An Alternative Procedure for

Assessing
Convergent and Discriminant Validity
Donald R. Lehmann
Columbia University

This paper presents a simple procedure for estab- dent variable. Illustrations indicate that the method
lishing convergent and discriminant validity. The provides reasonable answers. Index terms: conver-
method uses ordinary least-squares regression (OLS) gent validity, discriminant validity, multiple regres-
with the correlations between measures as the depen- sion.

Establishing the validity of measures is a of research. Essential to establishing validity

major focus
with multi-item measures are notions of convergent and discriminant validity (Anastasi, 1968; Bohmstedt,
1970; I~u~nally9 1978). Many approaches to convergent and discriminant validity assessment are derived
from the multitrait-multimethod approach (Campbell & Fiske, 1959). This paper presents an alternative
to the LISREL-based nested model procedure used by Bagozzi (1978) and Widaman (1985).

Background
The most common approach to establishing convergent and discriminant validity is to demonstrate
that multiple measures of a construct are ( 1 ) related, and (2) more related to each other than to measures
of other constructs, even when the two measurement methods are similar (Carr~pbell ~ p’iske9 1959). The
second criterion suggests that a construct should have more impact on a relation than does a common
measurement method.
The assessment of validity typically has focused on the Pearson product-moment correlation coef-
ficient. This requires both intervally scaled data and linear association, although a monotonic relation
will often suffice. Although it is relatively easy to account for non-linearities in the relation (Lehmann,
1974), here the focus is on the linear correlation coefficient as the measure of association.
Convergent and discriminant validity require that multiple measures of a construct relate highly to
each other and less highly to measures of other constructs. Consider the hypothetical two-construct, three-
measure-per-construct model in Figure 1. In the case of independent constructs (Figure 1 a), correlations
of variable pairs (1,2), (1,3), and (2,3) should be high (Block I in Figure 2). Similarly, correlations of
variable pairs (4,5), (4,6), and (5,6) should be high (Block III in Figure 2). By contrast, correlations of
variable pairs in Block II such as ( 1,4) should be zero.
When some measures share a common measurement method (Figure then the correlation between
them should be higher than among variables that do not use a common method. Thus, in Figure lb if F,

411

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 412

Figure 1
Two Construct Models

and Y, share a common measurement base, then rJ4 should be potentially higher than, say, r,,. However,
even when variables 4 and 5 are maximally different measures, rl5 should be greater than r14. If a variable
.A shares more variance with its construct than another variable .~ then the correlation between A and (1)
all other indicators of the construct and (2) any indicator of a related construct should be larger.
Examinations of convergent and discriminant validity have followed three major approaches. The
first approach involves examining some form of the multitrait-multimethod matrix (MTMM) approach to
analyze the correlation matrix (i.e., Sullivan & Feldman, 1979) for internal and external consistency
(Anderson & Gerbing, 1982).
Kalleberg and Klugel (1975) presented a path-analytic decomposition of the MTMM matrix, which
they found inadequate due to assumptions that traits and methods are uncorrelated and that methods are
minimally correlated with each other (Jackson, 1969). They also found the method &dquo;basically qualitative
in nature&dquo; and turned to confirmatory factor analysis to assess validity. Kavanaugh, MacKinney, and
Wolins (1971) used analysis of variance to decompose the observed data based on person, trait, and
method variables.
A second approach to assessing validity is factor analysis. When exploratory factor analysis is used,
the results are often largely-but not completely-consistent with prior theory. For example, if there are
seven constructs with three indicators of each, a principal components analysis must produce seven

eigenvalues greater than 1.0 (or possibly 10 if each method also had an effect) and the rotated factor
matrix must have seven factors, each with three indicators loading on it. Because this almost never occurs,
the researcher can only conclude that the results &dquo;generally support the prior theory.&dquo; Thus, because
exploratory factor analysis is imprecise and orthogonal factor analysis is not appropriate for correlated
constructs, confirmatory factor analysis often is used. This approach (Anderson & Gerbing, 1982; 3®resl~og9

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 413

Figure 2
Indicator Variable Correlation Matrix

1971; Long, 1983) is reasonable and widely recommended (Schmitt & Stults, 1986) but requires an
iterative procedure to arrive at a solution.
A number of factor-analytic procedures have been suggested for assessing convergent and discriminant
validity. In a comparison of several of these, Schmitt, Coyle, and Saari ( 1977) suggested that the approach
of Jackson (1969, 1975) has the advantage of requiring that the researcher explicitly state assumptions
about the hypothesized structure.
The third approach used to assess convergent and discriminant validity jointly examines both the
measurement model (convergent and discriminant validity) and the structural model (nomological validity),
using LISREL (Bagozzi, 1980; J6reskog & Sorbom, 1979) or PLS (Fomell & Larker, 1981; Wold, 1980).
This approach requires some prior theory, so that the relations among the constructs must be specified
exactly. Moreover, because the measurement and structural models are estimated simultaneously, errors
in specification of the structural model may affect the results for the measurement model (Burt, 1976;
Kumar & Dillon, 1986). Consequently, a preliminary analysis of convergent and discriminant validity
seems desirable before attempting to estimate a full structural equations model.

Recent attempts to demonstrate discriminant validity have focused on using LISREL on a series of
nested models, as recommended by Widaman (1985). The procedure begins with a parsimonious null
model assuming that all measures are identical and gradually relaxes the assumption, allowing multiple
traits (constructs) and multiple methods to exist. When the model that allows multiple traits fits better
than the model that assumes a single trait, discriminant validity is said to exist.
Two possible weaknesses exist in this procedure. First, it requires the estimation of several LISREL
models and hence is fairly computationally intense. Second, the procedure tests sets of coefficients
simultaneously. Thus, if in a three-construct model, construct A is distinct from constructs B and C, and
B and C are indistinguishable, the fit of the model allowing for separate traits will be significantly better
than the model assuming a single trait. This will indicate that discriminant validity exists whereas it does
not in the case of constructs B and C. Of course, adding separate traits one at a time eliminates this
problem, but with a large number of traits this procedure becomes rather unwieldy.
The method proposed here differs from all three of these general approaches, yet incorporates elements
of all. As in the first approach, it uses an analysis-of-variance procedure where the dependent variable

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 414

is the correlation between items, rather than the item score. In addition, it decomposes correlations based
on trait and method variables. As in the second approach, this method requires assumptions, but only
about ( 1 ) which items measure which construct and (2) which items share a common measurement method.
In contrast to the third approach, this method, although easily adaptable to the nested-model approach,
produces a single-stage least-squares estimate of a &dquo;null&dquo; model (separate traits and methods), which
allows examination of discriminant validity from a single output.

Proposed Method

The approach followed here &dquo;explains&dquo; correlations using a dummy variable regression analysis in
terms of the measure-to-construct relations, the construct-to-construct relations, and the impact of common
measurement bases. correlation rij, where is a measure of construct p and j is a
Every measure of
construct q, is assumed to be affected by several factors:
1. The closeness of the relation between constructs p and q.
2. Any common method that i and j might share.
3. The relation of each measure to its construct.
The procedure estimates the impact of each of these on correlations, thus allowing an explicit test of the
desired conditions.
Decomposing the correlations is directly related to the analysis-of-variance approach (see the Ap-
pendix). In essence, the correlations are decomposed as follows:

where ~p9 = 1 if p and q the two constructs, 0 otherwise;

are

Zmn =
1 if m the two methods, 0 otherwise; and
and n are

Z~=lift=~orj=~,0 otherwise.
Due to collinearity among the dummy variables, it is necessary to estimate a reduced form of Equation
1. Two models are of special interest. The first assumes that all the Bks equal 0 (which implies all are
equally good measures of their respective constructs):

This strong assumption allows estimation of,for example, correlated methods of measurement (B_,s). A
more reasonable initial assumption allows for the unequal quality of variables and deletes the variables
relating to two measures of the same construct (Z,,s) and the variables relating to measures by different
methods (Z,,,,,s). This model becomes

The interpretation of the coefficients in Equation 3 is as follows (assuming deletion of one of the Z,
variables).

Coefficients
Constant. This represents the average correlation of the (dummy) variable omitted from the equation
(in order to permit estimation) with other measures of the same construct.
Individual variable coefficients (Bis). These indicate how the average correlation for a particular
variable differs from the omitted (dummy) variable’s average correlation. Thus, B¡ = .2 means that
correlations involving variable i average .2 larger than those involving the omitted variable. Ignoring

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 415

common method effects, the correlation among two variables which measure the same construct is
predicted to be Bo + B; + Bj. Therefore, in order for a construct p to have convergent validity, all
Bo + Bi + Bj must be greater than 0 for all (ij) pairs, where z and j are measures of the particular
construct p.
Cross-construct coefficients (Bps). These coefficients indicate how the average correlation between
a measure of construct p and a measure of construct q differs from the average correlation between two
measures of the same construct (i.e., two measures of p, two measures of q, ... ). In order to establish
discriminant validity between constructs p and q, ~Pq should be significantly less than 0. Further, correlated
constructs exist if Bo + ~; + ~~ + B pq is greater than 0 for i (a measure of construct p) and j (a measure
of construct q).
Common method coefficients (Bmms). These coefficients measure the difference between the average
correlation of two measures using method m and two measures using different methods. When ~&dquo;,m is
significantly positive, this indicates a significant method effect for method rra.
One useful feature of Equation 3 is that it provides an estimate of the magnitude of the common
method effect. This magnitude is interesting both in its own right and in comparison to the magnitude
of the effects of the traits. If the method effects are larger than the trait effects, then a problem may exist
in isolating the effects of the traits, and the strength of the constructs is relatively weak.
Widaman (1985) suggested that the procedures of Campbell and Fiske (1959) have three main
problems: (1) the correlations are not independent, (2) estimates of the trait-related and method-related
variance are not obtainable, and (3) the procedures are distorted when the variables differ in reliability.
Here the dependence of the correlations is explicitly modeled in Equation 3. The proposed procedure
specifically separates trait-related and method-related sources of the correlations (the Bpqs and Bmns), which
is one of the advantages of using LISREL for analyzing validity. Furthermore, Equation 3 explicitly allows
for unequal reliabilities (Bks) and provides a framework in which to test the constraint of equal reliabilities.

Estimation
The statistical estimation of the model depends on the distribution of e,~. The variance of a correlation,
and hence eij, is approximately (1 - r~)/(~ - 2). This suggests that a weighted least-squares (wLs) pro-
cedure will be more efficient than ordinary least-squares (OLS), and that WLS will tend to weight the
larger correlations (typically between measures of the same construct) more heavily. The co variances of
the eiis are assumed to be 0. Note that this does not imply that the correlations themselves are unrelated.
If all correlations come from the same dataset, constraints exist on r23 if r2 and r13 are known, although
unless r°,z and r°13 are large the constraint is not very tight. Here it is assumed that the deviations in the
correlations from some &dquo;true&dquo; value are uncorrelated.
An alternative approach is to use Fisher’s r to z transformation on the correlations in Equation 3 to
normalize the variances. This method is not recommended because the coefficients are difficult to interpret
in terms of the magnitude of correlations, and because this procedure minimizes errors in predicting Z,
which does not necessarily minimize error in predicting r.

Variable Coding
Coding the variables for the analysis of the model in Equation 3 is straightforward. For each correlation
(i,j) the dummy variables Di and D, are coded as 1, and the rest of the dummy variables are coded as
0. If variable i is a measure of construct p and j is a measure of construct q, then Dpq is coded as 1 and

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 416

the other &dquo;cross-construct&dquo; dummy variables are coded as 0. Finally, if variables and j share method
~c, the D,~~ is coded as 1 and the rest of the common method dummy variables are coded as 0.
As an example of coding, consider the variables analyzed by Lawler ( 1967) which consisted of nine
variables representing all combinations of ratings by supervisors, peers, and persons themselves (methods)
of ability, quality, and effort on the job (traits):
1. Supervisor rating of quality (At),
2. Peer rating of quality (A2),
3. Self rating of quality (A3),
4. Supervisor rating of ability (Bl),
5. Peer rating of ability (B2),
6. Self rating of ability (B3),
7. Supervisor rating of effort (ci),
8. Peer rating of effort (C2), and
9. Self rating of effort (C3).
These nine variables produced 36 pairwise correlations, which are shown, along with the coding, in Table 1.

~~ pl~ Applications
Discriminant Validity
This example examines discriminant validity using the superior, peer, and self ratings of performance
on three criteria (quality, ability, and effort) presented by Lawler (1967). These correlations were analyzed
using Equation 3 and OLS.
The data in Table 2 lead to some obvious conclusions:
1. Bo + ~; + ~~ > 0 for all i, j in the same construct, indicating convergent validity, though only weakly
when self ratings are involved. (The negative values for the self-rating measures indicate that these
correlations tend to be somewhat smaller.)
2. There is clear discriminant validity between ability and effort (B = ―.22) and significant discriminant
validity between quality and effort (~ = ―.11). There is no discriminant validity between quality
and ability (B = - .05, not significantly different from 0).
3. The correlations involving supervisor or peer ratings are larger than those involving self ratings in
general (Bis range from .27 to .35 vs. - .07 and - .08 for self ratings).
4. Bo + + Bj + ~p9 is generally positive, indicating correlated constructs, although again the self
ratings tend to be different from the others.
5. There is massive method bias in the self ratings (~ _ .64), a smaller but significant one in the
supervisor ratings (~ _ .13), and a r~®nsi~nific~a~t ~ in peer ratings (.10). This suggests that self
ratings are consistent across attributes (ability, quality, effort).
These results mirror those reported by Widaman (1985). His chosen model (Model 3C in his paper)
showed a strong correlation between the supervisor and peer method factors (related to the higher cor-
relations in their person variables reported here). He argued that effort is the most distinct trait (here
effort was discriminantly different from both quality and ability), but that the traits were not easily
distinguishable (here indicated by the relatively small coefficients of the trait variables). Thus, the one-
step procedure recommended here produces essentially the same interpretation as the nested model testing
procedure used by Widaman (1985).

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 DONALD
417

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 418

Table 2
B Weights from the Lawler
Data Analysis2(R 2 _.92,
Adjusted ~ - .87)

the Three-Component of Attitude

As a complex example of the procedure, an assessment of the convergent and discriminant
more

validity three-component attitude model was made. Specifically, the data of Ostrom (1969) and
of the
Kothandapani (1971), which have been analyzed using covariance structure methods (Bagozzi, 1978;
Widaman, 1985), were analyzed using a form of Equation 3.
Ostrom data. The model used for the Ostrom data was

This model provides a direct test for method variance (if B12, B14, and ~,5 are positive and
significant, this indicates common method variance). It also provides a direct test of discriminant validity
(if ~,69 B 17, and B 18 are significantly negative, there is discriminant validity). Convergent validity requires
that ~o and all ~o + j6, be significantly positive; in other words, that correlations between measures of
the same construct should be positive. A stronger test is possible if a certain minimum level of correlation
of measures to their construct (e.g., rs above .6) is desired. In such situations, Bo and B, + Bi must be
above that cutoff level.
The results from an OLS regression analysis of the model are shown in Table 3, which used 1-0
coding for the dummy variables. The conclusions that can be drawn from the Ostrom ( 1969) data are as
follows:
1. Measure-to-construct relations are remarkably similar for any pair of variables because Bis range only
fr®rn - .®~ to -i- .09.

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 419

2. Correlations involving Likert scales tend to be about I larger, suggesting that a stronger method
component exists in them.
3. A significant common method effect is evident for self-rating scales (~,5 =
.15) and a small but
effect exists for Likert scales (~,3=
.06).
4. The behavioral construct is significantly different from both the affective and cognitive component,
though only barely so. Moreover, the affective and cognitive components are not significantly different
at the .05 level, and the sizes of the discriminant variables ( - .03, - .04, - .02) suggest that the
constructs are highly related.
5. Because all ~o + ~; + Bj 5= .46, there is reasonable convergent validity.

I~®th~~adc~p~ni ~catca. The Kothandapani (1971) data were analyzed using the same version of Equa-
tion 3 that was used for the Ostrom data, except that the three components were Feeling, Belief, and
Intention instead of Affect, Behavior, and Cognition. The results (also shown in Table 3) differed in
several ways:
1. Correlations involving Likert and Guilford scales tend to be higher (Bs range from -.07 to + .02
for these vs. - . 19 to - . 09 for the Thurstone and Guttman scales).
2. A significant and substantial common method effect exists for all four methods. The Bs =
.23, .25,
.29, and .3I , indicating, for example, that two Thurstone scales on average are correlated .23 higher
than correlations of variables which share no common method.
3. All the constructs are significantly distinct, with Feeling (Affect) and Intention (Behavior) the most
distinct. Also, the coefficients of the discriminant variables are as large as the coefficients of the
method factors (~s = .22, - .26, - .42).

Table 3
B Weights and Si~~abf3c~~nce Levels (p) From OLS
Decomposition of Ostrom and Kothandapani Correlation Matrices

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 420

4. Because B, = .62 and all j8 o + B; + ~J : .33, there is convergent validity.

Use of Fisher’s r to z transformation is often recommended to normalize the errors. Even though the
transformation is problematic for correlations above .7 (which often occur in MTMM data), the data were
reanalyzed using the transformation. Specifically, the dependent variable used was 1/2 In [(1 + rij)l(l1 ---
r,j)]. All coefficients retained the same sign and the R2 values were comparable.
Overall, the Kothandapani (1971) data give much stronger evidence of discriminant validity than
the Ostrom (1969) data. Hence, this procedure seems to provide essentially the same interpretation as
the nested model tests of Widaman (1985), in contrast to that of Bagozzi (1978). Notice also that formal
nested model tests are also not only possible but relatively simple with the proposed method. However,
running the complete model makes them unnecessary.

Limitations
The method as proposed has two major limitations. First, because the model in Equation 3 is only
a linear approximation of the correlation generating process, the analysis contains some model (specifi-
cation) error which can obscure differences. Based on initial investigations of well-behaved datasets, this
appears to be a relatively minor problem.
Second, the tests do not explicitly take into account the sample size involved in the correlations.
The model which is estimated is thus:
MeasuredCorrelationi,j = True ~®rr~lati®~;.a + Sampling Err®ri,~ _ 2lBD + Model Error,, , (5)
where SBD is the model in Equation 3. Hence the model is essentially
True Correlation,, =XM)+c,, , , (6)
where ~1~; is a combination of sampling and model error.

This suggests that for a smaller sample size, sampling error will be larger. Hence, assuming that
sampling error and model error are uncorrelated, small samples will produce more error in the regression
estimates, leading to fewer significant coefficients and hence less evidence of convergent and discriminant
validity. Future developments may make it possible to explicitly include sample size in the analysis of
the correlations.

Appendix
of of to the
Analysis-of-Variance Decomposition of ~~~ Scores

Consider a P construct model where each construct is measured by the same M methods. The
observed score for person k on the ith measured variable which uses method m to assess construct p is
X;~ =
B,+B,+C,+D~+~+~+~+~+~ , , (A 1)
where Bo is the average score on the items,
Bk is the average score of person k minus the average score of all persons on all items (person
effect),
9

Cp is the average score of items on the pth construct minus the average score of all items (construct
effect),
9

D~ is the average score on the mth method minus the average score on all methods (method
effect), 9

Ipm is the interaction of the pth construct and the mth method,

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 421

hk is the interaction effect of construct p and person k,

~mk is the interaction effect of method rra and person k,
h~&dquo;~k is the interaction of the pth construct, the mth method, and the kth person, and
eik is the random component.
Rewriting, this gives

where 2’L is the mean value on item i. Similarly,

If two constructs p and q are correlated, fpk and f9k will be correlated. Similarly, if two methods are

correlated, gm~ and g&dquo;k will be correlated. Also,

the correlation between two measures is positively related to

1. Large construct correlation: Corr(p,~)(Var~)~(Var~)’~
2. Large method correlation: Corr(m,n)(Var gm)1I2(Var gn)1I2
3. Large construct-method correlation: Corr(pm,qn)(Var hpm)1I2(Var hqn)1I2
4. Small variable variances: Var ei, Var ej
A linear approximation of this is given by

where ~Pq = 1 if p and q are the constructs, 0 otherwise;

Z,n,, I=
if rra and n are the methods, 0 otherwise;

Zpmqn I=
if Xi is a measure of construct p by method m and Xj is a measure of construct q by
method n, 0 otherwise; and
Z~=lif~’~~orj=~,0 otherwise.
The number of parameters to be estimated in Equation Allis as follows:

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 422

For three constructs and three methods, this becomes

Because there are only (9 x 8)/2 36 correlations, the full model is under-identified. In order to
=

have an identified model, it is necessary to assume that the correlation of hn&dquo;,k and hy&dquo;~ 0. This reduces
=

the model to a manageable size. The assumption means that an individual’s response to a combination
of construct and method is explainable by a main-effects-only model. (Note that selected interactions
could be estimated if a strong a priori reason existed for examining them.)
The model in Equation All thus reduces to

Various nested versions of Equation A17 are also of interest. For example, if it is assumed that there
is correlation between the effects of methods m and rc (except the common method effect when m
no
_
n), the number of parameters reduces. Comparison of the fit of this model to the fuller model in
Equation A17 provides an approximate test of the reasonableness of this assumption.

References

Anastasi, A. (1968). Psychological testing (3rd ed.). Jackson, D. N. (1975). Multimethod factor analysis: A
New York: Macmillan. reformulation. Psychological Bulletin, 75, 421-423.
Anderson, J. C., & Gerbing, D. W. (1982). Some meth- Jöreskog, K. G. sets of
(1971). Statistical analysis of
ods for respecifying measurement models to obtain congeneric tests. 109-133.
Psychometrika, 36,
unidimensional construct measurement. Journal of Jöreskog, K. G., & Sörbom, D. (1979). Advances in
Marketing Research, 19, 453-460. factor analysis and structural equation models. Cam-
Bagozzi, R. P. (1978). The construct validity of affec- bridge MA: Abe Books.
tive, behavioral, and cognitive components of attitude Kalleberg, A. L., & Klugel, J. R. (1975). Analysis of
by analysis of covariance structures. Multivariate Be- the multitrait-multimethod matrix: Some limitations
havioral Research, 13, 9-31. and an alternative. Journal of Applied Psychology,
Bagozzi, R. P. (1980). Causal models in marketing. 60, 1-9.
New York: Wiley. Kavanaugh, M. J., MacKinney, A. C., & Wolins, L.
Bohrnstedt, G. W. (1970). Reliability and validity as- (1971). Issues in managerial performance: Multitrait-
sessment in attitude measurement. In G. F. Summers multimethod analyses of ratings. Psychological Bul-
(Ed.), Attitude measurement. Chicago: Rand-Mc- letin, 75, 34-39.
Nally. Kothandapani, V. (1971). Validation of feeling, belief,
Burt, R. S. (1976). Interpretation confounding of unob- and intention to act as three components of attitude
served variables in structural equation models. Soci- and their contribution to prediction of contraceptive
ological Methods and Research, 5, 33-52. behavior. Journal of Personality and Social Psychol-
Campbell, D. T., & Fiske, D. W. (1959). Convergent ogy, 19, 321-333.
and discriminant validity by the multitrait-multi- Kumar, A., & Dillon, W. (1986). The interaction of
method matrix. Psychological Bulletin, 56, 1-105. measurement and structure in simultaneous equation
Fomell, C., & Larker, D. F. (1981). Evaluating struc- models with unobservable variables. Unpublished
tural equation models with unobservable variables and working paper.
measurement error. Journal of Marketing Research, Lawler, E. E. (1967). The multitrait-multirater approach
18, 39-50. to measuring managerial performance. Journal of Ap-
Jackson, D. N. (1969). Multimethod factor analysis in plied Psychology, 51, 369-381.
the evaluation of convergent and discriminant validity. Lehmann, D. R. (1974). Some alternatives to linear fac-
Psychological Bulletin, 72, 30-49. tor analysis for variable grouping applied to buyer

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/
 423

behavior models. Journal of Marketing Research, 11, Sullivan, J. L., & Feldman, S. (1979). Multiple indi-
206-213. cators. Beverly Hills CA: Sage.
Long, J. S. (1983). Confirmatory factor analysis. Bev- Widaman, K. F. (1985). Hierarchically nested co-

erly Hills CA: Sage. variance structure models for multitrait-multimethod

Nunnally, J. C. (1978). Psychometric theory (2nd ed.). data. Applied Psychological Measurement, 9, 1-26.
New York: McGraw-Hill. Wold, H. (1980). Soft modeling: Intermediate between
Ostrom, T. M. (1969). The relationship between the traditional model building and data analysis. Multi-
affective, behavioral and cognitive components of at- variate Statistics, 6, 333-346.
titude. Journal of Experimental Social Psychology, 5,
12-30.
Schmitt, N., Coyle, B. W., & Saari, B. B. (1977). A
review and critique of analyses of multitrait-multi-
method matrices. Multivariate Behavioral Research, Author’s Address
12, 447-478.
Schmitt, N., & Stults, D. M. (1986). Methodology re- Send requests for reprints or further information to Don-
view : Analysis of multitrait-multimethod matrices. ald R. Lehmann, 507 Uris Hall, Columbia University,
Applied Psychological Measurement, 10, 1-22. New York NY 10027, U.S.A.

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227.

May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction
requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/