Documente Academic
Documente Profesional
Documente Cultură
A review of the literature suggests that few studies use formative indicator mea-
surement models, even though they should. Therefore, the purpose of this research
Thus, the underlying latent construct causes the observed the same way that Churchill (1979) did for reflective in-
variation in the measures (Bollen 1989; Nunnally 1978). dicator constructs.
This assumed direction of causalityfrom the latent vari- Therefore, the objectives of this article are to (a) develop
able to its measuresis conceptually appropriate in many a set of conceptual criteria that can be used to determine
instances, but not all. Indeed, it was recognized very early whether a construct should be modeled as having formative
on that, for some constructs, it makes more sense concep- or reflective indicators, (b) determine the extent of measure-
tually to view causality flowing from the measures to the ment model misspecification by comprehensively reviewing
construct, rather than vice versa (Bagozzi 1981, 1984; Blal- measurement model specifications in four top-tier marketing
ock 1964; Fornell and Bookstein 1982). For example, For- journals, (c) conduct a Monte Carlo simulation designed to
nell and Bookstein (1982, p. 441) noted that the unobserved examine the severity of the estimation bias due to measure-
constructs can be viewed either as underlying factors or as ment model misspecification, and (d) provide recommenda-
indices produced by the observable variables. That is, the tions for modeling formative indicator constructs. We begin
observed indicators can be treated as reflective or formative. with a review of the conceptual distinctions between reflective
Reflective indicators are typical of classical test theory and and formative measurement models. These distinctions then
factor analysis models; they are invoked in an attempt to are generalized to multidimensional second-order constructs.
account for observed variances or covariances. Formative Next, we document the use of reflective and formative models
indicators, in contrast, are not designed to account for ob- in the marketing literature and consider the appropriateness
served variables. . . . The choice between formative and of these specifications for the theoretical constructs repre-
FIGURE 1
purchase intention. Attitudes are generally viewed as pre- referred to as causal (Bollen and Lennox 1991) or formative Downloaded from http://jcr.oxfordjournals.org/ by guest on January 7, 2017
dispositions to respond in a consistently favorable or unfa- (Fornell and Bookstein 1982) indicators. Unlike the reflec-
vorable manner toward an object and are usually measured tive model, this model does not assume that the measures
on multi-item scales with endpoints such as good-bad, like- are all caused by a single underlying construct. Rather, it
dislike, and favorable-unfavorable; purchase intentions are assumes that the measures all have an impact on (or cause)
typically measured using subjective estimates of how likely- a single construct. That is, the direction of causality flows
unlikely, probable-improbable, and/or possible-impossible from the indicators to the latent construct, and the indicators,
future purchases are perceived to be (e.g., MacKenzie, Lutz, as a group, jointly determine the conceptual and empirical
and Belch 1986). meaning of the construct (see col. 2 in fig. 1).
In contrast, in the composite latent variable model, Because some have hypothesized that formative measures
changes in the measures are hypothesized to cause changes influencerather than are influenced bythe latent construct,
in the underlying construct. Thus, this models measures are they may be correlated, but the model does not assume or
202 JOURNAL OF CONSUMER RESEARCH
require this. Indeed, it would be entirely consistent for for- An example of a collection of measured variables used to
mative indicators to be completely uncorrelated. This might indicate a composite latent construct might be Singhs (1988)
be the case where a composite latent construct is represented construct called consumer complaint behaviors, which has
by mutually exclusive types of behavior. For example, some indicators such as the likelihood of complaining to the store
of the indicators used by Crosby and Stephens (1987) to manager, telling friends and relatives about a bad service
measure the personal contact of life insurance agents (I was experience, reporting the company to a consumer agency, or
contacted by my agent who wanted to make changes in this pursuing legal action against the company. In this case, a high
policy to better serve my needs; I was contacted by my likelihood of one particular behaviorsay, a complaint to a
agent who wanted to sell me more life insurance; I was store manager about poor servicewould influence the level
contacted by my agent who wanted to describe new types of of the latent construct, but would not necessarily have an
policies that had become available; and My agent explained effect on the other measures. One could also conceive of the
why it was a good idea to keep this whole-life policy in beliefs construct in the Fishbein and Ajzen (1975) model as
force), may be mutually exclusive. For example, an insur- having a similar formative measurement structure with each
ance agent might encourage a customer to keep the current individual belief x evaluation component as causing the over-
life insurance policy or try to sell the customer a different all belief construct. The belief construct would be more than
policy to replace it, but the agent would not do both. simply the sum of the belief by evaluation products because
Therefore, internal consistency reliability is not an ap- one may not have measured all salient beliefs.
propriate standard for evaluating the adequacy of the mea- The two types of measurement models have some simi-
Winklhofer 2001; Edwards and Bagozzi 2000), to the best for the formative indicator model, because the measures do
of our knowledge, no comprehensive list of criteria exists not necessarily capture the same aspects of the constructs
to help guide researchers who are struggling with this is- domain and are therefore not necessarily interchangeable,
sue. The criteria are summarized in table 1 in the form of there is no reason to expect them to have the same ante-
questions that researchers can ask themselves in order to cedents and consequences.
determine what the appropriate relationship is between More specifically, a construct should be modeled as hav-
their measures and their constructs. ing formative indicators if the following conditions prevail:
Four sets of questions should be used in combination to (a) the indicators are viewed as defining characteristics of
determine the appropriate measurement model. The first set the construct, (b) changes in the indicators are expected to
of questions relate to the direction of causality between the cause changes in the construct, (c) changes in the construct
construct and its indicators. For formative measurement are not expected to cause changes in the indicators, (d) the
models, the direction of causality flows from the measures indicators do not necessarily share a common theme, (e)
to the construct, and it flows from the construct to the mea- eliminating an indicator may alter the conceptual domain of
sures for reflective measurement models. The second set of the construct, ( f ) a change in the value of one of the in-
questions relates to the interchangeability of the indicators. dicators is not necessarily expected to be associated with a
The indicators need not be interchangeable for formative change in all of the other indicators, and (g) the indicators
measurement models but should be for reflective measure- are not expected to have the same antecedents and conse-
ment models. The third criteria relates to the issue of whether quences. On the other hand, a construct should be modeled
TABLE 1
1. Direction of causality from construct to measure implied Direction of causality is from items to Direction of causality is from con-
by the conceptual definition construct struct to items
Are the indicators (items) (a) defining characteristics or Indicators are defining characteristics Indicators are manifestations of the
(b) manifestations of the construct? of the construct construct
Would changes in the indicators/items cause changes in Changes in the indicators should Changes in the indicator should not
the construct or not? cause changes in the construct cause changes in the construct
Would changes in the construct cause changes in the Changes in the construct do not Changes in the construct do cause
indicators? cause changes in the indicators changes in the indicators
2. Interchangeability of the indicators/items Indicators need not be interchangeable Indicators should be interchangeable
Should the indicators have the same or similar content? Indicators need not have the same or Indicators should have the same or
Do the indicators share a common theme? similar content/indicators need not similar content/indicators should
share a common theme share a common theme
Would dropping one of the indicators alter the conceptual Dropping an indicator may alter the Dropping an indicator should not al-
domain of the construct? conceptual domain of the construct ter the conceptual domain of the
construct
3. Covariation among the indicators Not necessary for indicators to covary Indicators are expected to covary
with each other with each other
Should a change in one of the indicators be associated Not necessarily Yes
with changes in the other indicators?
4. Nomological net of the construct indicators Nomological net for the indicators Nomological net for the indicators
may differ should not differ
Are the indicators/items expected to have the same ante- Indicators are not required to have Indicators are required to have the
cedents and consequences? the same antecedents and con- same antecedents and conse-
sequences quences
204 JOURNAL OF CONSUMER RESEARCH
MULTIDIMENSIONAL FORMATIVE AND cators like Type I models, but the first-order dimensions
REFLECTIVE INDICATOR CONSTRUCTS themselves have formative rather than reflective indicators.
Although this kind of second-order factor model has not
been explicitly recognized in the literature, Reillys (1982)
The above criteria focus on the relationships between mea- family social status construct may be an example of this
sures and first-order latent constructs. However, it is important type of model. In Reillys research, four well-established
to note that conceptual definitions of constructs are often scales for measuring social status were used as reflective
specified at a more abstract level, which sometimes include indicators of an underlying factor called family social status.
multiple formative and/or reflective first-order dimensions. However, each of these social status scales is composed of
For example, a single multidimensional construct might have several formative indicators. For example, one of the re-
one type of measurement model relating its measures to its flective indicators of the second-order family social status
first-order components and a different measurement model construct was Warners Index of Status Characteristics,
relating its components to the underlying second-order factor. which is itself a first-order factor composed of ratings of
Of course, some researchers might argue that a construct must occupation, source of income, dwelling type, and neigh-
be conceptually and empirically unidimensional to be mean- borhood quality. Since occupation, source of income, dwell-
ingful. However, such a view is often inconsistent with the ing type, and neighborhood quality are clearly not inter-
way constructs are defined in the field. We would argue that changeable, have potentially different antecedents and
whether a construct is viewed as unidimensional or multi- consequences, and would not necessarily covary with each
FIGURE 2
REVIEW OF THE MARKETING cide which model is appropriate to use in a specific instance.
LITERATURE Thus, a logical next step would be to review the research
literature, apply the criteria, and thereby determine how
Methodology prevalent measurement model misspecification is in the mar-
keting literature. To our knowledge, no one has ever at-
Our review up to this point has attempted to clarify the tempted to systematically evaluate the appropriateness of
distinctions between formative and reflective indicator mea- measurement model specifications used in our field. Such a
surement models, and the criteria that could be used to de- review would provide insights into not only the extent of
206 JOURNAL OF CONSUMER RESEARCH
TABLE 2
Overall JMR JM
Should be Should be Should be Should be Should be Should be
reflective formative Total reflective formative Total reflective formative Total
Modeled as reflective 810 336 1,146 319 120 439 368 187 555
(68) (28) (96) (70) (26) (96) (63) (32) (95)
Modeled as formative 17a 29b 46 7a 10b 17 10a 18b 28
(1) (3) (4) (2) (2) (4) (2) (3) (5)
Total 827 365 1,192 326 130 456 378 205 583
(69) (31) (100) (72) (28) (100) (65) (35) (100)
JCR MS
Should be Should be Should be Should be
reflective formative Total reflective formative Total
Modeled as reflective 107 22 129 16 7 23
(82) (17) (99) (70) (30) (100)
measurement model misspecification but also would reveal criteria in table 1, each construct was classified as formative
which constructs have been most frequently misspecified. or reflective. A construct was classified as formative if it
Therefore, in this section, we will report the results of a clearly met the majority of the criteria in the second column
review of measurement model specifications in the top four of this table and was classified as reflective if it met most
marketing journals by examining every construct for which of the criteria in the third column of this table. In those
a confirmatory factor analysis has been reported. cases where all three coders agreed that the construct met
The Journal of Consumer Research (JCR), Journal of the criteria for either a formative or a reflective measurement
Marketing (JM), Journal of Marketing Research (JMR), and model, the construct was assigned to that measurement
Marketing Science (MS) were selected as representative of model category. In instances where the coders disagreed
the best journals in the marketing literature. These four jour- about the extent to which the construct met the various
nals were searched for the 24-year period from 1977 through criteria, the points of disagreement were discussed until a
2000 (19822000 for MS) to identify all empirical appli- consensus was reached. In approximately 14% of the cases,
cations of latent variable SEM or confirmatory factor anal- it was difficult, if not impossible, to tell whether the con-
ysis. Methodological papers in which only simulated data struct should have been modeled as a formative or as a
were analyzed, or actual data were analyzed for illustrative reflective construct. In the majority of these instances (12%),
purposes only, were not considered. Similarly, conventional this happened because the authors failed to provide a com-
exploratory factor analysis models, path analysis, and other plete set of items or construct definition, although there were
structural models estimated by regression methods (e.g., a few instances (2%) when the construct of interest met
models estimated by two-stage least squares), nonlinear some of the criteria for a formative scale but other criteria
structural models, and observed variable models were ex- for a reflective scale. However, in all of these cases the
cluded from the analysis. Therefore, the database analyzed benefit of the doubt was given to the authors, and the con-
here consists of articles incorporating either confirmatory struct was categorized as being correctly modeled as spec-
factor models or latent variable SEM. Using these criteria, ified by the author. This procedure was adopted with the
we identified 178 articles containing 1,192 constructs mod- goal of making the estimates of misspecification as conser-
eled as latent factors with multiple indicators. vative as possible.
The classification of the constructs followed a multistep
procedure. The first step was for each of the three coders Results
(the authors) to independently read the articles, identify
those constructs with multiple measures, and determine how Table 2 summarizes our findings across all journals. Sev-
their measurement models were specified. Next, using the eral interesting patterns emerge. First, the results indicate
MEASUREMENT MODEL MISSPECIFICATION 207
that, overall, 71% (68% + 3%) of the latent constructs with of psychological constructs (e.g., brand attitude, purchase
multiple measures found in the top-four marketing journals intentions, and ad-evoked feelings).
during the past 24 years were correctly modeled, and 29%
(28% + 1%) were incorrectly modeled. By far, most of those
that were correctly modeled (810 out of a total of 839) were
Exemplars of Constructs with Formative
reflective constructs correctly modeled as having reflective Indicators
measures, while the remainder (29 of 839) were formative Table 3 reports a sample of constructs from our review
constructs correctly modeled as having formative measures. of the literature that were identified as first-order or second-
In contrast, the majority of constructs that were incorrectly order constructs with formative indicators. The columns in
modeled (336 out of a total of 353) were formative con- the table contain the construct names, the nature of the con-
structs incorrectly modeled as having reflective measures. struct (i.e., first-order or second-order), the study in which
Table 2 also reports the results separately for each of the the construct was examined, and examples of the indicators
four journals reviewed. The journals differed on the extent used to measure the construct. In the case of the second-
to which multiple indicator latent variables were examined. order constructs, the indicators were typically scale scores
The JM had the greatest number of latent variables (583), representing first-order dimensions of the construct. All of
and MS had the least (23). Although this is partially due to the constructs in this table satisfy the decision rules for
the fact that MS only began publication in 1982, it cannot formative constructs discussed in table 1.
account for a discrepancy of this magnitude. It seems more
Nature of construct
Construct name (first or second order) Studies Examples of indicators used to measure construct
TABLE 3 (Continued)
Nature of construct
Construct name (first or second order) Studies Examples of indicators used to measure construct
Perceived risk 2d Srinivasan and Ratch- Types of perceived risk:
ford (1991) Financial
Performance
Physical
Convenience
Positive emotion 2d Murry and Dacin (1996) Contented
Happy
Positive experience 2d Srinivasan and Ratch- Experience with previous manufacturer or dealer
ford (1991) Experience with previous car
Qualitative power 2d Gaski (1986) Types of qualitative power source:
source Expert
Legitimate
Referent
Role ambiguity 1st Michaels, Day, and With respect to yourself and your job, please indicate your
Joachimsthaler (1987) agreement with each statement listed below:
Clear, planned goals and objectives exist for my joba
or incorrectly specified as having reflective indicators (as quences of misspecification. Manipulating the focal con-
indicated in the misspecified version of models 1A1C and structs position within the model permitted us to test the
2A2C in fig. 3), (2) the focal construct was an exogenous effects of misspecification on structural parameters leading
construct (correctly specified model 1) or an endogenous to the misspecified construct, as well as those structural paths
construct (correctly specified model 2) in the structural emanating both directly and indirectly from the misspecified
model, and (3) the item intercorrelations of the focal con- construct. Finally, the manipulation of the magnitude of the
struct were weak (.1), moderate (.4), or strong (.7). Manip- correlations among the indicators of the formative construct
ulating the measurement model specification for the focal allowed us to test the significance of the effects of misspe-
construct obviously allowed us to investigate the conse- cification across a variety of situations, including at least
210 JOURNAL OF CONSUMER RESEARCH
FIGURE 3
FIGURE 4
(model 1, top panel in fig. 4), measurement model misspe- g11 are inflated by 492% in model 1A (formative item
cification positively biases estimates of g11 and g31, but has intercorrelation p .1), 385% in model 1B (formative item
no effect on b 21 or b 41. The degree of bias in the affected intercorrelation p .4), and 337% in model 1C (formative
parameters is negatively related to the magnitude of the item intercorrelation p .7). Estimates of g31 are positively
formative item intercorrelations. For example, estimates of biased 488% in model 1A, 384% in model 1B, and 335%
212 JOURNAL OF CONSUMER RESEARCH
in model 1C. Analysis of variance shows that the differences fit indices [CFI] 1 .95, standardized root mean square re-
between the average unstandardized parameter estimates and sidual [SRMR] ! .08, root mean square error of approxi-
the population values when the focal construct is exogenous mation [RMSEA] ! .05). So did all of the misspecified mod-
are statistically significant for both g11 (F p 36, 343.61, els, according to the CFI, SRMR, and the RMSEA indices.
df p 1, p ! .001) and g31 (F p 36, 461.06, df p 1, p ! Only the chi-square and the GFI were able to detect the
.001), but not b 21 (F p 1.09, df p 1, p 1 .05) or b 41 measurement model misspecification. More specifically, the
(F p .01, df p 1, p 1 .05). Moreover, it is worth noting chi-square and GFI identified all but one (model 1C) of the
that all of the individual Bonferroni pairwise comparisons six misspecified models. This suggests that when the mea-
across the three item intercorrelation conditions are signif- surement model is misspecified, researchers may have dif-
icant for g11 and g31 (all ps ! .001), but not for b 21 or b 41 ficulty detecting it based on the overall goodness-of-fit in-
(all ps 1 .05). dices. Although the chi-square statistics are significant, they
When the formatively indicated construct is in an endog- probably would have been discounted with a sample size
enous position in the model (correctly specified model 2 in of 500 based on the well-recognized dependence of the chi-
fig. 3), measurement model misspecification suppresses es- square on sample size. Therefore, only the GFI would have
timates of g11, negatively biasing the estimates by 88% in indicated a lack of fit, and the overall pattern would have
model 2A (formative item intercorrelation p .1), 89% in
suggested that the models fit the data adequately. Indeed, if
model 2B (intercorrelation p .4), and 93% in model 2C
researchers relied on only the two indices (i.e., the CFI and
(intercorrelation p .7), but has no effect on estimates of
TABLE 4
1A correctly specified Exogenous .10 163.50 160 .44 .94 .99 .025 .007
1A misspecified Exogenous .40 282.20 166 .00 .20 .98 .042 .037
1B correctly specified Exogenous .70 163.77 160 .45 .94 .99 .022 .007
1B misspecified Exogenous .10 212.26 166 .05 .81 .99 .031 .023
1C correctly specified Exogenous .40 167.74 160 .46 .94 .99 .09 .007
1C misspecified Exogenous .70 177.35 166 .33 .93 .99 .022 .010
2A correctly specified Endogenous .10 163.56 160 .44 .94 .99 .033 .008
2A misspecified Endogenous .40 289.19 166 .00 .27 .98 .046 .038
2B correctly specified Endogenous .70 162.38 160 .46 .94 .99 .031 .007
2B misspecified Endogenous .10 242.70 166 .00 .47 .99 .043 .030
2C correctly specified Endogenous .40 164.03 160 .43 .94 .99 .031 .008
2C misspecified Endogenous .70 232.17 166 .01 .80 .99 .045 .028
NOTE.Goodness-of-fit indices shown in bold indicate a lack of model fit.
MEASUREMENT MODEL MISSPECIFICATION 213
RECOMMENDATIONS FOR SPECIFYING that of model identification. Unfortunately, a model like the
FORMATIVE MODELS one shown in panel 1 of figure 5 is not identified due to
indeterminacies associated with the scale of measurement
Although it is difficult to know precisely why measure- for the latent construct and the construct level error term.
ment model misspecification has become so pervasive, it is MacCallum and Browne (1993) note two conditions nec-
possible to speculate about the potential reasons. First, it essary for the identification of formative indicator con-
may result from the fact that many marketing researchers structs. First, the scale of measurement for the latent con-
simply do not think of measurement model relationships as struct must be established either by constraining a path from
hypotheses to be tested with differing theoretical implica- one of the constructs indicators to be equal to one or by
tions. One indication of the failure to think about measure- constraining the residual error variance for the construct to
ment relationships is that most articles devote little or no be equal to one. Second, to resolve the indeterminacy as-
attention to this issue when describing the specification of
sociated with the construct level error term, formative con-
their models. Another possible explanation is that research-
structs must emit paths to (a) at least two unrelated latent
ers are simply unaware of the conceptual distinctions be-
constructs with reflective indicators (see panel 2, fig. 5), (b)
tween formative and reflective measurement models. Al-
at least two theoretically appropriate reflective indicators
though this is somewhat surprising since the distinction was
(see panel 3, fig. 5), or (c) one reflective indicator and one
made in the field over 20 years ago (e.g., Fornell and Book-
latent construct with reflective indicators (see panel 4, fig.
stein 1982), it is still relatively uncommon to see references
5). Although this indeterminacy could also be resolved by
FIGURE 5
priate items might be Overall, how satisfied are you with hypothesized (as opposed to nonhypothesized) relationships.
this product? and All things considered, I am very pleased However, as noted by MacCallum and Browne (1993), fix-
with this product. Since these new indicators capture a ing these covariances at zero will typically result in large
consumers overall level of satisfaction, they are reflective blocks of zeros in the predicted covariance (S-hat) matrix.
in nature. Therefore, with the addition of these two reflective This is a very strong theoretical statement. It assumes that
indicators, the consumer satisfaction construct would now the variables are perfectly uncorrelated. Thus, any common
have two paths emanating from it and be identified. It is causes of these variables that are outside of the system of
important to note that this procedure results in measurement relationships represented in the model, any causal relation-
error terms (e1 and e 2 ) for the reflective indicators (Y1 and ships among these variables that have been omitted from
Y2 ) plus a combined measurement error term (zeta 1) for the model, or even methodological artifacts that are shared
the five formative indicators (X1X 5). A third option illus- by these variables, will contribute to the lack of fit of the
trated in panel 4 in figure 5 represents a compromise be- proposed model. For these reasons, this does not appear to
tween these alternatives. be an acceptable solution.
Of course, to implement any of these solutions, the re- Another method of handling the covariances is to fol-
searcher must be familiar enough with the identification low standard practice and estimate the covariances among
problem to anticipate the need for either additional items or all exogenous latent constructs and manifest variables
constructs. The construct conceptualization and nature of (MacCallum and Browne 1993). The advantage of this
the indicators to be used should be determined in the ques- procedure is that the overall fit of the model is not un-
(RPNFI) could be used to evaluate the fit of the theoretical variance explained in many criterion variables (Peterson,
model relative to any of the baseline models that precede Albaum, and Beltramini 1985). In addition, as demonstrated
it in the model hierarchy. In our view, freeing up the co- by our simulation, the failure to correctly specify the mea-
variances among all exogenous constructs and measures is surement model can lead to different conclusions about the
the best approach for dealing with this issue when coupled empirical relationships between latent constructs. Thus by
with full disclosure of the impact of these nonhypothesized implication, a substantial proportion of the empirical results
relationships on the fit of the model. in the literature may be potentially misleading. Therefore,
it is imperative for the field as a whole to think more care-
CONCLUSIONS fully about measurement model relationships and do a better
job of making sure that the measurement models used match
The goals of this research were to draw attention to the that conceptualization. Hopefully, this research will aid pro-
theoretical distinctions between formative and reflective gress in this area.
measurement models, provide a set of guidelines for decid-
ing on the appropriate measurement model, determine the [David Glen Mick served as editor and William O. Bear-
extent to which measurement model misspecification has den served as associate editor for this article.]
occurred in the top-tier marketing journals, examine the ef-
fects of measurement model misspecification with a Monte REFERENCES
Carlo simulation, and recommend procedures for correctly
Benefit Congruency Framework of Sales Promotion Effec- Model with Multiple Indicators and Multiple Causes of a
tiveness, Journal of Marketing, 64 (October), 6584. Single Latent Variable, Journal of the American Statistical
Churchill, G. A., Jr. (1979), A Paradigm for Developing Better Association, 70 (September), 631639.
Measures of Marketing Constructs, Journal of Marketing Kohli, Ajay K., Bernard J. Jaworski, and Ajith Kumar (1993),
Research, 16 (February), 6473. MARKOR: A Measure of Market Orientation, Journal of
Cote, Joseph and M. Ronald Buckley (1987), Estimating Trait, Marketing Research, 30 (November), 467477.
Method, and Error Variance: Generalizing across 70 Construct Law, Kenneth, and Chi-Sum Wong (1999), Multidimensional
Validation Studies, Journal of Marketing Research, 24 (Au- Constructs in Structural Equation Analysis: An Illustration
gust), 315318. Using the Job Perception and Job Satisfaction Constructs,
Cravens, David W., Thomas N. Ingram, Raymond W. LaForge, Journal of Management, 25 (2), 143160.
and Clifford E. Young (1993), Behavior-Based and Outcome- Lichtenstein, Donald R., Nancy M. Ridgway, and Richard G. Ne-
Based Salesforce Control Systems, Journal of Marketing, 57 temeyer (1993), Price Perceptions and Consumer Shopping
(October), 4759. Behavior: A Field Study, Journal of Marketing Research,
Crosby, Lawrence A., Kenneth R. Evans, and Deborah Cowles 30 (May), 234245.
(1990), Relationship Quality in Services Selling: An Inter- MacCallum, Robert C. and Michael W. Browne (1993), The Use
personal Influence Perspective, Journal of Marketing, 54 of Causal Indicators in Covariance Structure Models: Some
(July), 6881. Practical Issues, Psychological Bulletin, 114 (3), 533541.
Crosby, Lawrence A. and Nancy Stephens (1987), Effects of Re- MacCorquodale, K. and Paul E. Meehl (1948), On a Distinction
lationship Marketing on Satisfaction, Retention and Prices in between Hypothetical Constructs and Intervening Variables,
Psychological Review, 55(2), 307321.
bles, Journal of Consumer Research, 9 (December), 263 Distributor Perspective, Journal of Marketing, 62 (July),
278. 99111.
Shimp, Terence A. and Alician Kavas (1984), The Theory of Singh, Jagdip (1988), Consumer Complaint Intentions and Be-
Reasoned Action Applied to Coupon Usage, Journal of Con- havior: Definitional and Taxonomical Issues, Journal of Mar-
sumer Research, 11 (December), 795809. keting, 52 (January), 93107.
Siguaw, Judy A., Penny M. Simpson, and Thomas L. Baker Srinivasan, Narasimhan and Brian T. Ratchford (1991), An Em-
(1998), Effects of Supplier Market Orientation on Distrib- pirical Test of a Model of External Search for Automobiles,
utor Market Orientation and the Channel Relationship: The Journal of Consumer Research, 18 (September), 233242.