit. In K. A.
Bollen & J. S. Long (Eds.), Testing structural equation models (pp. 136−162). New-
bury Park, CA: Sage.
Browne, M. W., & Mels, G. (1992). RAMONA user's guide. The Ohio State University:
Department of Psychology.
Burnham, K. P., & Anderson, D. R. (1998). Model selection and inference: A practical
information-theoretic approach. New York: Springer.
Chen, F., Bollen, K., Paxton, P., Curran, P. J., & Kirby, J. (2001). Improper solutions
in structural equation models: Causes, consequences, and strategies. Sociological
Methods and Research, 29, 468−508.
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 65
Chou, C.-P., & Bentler, P. M. (1995). Estimation and tests in structural equation mod-
eling. In R. H. Hoyle (Ed.), Structural equation modeling: Concepts, issues, and ap-
plications (pp. 37–55). Thousand Oaks, CA: Sage.
Chou, C.-P., Bentler, P. M., & Satorra, A. (1991). Scaled test statistics and robust
standard errors for non-normal data in covariance structure analysis: A Monte Carlo
study. British Journal of Mathematical and Statistical Psychology, 44, 347−357.
Curran, P. J., West, S. G, & Finch, J. F. (1996). The robustness of test statistics to
nonnormality and specification error in confirmatory factor analysis. Psychological
Methods, 1, 16−29.
Efron, B., & Tibshirani, R. (1993). An introduction to the bootstrap. New York: Chap-
man & Hall/CRC.
Fan, X., Thompson, B., & Wang, L. (1999). Effects of sample size, estimation method,
and model specification on structural equation modeling fit indexes. Structural
Equation Modeling, 6, 56−83.
Field, A. P. (2000). Discovering statistics using SPSS for Windows: Advanced techniques
for the beginner. London: Sage.
Hayduk, L. A. (1989). Structural equation modeling with LISREL. Baltimore: Johns
Hopkins University Press.
Hayduk, L. A. (1996). LISREL issues, debates, and strategies. Baltimore: Johns Hopkins
University Press.
Hoogland, J. J. (1999). The robustness of estimation methods for covariance structure
analysis. Unpublished doctoral dissertation, University of Groningen, The Nether-
lands.
Hoogland, J. J., & Boomsma, A. (1998). Robustness studies in covariance structure
modeling: An overview and a meta-analysis. Sociological Methods & Research, 26,
329–367.
Hoyle, R. H. (1999). Statistical strategies for small sample research. Thousand Oaks,
CA: Sage.
Hoyle, R. H., & Kenny, D. A. (1999). Sample size, reliability, and tests of statistical
mediation. In R. H. Hoyle (Ed.), Statistical strategies for small sample research (pp.
195-222). Thousand Oaks, CA: Sage.
Hoyle, R. H., & Panter, A. T. (1995). Writing about structural equation models in
structural equation modeling. In R. H. Hoyle (Ed.), Structural equation modeling:
Concepts, issues, and applications (pp. 158−176). Thousand Oaks, CA: Sage.
66 MPR-Online 2003, Vol. 8, No. 2
Hu, L., & Bentler, P. (1995). Evaluating model fit. In R. H. Hoyle (Ed.), Structural
equation modeling. Concepts, issues, and applications (pp. 76−99). London: Sage.
Hu, L., & Bentler, P. M. (1998). Fit indices in covariance structure analysis: Sensitivity
to underparameterized model misspecification. Psychological Methods, 3, 424−453.
Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure
analysis: Conventional criteria versus new alternatives. Structural Equation Model-
ing, 6, 1−55.
Hu, L., Bentler, P. M., & Kano, Y. (1992). Can test statistics in covariance structure
analysis be trusted? Psychological Bulletin, 112, 351−362.
James, L. R., Mulaik, S. A., & Brett, J. M. (1982). Causal analysis: Assumptions, mod-
els, and data. Beverly Hills, CA: Sage.
Jöreskog, K. G. (1993). Testing structural equation models. In K. A. Bollen & J. S.
Long (Eds.), Testing structural equation models (pp. 294−316). Newbury Park, CA:
Sage.
Jöreskog, K. G., & Sörbom, D. (1981). LISREL V: Analysis of linear structural relation-
ships by maximum likelihood and least squares methods (Research Report 81−8).
Uppsala, Sweden: University of Uppsala, Department of Statistics.
Jöreskog, K. G., & Sörbom, D. (1989). LISREL 7 user's reference guide. Chicago: SPSS
Publications.
Jöreskog, K. G., & Sörbom, D. (1993). Structural equation modeling with the SIMPLIS
command language. Chicago: Scientific Software.
Jöreskog, K. G., & Sörbom, D. (1996). LISREL 8 user's reference guide. Chicago: Scien-
tific Software.
Kaplan, D. (2000). Structural equation modeling: Foundation and extensions. Thousand
Oaks, CA: Sage Publications.
Kumar, A., & Sharma, S. (1999). A metric measure for direct comparison of competing
models in covariance structure analysis. Structural Equation Modeling, 6, 169−197.
Lee, S., & Hershberger, S. (1990). A simple rule for generating equivalent models in co-
variance structure modeling. Multivariate Behavioral Research, 25, 313−334.
MacCallum R. C. (1986). Specification searches in covariance structure modelling. Psy-
chologival Bulletin, 100, 107–20.
MacCallum, R. C., Browne, M. W., and Sugawara, H. M. (1996). Power analysis and
determination of sample size for covariance structure modeling. Psychological Meth-
ods 1, 130−149.
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 67
MacCallum, R. C., Roznowski, M., & Necowitz, L. B. (1992). Model modification in
covariance structure analysis: The problem of capitalization on chance. Psychologi-
cal Bulletin, 111, 490−504.
MacCallum, R. C., Wegener, D. T., Uchino, B. N., & Fabrigar, L. R. (1993). The prob-
lem of equivalent models in applications of covariance structure analysis. Psycho-
logical Bulletin, 114, 185−199.
Marsh, H. W., & Grayson, D. (1995). Latent variable models of multitrait-multimethod
data. In R. Hoyle (Ed.), Structural equation modeling: Concepts, issues and applica-
tions (pp. 177−198). Thousand Oaks, CA: Sage.
Marsh, H. W., & Hau, K.-T. (1999). Confirmatory factor analysis: Strategies for small
sample sizes. In R. H. Hoyle (Ed.), Statistical strategies for small sample size (pp.
251–306). Thousand Oaks, CA: Sage.
Marsh, H. W., Hau, K-T., Balla, J. R., & Grayson, D. (1998). Is more ever too much?
The number of indicators per factor in confirmatory factor analysis. Multivariate
Behavioral Research, 33, 181−220.
McDonald, R. P., & Marsh, H. W. (1990). Choosing a multivariate model: Noncentral-
ity and goodness-of-fit. Psychological Bulletin, 103, 391−411.
Mueller, R. O. (1996). Basic principles of structural equation modeling: An introduction
to LISREL and EQS. New York: Springer.
Mulaik, S. A., James, L. R., Van Alstine, J., Bennett, N., Lind, S., & Stilwell, C. D.
(1989). Evaluation of goodness-of-fit indices for structural equation models. Psycho-
logical Bulletin, 105, 430−445.
Muthén, B. O., & Kaplan, D. (1985). A comparison of some methodologies for the fac-
tor analysis of non-normal Likert variables. British Journal of Mathematical and
Statistical Psychology, 38, 171–189.
Muthén, B. O., & Kaplan, D. (1992). A comparison of some methodologies for the fac-
tor analysis of non-normal Likert variables: A note on the size of the model. British
Journal of Mathematical and Statistical Psychology, 45, 19–30.
Muthén, L. K., & Muthén, B. O. (1998). Mplus: The comprehensive modeling program
for applied researchers. Los Angeles, CA: Muthén & Muthén.
Muthén, L. K., & Muthén, B. O. (2002). How to use a Monte Carlo study to decide on
sample size and determine power. Structural Equation Modeling, 4, 599−620.
Olsson, U. H., Foss, T., Troye, S. V., & Howell, R. D. (2000). The performance of ML,
GLS, and WLS estimation in structural equation modeling under conditions of mis-
specification and nonnormality. Structural Equation Modeling, 7, 557–595.
68 MPR-Online 2003, Vol. 8, No. 2
Raykov, T., & Penev, S. (1998). Nested structural equation models: Noncentrality and
power of restriction test. Structural Equation Modeling, 5, 229−246.
Rigdon, E. E. (1996). CFI versus RMSEA: A comparison of two fit indexes for struc-
tural equation modeling. Structural Equation Modeling, 3, 369−369.
Rigdon, E. E. (1999). Using the Friedman method of ranks for model comparison in
structural equation modeling. Structural Equation Modeling, 6, 219−232.
Satorra, A. (2000). Scaled and adjusted restricted tests in multi-sample analysis of mo-
ment structures. In R. D. H. Heijmans, D. S. G. Pollock, & A. Satorra, (Eds.), In-
novations in multivariate statistical analysis. A Festschrift for Heinz Neudecker (pp.
233−247). London: Kluwer Academic Publishers.
Satorra, A., & Bentler, P. M. (1994). Corrections to test statistics and standard errors
in covariance structure analysis. In A. von Eye & C. C. Clogg (Eds.), Latent vari-
able analysis: Applications for developmental research (pp. 399–419). Thousand
Oaks, CA: Sage.
Satorra, A., & Bentler, P. M. (2001). A scaled difference chi-square test statistic for
moment structure analysis. Psychometrika, 66, 507−514.
Schumacker, R. E., & Lomax, R. G. (1996). A beginner's guide to structural equation
modeling. Mahwah, NJ: Lawrence Erlbaum Associates.
Shipley, B., 2000. Cause and correlation in biology. University Press, Cambridge.
Steiger, J. H. (1989). EzPATH: A supplementary module for SYSTAT and SYGRAPH.
Evanston, IL: SYSTAT.
Steiger, J. H. (1990). Structural model evaluation and modification: An interval estima-
tion approach. Multivariate Behavioral Research, 25, 173−180.
Steiger J. (1995). SEPATH structural equation modeling, Statistica, 3, 3539−3689.
Steiger, J. H., Shapiro, A., & Browne, M. W. (1985). On the multivariate asymptotic
distribution of sequential chi-square statistics. Psychometrika, 50, 253−264.
Stelzl, I. (1986). Changing causal relationships without changing the fit: Some rules for
generating equivalent LISREL models. Multivariate Behavioral Research, 21,
309−331.
Takane, Y., & Bozdogan, H. (1987). Introduction to special section [on AIC]. Psycho-
metrika, 52, 315.
Tanaka, J. S. (1993). Multifaceted conceptions of fit in structural equation models. In
K. A. Bollen & J. S. Long (Eds.), Testing structural equation models (pp. 10−40).
Newbury Park, CA: Sage.
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 69
Tanaka, J. S., & Huba, G. J. (1984). Confirmatory hierarchical factor analyses of psy-
chological distress measures. Journal of Personality and Social Psychology, 46,
621−635.
Tucker, L. R., & Lewis, C. (1973). A reliability coefficient for maximum likelihood fac-
tor analysis. Psychometrika, 38, 1−10.
West, S. G, Finch, J. F., & Curran, P. J. (1995). Structural equation models with non-
normal variables: Problems and remedies. In R. H. Hoyle (Ed.), Structural equation
modeling: Concepts, issues, and applications (pp. 56−75). Thousand Oaks, CA: Sage.
Yang-Wallentin, F., & Jöreskog, K. G. (2001). Robust standard errors and chi-squares
for interaction models. In G. A. Marcoulides and R. E. Schumacker (Eds.), New de-
velopments and techniques in structural equation modeling (pp. 159-171). Mahwah,
NJ: Lawrence Erlbaum.
70 MPR-Online 2003, Vol. 8, No. 2
APPENDIX A4
A1 LISREL Input File for Model A (Severely Misspecified Model)
!Model A: GA(1,2) and GA(2,2) fixed to zero
DA NI=8 NO=200 MA=CM
CM SY
1.429
1.069 1.369
0.516 0.536 1.681
0.436 0.425 1.321 1.621
0.384 0.485 0.192 0.183 1.021
0.494 0.424 0.181 0.191 0.640 0.940
0.021 0.045 -0.350 -0.352 0.325 0.319 1.032
0.035 0.013 -0.347 -0.348 0.324 0.320 0.642 0.941
LA
Y1 Y2 Y3 Y4 X1 X2 X3 X4
MO NX=4 NK=2 NY=4 NE=2 LX=FU,FI LY=FU,FI TD=DI,FR TE=DI,FR GA=FU,FR C
BE=FU,FI PH=FU,FR
LK
KSI1 KSI2
LE
ETA1 ETA2
VA 1 LX(1,1) LX(3,2)
VA 1 LY(1,1) LY(3,2)
FR LX(2,1) LX(4,2)
FR LY(2,1) LY(4,2)
FI GA(1,2) GA(2,2)
FR BE(2,1)
PATH DIAGRAM
OU ND=3 ME=ML RS SC MI
A2 LISREL Input File for Model B (Misspecified Model)
!Model B: GA(1,2) fixed to zero
DA NI=8 NO=200 MA=CM
CM SY
1.429
1.069 1.369
0.516 0.536 1.681
0.436 0.425 1.321 1.621
0.384 0.485 0.192 0.183 1.021
0.494 0.424 0.181 0.191 0.640 0.940
0.021 0.045 -0.350 -0.352 0.325 0.319 1.032
0.035 0.013 -0.347 -0.348 0.324 0.320 0.642 0.941
LA
Y1 Y2 Y3 Y4 X1 X2 X3 X4
MO NX=4 NK=2 NY=4 NE=2 LX=FU,FI LY=FU,FI TD=DI,FR TE=DI,FR GA=FU,FR BE=FU,FI
PH=FU,FR
LK
KSI1 KSI2
LE
ETA1 ETA2
VA 1 LX(1,1) LX(3,2)
VA 1 LY(1,1) LY(3,2)
FR LX(2,1) LX(4,2)
4 Appendix A is stored as an ASCII file next to this article at http://www.mpr-online.de
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 71
FR LY(2,1) LY(4,2)
FI GA(1,2)
FR BE(2,1)
PATH DIAGRAM
OU ND=3 ME=ML RS SC MI
A3 LISREL Input File for Model C (Correctly Specified Model)
!Model C: correctly specified model
DA NI=8 NO=200 MA=CM
CM SY
1.429
1.069 1.369
0.516 0.536 1.681
0.436 0.425 1.321 1.621
0.384 0.485 0.192 0.183 1.021
0.494 0.424 0.181 0.191 0.640 0.940
0.021 0.045 -0.350 -0.352 0.325 0.319 1.032
0.035 0.013 -0.347 -0.348 0.324 0.320 0.642 0.941
LA
Y1 Y2 Y3 Y4 X1 X2 X3 X4
MO NX=4 NK=2 NY=4 NE=2 LX=FU,FI LY=FU,FI TD=DI,FR TE=DI,FR GA=FU,FR C
BE=FU,FI PH=FU,FR
LK
KSI1 KSI2
LE
ETA1 ETA2
VA 1 LX(1,1) LX(3,2)
VA 1 LY(1,1) LY(3,2)
FR LX(2,1) LX(4,2)
FR LY(2,1) LY(4,2)
FR BE(2,1)
PATH DIAGRAM
OU ND=3 ME=ML RS SC MI
A4 LISREL Input File for Model D
(Correctly Specified Parsimonious Model)
!Model D: equality constraints on LX and LY
DA NI=8 NO=200 MA=CM
CM SY
1.429
1.069 1.369
0.516 0.536 1.681
0.436 0.425 1.321 1.621
0.384 0.485 0.192 0.183 1.021
0.494 0.424 0.181 0.191 0.640 0.940
0.021 0.045 -0.350 -0.352 0.325 0.319 1.032
0.035 0.013 -0.347 -0.348 0.324 0.320 0.642 0.941
LA
Y1 Y2 Y3 Y4 X1 X2 X3 X4
MO NX=4 NK=2 NY=4 NE=2 LX=FU,FI LY=FU,FI TD=DI,FR TE=DI,FR GA=FU,FR C
BE=FU,FI PH=FU,FR
LK
KSI1 KSI2
LE
ETA1 ETA2
VA 1 LX(1,1) LX(3,2)
VA 1 LY(1,1) LY(3,2)
VA 1 LX(2,1) LX(4,2)
VA 1 LY(2,1) LY(4,2)
FR BE(2,1)
PATH DIAGRAM
OU ND=3 ME=ML RS SC MI
72 MPR-Online 2003, Vol. 8, No. 2
APPENDIX B
B1 Modification Indices for Model A (Severely Misspecified Model)
No Non-Zero Modification Indices for BETA
Modification Indices for GAMMA
KSI1 KSI2
-------- --------
ETA1 - - 13.505
ETA2 - - 24.873
Expected Change for GAMMA
KSI1 KSI2
-------- --------
ETA1 - - -0.440
ETA2 - - -0.715
Standardized Expected Change for GAMMA
KSI1 KSI2
-------- --------
ETA1 - - -0.344
ETA2 - - -0.448
B2 Modification Indices for Model B (Misspecified Model)
Modification Indices for BETA
ETA1 ETA2
-------- --------
ETA1 - - 10.533
ETA2 - - - -
Expected Change for BETA
ETA1 ETA2
-------- --------
ETA1 - - 0.488
ETA2 - - - -
Standardized Expected Change for BETA
ETA1 ETA2
-------- --------
ETA1 - - 0.406
ETA2 - - - -
Modification Indices for GAMMA
KSI1 KSI2
-------- --------
ETA1 - - 10.533
ETA2 - - - -
Expected Change for GAMMA
Schermelleh-Engel et al.: Evaluating the fit of Structural Equation Models 73
KSI1 KSI2
-------- --------
ETA1 - - -0.389
ETA2 - - - -
Standardized Expected Change for GAMMA
KSI1 KSI2
-------- --------
ETA1 - - -0.304
ETA2 - - - -
B3 Modification Indices for Model C (Correctly Specified Model)
No Non-Zero Modification Indices for BETA
No Non-Zero Modification Indices for GAMMA
Modification Indices for THETA-DELTA-EPS
Y1 Y2 Y3 Y4
-------- -------- -------- --------
X1 11.327 9.719 0.005 0.040
X2 10.705 9.216 0.942 0.812
X3 0.537 0.768 0.126 0.198
X4 0.517 0.734 0.251 0.197
Expected Change for THETA-DELTA-EPS
Y1 Y2 Y3 Y4
-------- -------- -------- --------
X1 -0.145 0.133 -0.003 0.008
X2 0.137 -0.126 -0.038 0.035
X3 -0.030 0.035 0.016 -0.019
X4 0.028 -0.032 0.022 -0.019
Completely Standardized Expected Change for THETA-DELTA-EPS
Y1 Y2 Y3 Y4
-------- -------- -------- --------
X1 -0.120 0.112 -0.002 0.006
X2 0.119 -0.111 -0.030 0.028
X3 -0.025 0.029 0.012 -0.015
X4 0.024 -0.028 0.017 -0.015
74 MPR-Online 2003, Vol. 8, No. 2
B4 Modification Indices for Model D
(Correctly Specified Parsimonious Model)
No Non-Zero Modification Indices for BETA
No Non-Zero Modification Indices for GAMMA
Modification Indices for THETA-DELTA-EPS
Y1 Y2 Y3 Y4
-------- -------- -------- --------
X1 11.327 9.719 0.005 0.040
X2 10.705 9.216 0.942 0.812
X3 0.537 0.768 0.126 0.198
X4 0.517 0.734 0.251 0.197
Expected Change for THETA-DELTA-EPS
Y1 Y2 Y3 Y4
-------- -------- -------- --------
X1 -0.145 0.133 -0.003 0.008
X2 0.137 -0.126 -0.038 0.035
X3 -0.030 0.035 0.016 -0.019
X4 0.028 -0.032 0.022 -0.019
Completely Standardized Expected Change for THETA-DELTA-EPS
Y1 Y2 Y3 Y4
-------- -------- -------- --------
X1 -0.120 0.112 -0.002 0.006
X2 0.119 -0.111 -0.030 0.028
X3 -0.025 0.029 0.012 -0.015
X4 0.024 -0.028 0.017 -0.015