Vous êtes sur la page 1sur 8

Does Satisfaction Moderate

the Association between Alternative


Attractiveness and Exit Intention
in a Marketing Channel?

Robert A. Ping, Jr.


Wright State University

This study investigated the predicted moderating effect of ings by providing a deeper understanding of the major
satisfaction on the association between the attractiveness antecedents of exit intention. They also provide empirical
of the alternative relationship and exit intention in a mar- support for the long-term buyer-seller relationship frame-
keting channel. The study used a variation of the Kenny work proposed by Dwyer, Schurr, and Oh (1987) and
and Judd structural equation technique proposed by the others by showing an effect predicted by Dwyer, Schurr,
author. The results suggested that for channel customers and Oh (1987): that higher satisfaction attenuates the al-
ternative attractiveness-exit intention association.
with lower satisfaction, alternative attractiveness was
After summarizing the current knowledge about chan-
positively associated with exit intention. When satisfaction
nel relationship exit intention, a satisfaction-alternative
was higher, however, alternative attractiveness had no attractiveness interaction is proposed and tested using a
association with exit intention. The implications of these field survey and the Ping technique. The implications of
results are discussed. the results are then discussed.

CHANNEL RELATIONSHIP TERMINATION


Channel relationship termination has received some
attention in recent studies (Ping and Dwyer 1991; Ping In an investigation of generalized responses to channel
1993a). These studies generally contend that exit intention relationship problems, Ping (1993a) observed that the
intention to exit a channel relationship was negatively
in a channel relationship has several antecedents, among
associated with overall relationship satisfaction and posi-
them overall satisfaction and the attractiveness of the best
tively associated with the attractiveness of the best avail-
alternative relationship. The associations among these able alternative relationship. The specification of these
variables have been modeled using linear effects, and relationships involved a structural equation analysis of a
nonlinear effects have been assumed to be absent. This model that implicitly posited only linear relationships.
study investigates a plausible nonlinear effect involving In a conceptualization of the history of a buyer-seller
these variables: the interaction of overall satisfaction and relationship, Dwyer, Schurr, and Oh (1987) proposed that
alternative attractiveness in their effect on exit intention. these relationships pass through several phases (see also
It uses a field survey and a structural equation technique Ford 1980; Gadde and Mattsson 1987). Both parties to the
proposed by Ping (1993b) that estimates interaction and buyer-seller relationship, they argued, pass through aware-
quadratic effects for latent variables. ness, exploration, expansion, commitment, and, ulti-
The results fill a gap in the channel reactions-to- mately, dissolution phases of the relationship. They noted
dissatisfaction literature and extend Ping's (1993a) find- that exchange partners in the committed phase achieve a
level of satisfaction that precludes other primary exchange
Journal of the Academy of Marketing Science. partners. They stated that awareness of alternative relation-
Volume 22, No. 4, pages 364-371. ships is maintained, but without constant comparisons to
Copyright 9 1994 by Academy of Marketing Science. the current relationship. One plausible result of this pre-
Ping / ALTERNATIVE ATTRACTIVENESS & EXIT INTENTION 365

clusionary state is that for satisfied firms, increases in the the alternative would be (1) its fairness, (2) its products
attractiveness of alternative relationships would not affect and services, (3) its policies and, (4) in general, how
relationship exit intention. Ping (1993a) observed a posi- satisfied the firm would be with the alternative supplier.
tive association between alternative attractiveness and exit Exit intention, the intention to physically leave the
intention, however; this suggests that the association is relationship, was operationalized as the propensity to ter-
contingent on the level of overall relationship satisfaction. minate the primary supplier relationship (Ping 1993a). The
In particular, if Dwyer, Schurr, and Oh (1987) are correct, conceptualization taps the degree of intention to discon-
when overall satisfaction is lower, changes in alternative tinue the relationship with the economic exchange partner.
attractiveness should be positively associated with exit The six items in the exit intention measure concerned the
intention. At higher levels of satisfaction, however, this buyer firm's thinking of exiting, looking for a replacement
association should not be significant. Accordingly, supplier, considering a replacement (two items), and re-
solving to end the relationship (two items).
HI: Overall relationship satisfaction moderates These measures were combined into a self-administered
the association between alternative attrac- questionnaire that was mailed to a sample of hardware
tiveness and exit intention. retailers. The analysis of the resulting data was conducted
using structural equation analysis with a satisfaction-alternative
Specifically, attractiveness latent variable interaction specified using
the Ping technique. Before describing the study, however,
some background on the estimation of nonlinear effects in
H2a: At higher levels of overall relationship satis- structural equations is appropriate.
faction there is no association between alter-
native attractiveness and exit intention,
NONLINEAR LATENT VARIABLES
and
Kenny and Judd (1984) proposed that interaction and
H2b: At lower levels of overall relationship satis- quadratic latent variables could be specified using indica-
faction alternative attractiveness is positively tors that are products of observed variables. They proposed
associated with exit intention. that products of the indicators for the linear latent variables
X and Z, for example, would specify the latent interaction
The remainder of the article reports a test of these hypothe- variable XZ. Specifically, if X had indicators x~ and x2 and
ses using data provided by a field survey. Z had indicators z~ and z2, respectively, XZ is specified
using the indicators XlZl, XlZ2, X2Zl, and x2z2.
In addition, they showed that under certain conditions,
MEASUREMENT the variance of these indicator products is determined by
their constituent indicators. They showed that, for exam-
Satisfaction, the global evaluation of relationship ful- ple, the variance of the indicator x~zz depends on ~1, ~,,1,
fillment by the firm (Dwyer and Oh 1987), was measured Var(X), Var(Z), 0~xl, and 0~zl, where Var(X) and Var(Z) are
using a modification of the Dwyer and Oh satisfaction the variances of the latent variables X and Z, ~1 and ~1 are
scale inspired by Gaski and Nevin (1985). The domain of the loadings ofxl on X and z~ on Z, and 0~xl and 0~zl are the
satisfaction includes all the characteristics of the buyer-
variances of the error terms exl and ez~. Specifically, assum-
seller relationship that a firm deems "rewarding, profit-
ing the latent variables X and Z are independent of the error
able, or instrumental" (Rukert and Churchill 1984), or
terms exl and ~1; the error terms are independent of each
costly, unfair, or frustrating (Ping 1993a). The five items
in the satisfaction measure assessed the buyer firm's over- other; and x~, Zl, ~1, and e,l are normally distributed, they
all satisfaction with the relationship (two items), fairness showed the variance of XlZl is given by
in the exchange relationship (two items), and the degree to
Var(xlzl) = Var[(k, lX + gxl)(~.zlZ+ ez0]
which the seller firm was a good company with which to
= ~xl2 ~zl2 Var(XZ) + ~xl2 VaF(X)0ez1 (1)
do business.
The attractiveness of the best alternative relationship, a + ~.,12Var(Z)0exl+ 0Exl0c,1.
firm's estimate of the satisfaction available in the best They then specified latent variables such as X Z with
available alternative relationship, was operationalized as indicators such as XlZl by constraining the indicator loading
the buyer firm's perception of the overall fulfillment avail- and error term for XlZl (Xxlzl and 0~xlzl)to be the combina-
able from the best alternative supplier, in addition to the tions of the parameters shown in equation (1), that is,
overall fulfillment available in the existing buyer-seller
relationship (Ping 1993a). The conceptualization encom- ~'xlzl = ~xl~'zl (2)
passes a firm's generalized perceptions of the rewards and
costs available in the most salient available relationship and
alternative. The four items in this measure dealt with the
buyer firm's evaluation of how good a supplier company 0exlzl = ~xl 2 War(X)0ez1 + ~12 Var(Z)0ex 1 + 0exl0ezl. (3)
366 JOURNALOF THE ACADEMYOF MARKETINGSCIENCE FALL1994

They also used COSAN (currently a subprocedure of the of product indicators can be determined using a measure-
SAS procedure CALIS), which is particularly suited to ment model and equations (2) and (3), then specified as
modeling structural equations with nonlinear terms such constants in the structural model.
as those in equations (2) and (3). Ping also pointed out that the sufficient unidimension-
Although the Kenny and Judd technique is an important ality assumption enables the omission of the nonlinear
theoretical contribution, it has proven difficult for re- latent variables from the measurement model with no
searchers to implement (Aiken and West 1991). The num- effect on the parameter estimates for the linear latent
ber of dummy variables required to specify each indicator variables. By the definition of unidimensionality, unidi-
of a nonlinear variable using the Kenny and Judd technique mensional constructs are unaffected by the presence or
(e.g., one for each term in equations [2] and [3]) can absence of other latent variables in the measurement
become tedious for models with many indicators or several model. Similarly, adding or deleting unidimensional con-
nonlinear latent variables (Ping 1993b). 1 structs in the structural model does not affect the measure-
Hayduk (1987) and others have proposed a variation of ment parameter estimates for the added or other latent
the Kenny and Judd (1984) technique that can be imple- variables in the structural model.
mented using LISREL (Jtreskog and Strbom 1989) and The Ping technique involves
EQS (Bentler 1989). The technique is difficult to summa-
rize, however, and the interested reader is directed to 1. verifying indicator normality,3
chapter 7 of Hayduk (1987) for details. Unfortunately, the 2. assuming the latent variables are independent of
technique also requires the specification and estimation of the error terms, and the error terms are inde-
many dummy variables. pendent of each other (a standard structural equa-
As a result, Ping (1993b) proposed an additional vari- tion assumption),
ation of the Kenny and Judd technique that requires no 3. unidimensionalizing each linear latent variable,
dummy variables and can be implemented in LISREL and 4. centering the observed variables at zero by sub-
EQS. The technique is carried out in two steps, paralleling tracting the mean of a variable from each case
the two-step estimation approach for structural equation value for that variable (see Bollen 1989),
suggested by Anderson and Gerbing (1988)--estimate the 5. estimating loadings and error variances for the
measurement model before estimating the structural linear independent variable indicators using a
model. measurement model,
The measurement parameters for the linear latent vari- 6. using these estimates to calculate equation (2)
ables are estimated in a measurement model. These esti- and (3) estimates of the loadings and error vari-
mates are used to calculate the loadings and error variances ances for the nonlinear latent variable indicators,
for the indicators of the interaction variables in equations 7. specifying equation (2) and (3) estimates as fixed
(2) and (3). The interaction variables are added to the values in a structural model, then estimating that
model, and the calculated loadings and error variances for model, and
the indicator products are specified as fixed rather than free 8. repeating steps 6 and 7 as required to obtain
minimal change in the measurement parameters
variables in the structural model.
Under the Kenny and Judd normality assumptions, and between two structural model estimates (fre-
assuming the unidimensionality of each linear latent vari- quently not necessary for strongly unidimen-
able, Ping suggested that the loadings and error variances sional latent variables).
for the Kenny and Judd product indicators of an interaction
The balance of this article describes the field survey.
or quadratic latent variable need not be estimated in the
structural model. 2 Specifically, he proposed that parameter
estimates from the measurement model be used to compute
the loading and error variance for product indicators such
METHOD
as xlz~ using equations (2) and (3). Then a structural model
To test the hypotheses, satisfaction, alternative attrac-
with XZ can be estimated with these calculated interaction
tiveness, and exit intention were measured with balanced
indicator loadings and error variances (e.g., ~1zl and 0~xlzl) 5-point Likert-type scales. The survey population was
specified as fixed values (set equal to the equation [2] and hardware retailers, and the sampling frame was the sub-
[3] values for ~lzl and 0~xl~l).This is possible because with scription list of a popular hardware trade publication.
sufficient unidimensionality, the measurement parameters Sampling involved nth name selections of 100 pretest
for a latent variable's indicators (e.g., ~1, ~,~1, Var[X], names and 600 final test names. The pretest responses were
Var[Z], 0~xl, and 0 ~ ) are trivially variant between the used to verify the psychometric properties of the measures.
measurement and structural models (Anderson and Gerbing The resulting measures appeared to be content valid, uni-
1988). In other words, the measurement parameter esti- dimensional, and internally and externally consistent (see
mates for the indicators of a sufficiently unidimensional Gerbing and Anderson 1984); they each had a coefficient
latent variable change very little between the measurement alpha of .8 or above. The final test mailing yielded 288
and structural models (frequently only in the third decimal responses by firms after two postcard follow-ups. The
place). As a consequence, the loadings and error variances psychometric properties of the measures were reexamined
Ping / ALTERNATIVE ATTRACTIVENESS & EXIT INTENTION 367

FIGURE 1 TABLE 1
Linear-Terms-Only Measurement Model Measurement Model Results
Parameter Estimate Parameter Estimate

~'sl 0.79 es4 0.12


~'sz 0.88 es5 0.10
~'S ~
SAT~ / ~ ALT
~ ~ EXI ~s3 1.00 Ca1 0.27
~'s4 0.87 Ca2 0.24
~5 0,94 Ca3 0.07
~al 0.92 Ca4 0.24
tt I ltl It] \\ Xa2 0.90 e~i 0.52
~51 ~s2 ~s3 ~s4 ~s5 ~al ~a2 ~a3 ~a4 ~el ~e2 ~e3 ~el ~e5 ~e6 9~a3 1.00 s 0.19
~'a4 0.78 s 0.11
~i 0.84 ee4 0.32
NOTE: SAT = satisfaction, ALT = alternative attractiveness, EXI =
~z 0.83 ee5 0.09
exiting.
~'e3 1.00 Ee6 0.12
using these responses and item-to-total correlations, coef- ~e4 0,94 ~$AT 0.51
ficient alpha calculations, ordered similarity coefficients ~'e5 0.96 (PSAT,~LT --0.37
(Hunter 1973), multiple group analysis (Anderson, Gerbing, ~e6 0.92 qbSAT,EXl -0.43
and Hunter 1987), and LISREL single factor analysis esl 0.16 ~ALT 0.85
(J6reskog 1993, pp. 297, 313). The measures were content es2 0.13 ~ALT,EX~ 0.51
valid, unidimensional, and internally and externally con- es3 0.10 ~xI 0.78
sistent; they had latent variable reliabilities of .9 or above Equation 2 and 3 estimates
and average extracted variances of .7 or above (Fornell and ~'slal .533 0eslaI .257
Larker 1981). ~'sla2 .509 0esla2 ,239
The indicators for satisfaction and alternative attrac- IsXa3 .625 0~Xa3 .180
tiveness were then zero centered, the normality of all the ~'s144 .384 0~sxa4 .207
indicators was assessed, and the measurement model for ~'s2al .669 0es2a1 .242
the linear-terms-only model shown in Figure 1 was esti- ~LsZa2 .639 0esza2 .224
mated using EQS and maximum likelihood. The resulting ~243 .784 0es2a3 .151
measurement parameter estimates for satisfaction and al- ~'s2a4 .482 0~sza4 .198
ternative attractiveness are shown in Table 1. ~'s3al .852 0~3al .251
Because the latent variable intercorrelations were above ~s342 .815 0es3a2 .231
.5, latent variable discriminant validity was tested using ks3a3 1.000 0es3a3 .140
two-group measurement models. In this test, a measure- ~'s344 .615 0es3a4 .208
ment model with, for example, the satisfaction-exit inten- ks4al .657 0es4al .231
tion path fixed at 1 was nested in an identical model with ~'s442 .628 0es4a2 .213
that path freed. One degree of freedom tests of the differ- ~'s443 .770 0e.s4a3 .142
ences between these nested models suggested that satisfac- ~ks4a4 .473 0es4a4 .189
tion, alternative attractiveness, and exit intention were ~5al .755 0esSal .227
distinct. ~saZ .722 0essa2 .209
The measurement model parameter estimates were ~5a3 .885 0esSa3 .128
combined to produce equation (2) and (3) loadings and ~'s5a4 .544 Oes5a4 .188
error variances for the satisfaction-alternative attractive-
ness interaction indicators shown in the Figure 2 structural
model. This structural model was then estimated using
EQS and m a x i m u m likelihood by fixing the loadings and a result, contains the estimates from the second structural
error variances for the product indicators at the Table 1 model estimation. The Table 2 degrees of freedom for the
values. structural model has been reduced by the number of pre-
Because there were slight differences in the measure- viously estimated product indicator loadings and error
ment and structural model estimates for the measurement variances (40; see Table 1). 4
parameters of satisfaction and alternative attractiveness, an Because the use of product indicators in a structural
iterative approach was used to produce the Table 2 results. equation model renders the model formally nonnormal and
This was accomplished by recomputing equation (2) and maximum likelihood estimation assumes multivariate nor-
(3) values using the structural equation estimates of the mality, maximum likelihood standard errors for the struc-
measurement parameters for satisfaction and alternative tural effect coefficients are not formally appropriate
attractiveness, then reestimating the structural model with (Bollen 1989; see Hu, Bentler, and Kano 1992). The effect
these revised equation (2) and (3) values fixed. Table 2, as estimates appear to be robust to departures from normality,
368 JOURNALOF THE ACADEMYOF MARKETINGSCIENCE FALL1994

FIGURE 2 DISCUSSION
Structural Model
The Ping, Kenny and Judd/COSAN, and regression
results were directionally similar, and the t values for the
Eel ~=2 E.3 ~s4 ~s5 Ping technique and the Kenny and Judd/COSAN approach
generally lead to the same inferences. The maximum like-
~I $2 ~3 $4 ~5 lihood and Robust standard error estimates were also simi-
lar, whereas the chi-square estimates were not. Because the
Robust estimator does not assume multivariate normality
S A T " ...
~ei ! S as the maximum likelihood estimator does, the Robust
el 9 " ~el estimator results will be interpreted.
ALT =- E X I :...r.,,e3 ",,----~.3 Based on the Robust estimator, the model fit the data,
es 9 ' ees
and the hypothesized associations were supported. Satis-
e 6 .,L (~e6 faction moderated the alternative attractiveness-exit inten-
a I a 2 a 3 a~
tion association, as shown by the significant effect
coefficient for the alternative attractiveness-exit intention
~al ~a2 ~a3 ~a4 interaction (see Table 2). In addition, satisfaction attenu-
ated the alternative attractiveness-exit intention associa-
tion as the level of satisfaction increased. Table 3 shows
SAT*ALT
the alternative attractiveness-exit intention effect coeffi-
cient at selected levels of satisfaction. As shown in this
s 1 a I S l a 2. . . sla 4 s 2 a l . . . $5a4
table, at higher levels of measured satisfaction (e.g., above
4, which corresponded to "agree"), the alternative attrac-
/t tf
~slal ~sla2" 9 "~sla4 ~s2al" 9 .~s5a4
t tiveness-exit intention association was not significant. At
lower levels of satisfaction, the alternative attractiveness-
exit intention association was statistically significant.
NOTE: SAT= satisfaction,ALT= alternativeattractiveness,SAT*ALT These results may have implications for practitioners.
= satisfaction-alternativeattractivenessinteraction,EXI = exiting.
The observed lack of association between exit intention
and alternative attractiveness at higher levels of satisfac-
tion in this study suggests that increases in the attractive-
but it is unknown whether the standard errors and chi- ness of alternatives (competition) may not necessarily
square statistics are robust to departures from normality. tempt satisfied customer firms to consider exiting their
The structural model was therefore reestimated using a less current buyer-seller relationship. In other words, an alter-
distributionally dependent estimator, EQS's Robust esti- native's efforts to be more attractive to a target set of firms
mator, to produce more distributionally appropriate may not, by themselves, increase exit intentions in these
standard errors and chi-square statistics (Satorra and firms; decreased satisfaction in the target firms may also
Bentler 1988; see Bentler 1989, p. 217 et seq.; Hu, Bentler, be required. Only for the less satisfied firms in this study
and Kano 1992). were changes in alternative attractiveness associated with
The structural model was then reestimated in two steps changes in exit intentions. For satisfied firms, changes in
to assess the improvement in fit of adding the interaction alternative attractiveness were not associated with exit
term. The result suggested that adding the interaction term intention.
significantly improved model fit (see the incremental fit This finding suggests that less satisfied firms may be
index in Table 2). The Figure 2 model was also estimated vulnerable to competitive moves aimed at increasing com-
using the Kenny and Judd/COSAN approach for compari- petitor attractiveness, whereas more satisfied firms may
son. Because COSAN provides no distribution-free esti- not be so vulnerable. This in turn may provide some insight
mators, only maximum likelihood estimates were into a basic tenet in the customer service literature, that
produced. superior customer service creates satisfaction, which leads
In addition, Table 2 shows OLS regression estimates for to a competitive advantage (see Davidow and Uttal 1989).
the Figure 2 model as a directional point of reference.5 These results suggest not only that satisfaction may reduce
Regression estimates are known to be biased and ineffi- exit intention but also that it may reduce the attractiveness
cient for variables measured with error (Busemeyer and of the competition. Hence superior customer service may
Jones 1983). As a result, they provide only an indication also reduce exit intention and the attractiveness of the
of the direction (positive or negative) of the Figure 2 competition's superior service.
effects. These regression estimates were produced by av- The model explained 65 percent of the variance in exit
eraging the indicators for satisfaction, alternative attrac- intention, which is notable for marketing studies. This
tiveness, and exit intention, and creating the interaction term result suggests that whereas unmodeled antecedents of exit
by forming the product variable satisfaction-alternative intention may remain to be identified, their combined
attractiveness in each case. contribution to explaining variance in exit intention in this
Ping / ALTERNATIVE ATTRACTIVENESS & EXIT INTENTION 369

TABLE 2
Structural Model Estimation Results
Structural Equation Analysis Estimates t Value
Parameter Estimate Parameter Estimate ML a ROBUST b

~sl 0.79 es5 0.10


~s2 0.88 Eal 0.27
~s3 1.00 Ea2 0.24
~bs4 0.87 ea3 0.08
Xs5 0.94 ea4 0.23
~Lal 0.92 Eel 0.53
~'a2 0.90 ee2 0.19
La3 1.00 ee3 0.11
~a4 0.78 Ee4 0.32
~'e] 0.84 ee5 0.09
Xe2 0.83 ee6 0.12
Le3 1.00 (~SAT 0.51
~4 0.94 ~SAT,ALT -0.37
~5 0.96 (~ALT 0.84
~6 0.92 (~SAT,ALr 17.95
esl 0.16 ~EXI 0.34
es2 0.13 ~EXI, SAT --0.35 --4.59 -4.79
Es3 0.10 [~EXI,ALT 0.65 10.58 9.61
es4 0.12 ~EXI, SAT,ALT -0.05 -5.64 -5.45

Fit indices
Chi-square 88,248 380
p value of chi-square value .000 1.000
Bender (1989) comparative fit index .000 1.000
Null model Z 2 25,884 18,831
Chi-square degrees of freedom 555 555
Squared multiple correlation for EXI .65
Total coefficient of determination .65
Incremental fit index: chi-square difference 3.91
p value .04

Kenny and Judd/COSAN estimates


~EXI, SAT -0.34 4.42
~EXI,ALT 0.67 10.67
[~EXI,SAT,ALT -0.04 5.41

OLS regression estimates


Dependent variable Independent variable b coefficient t Value F value and p R2
EXI SAT -.287 -7.3 79.52 (.00) .52
ALT .745 6.0
SATALT -. 100 -1.8

a. ML = maximum likelihood estimate.


b. ROBUST = ML estimate with EQS Robust option.

context may not be large. For the busy channel manager, are less distributionally dependent should be used for the
this in turn suggests that satisfaction maintenance may be chi-square statistic with the Ping technique.
sufficient for relationship maintenance.
The t values for the maximum likelihood coefficient
estimates and those for the Robust option were similar, SUMMARY
whereas the chi-square estimates were not (see Table 2).
This suggests that standard errors produced by the Ping The study investigated a hypothesized moderating ef-
technique may be robust to product indicator nonnormality fect of overall relationship satisfaction on the association
but that the chi-square statistic is not, and estimators that between exit intention and the attractiveness of an alterna-
370 JOURNAL OF THE ACADEMY OF MARKETING SCIENCE FALL 1994

TABLE 3 H a r l o w 1985; Sharma, D u r v a s u l a , a n d Dillon 1989;


Satisfaction Moderation of the Alternative Tanaka 1984), and strict multivariate normality may not be
Attractiveness-Exit Intention Association required.
4, This adjustment to the degrees of freedom was suggested
SAT Valuea ALT Coefficientb SE of the Coefficient c t Value by Leslie Hayduk (personal communication).
5. The Kenny and Judd approach apparently has been ac-
1 0.596 0.1239 4.80 cepted without any empirical justification. Kenny and Judd
2 0.539 0.1859 2.89 (1984) reported estimates for one synthetic data set for an
3 0.482 0.2510 1.91 interaction and one for a quadratic. There have been no
4 0.425 0.3173 1.33 further investigations to justify or evaluate their technique,
5 0.368 0.3841 0.95 and its performance under varying conditions is unknown.

a. Satisfaction ranged from 1 (low) to 5 in the study."


b. The coefficient of ALT is given by .65ALT - .05SAT*ALT =
(.65 - .05SAT)ALT. REFERENCES
c. The standard error of the ALT coefficient is given by
"~Var(bALT-bsATSAT) Aiken, Leona S. and Stephen G. West. 1991. Multiple Regression: Testing
= ~/Var(bALT)+ SAT2VaI(bsAT)+ 2SATCov(bALT,bsAT). and Interpreting Interactions. Newbury Park, CA: Sage.
Anderson, James C. and David W. Gerbing. 1988. "Structural Equation
Modeling in Practice: A Review and Recommended Two-Step Ap-
tive relationship in a marketing channel. The results sug- proach." Psychological Bulletin 103 (May): 411-423.
Anderson, James C., David W. Gerbing, and John E. Hunter. 1987. "On
gest that the hypothesized moderating effect is present in
the Assessment of Unidimensional Measurement: Internal and Exter-
the study context, and that at higher levels of satisfaction, nal Consistency, and Overall Consistency Criteria." Journal of Mar-
there was no alternative attractiveness-exit intention effect keting Research 24 (November): 432-437.
in the study population. Anderson, Theodore W. and Yasuo Amemiya. 1985. "The Asymptotic
The investigation used the Ping technique and EQS to Normal Distribution of Estimators in Factor Analysis under General
estimate nonlinear latent variable effects. The results were Conditions." Technical Report 12. Econometric Workshop. Stanford
equivalent to those of Kenny and Judd/COSAN and were University.
generally plausible when compared to OLS regression .1986. "Asymptomatic Distribution in Factor Analysis and Lin-
results. In addition, maximum likelihood and Robust ear Structural Relations." Technical Report 18. Econometric Work-
standard error estimates using the Ping technique were shop. Stanford University.
similar, suggesting that maximum likelihood estimates Bentler, Peter M. 1989. EQS Structural Equations Program Manual. Los
Angeles: BMDP Statistical Software.
from the Ping technique may be robust to the departure
Bollen, Kenneth A. 1989. Structural Equations with Latent Variables.
from normality inherent when product indicators are used. New York: Wiley.
The maximum likelihood and Robust chi-square esti- Boomsma, Anne. 1983. On the Robustness of LISREL (Maximum Likeli-
mates, however, were very different. This suggests that hood Estimation) against Small Sample Size and Nonnormality. Dis-
maximum likelihood chi-square estimates are affected by sertation. University of Groningen.
product indicator nonnormality and that chi-square esti- Browne, Michael W. 1987. "Robustness of Statistical Inference in Factor
mates based on less distributionaUy dependent estimators Analysis and Related Models." Biometrika 74 (June): 375-384.
should be used with the Ping technique. Busemeyer, Jerome R. and Lawrence E. Jones. 1983. "Analysis of
Multiplicative Combination Rules When the Causal Variables Are
Measured with Error." Psychological Bulletin 93 (May): 549-562.
Davidow, William H. and Bro Uttal. 1989. Total Customer Service. New
ACKNOWLEDGMENTS York: Harper.
Dwyer, E Robert and Sejo Oh. 1987. "Output Sector Munificence Effects
The author wishes to thank E Robert Dwyer, Leona Aiken, on the Internal Political Economy of Marketing Channels." Journal of
Leslie Hayduk, Robert Peterson, and two anonymous reviewers Marketing Research 24 (November): 347-358.
for their contributions to this article. Dwyer, E Robert, Paul H. Schurr, and Sejo Oh. 1987. "Developing
Buyer-Seller Relationships." Journal of Marketing 51 (April): 11-27.
Ford, David I. 1980. "The Development of Buyer-Seller Relations in
NOTES Industrial Markets." European Journal of Marketing 14(5/6): 339-353.
Fornell, Claes and David E Larker. 1981. "Evaluating Structural Equa-
1. LISREL 8 reduces the specification effort using equations tion Models with Unobservable Variables and Measurement Error."
but generates the d m n m y variables for equations (2) and Journal of Marketing Research 18 (February): 39-50.
(3) using partial derivatives. Larger models, however, still Gadde, Lars-Erik and Lars-Gunnar Mattsson. 1987. "Stability and
require considerable specification effort, and the many Change in Network Relationships." International Journal of Research
d u m m y variables that result produce large matrices that in Marketing 4(1): 29-41.
can create estimation problems (Ping 1993b). Gaski, John and John R. Nevin. 1985. "l'he Differential Effects of
2. The Kenny and Judd normality assumptions were stated Exercised and Unexercised Power Sources in a Marketing Channel."
above equation (2). Journal of Marketing Research 12 (May): 130-142.
3. Maximum likelihood and generalized least squares esti- Garbing, David W. and James C. Anderson. 1984. "An Updated Paradigm
mates are robust to departures from normality (Anderson for Scale Development Incorporating Unidimensionality and Its As-
and Amemiya 1985, 1986; Boomsma 1983; Browne 1987; sessment." Journal of Marketing Research 25 (May): 186-192.
Ping / ALTERNATIVE ATTRACTIVENESS & EXIT INTENTION 371

Harlow, Lawrence L. 1985. Behavior of Some Elliptical Theory Estima- Ping, Robert A. and E Robert Dwyer. 19919"A Preliminary Model of
tors with Nonnormal Data in a Covariance Structures Framework: A Relationship Termination in Marketing Channels9 In Advances in
Monte Carlo Study. Dissertation. University of California, Los Angeles. Distribution Channel Research. Ed. Gary Frazier. Greenwich, CT: JAI,
Hayduk, Leslie A. 1987. Structural Equation Modeling with LISREL: 215-233.
Essential and Advances. Baltimore, MD: Johns Hopkins University Rukert, Robert W. and Gilbert A. Churchill. 1984. "Reliability and
Press. Validity of Alternative Measures of Channel Member Satisfaction."
Hu, Li-tze, Peter M. Bentler, and Yutaka Kano. 1992. "Can Test Statistics Journal of Marketing Research 21 (May): 226-2339
in Covariance Structure Analysis Be Trusted?" Psychological Bulletin Satorra, Albert and Peter M. Bentler. 1988. "Scaling Corrections for
112 (September): 351-3629 Chi-squared Statistics in Covariance Structure Analysis." Proceedings
Hunter, John E. 1973. "Methods for Reordering the Correlation Matrix of the American Statistical Association. Alexandria, VA.
to Facilitate Visual Inspection and Preliminary Cluster Analysis." Sharma, Subhash, Srinivas Durvasula, and William R. Dillon. 1989.
Journal of Educational Measurement 10 (Spring): 51-61. "Some Results on the Behavior of Alternative Covariance Structure
Jtireskog, Karl G. 1993. "Testing Structural Equatton Models." In Testing Estimation Procedures in the Presence of Nonnormal Data." Journal
Structural Equation Models. Eds. Kenneth A. Bollen and J. Scott Long. of Marketing Research 26 (May): 214-221.
Newbury Park, CA: Sage, 294-3169 Tanaka, Jeffrey S. 1984. Some Results on the Estimation of Covariance
J/Sreskog, Karl G. and Dag S6rbom. 19899LISREL 7: A Guide to the Structure Models. Dissertation. University of California, Los Angeles.
Program and Applications. Second Edition. Chicago: SPSS Inc.
Kenny, David and Charles M. Judd. 1984. "Estimating the Nonlinear and
Interactive Effects of Latent Variables." Psychological Bulletin 96 ABOUT THE AUTHOR
(July): 201-210.
Ping, Robert A., Jr. 1993a. "The Effects of Satisfaction and Structural R o b e r t A. Ping, Jr. is an assistant professor in the Department
Constraints on Retailer Exiting, Voice, Loyalty, Opportunism, and of Marketing at Wright State University, Dayton, OH. He ob-
Neglect." Journal of Retailing 69 (Fail): 320-352. tained his Ph.D. at the University of Cincinnati. His research
9 1993b. "Latent Variable Interaction and Quadratic Effect Estima- interests include the political economy of marketing channels and
tion: A Suggested Approach." Technical Report. Wright State University. nonlinear estimation for latent variables.

Vous aimerez peut-être aussi