STRUCTURAL EQUATION MODELING, 9(2), 233–255

Copyright © 2002, Lawrence Erlbaum Associates, Inc.

Evaluating Goodness-of-Fit Indexes for

Testing Measurement Invariance
Gordon W. Cheung
Department of Management
The Chinese University of Hong Kong

Roger B. Rensvold
Department of Management
City University of Hong Kong

Measurement invariance is usually tested using Multigroup Confirmatory Factor

Analysis, which examines the change in the goodness-of-fit index (GFI) when
cross-group constraints are imposed on a measurement model. Although many stud-
ies have examined the properties of GFI as indicators of overall model fit for sin-
gle-group data, there have been none to date that examine how GFIs change when be-
tween-group constraints are added to a measurement model. The lack of a consensus
about what constitutes significant GFI differences places limits on measurement
invariance testing.
We examine 20 GFIs based on the minimum fit function. A simulation under the
two-group situation was used to examine changes in the GFIs (∆GFIs) when
invariance constraints were added. Based on the results, we recommend using
∆comparative fit index, ∆Gamma hat, and ∆McDonald’s Noncentrality Index to eval-
uate measurement invariance. These three ∆GFIs are independent of both model
complexity and sample size, and are not correlated with the overall fit measures. We
propose critical values of these ∆GFIs that indicate measurement invariance.

Social science researchers are increasingly concerned with testing for measure-
ment invariance; that is, determining if items used in survey-type instruments
mean the same things to members of different groups. Measurement invariance is
critically important when comparing groups. If measurement invariance cannot be
established, then the finding of a between-group difference cannot be unambigu-

Requests for reprints should be sent to Gordon W. Cheung, Department of Management, The Chi-
nese University of Hong Kong, Shatin, Hong Kong. E-mail: [email protected]
234 CHEUNG AND RENSVOLD

ously interpreted. One does not know if it is due to a true attitudinal difference, or
to different psychometric responses to the scale items. This is of particular concern
in cross-cultural research when the cultures speak different languages, and re-
searchers use translated versions of a survey instrument (Janssens, Brett, & Smith,
1995; Reise, Widaman, & Pugh, 1993; Riordan & Vandenberg, 1994; Steenkamp
& Baumgartner, 1998). Other examples include groups having different levels of
academic achievement (Byrne, Shavelson, & Muthén, 1989), working in different
industries (Drasgow & Kanfer, 1985), of different genders (Byrne, 1994), and in
experimental versus control groups (Pentz & Chou, 1994).
Measurement invariance is a general term that can be applied to various compo-
nents of measurement models. Little (1997) identified two types of invariance.
Category 1 invariance has to do with the psychometric properties of the measure-
ment scales, and includes configural invariance (e.g., Buss & Royce, 1975; Irvine,
1969; Suzuki & Rancer, 1994), metric invariance (Horn & McArdle, 1992; also
called weak factorial invariance by Meredith, 1993), measurement error invariance
(Mullen, 1995; Singh, 1995), and scalar invariance (Meredith, 1993; Steenkamp &
Baumgartner, 1998; Vandenberg & Lance, 2000). Category 2 invariance has to do
with between-group differences in latent means, variances, and covariances. Gen-
erally speaking, Category 1 invariance is a prerequisite for the interpretation of
Category 2 differences, whereas Category 2 differences are usually the data having
substantive research interest.
Structural equation modeling (SEM) is widely used in the social sciences. The
suitability of a single-group measurement model is usually assessed using an SEM
procedure known as confirmatory factor analysis (CFA). A model is considered
suitable if the covariance structure implied by the model is similar to the
covariance structure of the sample data, as indicated by an acceptable value of
goodness-of-fit index (GFI).
The most commonly used GFI of SEM is the χ2 statistic, defined as

χ2 = ( N - 1)Fˆmin (1)

where N is the sample size, and F$min is the minimum value of the empirical fit
function, estimated using an iterative procedure under the assumption that the data
have multivariate normal distribution. A nonsignificant value of χ2 indicates fail-
ure to reject the null hypothesis that the hypothesized covariance matrix is identi-
cal to the observed covariance matrix, which is usually accepted as evidence of ad-
equate fit. A problem arises because of the statistic’s functional dependence on N.
For large sample sizes, the χ2 statistic provides a highly sensitive statistical test, but
not a practical test, of model fit. Owing to this and other considerations, many GFIs
have been proposed as alternatives to χ2. Some in common use include the compar-
ative fit index (CFI; Bentler, 1990), Tucker–Lewis Index (TLI; Tucker & Lewis,
1973), Normed Fit Index (NNFI; Bentler & Bonett, 1980), and root mean squared
 GFIs FOR MEASUREMENT INVARIANCE 235

error of approximation (RMSEA; Steiger, 1989). Due to the fact that most of the
“practical” GFIs do not have known sampling distributions, researchers have pro-
posed many criterion values indicative of satisfactory model fit; examples include
.90 or above for TLI and CFI. It is common practice to use multiple GFIs when
evaluating and reporting overall model fit.
An extension of CFA, Multigroup Confirmatory Factor Analysis (MGCFA),
tests the invariance of estimated parameters of two nested models across groups.
The degree of invariance is most frequently assessed by the Likelihood Ratio Test
(differences in χ2 between two nested models), although researchers have demon-
strated that differences in χ2 are also dependent on sample size (Brannick, 1995;
Kelloway, 1995). Reliance on the Likelihood Ratio Test is probably due to the lack
of sampling distributions for GFI differences. In contrast to the CFA test for overall
fit, there are no generally accepted criteria in MGCFA for determining if changes
in the “practical” GFIs are meaningful when measurement invariance constraints
are added. For example, there is no standard against which a researcher can com-
pare changes in CFI when measurement invariance constraints are added, in order
to determine if the constrained model fits the data less well than the less-con-
strained model (Vandenberg & Lance, 2000). Hence, the objective of this article is
to assess the effects of sampling error and model characteristics on MGCFA out-
comes; that is, differences in GFIs (∆GFIs) obtained when an unconstrained model
is compared with one having measurement invariance constraints, under the null
hypothesis of invariance. We propose critical values of ∆GFIs that are independent
from model characteristics, basing our proposals on sampling distributions of
∆GFIs obtained using simulations.

MEASUREMENT INVARIANCE

The series of tests that constitutes invariance testing through MGCFA are covered
at length elsewhere (Bollen, 1989b; Byrne et al., 1989; Cheung & Rensvold, 2000;
Drasgow & Kanfer, 1985; Jöreskog & Sörbom, 1993; Little, 1997; Steenkamp &
Baumgartner, 1998; Vandenberg & Lance, 2000). For the reader’s convenience we
briefly review them. There are eight invariance hypotheses that are frequently ex-
amined, as summarized in Table 1. The sequence of the invariance tests in Table 1
is only one of the many possible sequences that have been proposed, based on the
substantive research questions at hand. The five hypotheses at the top of the table
relate to measurement level invariance, whereas the three in the lower portion re-
late to construct level invariance (Little, 1997; Steenkamp & Baumgartner, 1998;
Vandenberg & Lance, 2000).
Hypothesis Hform postulates configural invariance; that is, participants belong-
ing to different groups conceptualize the constructs in the same way (Riordan &
Vandenberg, 1994). If configural invariance exists, then data collected from each
236
TABLE 1
Hypotheses of Measurement Invariance

Model Hypothesis Hypothesis Testa Hypothesis Name Symbolic Statementb Conceptual Meanings of Hypotheses

1 Hform Overall fit Configural invariance Λ(form

1)
= Λ(form
2)
Both groups associate the same subsets of items with the
same constructs (the cognitive domains are the same).
2 HΛ 2–1 Construct-level metric Λ(1) = Λ( 2 ) Overall, the strength of the relationships between items and
invariance their underlying constructs are the same for both groups.
(The constructs are manifested in the same way across
groups.)
3 Hλ 3–1 Item-level metric λ(ij1) = Λ(ij2 ) The strength of the relationship between each item and its
invariance underlying construct is the same for both groups.
4 HΛ,Θ(δ) 4–2 Residual variance Θ (1) = Θ ( 2 ) Items have the same internal consistency for both groups.
δ δ
invariance Alternatively: For both groups, items have the same
quality as measures of the underlying construct.
5 HΛ,ν 5–2 Intercept invariance τ (i1) = τ (i 2 ) The cross-cultural differences latent means indicated by the
items are the same across items. Alternatively: All items
indicate the same cross-cultural differences.
6 HΛ,Φ(jj) 6–2 Equivalence of construct Φ (jj1) = Φ (jj2 ) The range of responses given to each item is the same
variance across groups. Alternatively: the variability/range of
diversity with respect to the constructs are the same
across groups.
7 HΛ,Φ(jj′) 7–2 Equivalence of construct Φ (1) = Φ ( 2 ) The relationships among constructs (e.g., covariances) are
jj ¢ jj ¢
covariance the same across groups.
8 HΛ,ν,κ 8–5 Equivalence of latent κ (j1) = κ (j2 ) The mean level of each construct is the same across groups.
means
aIn each case, the statistic for testing the hypothesis is the difference between the fit of the constrained model and that of a less constrained model. For exam-

ple, the test statistic for HΛ is the fit difference between Model 2 (with construct-level constraints) and Model 1 (with no constraints). This is indicated in the table
as “2 – 1.”
bParenthetical superscripts indicate groups. For brevity, only the two-group case is shown, but each hypothesis generalizes to K groups. The general state-

ment of Hypothesis 1, for example, is Λ(form

1)
= Λ(form
2)
= Λ(form
3)
= L = Λ(form
k)
.
 GFIs FOR MEASUREMENT INVARIANCE 237

group decompose into the same number of factors, with the same items associated
with each factor (Meredith, 1993). Configural invariance may fail when, for exam-
ple, the concepts are so abstract such that participants’ perceptions of the con-
structs depend on their cultural context (Tayeb, 1994), or when participants from
different groups use different conceptual frames of reference and attach different
meanings to constructs (Millsap & Everson, 1991; Millsap & Hartog, 1988;
Riordan & Vandenberg, 1994). Configural invariance may also fail owing to a host
of other reasons, including data collection problems, translation errors, and so
forth.
Hypothesis HΛ postulates metric invariance (i.e., that all factor loading parame-
ters are equal across groups). Samples drawn from two populations may provide
data that indicate conceptual agreement in terms of the type and number of under-
lying constructs, and the items associated with each construct. Despite this, the
strengths of the relations between specific scale items and the underlying con-
structs may differ. The data may indicate disagreement concerning how the con-
structs are manifested. Metric invariance is important as a prerequisite for mean-
ingful cross-group comparison (Bollen, 1989b).
Hypothesis Hλ posits the invariance of factor loadings associated with a particu-
lar item. Due to the fact that the metric invariance requirement is usually difficult
to satisfy, some researchers (Byrne, et al., 1989; Marsh & Hocevar, 1985) propose
relaxing it. They suggest that if the noninvariant items constitute only a small por-
tion of the model, then cross-group comparisons can still be made because the
noninvariant items will not affect the comparisons to any meaningful degree. Be-
fore settling on this course of action, however, one must first identify the
noninvariant items. Hence, if the metric invariance hypothesis HΛ is rejected, a se-
ries of Hλ hypotheses are often tested in an attempt to locate items responsible for
overall noninvariance.
Hypothesis HΛ,Θ(δ) states that residual variance is invariant across groups. Re-
sidual variance is the portion of item variance not attributable to the variance of the
associated latent variable. Therefore, testing for the equality of between-group re-
sidual variance determines if the scale items measure the latent constructs with the
same degree of measurement error. Residual invariance may fail when participants
belonging to one group, compared with those of another, are unfamiliar with a
scale and its scoring formats, and therefore respond to it inconsistently (Mullen,
1995). In addition, differences in vocabulary, idioms, grammar, syntax, and the
common experiences of different cultures may produce residual noninvariance
(Malpass, 1977).
Hypothesis HΛ,ν proposes that in addition to metric invariance (HΛ), the vectors
of item intercepts are also invariant. The item intercepts are the values of each item
corresponding to the zero value of the underlying construct. Support for HΛ,ν indi-
cates the existence of strong factorial invariance (Meredith, 1993), also referred to
as scalar equivalence (Mullen, 1995). Strong factorial invariance is a prerequisite
238 CHEUNG AND RENSVOLD

for the comparison of latent means, because it implies that the measurement scales
have the same operational definition across groups (i.e., have the same intervals
and the zero points). In the absence of strong factorial invariance, the comparison
of latent means is ambiguous, because the effects of a between-group difference in
the latent means are confounded with differences in the scale and origin of the la-
tent variable. Under these circumstances, Byrne et al. (1989) proposed to compare
latent means under partial intercept invariance, assuming that the noninvariant
item will not affect the latent means comparison at a great extent.

CONSTRUCT LEVEL INVARIANCE

The remaining hypotheses in Table 1 are concerned with construct-level

invariance; that is, with the means and variances of constructs, and their equiva-
lence across groups. Hypothesis HΛ,Φ(jj) states that the variances of constructs (i.e.,
latent variables) are invariant across groups. This hypothesis may be rejected, for
example, when a construct representing an attitude having different levels of diver-
sity across groups is assessed. Hypothesis HΛ,Φ(jj) must be supported before re-
searchers can compare the correlations of constructs across groups (Byrne, 1994;
Jackson, Wall, Martin, & Davids, 1993; Marsh, 1993). Hypothesis HΛ,Φ(jj′) posits
that the covariances among the constructs are invariant across groups, whereas Hy-
pothesis HΛ,ν,κ posits that the latent means are invariant across groups. These two
hypotheses address substantive research questions, and are usually based on theo-
retical considerations.

FIT STATISTICS

When testing measurement invariance under MGCFA, a series of models are esti-
mated, and invariance is tested by comparing the GFI statistics of particular mod-
els with a model having additional between-group constraints. For example, test-
ing construct-level metric invariance involves comparing the fit of an
unconstrained model, which places no restrictions on model parameters,1 with a
constrained model in which all factor loadings associated with a particular con-
struct are constrained to be equal across groups. If the imposition of additional
constraints results in a signignicantly lower value of the fit statistic, then the con-
straint is “wrong.” The parameters constrained to be equal across groups should
not be constrained, because they are noninvariant. The fit differences used to test
various invariance hypotheses under MGCFA are shown in Table 1.
1Except for the referent associated with each construct, which is set equal to unity across groups.
 GFIs FOR MEASUREMENT INVARIANCE 239

Model fit differences are often determined using the likelihood-ratio (LR) test,
also known as the chi-square difference test (Bollen, 1989b). The chi-square dif-
ference (∆χ2) is calculated as

∆χ2 = χ2 c - χ2uc (2)

where χ2c and χ2uc are the values for the constrained model and the unconstrained
(or less constrained) model, respectively. Significance is evaluated with ∆df de-
grees of freedom, where
∆df = dfc – dfuc (3)
The LR test, like the usual chi-square test, is a null-hypothesis significance test
for a difference between the two groups. If there is no difference in fit, that is, if
( F$min ) c = ( F$min ) uc , then ∆χ2 = 0. If the sample sizes are large, however, even a
small difference between ( F$min ) c and ( F$min ) uc may result in a significant value of
∆χ2, indicating that the null hypothesis of no difference should be rejected even
when the difference is trivial (Brannick, 1995; Kelloway, 1995). The question then
becomes one of statistical significance versus practical significance.
When assessing overall model fit using CFA, it is common for researchers to
use other GFIs in preference to the chi-square statistic because the models rarely
seem to fit by that criterion due to its well-known dependence on sample size. On
the other hand, the same researchers tend to adopt the LR test when testing
invariance hypotheses using MGCFA (e.g., Byrne et al., 1989; Reise et al., 1993;
Steenkamp & Baumgartner, 1998). Brannick (1995) and Kelloway (1995) warned
against this double standard, and suggested that researchers should consistently
use a single standard for all GFI tests, whether applied to tests of overall model fit,
or to differences in fit between constrained and unconstrained models.
An alternative to ∆χ2 is ∆GFI, defined as
∆GFI = GFIc – GFIuc (4)
where GFIc and GFIuc are the values of some selected GFI estimated with respect
to the constrained and unconstrained model. As opposed to ∆χ2, however, there is
in general no statistically based criterion (controlling for sampling error) for deter-
mining if the invariance hypothesis ought to be rejected based on a particular value
of ∆GFI. Although many simulation studies have examined GFIs as indicators of
overall model fit for single-group data, there have been no studies that examine
various ∆GFI as indicators of measurement invariance. One exception was Little
(1997), who proposed four criteria for assessing the relative fit indexes of two
nested models: (a) the overall fit should be acceptable, (b) the difference in
Tucker–Lewis Index (TLI) should be less than or equal to 0.05, (c) indexes of local
misfit are uniformly and unsystematically distributed with respect to the con-
strained parameters, and (d) the constrained model is substantively more meaning-
240 CHEUNG AND RENSVOLD

ful and parsimonious than the unconstrained model. However, the 0.05 criterion
has neither strong theoretical nor empirical support, and it is not widely used. In
short, researchers to date have enjoyed considerable latitude in deciding whether a
particular value of ∆GFI indicates noninvariance (e.g., Drasgow & Kanfer, 1985).
We review the properties of 20 different ∆GFIs when invariance constraints are
added to two-group measurement models when the null hypotheses of invariance
are true. The GFIs can be classified into six categories, as follows:

1. GFIs based on the minimum sample discrepancy, including chi-square (χ2)

and normed chi-square ( χ2/df; Wheaton, Muthen, Alwin, & Summers, 1977).
2. GFIs based on the population discrepancy, including the noncentrality pa-
rameter (NCP; Steiger, Shapiro, & Browne, 1985) and RMSEA (Steiger, 1989).
3. Information-theoretic GFIs, including the Akaike’s Information Criterion
(Akaike, 1987), Browne and Cudeck’s Criterion (1989), and the Expected
Cross-Validation Index (Browne & Cudeck, 1993).
4. Incremental GFIs, including the Normed Fit Index (Bentler & Bonett,
1980), Relative Fit Index (RFI; Bollen, 1986), Incremental Fit Index (IFI; Bollen
1989a), TLI (Tucker & Lewis, 1973), CFI (Bentler, 1990), and Relative
Noncentrality Index (RNI; McDonald & Marsh, 1990).
5. Parsimony adjusted GFIs, including the parsimony-adjusted NFI (James,
Muliak, & Brett, 1982) and parsimonious CFI (Arbuckle & Wothke, 1999).
6. Absolute GFIs, including gamma hat (Steiger, 1989), a rescaled version of
Akaike’s Information Criterion (Cudeck & Browne, 1983), cross-validation index
(Browne & Cudeck, 1989), McDonald’s (1989) Non-Centrality Index, and critical
N (Hoelter, 1983). Algebraic definitions and properties of these GFIs, plus rules of
thumb for using them to evaluate overall model fit, can be found in Arbuckle and
Wothke (1999) and Hu and Bentler (1998).

The effect of sampling error is a critical concern. Knowledge of the sampling

distribution is a prerequisite to deciding whether a value of ∆GFI should be attrib-
uted to noninvariance or to sampling error. In addition to sampling error, it seems
likely that sample size, the characteristics of a model (number of factors, etc.) and
the particular GFI being used should also affect the ∆GFI distribution. We con-
ducted a simulation to address these issues.

SIMULATION

We used a Monte Carlo procedure to generate ∆GFI for the 20 GFIs described pre-
viously. The sampling distributions of the ∆GFIs were examined under the null hy-
pothesis of invariance. A total of 48 different models were generated by varying
the parameters shown in Table 2: the number of Factors F (either 2 or 3), the factor
 GFIs FOR MEASUREMENT INVARIANCE 241

TABLE 2
Model Parameters for Simulation

Code Variable Values

F Number of factors (2) 2 or 3

V Factor variance (2) .36 or .81
C Correlations of factors (2) .3 or .5
I Number of items per factor (3) 3, 4, or 5
L Factor loadings (2 for each value of I) If I = 3: 1, 1, 1
1, 1.25, 1.5
If I = 4: 1, 1, 1, 1
1, 1.25, 1.25, 1.5
If I = 5: 1, 1, 1, 1, 1
1, 1, 1.25 1.5, 1.5
N Sample size per group (2) 150 or 300

variance V (.36 or .81), the correlations between Factors C (.3 or .5), the number of
items per Factor I (3, 4, or 5), and the factor loadings L (two patterns for each value
of I). This range of model parameters, which follows Anderson and Gerbing’s
(1984) simulation, was an attempt to represent a range of models encountered in
practice, whereas keeping the size of the simulation within manageable limits.
Two different sample sizes N per group (150 or 300) were also examined. All
model parameters were estimated with maximum likelihood (ML) method.
The simulation procedure is shown schematically in Figure 1. It was assumed
that the latent factors would not explain all the variance in the items; therefore, the
reliability of each factor was set to 0.80. Appropriate item residual variances were
calculated using Fornell and Larcker’s (1981) equation. These, together with the
model parameters F, C, V, I, and L, were used to calculate a population covariance
matrix Σ for each model with χ2 equal to zero.
Instead of running the simulation using models with perfect fit, we used models
containing approximation errors (Cudeck & Henly, 1991). A residual Matrix E
consisting of random terms was generated, and the random off-diagonal elements
were added to Σ to produce a population covariance Matrix S. This matrix was
transformed using Cudeck and Browne’s (1992) procedure into a population
covariance matrix, having a model χ2 equal to the number of degrees of freedom
with sample size equals to 300. The choice of χ2 = df in the population produced
samples with unconstrained fit statistics representative of actual cases; for exam-
ple, the mean values of CFI, TLI, and RMSEA for samples generated during this
study were .97, .96 and .057, respectively. These were less than the perfect values
(e.g., CFI = 1.0), which are seldom encountered in practice. On the other hand,
they were high enough to justify invariance testing. If a test of overall fit produced
a value of CFI less than .90, for example, it is unlikely that the model would receive
further consideration.
242 CHEUNG AND RENSVOLD

FIGURE 1 Simulation flowchart.

The Cudeck and Browne (1992) procedure warrants a brief digression. For a
hypothesized Model K and a specified value of the fit function Fmin, there exists a
corresponding covariance matrix Σk(Fmin) and a corresponding covariance matrix
of measurement errors, E(Fmin), such that

Σ( Fmin ) = Σˆ k + E ( Fmin ) (5)

 GFIs FOR MEASUREMENT INVARIANCE 243

For any value of a scalar multiplier κ, there exists a specific degree of model fit.

Σ( Fmin ) = Σˆ k + κE ( Fmin ) (6)

If κ = 0, then Σ( Fmin ) = Σˆ k ; that is, the fit is perfect. If κ is greater than zero, then
the fit is less than perfect. In the context of a particular model, there exists a
one-to-one correspondence between the value of κ and model fit, as represented by
the value of some GFI. Cudeck and Browne (1992) demonstrated that finding κ for
any value of an GFI is straightforward when the generalized least squares function
is used to estimate the model. If the ML function is used, as in this study, then an it-
erative method for finding κ is required.
Amos 3.6 was used to generate two multivariate-normal samples of size N, rep-
resenting samples from two different groups that were completely invariant except
for the effects of sampling error. These two samples were tested for the eight types
of invariance represented by Hypotheses 1 through 8, and the relevant statistics
were appended to an output file. The portion of the process beginning with the gen-
eration of the two data sets was repeated 1,000 times for each combination of F, C,
V, I, L, and N. The same population covariance matrices were used for models with
sample sizes of 150 and 300.

SIMULATION RESULTS

The quality of the simulation was assessed by examining the distribution of the
∆χ2 statistics. As expected, these closely followed the distributions of χ2 having
∆df degrees of freedom (Steiger, Shapiro, & Browne, 1985).

Configural Invariance (Hypothesis Hform)

The effects of model parameters on the GFIs were assessed by performing a
separate six-way analysis of variance (ANOVA) for each index. The factor load-
ings, factor variances, and factor correlations had no effect on any of the GFIs for
the configural invariance (Hform) model. The percentages of total variance ex-
plained by number of items (I), number of factors (F), sample size (N), and Item ×
Factor interaction (I × F) are shown in Table 3. The number of items, number of
factors, and Item × Factor interactions had significant effects on the values of the
GFIs, showing that they were affected by the complexity of the model. The result
for χ2 is an artifact of the design of the simulation that set χ2 equals to the degrees
of freedom in the population. This procedure generates consistent theoretical p
values of the χ2 statistics, which fall between .45 to .48 across various levels of
model complexity considered in this study.
244 CHEUNG AND RENSVOLD

TABLE 3
Factors Affecting Goodness-of-Fit Indexes (GFIs) in Testing Configural
Invariance

Percent Variance Due to:

GFIs I F I×F N

GFIs based on the minimum sample discrepancy

Chi-square (χ2) 45.6 38.9 8.0 3.4
Normed chi-square (χ2/df) 33.4
GFIs based on the population discrepancy
Noncentrality parametera 32.9 27.9 5.8 12.9
Root mean square error of approximation
Information-theoretic GFIs
Akaike’s information criterion 45.7 44.3 6.4
Browne and Cudeck criterion 44.7 43.3 6.3
Expected value of the cross-validation index 35.7 34.8 5.0 18.1
Incremental GFIs
Normed fit index 40.9 26.1 10.7
Relative fit index 32.1 23.4 15.1
Incremental fit index 34.2 21.9
Tucker-Lewis index 24.8 18.1
Comparative fit index 34.0 21.9
Relative noncentrality index 34.0 21.9
Parsimony adjusted GFIs
Parsimonious normed fit index 67.7 21.8 6.1
Parsimonious comparative fit index 70.0 25.6
Other absolute GFIs
Gamma hat 11.7 6.3
Rescaled Akaike’s information criterion 35.8 34.8 5.0 18.0
Cross-validation indexb 27.0 23.6 10.5 17.9
McDonald’s (1989) noncentrality index 39.1 33.4 6.3
Critical N 10.0 7.0 35.0

Note. I = number of items (manifest variables); F = number of factors (latent variables); N = sam-
ple size.
aI × N (3.0%) also significantly affects NCP. bI × F × N (5.2%), I x N (8.6%), and F × N (6.8%) also

significantly affect cross-validation index.

Although most GFIs attempt to accommodate the effects of model complexity

in one way or another, the results show that except for RMSEA, all the GFIs fail to
fully do so. The findings are consistent with previous suggestions that the same
cutoff values should not be applied to evaluate models having different levels of
complexity (Anderson & Gerbing, 1984; Bearden, Sharma, & Teel, 1982;
Boomsma, 1982; Gerbing & Anderson, 1993; LaDu & Tanaka, 1989; Marsh,
Balla, & McDonald, 1988). Table 4 lists the means, first (or 99th) percentiles, and
standard deviation of the GFIs for the test of configural invariance for each combi-
nation of I and F. None of the factors had a significant effect on RMSEA. The mean
 TABLE 4
Summary Statistics of Fit Indexes for Testing Configural Invariance

CFI Gamma Hat Mc

M SD 1% M SD 1% M SD 1%

Overall .9728 .0180 .9196 .9980 .0009 .9957 .9385 .0474 .8077
I*F
3, 2 .9899 .0085 .9613 .9987 .0011 .9956 .9865 .0110 .9547
3, 3 .9793 .0109 .9449 .9982 .0008 .9958 .9598 .0182 .9106
4, 2 .9824 .0094 .9549 .9982 .0009 .9956 .9681 .0164 .9229
4, 3 .9672 .0123 .9306 .9977 .0007 .9958 .9153 .0257 .8467
5, 2 .9715 .0116 .9389 .9979 .0008 .9956 .9432 .0216 .8856
5, 3 .9468 .0156 .9017 .9974 .0006 .9957 .8583 .0317 .7733

IFI RNI

M SD 1% M SD 1%

Overall .9733 .0178 .9212 .9729 .0181 .9196

I*F
3, 2 .9903 .0087 .9625 .9902 .0089 .9613
3, 3 .9797 .0107 .9461 .9793 .0110 .9449
4, 2 .9826 .0094 .9556 .9824 .0095 .9549
4, 3 .9676 .0121 .9320 .9672 .0123 .9305
5, 2 .9719 .0114 .9400 .9715 .0116 .9389
5, 3 .9476 .0152 .9042 .9468 .0156 .9017

χ2 χ2/df NCP

M SD 99% M SD 99% M SD 99%

Overall 131.95 97.09 389.08 1.77 .418 2.90 57.61 50.01 215.12
I*F
3, 2 28.21 9.79 55.75 1.76 .612 3.48 12.21 9.79 39.75
3, 3 84.73 19.18 133.60 1.77 .400 2.78 36.73 19.18 85.60
4, 2 67.14 16.59 109.70 1.77 .437 2.89 29.14 16.59 71.70
4, 3 181.00 33.17 258.21 1.77 .325 2.53 79.00 33.17 156.21
5, 2 120.32 24.69 180.40 1.77 .363 2.65 52.32 24.69 112.40
5, 3 310.29 50.74 419.13 1.78 .292 2.41 136.29 50.74 245.13

RMSEA AIC BCC

M SD 99% M SD 99% M SD 99%

Overall .0568 .0157 .0926 219.95 121.66 521.08 225.59 124.40 528.54
I*F
3, 2 .0533 .0256 .1075 80.21 9.79 107.75 82.11 9.54 109.05
3, 3 .0567 .0148 .0883 168.73 19.18 217.60 173.20 18.27 220.51
4, 2 .0561 .0171 .0918 135.14 16.59 177.70 138.38 15.98 179.84
4, 3 .0582 .0097 .0806 289.00 33.17 366.21 296.62 31.24 371.11
5, 2 .0574 .0123 .0844 204.32 24.69 264.40 209.27 23.57 267.61
5, 3 .0589 .0073 .0768 442.29 50.74 551.13 453.96 47.37 558.59

(continued)

245
 TABLE 4 (Continued)

ECVI NFI RFI

M SD 99% M SD 1% M SD 1%

Overall .538 .330 1.451 .938 .0358 .828 .916 .0409 .792
I*F
3, 2 .198 .062 .321 .977 .0118 .936 .956 .0222 .880
3, 3 .414 .119 .630 .951 .0174 .896 .927 .0261 .844
4, 2 .332 .096 .516 .958 .0134 .921 .939 .0197 .883
4, 3 .706 .191 1.037 .925 .0223 .857 .903 .0289 .815
5, 2 .500 .139 .753 .934 .0180 .886 .912 .0238 .849
5, 3 1.077 .278 1.542 .883 .0303 .804 .858 .0366 .763

TLI PNFI PCFI

M SD 1% M SD 1% M SD 1%

Overall .9633 .0219 .9012 .6595 .0726 .5096 .6855 .0843 .5201
I*F
3, 2 .9816 .0166 .9275 .5208 .0063 .4991 .5280 .0045 .5127
3, 3 .9690 .0165 .9174 .6343 .0116 .5973 .6529 .0073 .6299
4, 2 .9741 .0140 .9136 .6503 .0091 .6248 .6666 .0064 .6480
4, 3 .9575 .0160 .9101 .7147 .0172 .6622 .7473 .0095 .7191
5, 2 .9623 .0154 .9191 .7056 .0136 .6694 .7340 .0087 .7094
5, 3 .9358 .0188 .8814 .7313 .0251 .6662 .7844 .0129 .7471

CAIC CK Critical N

M SD 99% M SD 99% M SD 99%

Overall .5365 .3290 1.4459 .7796 .7883 3.2803 348.95 122.49 779.35
I*F
3, 2 .1978 .0616 .3195 .2158 .0722 .3488 452.96 188.99 1075.12
3, 3 .4133 .1185 .6282 .4873 .1656 .7512 347.26 96.64 601.84
4, 2 .3309 .0959 .5139 .3755 .1233 .5873 362.27 106.62 659.02
4, 3 .7042 .1896 1.0336 .9359 .3533 1.4329 310.45 75.72 482.35
5, 2 .4985 .1379 .7507 .5974 .2023 .9166 328.18 85.16 540.03
5, 3 1.0741 .2766 1.5368 2.0654 1.1121 3.3712 292.56 67.30 426.23

Note. CFI = comparative fit index; Mc = McDonald’s Noncentrality Index; IFI = Incremental Fit Index;
RNI = Relative Noncentrality Index; NCP = noncentrality parameter; RMSEA = root mean squared error of ap-
proximation; AIC = Akaike’s Information Criterion; BCC = Browne and Cudeck Criterion; ECVI = Expected
Cross-Validation Index; NFI = Normed Fit Index; RFI = Relative Fit Index; TLI = Tucker–Lewis Index; PNFI =
parsimony-adjusted NFI; PCFI = parsimonious CFI; CAIC = rescaled Akaike’s Information Criterion; CK =
cross-validation index.

246
 GFIs FOR MEASUREMENT INVARIANCE 247

RMSEA for all models was .0401, with a standard deviation of 0.011 and a 99th
percentile value of .0655. These values are consistent with Browne and Cudeck’s
(1993) 0.05 criterion for RMSEA.

Other Invariance Tests (Hypotheses HΛ Through HΛ,ν,κ)

As noted previously, the test statistics for hypotheses HΛ through HΛ,ν were the
∆GFIs. These were calculated for 1,000 simulated tests for each hypothesis, for
each GFI. Correlations were calculated between the overall GFI (obtained from
the test of hypothesis Hform: configural invariance) and the ∆GFI. One desirable
property of a ∆GFI is that there should be a nonsignificant correlation between
overall fit (GFI) and incremental fit (∆GFI), owing to the independence between
the configural invariance test and other invariance tests. Except for NCP, IFI, CFI,
RNI, Gamma hat, McDonald’s NCI, and critical N, the value of each ∆GFI was
correlated with the overall GFI. For example, the values of TLI obtained with test-
ing for configural invariance were signficantly correlated with values of ∆TLI ob-
tained when testing for metric invariance, residual invariance, and equivalence of
construct covariance (–.385, –.426, and –.408). These results indicate that the tests
of these three invariance hypotheses were dependent on configural invariance. The
correlations between RMSEA and ∆RMSEA varied from –.108 (equivalence of
construct variance) to –.257 (metric invariance). On the other hand, the correla-
tions between CFI and ∆CFI varied from .000 (equivalence of construct variance)
to –.021 (equivalence of construct covariance).
A second desirable property of a ∆GFI is that it should not be affected by model
complexity. The effects of the model parameters on ∆NCP, ∆IFI, ∆CFI, ∆RNI,
∆Gamma hat, ∆McDonald’s NCI, and ∆critical N were tested using a six-way
ANOVA. Except for ∆NCP and ∆critical N, which were affected by sample size,
none of the model parameters explained more than 5% of the variance in these sta-
tistics.
A third desirable property of a ∆GFI is that it should not be redundant with other
∆GFIs. Correlations among the overall fit statistics of IFI, CFI, RNI, Gamma hat,
and McDonald’s NCI were calculated. Correlations among IFI, CFI, and RNI were
at the .99 level, and correlations among ∆IFI, ∆CFI and ∆RNI for Hypotheses 2
through 8 were all above .98. This indicates that reporting all six of these statistics
would be redundant. Due to the fact that CFI is the most frequently used index of
this group, we recommend reporting only CFI and ∆CFI, and not IFI, ∆IFI, RNI, or
∆RNI. Gamma hat, ∆Gamma hat, McDonald’s NCI, and ∆McDonald’s NCI are
not redundant with CFI and ∆CFI; therefore, we recommend that these also be re-
ported when the results of measurement invariance tests are reported.
The means, standard deviations, and first (or 99th) percentiles of the ∆GFIs are
listed in Table 5. The mean values in Table 5 show that the expected value for
∆CFI, ∆Gamma hat, and ∆McDonald’s NCI are all zero. The percentile values
 TABLE 5
Summary Statistics of Test Statistics for Hypotheses 2 Through 8

∆CFI ∆Gamma Hat ∆Mc

M SD 1% M SD 1% M SD 1%

H2 –.0001 .0025 –.0085 –.0000 .0002 –.0008 –.0001 .0049 –.0160

H3 .0000 .0010 –.0039 –.0000 .0001 –.0004 .0000 .0018 –.0071
H4 –.0001 .0029 –.0094 –.0000 .0003 –.0009 –.0001 .0055 –.0180
H5 .0000 .0024 –.0082 –.0000 .0002 –.0008 –.0000 .0048 –.0160
H6 .0000 .0015 –.0056 –.0000 .0001 –.0006 –.0000 .0028 –.0100
H7 .0000 .0013 –.0048 –.0000 .0001 –.0005 –.0000 .0025 –.0094
H8 .0000 .0015 –.0055 –.0000 .0001 –.0005 –.0000 .0028 –.0100

∆IFI ∆RNI

M SD 1% M SD 1%

H2 –.0002 .0025 –.0086 –.0001 .0025 –.0085

H3 –.0000 .0010 –.0039 .0000 .0010 –.0040
H4 –.0002 .0029 –.0097 –.0001 .0029 –.0095
H5 –.0000 .0015 –.0056 –.0000 .0015 –.0056
H6 –.0000 .0013 –.0048 –.0000 .0013 –.0048
H7 –.0001 .0025 –.0084 –.0000 .0025 –.0083
H8 –.0000 .0015 –.0056 –.0000 .0015 –.0056

∆χ2 ∆χ2/df ∆NCP

M SD 99% M SD 99% M SD 99%

H2 7.6158 4.7728 22.3030 –.088 .1053 .1856 .1158 3.9876 12.605

H3 0.9901 1.4059 6.6320 –.017 .0456 .1378 –.0099 1.4059 5.632
H4 10.1180 5.3916 26.0820 –.089 .1013 .1758 .1180 4.5319 13.757
H5 2.5282 2.3153 10.5190 –.028 .0544 .1457 .0282 2.2612 7.846
H6 2.0178 2.2571 10.1330 –.020 .0410 .1241 .0178 2.0169 7.438
H7 7.5301 4.6801 21.7879 –.068 .0826 .1655 .0301 3.9248 12.209
H8 2.5101 2.2995 10.4890 –.022 .0453 .1261 .0101 2.2437 7.794

∆RMSEA ∆AIC ∆BCC

M SD 99% M SD 99% M SD 99%

H2 –.0034 .0050 .0126 –7.3842 4.7105 4.9360 –8.3574 5.0638 4.0818

H3 –.0007 .0024 .0074 –1.0099 1.4059 4.6320 –1.1283 1.4071 4.5253
H4 –.0037 .0053 .0127 –9.8820 5.3547 3.7919 –11.1629 5.7887 2.8599
H5 –.0011 .0031 .0091 –2.4718 2.3163 5.3710 –2.7795 2.3486 5.0888
H6 –.0008 .0023 .0072 –1.9822 2.2452 5.0940 –2.2422 2.3250 4.8233
H7 –.0028 .0044 .0117 –7.4699 4.6980 4.5940 –8.4431 5.0634 3.9208
H8 –.0038 .0052 .0127 –2.4899 2.2980 5.3009 –2.7976 2.3290 5.0018

(continued)

248
 TABLE 5 (Continued)

∆ECVI ∆NFI ∆RFI

M SD 99% M SD 1% M SD 1%

H2 –.019 .0139 .0123 –.0041 .0029 –.014 .0034 .0036 –.0073

H3 –.0025 .0038 .0122 –.0006 .0010 –.0047 .0006 .0015 –.0049
H4 –.025 .0164 .0093 –.0055 .0035 –.017 .0036 .0036 –.0070
H5 –.0062 .0065 .0141 –.0014 .0016 –.0075 .0010 .0019 –.0055
H6 –.005 .0062 .0135 –.0011 .0014 –.0063 .0008 .0016 –.0048
H7 –.019 .0139 .0113 –.0041 .0028 –.014 .0028 .0031 –.0065
H8 –.0062 .0065 .0139 –.0014 .0016 –.0075 .0008 .0017 –.0049

∆TLI ∆PNFI ∆PCFI

M SD 1% M SD 1% M SD 1%

H2 .0036 .0039 –.0078 .0815 .0268 .0398 .0875 .0256 .0502

H3 .0006 .0016 –.0052 .0137 .0092 .0029 .0145 .0093 .0036
H4 .0038 .0038 –.0073 .1114 .0427 .0486 .1206 .0417 .0630
H5 .0011 .0020 –.0059 .0309 .0172 .0093 .0331 .0173 .0117
H6 .0008 .0017 –.0051 .0212 .0104 .0080 .0229 .0107 .0093
H7 .0029 .0033 –.0069 .0808 .0266 .0390 .0875 .0256 .0502
H8 .0009 .0018 –.0052 .0307 .0171 .0090 .0331 .0173 .0116

∆CAIC ∆CK ∆CN

M SD 99% M SD 99% M SD 99%

H2 –.018 .0139 .0123 –.062 .0973 .0037 10.329 27.503 75.533

H3 –.0025 .0038 .0121 –.0068 .0087 .0090 1.869 13.445 28.289
H4 –.025 .0163 .0093 –.080 .1211 .0006 11.232 28.428 81.915
H5 –.0062 .0065 .014 –.018 .0246 .0083 3.282 16.642 40.205
H6 –.0050 .0062 .0134 –.016 .0252 .0084 2.530 11.702 24.038
H7 –.019 .0139 .0113 –.062 .0975 .0033 8.642 23.843 64.015
H8 –.0062 .0064 .0139 –.018 .0246 .0078 2.789 14.583 34.613

Note. CFI = comparative fit index; Mc = McDonald’s Noncentrality Index; IFI = Incremental Fit Index;
RNI = Relative Noncentrality Index; NCP = noncentrality parameter; RMSEA = root mean squared error of ap-
proximation; AIC = Akaike’s Information Criterion; BCC = Browne and Cudeck Criterion; ECVI = Expected
Cross-Validation Index; NFI = Normed Fit Index; RFI = Relative Fit Index; TLI = Tucker–Lewis Index; PNFI =
parsimony-adjusted NFI; PCFI = parsimonious CFI; CAIC = rescaled Akaike’s Information Criterion; CK =
cross-validation index. H2 = metric invariance (weak factorial invariance); H3 = partial metric invariance (par-
tial measurement invariance); H4 = metric invariance + invariance of residual variance; H5 = strong factorial
invariance (metric invariance + scalar invariance); H6 = metric invariance + invariance of construct variance;
H7 = metric invariance + invariance of construct covariance; H8 = strong factorial invariance + invariance of la-
tent means.

249
250 CHEUNG AND RENSVOLD

shown in Table 5 are the critical values for rejecting the null hypothesis of equiva-
lence, with an alpha of 0.01 and assuming multivariate normal distributions.

DISCUSSION OF SIMULATION RESULTS

Although the formulas for many GFIs (e.g., CFI and TLI) involve terms that adjust
for degrees of freedom, this study shows that number of items per factor and num-
ber of factors in the model affect most of the GFIs (except for RMSEA). When
overall fit is examined, models with more items and more factors can be expected
to yield smaller values of these GFIs. This is due to the omission of small, theoreti-
cally insignificant factor loadings and correlated error terms in the model (Hall,
Snell, & Foust, 1999; Hu & Bentler, 1998). In exploratory factor analysis, these
terms are usually ignored; in CFA, and other SEM, they are hypothesized to be
zero. This assumption has a negative impact on overall fit. This should serve as a
warning to researchers who judge model fit in accordance with some generally ac-
cepted criterion (e.g., CFI = .90) while ignoring the effects of model complexity.
RMSEA was not affected by any of the model parameters examined in this study,
but its standard error was affected. Models with fewer items and factors were asso-
ciated with larger standard errors in RMSEA.
Unlike the LR test in which ∆χ2 is always greater than or equal to zero, many
difference indexes (e.g., ∆RFI and ∆TLI) can assume both positive and negative
values. This is because the underlying GFIs are functions of the number of degrees
of freedom in the model. If the null hypothesis of invariance is true, and there is no
sampling error, then decreasing the number of degrees of freedom can produce a
value of ∆GFI greater than zero. If a ∆GFI is less than zero, then its value does not
represent a change from a baseline value of zero, but rather a change from its hypo-
thetical positive value. Hence, a slight reduction in the difference indexes may in-
dicate a substantial change in the minimum value of the fit function.
Many ∆GFIs are superior to ∆χ2 as tests of invariance because they are not af-
fected by sample size. In many cases, however, ∆GFI is correlated with the GFI of
the overall model. This implies that less accurately specified models produce
larger values of difference statistics when measurement invariance constraints are
added. As shown previously, the only difference statistics not having this undesir-
able characteristic are ∆CFI, ∆Gamma hat, ∆McDonald’s NCI, ∆NCP, ∆IFI,
∆RNI, and ∆critical N.
The aforementioned results show that ∆CFI, ∆Gamma hat, and ∆McDonald’s
NCI are robust statistics for testing the between-group invariance of CFA models.
These results are unexpected because many simulation studies have demonstrated
that GFIs are affected by model complexity when evaluating overall model fit.
Although the standard errors and critical values differ for the different
invariance models, the between-model variations are so small that a general crite-
 GFIs FOR MEASUREMENT INVARIANCE 251

rion for all hypotheses can be proposed. A value of ∆CFI smaller than or equal to
–0.01 indicates that the null hypothesis of invariance should not be rejected. For
∆Gamma hat and ∆McDonald’s NCI, the critical values are –.001 and –.02, re-
spectively.

Limitations of the Simulation

One limitation of this study is that only GFIs based on the minimum value of the fit
function are examined. Future studies are required to review other GFIs, such as
the Jöreskog–Sörbom GFI (Jöreskog & Sörbom, 1993), adjusted GFI (Jöreskog &
Sörbom, 1993), and root mean squared residual (Jöreskog & Sörbom, 1993).
Although 768,000 CFA models were estimated in the course of this study,
generalizability is limited by the fact that only ML estimation was used and only
96 combinations of model parameters and sample sizes were examined. Further
studies are needed to assess the effects of a wider range of parameters and other es-
timation methods. In addition, this study stipulated that the data distributions were
multivariate normal. Deviations from multivariate normality may affect the re-
sults.
Another limitation of this study is that we only examined the Type I error when
testing for measurement invariance. Although the Type II error or power may not
be a major concern in most CFA work because of the large sample sizes, future ef-
forts are needed to examine the power of the proposed criteria for determining
measurement invariance. However, before the power can be examined, we must
first meaningfully define the effect size of deviations from measurement
invariance. Otherwise, using any arbitrary non-invariant model to examine the
power will not be meaningful.
Finally, this simulation is limited to measurement models with two groups.
Suitability of the recommended GFIs for testing across three or more groups is an
interesting topic for future study.

CONCLUSION

The purpose of this article is to assist the decision-making processes of researchers

engaged in examining measurement invariance. The various tests for be-
tween-group equivalence are laid out in a hierarchical sequence that researchers
should find useful.
This study contributes to the literature in two ways. First, it invites the attention
of researchers to different forms of measurement invariance that can be detected
using structural equation modeling. MGCFA provides an elegant way to examine a
large number of issues via a single procedure, rather than many separate proce-
dures. This should facilitate empirical tests of many conceptual models having
252 CHEUNG AND RENSVOLD

cross-group differences. In addition, SEM provides a direct measure of how much

a measurement model is improved or degraded by various between-group con-
straints, which offers a clear advantage over other techniques currently in use.
MGCFA is an extremely powerful procedure for detecting a range of
cross-group differences, particularly when criterion values of ∆CFI, ∆Gamma hat,
and ∆McDonald’s NCI criteria are utilized. These cross-group differences, rather
than being viewed as impediments to research, should be considered valid objects
of research. Metric invariance, for example, need not be seen merely as an obstacle
that must be surmounted before the equality of latent means can be assessed;
rather, it should seen as a source of potentially interesting and valuable information
about how different groups view the world. The same comment can be made with
respect to any one of the measurement invariance failures considered.
Second, this article provides practical guidelines for determining when mea-
surement invariance should be rejected. Currently, measurement invariance is
most commonly examined with the LR test, which is based on the chi-square sta-
tistic. However, it makes no sense to argue against the usefulness of the chi-square
and rely on various GFIs to evaluate the overall model fit and then argue for the
usefulness of the chi-square instead of various GFIs to test for measurement
invariance. We proposed critical values of ∆CFI, ∆Gamma hat, and ∆McDonald’s
NCI, that are independent of model parameters and sample size for testing mea-
surement invariance and more generally, two nested models.

ACKNOWLEDGMENTS

Preparation of this article was supported in part by a Competitive Earmarked Re-

search Grant funded by the Research Grants Council of the Hong Kong University
Grants Committee (Project Code No. CUHK 4032/00H) to Gordon W. Cheung.
We thank the editor and three anonymous reviewers for their invaluable help
and input.
An earlier version of this article was presented at the 1999 Annual Meeting of
the Academy of Management, Chicago, Illinois.

REFERENCES

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 solu-
tions, and goodness-of-fit indices for maximum likelihood confirmatory factor analysis.
Psychometrika, 49, 155–173.
Arbuckle, J. L., & Wothke, W. (1999). Amos 4.0 user’s guide. Chicago: SmallWaters.
Bearden, W. O., Sharma, S., & Teel, J. R. (1982). Sample size effects on chi-square and other statistics
used in evaluating causal models. Journal of Marketing Research, 19, 425–430.
 GFIs FOR MEASUREMENT INVARIANCE 253

Bentler, P. M. (1990). Comparative fit indexes in structural models. Psychological Bulletin, 107,
238–246.
Bentler, P. M., & Bonett, D. G. (1980). Significance tests and goodness of fit in the analysis of
covariance structures. Psychological Bulletin, 88, 588–606.
Bollen, K. A. (1986). Sample size and Bentler and Bonett’s nonnormed fit index. Psychometrika, 51,
375–377.
Bollen, K. A. (1989a). A new incremental fiIt index for general structural equation models. Sociologi-
cal Methods and Research, 17, 303–316.
Bollen, K. A. (1989b). Structural equations with latent variables. New York: Wiley.
Boomsma, A. (1982). The robustness of LISREL against small sample sizes in factor analysis models.
In K. G. Jöreskog & H. Wold (Eds.), Systems under indirect observation: Causality, structure, pre-
diction (Part 1, pp. 149–173). Amsterdam: North-Holland.
Brannick, M. T. (1995). Critical comments on applying covariance structure modeling. Journal of Or-
ganizational Behavior, 16, 201–213.
Browne, M. W., & Cudeck, R. (1989). Single sample cross-validation indices for covariance structures.
Multivariate Behavioral Research, 24, 445–455.
Browne, M. W., & Cudeck, R. (1993). Alternative ways of assessing model fiIt. In K. A. Bollen & J. S.
Long (Eds.), Testing structural equations models (pp. 136–162). Newbury Park, CA: Sage.
Buss, A. R., & Royce, J. R. (1975). Detecting cross-cultural commonalties and differences: Intergroup
factor analysis. Psychological Bulletin, 82, 128–136.
Byrne, B. M. (1994). Testing for the factorial validity, replication, and invariance of a measurement in-
strument: A paradigmatic application based on the Maslach Burnout Inventory. Multivariate Behav-
ioral Research, 29, 289–311.
Byrne, B. M., Shavelson, R. J., & Muthén, B. (1989). Testing for the equivalence of factor covariance
and mean structures: The issue of partial measurement invariance. Psychological Bulletin, 105,
456–466.
Cheung, G. W., & Rensvold, R. B. (2000). Assessing extreme and acquiescence response sets in
cross-cultural research using structural equation modeling. Journal of Cross-Cultural Psychology,
31, 187–212.
Cudeck, R., & Browne, M. W. (1983). Cross-validation of covariance structures. Multivariate Behav-
ioral Research, 18, 147–167.
Cudeck, R., & Browne, M. W. (1992). Constructing a covariance matrix that yields a specified mini-
mizer and a specified minimum discrepancy function value. Psychometrika, 57, 357–369.
Cudeck, R., & Henly, S. J. (1991). Model selection in covariance structures analysis and the “problem”
of sample size: A clarification. Psychological Bulletin, 109, 512–519.
Drasgow, F., & Kanfer, R. (1985). Equivalence of psychological measurement in heterogeneous popu-
lations. Journal of Applied Psychology, 70, 662–680.
Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservable variables
and measurement error. Journal of Marketing Research, 18, 39–50.
Gerbing, D. W., & Anderson, J. C. (1993). Monte Carlo evaluations of goodness-of-fit indices for struc-
tural equation models. In K. A. Bollen & J. S. Long (Eds.), Testing structural equation models (pp.
40–65). Newbury Park, CA: Sage.
Hall, R. J., Snell, A. F., & Foust, M. S. (1999). Item parcelling strategies in SEM: Investigating the sub-
tle effects of unmodeled secondary constructs. Organizational Research Methods, 2, 233–256.
Hoelter, J. W. (1983). The analysis of covariance structures: Goodness-of-fit indices. Sociological
Methods and Research, 11, 325–344.
Horn, J. L., & McArdle, J. J. (1992). A practical and theoretical guide to measurement invariance in ag-
ing research. Experimental Aging Research, 18, 117–144.
Hu, L. T., & Bentler, P. M. (1998). Fit indices in covariance structure modeling: Sensitivity to
underparameterized model misspecification. Psychological Methods, 3, 424–453.
254 CHEUNG AND RENSVOLD

Irvine, S. H. (1969). Contributions of ability and attainment testing in Africa to a general theory of in-
tellect. Journal of Biosocial Science, 1, 91–102.
Jackson, P., Wall, T., Martin, R., & Davids, K. (1993). New measures of job control, cognitive demand,
and production responsibility. Journal of Applied Psychology, 78, 753–762.
James, L. R., Mulaik, S. A., & Brett, J. M. (1982). Causal analysis: Assumptions, models and data.
Beverly Hills: Sage.
Janssens, M., Brett, J. M., & Smith, F. J. (1995). Confirmatory cross-cultural research: Testing the via-
bility of a corporation-wide safety policy. Academy of Management Journal, 38, 364–382.
Jöreskog, K., & Sörbom, D. (1993). LISREL 8: User’s reference guide. Chicago: Scientific Software
International.
Kelloway, E. K. (1995). Structural equation modelling in perspective. Journal of Organizational Be-
havior, 16, 215–224.
La Du, T. J., & Tanaka, J. S. (1989). The influence of sample size, estimation method, and model speci-
fication on goodness-of-fit assessment in structural equation models. Journal of Applied Psychology,
74, 625–635.
Little, T. D. (1997). Mean and covariance structures (MACS) analyses of cross-cultural data: Practical
and theoretical issues. Multivariate Behavioral Research, 32, 53–76.
Malpass, R. S. (1977). Theory and method in cross-cultural psychology. American Psychologist, 32,
1069–1079.
Marsh, H. W. (1993). The multidimensional structure of academic self-concept: Invariance over gender
and age. American Educational Research Journal, 30, 841–860.
Marsh, H. W., Balla, J. R., & McDonald, R. P. (1988). Goodness-of-fit indexes in confirmatory factor
analysis: The effect of sample size. Psychological Bulletin, 103, 391–410.
Marsh, H. W., & Hocevar, D. (1985). Application of confirmatory factor analysis to the study of
self-concept: First- and higher order factor models and their invariance across groups. Psychological
Bulletin, 97, 562–582.
McDonald, R. P. (1989). An index of goodness-of-fit based on noncentrality. Journal of Classification,
6, 97–103.
McDonald, R. P., & Marsh, H. W. (1990). Choosing a multivariate model: Noncentrality and goodness
of fit. Psychological Bulletin, 107, 247–255.
Meredith, W. (1993). Measurement invariance, factor analysis and factorial invariance. Psychometrika,
58, 525–543.
Millsap, R. E., & Everson, H. (1991). Confirmatory measurement model comparisons using latent
means. Multivariate Behavioral Research, 26, 479–497.
Millsap, R. E., & Hartog, S. B. (1988). Alpha, beta, and gamma changes in evaluation research: A struc-
tural equation approach. Journal of Applied Psychology, 73, 574–584.
Mullen, M. (1995). Diagnosing measurement equivalence in cross-national research. Journal of Inter-
national Business Studies, 3, 573–596.
Pentz, M. A., & Chou, C. (1994). Measurement invariance in longitudinal clinical research assuming
change from development and intervention. Journal of Consulting and Clinical Psychology, 62,
450–462.
Reise, S. P., Widaman, K. F., & Pugh, R. H. (1993). Confirmatory factor analysis and item response the-
ory: Two approaches for exploring measurement invariance. Psychological Bulletin, 114, 552–566.
Riordan, C. M., & Vandenberg, R. J. (1994). A central question in cross-cultural research: Do employ-
ees of different cultures interpret work-related measures in an equivalent manner? Journal of Man-
agement, 20, 643–671.
Singh, J. (1995). Measurement issues in cross-national research. Journal of International Business
Studies, 26, 597–619.
Steenkamp, J. E. M., & Baumgartner, H. (1998). Assessing measurement invariance in cross- national
consumer research. Journal of Consumer Research, 25, 78–90.
 GFIs FOR MEASUREMENT INVARIANCE 255

Steiger, J. H. (1989). EzPATH: Causal modeling. Evanston, IL: SYSTAT.

Steiger, J. H., Shapiro, A., & Browne, M. W. (1985). On the multivariate asymptotic distribution of se-
quential chi-square statistics. Psychometrika, 50, 253–263.
Suzuki, S., & Rancer, A. S. (1994). Argumentativeness and verbal aggressiveness: Testing for concep-
tual and measurement equivalence across cultures. Communication Monographs, 6, 256–279.
Tayeb, M. (1994). Organizations and national culture: Methodology considered. Organization Studies,
15, 429–446.
Tucker, L. R., & Lewis, C. (1973). A reliability coefficient for maximum likelihood factor analysis.
Psychometrika, 38, 1–10.
Vandenberg, R. J., & Lance, C. E. (2000). A review and synthesis of the measurement invariance litera-
ture: Suggestions, practices, and recommendations for organizational research. Organizational Re-
search Methods, 3, 4–69.
Wheaton, B., Muthén, B., Alwin, D. F., & Summers, G. F. (1977). Assessing reliability and stability in
panel models. In D. R. Heise (Ed.), Sociological methodology (pp. 84–136). San Francisco:
Jossey-Bass.
Windle, M., Iwawaki, S., & Lerner, R. M. (1988). Cross-cultural comparability of temperament among
Japanese and American preschool children. International Journal of Psychology, 23, 547–567.