An Updated Paradigm for Scale Development Incorporating Unidimensionality and Its

Assessment
Author(s): David W. Gerbing and James C. Anderson
Source: Journal of Marketing Research , May, 1988, Vol. 25, No. 2 (May, 1988), pp. 186-
192
Published by: Sage Publications, Inc. on behalf of American Marketing Association

Stable URL: https://www.jstor.org/stable/3172650

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

American Marketing Association and Sage Publications, Inc. are collaborating with JSTOR to
digitize, preserve and extend access to Journal of Marketing Research

This content downloaded from

182.72.243.45 on Fri, 01 Mar 2024 06:14:07 +00:00
All use subject to https://about.jstor.org/terms
 DAVID W. GERBING and JAMES C. ANDERSON*

The authors outline an updated paradigm for scale development that incorporates
confirmatory factor analysis for the assessment of unidimensionality. Under this par-
adigm, item-total correlations and exploratory factor analysis are used to provide
preliminary scales. The unidimensionality of each scale then is assessed simulta-
neously with confirmatory factor analysis. After unidimensional measurement has
been acceptably achieved, the reliability of each scale is assessed. Additional evi-
dence for construct validity beyond the establishment of unidimensionality then can
be provided by embedding the unidimensional sets of indicators within a nomolog-
ical network defined by the complete structural model.

An Updated Paradigm for Scale Development

Incorporating Unidimensionality and Its
Assessment

McDonald 1981).' The importance of unidimensionality

The purpose of measurement in theory testing and de-
has been stated succinctly by Hattie (1985 p. 49): "That
velopment research is to provide an empirical estimate
of each theoretical construct of interest. Because of the a set of items forming an instrument all measure just one
limitations inherent in single-item measures (cf. Chur-thing in common is a most critical and basic assumption
chill 1979), respondents usually are administered two orof measurement theory."
more measures, often referred to as a scale, that are in- Because the meaning of a measure intended by the
tended to be alternative indicators of the same underly- researcher may not be the same as the meaning imputed
ing construct. A composite score defined by the respon- to it by the respondents, the scale development process
dent's scores on these measures, generally calculated asmust include an assessment of whether the multiple mea-
an unweighted sum, provides an estimate of the corre-sures that define a scale can be acceptably regarded as
sponding construct. Our central thesis is that the com-alternative indicators of the same construct. Building on
putation of this composite score is meaningful only ifthe earlier work of Churchill (1979) and Peter (1979,
each of the measures is acceptably unidimensional. Uni-1981), we outline an updated paradigm for scale devel-
dimensionality refers to the existence of a single trait or opment that incorporates confirmatory factor analysis (cf.
construct underlying a set of measures (Hattie 1985; Bentler 1985; J6reskog and Sorbom 1984) for the eval-
uation of unidimensionality. The key aspect of this up-
dated paradigm is that confirmatory factor analysis af-
fords a stricter interpretation of unidimensionality than
can be provided by more traditional methods such as
*David W. Gerbing is Associate Professor of Business Adminis- coefficient alpha, item-total correlations, and explora-
tration, Department of Management, Portland State University. James tory factor analysis and thus generally will provide dif-
C. Anderson is the William L. Ford Professor of Marketing and
Wholesale Distribution, and Associate Professor of Behavioral Sci- ferent conclusions about the acceptability of a scale.
ence in Management, J. L. Kellogg Graduate School of Management,
Northwestern University.
Support for this research was provided in part by the McManus
Research Chair awarded to the second author and is gratefully ac- 'Measures that satisfy this definition of unidimensionality have been
knowledged. referred to as "congeneric" measurements (Joreskog 1971) or "point"
variables (Burt 1973).

186

Journal of Marketing Research

Vol. XXV (May 1988), 186-92

This content downloaded from

182.72.243.45 on Fri, 01 Mar 2024 06:14:07 +00:00
All use subject to https://about.jstor.org/terms
 UPDATED PARADIGM FOR SCALE DEVELOPMENT 187

As background for understanding, A PARADIGM we FORfirst

DEVELOPING
briefly re-
view the concept of unidimensionality, UNIDIMENSIONAL SCALES present the
then
confirmatory factor analysis model.
The development and evaluation of measurement scale
UNIDIMENSIONAL MEASUREMENT AND traditionally have relied on one or more of the followin
CONFIRMATORY FACTOR ANALYSIS analyses: coefficient alpha, item-total correlations, an
exploratory factor analysis. The use of confirmatory fa
The mathematical definition of unidimensionalitytor analysis for building and evaluating measureme
adopted here is based on the traditional common factor
scales is a relatively recent development. Because ea
model in which a set of indicators share only a ofsingle
these four methods is based on a different criterion,
underlying factor, t (McDonald 1981). Assumingdiferent
lin- conclusions may be drawn depending on t
earity, the model for a single indicator can be given as of method. Our key point is that, of these meth-
choice
ods, only a confirmatory factor analysis of a multip
(1) Xii = ki + 6, indicator measurement model directly tests unidime
where xi is the ith indicator from the set of unidimen-
sionality according to the definition provided in equatio
sional indicators, Xi is the corresponding factor loading, 1. Using the obtained parameter estimates, one can com-
and 8i is the corresponding regression residual, uncor-
pute the indicator correlations predicted by the mod
related with any factors or residuals (Joreskog and Srrbomwith the product rules for internal consistency and e
1984). Because most studies investigate posited rela- ternal consistency. The overall goodness of fit is eva
tionships among several constructs, they are character-
uated according to the similarity of the predicted wi
ized by several sets of postulated unidimensional mea-the actual correlations.
sures. The relationship of these measures to their
respective constructs is represented formally by a mul- An Illustrative Model
tiple-indicator measurement model in which each esti- For pedagogical clarity, we illustrate this updated par-
mated construct is defined by at least two indicators andadigm with synthetic data. The model in Figure 1 rep-
each indicator is intended as an estimate of only one con-resents the underlying known structure that generates these
struct (Anderson and Gerbing 1982; Hunter and Gerbing data. This model specifies two moderately correlated
1982).
factors (t, and (2). Each factor is indicated primarily by
Two criteria follow directly from equation 1, ex- five items, though one item in each set also indicates to
pressed as product rules; each represents a necessary a smaller degree the other factor. Each of the two ad-
condition for unidimensionality. The correlation of two
ditional factors (63 and t) provides a source of common
indicators, i and j, of the same construct, t, is described covariance for two pairs of items across the two sets.
by the product rule for internal consistency:2
Substantively, 63 and t might represent either unwanted
constructs or method factors (Gerbing and Anderson
(2) pij = PitgPik. 1984).
The correlation of two indicators, i and p, where p is an
indicator of another construct t*, is described by the
product rule for external consistency:
Figure 1
(3) pip = PikPS, Ppk,. MODEL UNDERLYING CORRELATION MATRIX IN TABLE 2

When t is the same as t*, equation 3 reduces to equation

2; that is, internal consistency represents a special case .30

of external consistency. A useful methodological con-

sideration follows from an analysis of external consis-
tency: indicators from other scales provide a means by 2

which to assess unidimensionality of the items that de-

fine the given scale. Simply stated, given equation 3,
measures that are truly alternate indicators of the same
underlying construct will evince a parallel pattern
.78of re-
.78 .78 .81 .78 .78 .81 .78 .78 .78
lationships, to within sampling error, with all other mea-
sures in the set.
x1 X2 6x X4 X5 X6 X7 X8 X9 X1

2Note that equation 2 represents a generalization of the parallel test

model from classical test theory (Lord and Novick 1968) in which

Pii t 2
where i and i' are parallel measures. Unidimensional measures are
more general in that they may have unequal true score variances and
unequal error variances (cf. Joreskog 1971).

This content downloaded from

182.72.243.45 on Fri, 01 Mar 2024 06:14:07 +00:00
All use subject to https://about.jstor.org/terms
 188 JOURNAL OF MARKETING RESEARCH, MAY 1988

This measurement model was chosen to illustrate two with total scores are the best items for a general-purpose
types of violations of multiple-indicator measurement. test" (p. 279). Kline (1983, p. 48, 49) maintains, "The
First, unidimensional measurement for each set of itemsstandard item-analytic method of producing a homoge-
neous test is to correlate each item with the total score
as formalized by a multiple-indicator measurement model
could not be achieved successfully until the items that on the item pool or, if several scales are being consid-
indicate both k, and 2, x4 and x7, are deleted from the ered simultaneously, to correlate each item with their to-
model. Second, deleting these items also means that the tal score." Though both authors recognize problems that
other factors, k3 and 4, are no longer common factors. may arise from an analysis of item-total correlations, they
Once x4 and x7 are removed, neither 63 nor 4 provides do not consider them so great as to diminish the use-
sources of common variance across two or more items, fulness of the analysis. Nunnally prefers the item-total
illustrating that the composition of indicator-specific analysis over alternative analyses such as factor analysis
for scale construction and though Kline (p. 49) recog-
variance is defined in relation to the particular set of in-
dicators in the analysis. The initial common factors k3 nizes the advantage of factor analysis in some situations,
and 4 would be redefined as specific factors that are he concludes that "there is little essential difference be-
tween the methods."
indistinguishable in the resulting factor analysis from the
random error variance of classical test theory (GerbingThe problem is that the item-total method does not
and Anderson 1984; Smith 1974). That is, deleting x4 account for external consistency. By not accounting for
and x7 from the model means that k3 and 4 will con- the relations of the posited alternate indicators with in-
tribute only to the uniquenesses (Harman 1976) of the dicators of different factors, an item-total analysis may
affected indicators, x2 and x9. Related to this, deleting
fail to discriminate between sets of indicators that rep-
resent different, though correlated, factors. This prob-
x4 and x7 means that k3 and 4 contribute only to the vari-
ances of x2 and x,. lem can be explicated within the context of multiple-in-
Using the product rules for internal consistency and dicator measurement models. Though an indicator, x, that
external consistency, in conjunction with the specified is consistent with a multiple-indicator measurement model
parameter values, we generated the population item cor-will correlate more highly with its own factor, k, than
relation matrix in Table 1 from the model in Figure 1. with another factor, k*, if the factors k and t* are highly
A preliminary examination of the patterning of the cor- correlated, then the correlation of x with the other factor,
relations, without knowledge of the underlying structure, Pxk*, may be substantial. Let C represent a composite
is consistent with the hypothesis that the 10 items po- arbitrarily formed by combining items that represent in-
dicators of distinct, though correlated, factors k and k*.
tentially could represent two distinct scales. The first five
indicators all correlate .61 with one another, as do theThe correlation of each factor with this composite, Ptc
last five indicators. The mean correlation between the and tktc, would also be high. As the following equation
indicators in these two sets is only .19. demonstrates, indicators from different factors as spec-
ified by the true underlying multiple-indicator measure-
Item-Total Correlations
ment model, especially indicators that are highly related
The use of item-total correlations in the construction to their own factor, may correlate highly with the arbi-
of unidimensional scales has been long advocated. Fortrarily formed composite:
example, Nunnally (1978, p. 274) first establishes the
need for unidimensionality by stating, "Items within a(4) PxC = PxkPkc-
measure are useful only to the extent that they share a As a specific example, consider falsely combin
common core-the attribute which is to be measured," 10 items whose correlations appear in Table 1 into
but then writes, ". . . the items that correlate most highlygle scale. For these data, the correlations of eac

Table 1
AN ILLUSTRATIVE CORRELATION MATRIX

Item 1 2 3 4 5 6 7 8 9 10
1 1.00 .61 .61 .61 .61 .18 .10 .18 .18 .18
2 .61 1.00 .61 .61 .61 .18 .48 .18 .18 .18
3 .61 .61 1.00 .61 .61 .18 .10 .18 .18 .18
4 .61 .61 .61 1.00 .61 .10 .01 .10 .48 .10
5 .61 .61 .61 .61 1.00 .18 .10 .18 .18 .18

6 .18 .18 .18 .10 .18 1.00 .61 .61 .61 .61
7 .10 .48 .10 .01 .10 .61 1.00 .61 .61 .61
8 .18 .18 .18 .10 .18 .61 .61 1.00 .61 .61
9 .18 .18 .18 .48 .18 .61 .61 .61 1.00 .61
10 .18 .18 .18 .10 .18 .61 .61 .61 .61 1.00

This content downloaded from

182.72.243.45 on Fri, 01 Mar 2024 06:14:07 +00:00
All use subject to https://about.jstor.org/terms
 UPDATED PARADIGM FOR SCALE DEVELOPMENT 189

with the 10-item total score gradient range from

after the second .64 to
factor. Accordingly, two as
prin-hig
as .71. Even with the correction cipal-axis factorsforwere the
extracted by iterating for
overlap of com-
ite
and total score (Nunnally munalities. 1978), The the
oblique (promax) rotation of the result-
correlations rang
from .58 to .67. Thus item-total correlations in the .60's ing two factors also shows a clean two-factor solution:3
can be obtained from the analysis of a scale defined byall loadings of the first five items on the first factor range
two sets of five items with an average intercorrelation ofin magnitude from .74 to .80, as do the loadings of the
only .19! second five items on the second factor, whereas the
magnitude of all other loadings is less than .10. Thus
Exploratory Factor Analysis the substantive researcher would be likely to form two
Exploratory factor analysis is a useful scale develop- 5-item scales on the basis of this exploratory analysis.
ment technique for reducing a large number of indicators However, because of the lack of external consistency
to a more manageable set. It is particularly useful as a for items 2 and 7, these 5-item scales do not conform to
preliminary analysis in the absence of sufficiently de- a multiple-indicator measurement model and, accord-
tailed theory about the relations of the indicators to the ingly, a confirmatory factor analysis using the LISREL
underlying constructs. As Churchill (1979, p. 69) wrote, program (J6reskog and S6rbom 1984) demonstrates a
". .. factor analysis can indeed be used to suggest di- decided lack of fit for the two-factor multiple-indicator
mensions, and the marketing literature is replete with ar- model of all 10 items.4 The goodness-of-fit index (GFI)
ticles reporting such use." A typical usage of exploratory value is .79 and the root mean square residual (RMR)
factor analysis within this context is to factor an overall value is .067. Moreover, two of the normalized residuals
set of items and then construct scales on the basis of the are 9.34 and seven additional normalized residuals are
resulting factor loadings. Scales are formed by assigning greater than 2.5.
to the same scale the items that load at least moderately The pattern of normalized residuals from the LISREL
on the same factor (e.g., .4) and do not load as highly analysis clearly indicates the need for a respecification:
on other factors. all the normalized residuals for items 4 and 9 with every
A primary conceptual difference between exploratory other item are greater than 2.5. The next largest nor-
and confirmatory analyses, though, is that the explora- malized residual for any other item is only .10. Unlike
tory analysis typically does not provide an explicit test the information provided by the exploratory analysis, this
of unidimensionality as defined by the application of pattern strongly suggests that each of these two items is
equation 1 to each indicator. Factors in an exploratory not tapping a single underlying construct as are the other
analysis do not correspond directly to the constructs rep- four items in their respective scales.5 In practice the
resented by each set of indicators because each factor measurement model would be respecified by eliminating
from an exploratory analysis is defined as a weighted these two items. Performing this respecification results
sum of all observed variables in the analysis. For ex- in a model that provides perfect goodness of fit, dem-
ample, even if the factors are orthogonal, the resulting onstrating that unidimensional measurement has been
scales generally are correlated. The construction of scales achieved.
from an analysis of the size of the factor loadings does Without specifically referring to confirmatory factor
not provide an evaluation of the unidimensionality of these analysis, Churchill (1979, p. 69) called for more theory
scales, as would be accomplished by a confirmatory fac- testing with factor analysis: "Though this application
tor analysis in which each factor is antecedent to a mu- [exploratory factor analysis] may be satisfactory during
tually exclusive subset of the indicators. The more rig- the early stages of research on a construct, the use of
orous specification that is required for a confirmatory factor analysis in a confirmatory fashion would seem better
factor analysis of a multiple-indicator measurement model, at later stages." The subsequent refinement of confir-
in turn, affords a more rigorous evaluation of unidimen- matory factor analysis programs (Bentler 1985; J6reskog
sionality according to the constraints imposed by internal
and external consistency.
Thus, exploratory factor analysis can be a useful pre-
3An oblique rotation was used because it more accurately reflects
liminary technique for scale construction but, if the def- the underlying structure of the data than that provided by the more
inition of unidimensionality in equation 1 is used, a sub- restrictive orthogonal solution. However, for completeness, an or-
sequent confirmatory analysis would be needed to thogonal rotation (varimax) also was performed. The conclusion that
evaluate, and likely refine, the resulting scales. As an the exploratory factor analysis supports the hypothesis of two 5-item
scales remained unchanged. The only practical difference in the pat-
illustration, consider the first three eigenvalues extracted
tern of loadings between the oblique and orthogonal rotations was
from the correlation matrix in Table 1: 4.33, 2.54, and that, in the orthogonal solution, the highest loading of an item on the
.77, respectively. Unlike the results obtained from using factor corresponding to the other scale was .21 instead of less than
.10.
item-total correlations, the resulting exploratory solution
does show a clear two-factor solution. Only two eigen- 4For purposes of estimation, a sample size of 1000 was assumed.
5This comparative example underscores the need to include all pos-
values are larger than 1.0 and according to the scree test ited scales within a single analysis so that an assessment of external
(Cattell 1978) the differences between successive eigen- consistency can be made. Separate confirmatory factor analyses of
values are 1.78, 1.77, and .15, indicating a steep each scale cannot provide this assessment.

This content downloaded from

182.72.243.45 on Fri, 01 Mar 2024 06:14:07 +00:00
All use subject to https://about.jstor.org/terms
 190 JOURNAL OF MARKETING RESEARCH, MAY 1988

The 10-item correlation

and S6rbom 1984) with the growing acceptancematrix in Table
of 1thisprovides an
methodology provides an almost
example oftailormade means
the distinction between reliabilityof
and di-
accomplishing this testing. mensionality. Each item in the matrix indicates at least
Coefficient Alpha and Reliability
one of two item
within-scale distinct factors of
correlations (k1.61
and k2), resulting
relative in high
to the mean
Unidimensionality alone is not sufficient
between-scale to ensure
item correlations of only .19, the
yet coeffi-
usefulness of a scale. According
cient alphatofor the scale
the 10 items develop-
considered as a single scale
ment paradigm advocated here, the.85.
is a respectable reliability
After the refinement ofof thethe initial
composite score should be assessed after
10-item scale with unidimension-
a confirmatory factor analysis, coef-
ality has been acceptably established.
ficient alpha for eachEven a perfectly
4-item scale is .86, a slightly higher
unidimensional (and otherwise valueconstruct valid)
despite each scale having scale
less thanwould
half as many
be of little or no practical use items.
if the resultant composite
score were determined primarily by inmeasurement
Moreover, error,
the computation of coefficient alpha, one
with the values of the scores widely fluctuating
assumes that (1) over
the items already form re-
a unidimensional
peated measurements. set and (2) the items have equal reliabilities (Nunnally
The three types of reliability coefficients
1978). Computing inunequal
alpha for items with Cron- reliabili-
bach's (194-7) classification are the coefficient of stabil-
ties will lead to the underestimation of the reliability of
ity, the coefficient of equivalence, and
the composite the
score coefficient
(cf. Smith 1974). Beginningof
with
stability and equivalence. Both
Mosierthe
(1943),coefficient
the unbiased estimateof stabil-
of this reliability
ity and the coefficient of stability
for the general and
case for equivalence in-
both weighted and unweighted
volve administering forms of the scale on two or more
composites has been given in a variety of forms.6 J6reskog
different occasions. Only the coefficient
(1971) of for
has provided an expression equivalence
reliability that does
can be estimated from the administration of the scale at
not assume equal item reliabilities within the context of
a single time; consequently it is the reliability coefficientconfirmatory factor analysis. For a single unweighted
most often used by marketing researchers (Peter 1979).
composite for standardized measures, A, this expression
Coefficient alpha (Cronbach 1951), the most widelycan be given as
used coefficient of equivalence, sometimes has been
misinterpreted as an index of unidimensionality rather
than reliability (cf. Danes and Mann 1984; Green, Lis-
sitz, and Mulaik 1977; McDonald 1981). Consistent with (6) () = 2
this misinterpretation, one widely employed practice for
developing scales has been to select items that maximize
reliability (e.g., Churchill 1979). This practice has been
encouraged by computer packages such as SPSSx, which where Kh represents t
has a specialized routine for accomplishing it. However, sponding factor. Beca
unidimensionality and reliability are distinct concepts all have the same inte
(Green, Lissitz, and Mulaik 1977). This distinction be- values for alpha and J
tween dimensionality and reliability can be made mean- identical. It is interes
ingfully in the following way. The dimensionality of a item scale the two valu
scale can be evaluated by examining the patterning of the fourth decimal poi
its component indicator correlations, whereas the reli- of items on the scale
ability of a scale is determined by the number of itemsliabilities are very disc
that define the scale and the reliabilities of those items. composite reliability b
In particular, as the following equation demonstrates, tical consequence.
coefficient alpha is an application of the Spearman-Brown
formula of classical test theory (Lord and Novick 1968),
the formula also used in the computation of split-half 6Mosier's formula accounts for the items' reliabilities as well as
reliability. The difference is that instead of two splits,
their intercorrelations, as can be seen in the following formula.
the number of splits for coefficient alpha is equal to the
EE p(x,x,) + E p(xx )
number of items on the scale (p): jk i j

ZE
j#k
p(xjx,) + p
(5) ? =
1 + (p - 1)F Heise and Bohrnstedt (1970) developed the same formula from a path
where f represents the average off-diagonal correlation. analytic context in which they proposed using factor analysis to es-
Hence, regardless of the dimensionality of the scale, its timate the item communalities and then, in turn, using these com-
munality estimates as approximations of the item reliabilities. The re-
reliability tends to increase as the average off-diagonal sulting expression was called coefficient omega, which also will
item correlation increases and/or the number of items underestimate reliability to the extent that item-specific variance is
increases. nonzero, though the underestimation will not be as large.

This content downloaded from

182.72.243.45 on Fri, 01 Mar 2024 06:14:07 +00:00
All use subject to https://about.jstor.org/terms
 UPDATED PARADIGM FOR SCALE DEVELOPMENT 191

SUBSTANTIVE CONSIDERATIONS factor analysis of the individual measures as specif

by a multiple-indicator measurement model. Coefficie
One consequence of the paradigm for scale develop-
ment outlined here is that the isolation of unidimensional alpha is important in the assessment of reliability, bu
scales with a confirmatory factor analysis of a multiple-
does not assess dimensionality. Though item-total c
indicator measurement model tends to delineate rather relations and exploratory factor analysis can provide u
specific domains of content. As an example, consider ful
an preliminary analyses, particularly in the absenc
sufficiently detailed theory, they do not directly asse
analysis of the Likert measure of Machiavellianism, the
Mach IV scale, which originally had been constructed unidimensionality. The reason is that a confirmatory
on the basis of the item-total correlational method. A tor analysis makes possible an assessment of the intern
confirmatory factor analysis revealed that the Machconsistency
IV and external consistency criteria of un
measure was not, itself, unidimensional though it com- mensionality implied by the multiple-indicator measu
ment model.
prises unidimensional subscales (Hunter, Gerbing, and
Boster 1982). One of the resulting scales, for example,Following the paradigm of scale development outlined
contained items pertaining only to "duplicity" and an- here, after the unidimensionality of a set of scales has
been acceptably established, one would assess its reli-
other scale contained items pertaining only to "flattery
of important others." Though these scales represented ability. Even a perfectly unidimensional scale will not
be useful in practice if the resultant scale score has un-
narrow content domains, because of their nature they were
acceptably low reliability. Because most measures in
themselves of substantive interest, as shown by their in-
corporation into a subsequent structural equation modelmarketing are administered at a single point in time,
(cf. Hunter, Gerbing, and Boster 1982). coefficient alpha or some other coefficient of equiva-
In some cases, however, the items are unidimensional lence reliability would probably be used for this assess-
ment.
with respect to a single common construct, but that con-
struct is itself not of general interest because the result-The goal of most research projects is not just to de-
ing content domain from which the items were sampled velop unidimensional and reliable measurement scales,
is too restrictive. An alternative is to embed the unidi- but to build and test theory. Essential to this undertaking
is the assessment of construct validity. A construct
mensional scales, as indicators themselves, within a higher
order factor structure (cf. Gerbing and Anderson 1984). achieves its meaning in two ways (Anderson 1987;
The concept of multiple-indicator measurement can be Cronbach and Meehl 1955): (1) through observed indi-
preserved, but the level of analysis shifts so that a morecators for which it is posited to be causally antecedent
broadly defined construct enters into a structural model(and through observed measures for which it is not) and
or other application. An advantage of this approach (2) is through the set of relationships of the construct with
other
that the first-order factors, estimated with multiple mea- constructs as specified by some theory (the nomo-
sures, have their estimated construct correlations im-logical network). Unidimensionality, then, is necessary
but not sufficient for construct validity. Not only should
plicitly corrected for attenuation due to measurement er-
ror (cf. Howell 1987). The primary analysis then shifts all the indicators that define a scale provide estimates of
exactly one factor, but the meaning of the underlying
from an analysis of individual items to an analysis of
factor should correspond to the construct of interest.
more reliable versions of those items, the first-order con-
structs. An example of developing such second-order The nomological network can be explored within the
context of the full structural equation model. One means
constructs is found in the analysis of impulsivity by
Gerbing, Ahadi, and Patton (1987).7 for accomplishing this is the approach developed by An-
derson and Gerbing (1988) that allows an assessment of
SUMMARY AND CONCLUSION nomological validity that is asymptotically independent
of the assessment of the measurement model. It is called
Contributing to the tradition of articles by Churchill
(1979) and Peter (1979, 1981), we outline ana updated
"two-step" approach because the measurement model
first is developed and evaluated separately from the full
paradigm for scale development that incorporates a more
structural
recent methodological development: confirmatory factor equation model that simultaneously models
measurement and structural relations. The measurement
analysis. In doing so, we attempt to provide a better un-
derstanding of the concept of unidimensional model in conjunction with the structural model makes
measure-
possible
ment and the ways in which it can be assessed and, in a comprehensive confirmatory assessment of
particular, to demonstrate that an explicit evaluation ofvalidity (Bentler 1978). Hence, the assessment
construct
of unidimensionality provided by a confirmatory factor
unidimensionality is accomplished with a confirmatory
analysis represents but a first step in the establishment
of meaning for the estimated factors.

7This analysis of impulsivity measures is a substantive example of

the scale development paradigm outlined here. Item-total correlations REFERENCES
and exploratory factor analysis were used to reduce a large number
of items to a more manageable set and to provide preliminary scales, Anderson, James C. (1987), "An Approach for Confirmat
which then were refined with a confirmatory factor analysis. Measurement and Structural Equation Modeling of Or

This content downloaded from

182.72.243.45 on Fri, 01 Mar 2024 06:14:07 +00:00
All use subject to https://about.jstor.org/terms
 192 JOURNAL OF MARKETING RESEARCH, MAY 1988

nizational Properties," Management Harman, Harry H. (1976), Modern

Science, Factor Analysis,
33 (April), 3rd ed.
525-
41. Chicago: The University of Chicago Press.
and David W. Gerbing (1982), "Some Methods for Hattie, John (1985), "Methodology Review: Assessing Uni-
Respecifying Measurement Models to Obtain Unidimen- dimensionality of Tests and Items," Applied Psychological
sional Construct Measurement," Journal of Marketing Re- Measurement, 9 (June), 139-64.
search, 19 (November), 453-60. Heise, David R. and George W. Bohmstedt (1970), "Validity,
- and - (1988), "Structural Equation Modeling in Invalidity, and Reliability," in Sociological Methodology,
Practice: A Review and Recommended Two-Step Ap- Edgar F. Borgatta and George W. Bohrnstedt, eds. San
proach," Psychological Bulletin, forthcoming. Francisco: Jossey-Bass, Inc., 104-29.
Bentler, Peter M. (1978), "The Interdependence of Theory, Howell, Roy D. (1987), "Covariance Structure Modeling and
Methodology, and Empirical Data: Causal Modeling as an Measurement Issues: A Note on 'Interrelations Among a
Approach to Construct Validation," in Longitudinal Drug Channel Entity's Power Sources'," Journal of Marketing
Research, D. B. Kandel, ed. New York: John Wiley & Sons,Research, 24 (February), 119-26.
Inc., 267-302. Hunter, John E. and David W. Gerbing (1982), "Unidimen-
(1985), "Theory and Implementation of EQS, A sional Measurement, Second Order Factor Analysis, and
Structural Equation Program." Los Angeles: BMDP Statis- Causal Models," in Research in Organizational Behavior,
tical Software. Vol. 4, B. M. Staw and L. L. Cummings, eds. Greenwich,
Burt, Ronald S. (1973), "Confirmatory Factor-Analytic Struc- CT: JAI Press, Inc., 267-99.
tures and the Theory Construction Process," Sociological 9 , and Frank J. Boster (1982), "Machiavellian
Methods & Research, 2 (November), 131-90. Beliefs and Personality: The Construct Invalidity of the Ma-
Cattell, Raymond B. (1978), The Scientific Use of Factor chiavellian Dimension," Journal of Personality and Social
Analysis in Behavioral and Life Sciences. New York: Plenum Psychology, 43 (December), 1293-1305.
Press. J6reskog, Karl G. (1971), "Statistical Analysis of Sets of Con-
Churchill, Gilbert A., Jr. (1979), "A Paradigm for Developing generic Tests," Psychometrika, 36 (June), 109-33.
Better Measures of Marketing Constructs," Journal of Mar- - and Dag Sorbom (1984), LISREL VI: Analysis of Lin-
keting Research, 16 (February), 64-73. ear Structural Relationships by the Method of Maximum
Cronbach, Lee J. (1947), "Test 'Reliability': Its Meaning and Likelihood. Chicago: National Educational Resources, Inc.
Determination," Psychometrika, 12 (March), 1-16. Kline, Paul (1983), Personality: Measurement and Theory. New
(1951), "Coefficient Alpha and the Internal Structure York: St. Martin's Press.
of Tests," Psychometrika, 16 (September), 297-334. Lord, Fred M. and Melvin R. Novick (1968), Statistical The-
- and Paul E. Meehl (1955), "Construct Validity in Psy- ories of Mental Test Scores. Reading, MA: Addison-Wesley
chological Tests," Psychological Bulletin, 52 (July), 281- Publishing Company, Inc.
302. McDonald, Roderick P. (1981), "The Dimensionality of Tests
Danes, Jeffery E. and 0. Karl Mann (1984), "Unidimensional and Items," British Journal of Mathematical and Statistical
Measurement and Structural Equation Models with Latent Psychology, 34 (May), 100-17.
Variables," Journal of Business Research, 12 (September), Mosier, Charles I. (1943), "On the Reliability of a Weighted
337-52. Composite," Psychometrika, 8 (September), 161-8.
Gerbing, David W. and James C. Anderson (1984), "OnNunnally, the Jum C. (1978), Psychometric Theory, 2nd ed. New
Meaning of Within-Factor Correlated Measurement Errors," York: McGraw-Hill Book Company.
Journal of Consumer Research, 11 (June), 572-80. Peter, J. Paul (1979), "Reliability: A Review of Psychometric
- , Stephan A. Ahadi, and Jim H. Patton (1987), "To- Basics and Recent Marketing Practices," Journal of Mar-
ward a Conceptualization of Impulsivity: Components Across keting Research, 16 (February), 6-17.
the Behavioral and Self-Report Domains," Multivariate- Be- (1981), "Construct Validity: A Review of Basic Issues
havioral Research, 22 (July), 357-79. and Marketing Practices," Journal of Marketing Research,
Green, Samuel B., Robert W. Lissitz, and Stanley A. Mulaik 18 (May), 133-45.
(1977), "Limitations of Coefficient Alpha as an Index Smith,
of Kent W. (1974), "On Estimating the Reliability of
Test Unidimensionality," Educational and Psychological Composite Indexes Through Factor Analysis," Sociological
Measurement, 37 (October), 827-38. Methods and Research, 2 (May), 485-510.

This content downloaded from

182.72.243.45 on Fri, 01 Mar 2024 06:14:07 +00:00
All use subject to https://about.jstor.org/terms