The Asymmetric Impact of Negative and Positive Attribute-Level Performance on Overall

Satisfaction and Repurchase Intentions

Author(s): Vikas Mittal, William T. Ross, Jr., Patrick M. Baldasare
Source: The Journal of Marketing, Vol. 62, No. 1 (Jan., 1998), pp. 33-47
Published by: American Marketing Association
Stable URL: http://www.jstor.org/stable/1251801
Accessed: 31/08/2010 19:38
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=ama.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Marketing Association is collaborating with JSTOR to digitize, preserve and extend access to The
Journal of Marketing.

http://www.jstor.org

VikasMittal,WilliamT. Ross, Jr.,& PatrickM.Baldasare

of Negative
and
Positive
Attribute-Level
Performance
on
Overall
Satisfaction
and
Intentions
Repurchase

The

Asymmetric
Impact

The relationship between attribute-level performance, overall satisfaction, and repurchase intentions is of critical
importance to managers and generally has been conceptualized as linear and symmetric. The authors investigate
the asymmetric and nonlinear nature of the relationship among these constructs. Predictions are developed and
tested empirically using survey data from two different contexts: a service (health care, n = 4517) and a product
(automobile, n = 9359 and n = 13,759). Results show that (1) negative performance on an attribute has a greater
impact on overall satisfaction and repurchase intentions than positive performance has on that same attribute, and
(2) overall satisfaction displays diminishing sensitivity to attribute-level performance. Surprisingly, results show that
attribute performance has a direct impact on repurchase intentions in addition to its effect through satisfaction.
Consider the following scenarios:
?In an effortto increasecustomersatisfaction,an automotive
manufacturer
consistentlyhas increasedperformanceon
engine power.However,resultsshow that, after a while,
increasesin the ratingsfor performance
on enginepowerdo
not yield correspondingchanges in satisfactionratings.
Designerswonderif theyshouldkeepinvestingresourcesto
enhanceenginepower.
*A physicianhistoricallyhas receivedhigh performance
ratsuchas politeness,timeliness,reliabilityof
ingson attributes
service,andcorrectdiagnosis.Inthecurrentsurvey,shefinds
thatperformance
ratingson timelinessaredown,butareup
on all the otherattributes.However,she also findsa sharp
declinein heroverallsatisfaction
ratings.She is puzzledas to
on a singleattributeis not offset
why negativeperformance
on a hostof otherattributes.
by positiveperformance
These scenarios highlight typical dilemmas that managers face at customer-drivenorganizationswhere the goals
are to design productsand services with attributesthat maximize customer satisfaction.These goals generally are operationalized by multiple-regressionmodels that identify key
attributes into which managers should invest resources to
enhance overall satisfaction (e.g., Bolton and Drew 1991;
Hanson 1992; Wittink and Bayer 1994). The assumption
underlyingsuch "key driver"models is that there is a symVikasMittal
is a visiting
assistantprofessor,
Graduate
Schoolof
Kellogg
Northwestern
William
T.Ross,Jr.is anassociate
Management,
University.
Patrick
M.Baldasare
is President
andChief
professor,
Temple
University.
Executive
Officer
of TheResponseCenter,a division
of TeleSpectrum
Inc.Theauthors
thehelpful
comments
andsugWorldWide,
acknowledge
Oliva,D.Raghavarao,
gestionsof EugeneAnderson,
Terry
SoumyaRoy,
andfouranonymous
reviewers.
Theauthors
thankFrank
Mark
G.
Forkin,
andAlanRogersof Consumer
Attitude
Inc./Research
Konkel,
Research,
DataAnalysis,
Inc.forproviding
theautomotive
data,andJamesCuthrell,
Kristine
Michael
andE.Monsoon
fortheir
Gerhart,
Gupta,
JerryKatrichis,
support.
Journal of Marketing
Vol. 62 (January 1998), 33-47

metric and linear relationshipbetween attribute-levelperformanceand dependentconstructssuch as overall satisfaction and purchase intentions. However, what if, at the
attribute level, performance and satisfaction were linked
asymmetrically?The answer to this question is important
for academics and managersalike. Although academics (cf.
Anderson and Sullivan 1993; Oliva, Oliver, and Bearden
1995) and practitioners(cf. Coyne 1989) recognize that the
overall satisfactionfunction might not be linearand/orsymmetric, no such conclusion can be offered at the attribute
level. Yet,it is at the attributelevel thatmanagersmust make
decisions.
Therefore,our objective is to investigate-theoretically
and empirically-the existence and natureof the asymmetric and nonlinearresponse of satisfaction to attribute-level
performance.In the first section, we discuss the need for an
attribute-levelconceptualization of the performance/satisfaction relationship.That is followed by a discussion of the
asymmetricand nonlinearnatureof the relationshipand the
development of hypotheses. Then we present results from
three studies that test the hypotheses.

Multi-AttributeProducts
and Satisfaction
There are several reasons-theoretical and managerial-to
use multi-attributemodels in the context of customer satisfaction (cf. LaTourand Peat 1979; Wilkie and Pessemier
1973). First, consumers are more likely to render evaluations of their postpurchaseexperiences of satisfaction at an
attributelevel ratherthan at the productlevel (Gardialet al.
1994). Second, an attribute-based approach enables
researchersto conceptualize commonly observed phenomena, such as consumersexperiencing mixed feelings toward
a productor service. A consumer can be both satisfied and
dissatisfied with different aspects of the same product.
Performance/ 33
Negativeand PositiveAttribute-Level

Although such phenomena are not easy to model in an

overall satisfaction approach, the attribute-level approach
provides a simple and elegant solution: Mixed feelings
towarda productexist because a consumer may be satisfied
with one attributebut dissatisfied with another.For example, in a restaurant,a customer may be highly satisfied with
the food but highly dissatisfied with the service at the same
time. Third, an attribute-level approach to satisfaction
affords researchersa higher level of specificity and diagnostic usefulness comparedwith the productlevel or "overall" approach(LaTourand Peat 1979). For example, Parasuraman,Zeithaml, and Berry (1988) show that measuring
various attributesor dimensions of a service provides a better understandingof global constructs,such as service quality. Similarly, Boulding and colleagues (1993) suggest that
quality is multidimensional and different dimensions of
quality are averaged together in some fashion to producean
overall assessment of quality. However, both works treat
the relationship between attributesand overall quality as
linear and symmetric. Fourth, the higher specificity and
diagnostic usefulness of multi-attributemodels also ensures
that academic research will find more welcome applications among managers who generally work at the attribute
level ratherthanat an overall level (e.g., Hanson 1992; Wittink and Bayer 1994). Finally, there is some evidence that
attribute-level performance/disconfirmation and overall
satisfaction are qualitatively different constructs (Oliva,
Oliver, and Bearden 1995), and, if treatedinterchangeably,
"global customer satisfaction responses may mask specific
product issues" (Oliva, Oliver, and Bearden 1995, p. 26).
Thus, studying satisfaction at the attributelevel can help
extend both conceptual and empirical understandingof the
phenomenon.

The Asymmetric Effect of Product

Performance and Satisfaction:
Theory and Hypotheses

When examining the theoretical and analytical importance

of the link between attribute-levelperformanceand overall
satisfaction, it is importantto recognize thatthe relationship
could be asymmetric (cf. Anderson and Sullivan 1993;
Oliva, Oliver, and Bearden 1995). One unit of negative performanceon an attributecould have a greatereffect on overall satisfaction or repurchaseintentions than a corresponding unit of positive performance.Similarly,in a given set of
attributes,negative performanceon a single attributecould
outweigh positive performance on many other attributes
combined. Oliver (1993) finds that (1) attribute-levelsatisfaction and dissatisfaction significantly affect overall satisfaction with a product(automobile)and a service (an undergraduate course offering) and (2) attribute dissatisfaction
has a larger weight than attributesatisfaction for the productI but not for the service. However,no theoreticalmotivation for the observed disparity between the impact of
attributesatisfaction and dissatisfaction is provided. In the
tStandardized
loadingswere as follows:attributesatisfaction
and dissatisfactionfor the automobile(.135 and -.175, respectively)andthecourse(.130 and-.059, respectively).
34 Journalof Marketing,January1998

next sections, the theoretical logic is developed along two

lines of reasoning.One is based on prospecttheory (Kahneman and Tversky 1979), and the other is rooted in cognitive
researchthat examines the memorabilityof positive versus
negative events.
Prospect Theory
Prospecttheory(Kahnemanand Tversky 1979) is a descriptive theoryin which all of the alternativesthata personfaces
are reducedto a series of prospects that are evaluatedindependently on the basis of an S-shaped value function (see
Figure 1).
As depicted in Figure 1, prospect theory postulatesthat
peoples'judgments display (1) referencedependence(carriers of value are gains and losses from a referencepoint) and
(2) loss aversion (the functionis steeper in the negative than
in the positive domain). In addition, evaluations display
diminishing sensitivity (marginal values of both gains and
losses decrease with their size). The two key propertiesof
the value function for our discussion are loss aversion and
diminishingsensitivity.
The loss aversionbuilt into prospecttheorysuggests that
losses loom largerthan gains (Einhom and Hogarth 1981).
Psychologically, a one-unit loss is weighted more than an
equal amountof gain. In a satisfactioncontext, negative outcomes on attributeperformanceshould carrymore weight in
the overall satisfactionjudgmentthanequal amountsof positive outcomes on attributeperformance.For example, if a
car's mileage were to decrease by 10 miles per gallon, it
would have a greaterimpacton the overall satisfactionjudgment than if the car's mileage were to increase by 10 miles
per gallon. Thus, negative performanceon an attributewill
loom largerthanpositive performanceon the same attribute.
On the basis of these theories, we propose
FIGURE 1
Asymmetric Impact of Attribute-Level
Performance

Negative
Performance
on Attribute

II

i
OverallSatisfaction/
RepurchaseIntentions

Positive
Performance
on Attribute

on an attributewill havea greater

H ,,:Negativeperformance
impacton overallsatisfactionthanpositiveperformance
on thesameattribute.
on an attributewill have a
Hlb:Negativedisconfirmation
greaterimpacton overallsatisfactionthanpositivedison thesameattribute.
confirmation
In addition, on the basis of prospect theory,overall satisfaction also should display diminishing sensitivity toward
attributeperformance.That is, at high (low) levels of performance, positive (negative) performanceon an attribute
should not affect satisfaction as dramaticallyas it does at
lower levels of performance.This hypothesis is similar to
the diminishing returns hypothesis in classical economics
and also is depicted in Figure 1. Therefore,
H2:Overallsatisfactionwill displaydiminishingsensitivityto
changesin the magnitudeof performancefor a given
In otherwords,at highlevelsof positiveor negattribute.
ativeperformance
on an attribute,
overallsatisfactionwill
be affectedless thanat intermediate
levelsof performance
on thatattribute.
Memory for Negative Versus Positive Instances
Researchers distinguish between transaction-based and
cumulativesatisfaction(Bitnerand Hubbert1994). Cumulative satisfaction typically is conceptualized as an overall
judgment based on several transactionswith a product or
service. To the extent that all aspects of the transactionsare
not equally accessible (Gardialet al. 1994), the overall satisfactionjudgment will vary on the basis of the accessibility
of a given aspect. Memory accessibility is a function of
stimulus salience, among other things (Taylor 1982). Evidence shows that negative informationis more perceptually
salient than positively valenced information,is given more
weight than positive information, and elicits a stronger
physiological response than positive information (Peeters
and Czapinski 1990).
Similar psychological operations should occur for customer satisfactionbecause satisfactionis linked to memorybased processing (Yi 1990). To the extent that attributes
with negative performancewill be more perceptuallysalient
than attributes with positive performance, attributes with
negative performanceshould have a greater impact on the
cumulative satisfactionjudgment. Thus, within a given set
of attributes,the relative impact of each attributewill be
asymmetric.Consequently,when combined, attributeswith
negative performanceshould have a greaterimpacton overall satisfactionthan theircorrespondingattributeswith positive performancecombined. On the basis of the previous
discussion, we hypothesize that
H3a: Cumulatively, attributeswith negative performancewill

have greaterimpactson overallsatisfactionthanattributeswithpositiveperformance.

Along similar lines, a diminishingsensitivity hypothesis

between overall satisfaction and performance on various
attributescan be proposed. In a given set of attributes,each
additionalinstance of positive performanceon an attribute
will have a smaller impact than the other attributes.Conversely, each additional instance of negative performance
should have a correspondinglysmaller negative impact on
overall satisfaction.Thus,

willdisplaydiminishing
sensitivityto
H3b:Overallsatisfaction

additionalinstances of negative or positive performance.

In otherwords,each additionalinstanceof positiveor

shouldhave a smallerimpacton
negativeperformance
overallsatisfaction.

Repurchase Intentions
Attribute-level performanceshould affect satisfaction and
repurchase intentions differently (e.g., Ostrom and
Iacobucci 1995). One reason could be that satisfaction and
repurchaseintentions are qualitatively different constructs.
Satisfaction may be merely a judgment with cognitive and
affective dimensions, whereas repurchase intentions also
have a behavioral component. Based on the consumer's
goals (e.g., Mittal et al. 1993), performanceon a certain
attributemay become crucial for repurchaseintentions but
not satisfaction. For example, consider the case of a patient
who is satisfied with all aspects of the service provided by
his or her primary care physician (PCP), except that the
patient has now relocated to another area far from the
PCP's office. When it comes time to renew and choose a
PCP, the patient might indicate high overall satisfaction
with the PCP but still might choose another PCP, because
performance on a critical attributehas changed. In other
words, though the performance on "distance of PCP's
office from home" had a negative impact on repurchase
intentions, it has a small or virtually no impact on overall
satisfaction. One reason for such a discrepancy can be
found in the attributionliterature(Folkes 1988). Here, the
patient may attribute"poor" performanceon "distance of
PCP's office from home" to him or herself or to causes
other than the PCP and thus not change his or her overall
satisfaction rating; nevertheless, the repurchaseintentions
of the patient change. Thus,
will
H4:Overallsatisfactionand attribute-level
performance
haveseparateanddistincteffectson repurchase
intentions.
However, the relative magnitudeof overall satisfaction
and attribute performance should be driven contextually,
and no specific predictionsare made in that regard.H4 simply states that attribute-levelperformanceand overall satisfaction will have separate and distinct impacts on repurchase intentions.
Next, we proposean asymmetriceffect of attribute-level
performance on repurchase intentions. Previous research
(e.g., Oliva, Oliver, and MacMillan 1992) shows that overall satisfaction and performanceare related nonlinearly to
repurchase intentions or loyalty. Feinberg and colleagues
(1990, p. 113) found that across several productcategories
"the probability of repurchase was not isomorphic with
either positive or negative service experiences." Research
also shows that satisfaction and dissatisfactionhave different affective consequences (Oliver 1993), which may be
related differentiallyto repurchaseintentions.Furthermore,
the same psychological principlesthatoperatein the context
of satisfaction also should operate in the context of repurchase intentions.Therefore,
intentionswill be affectedasymmetrically
H5:Repurchase
by
attribute-levelperformance.Negative performanceon
attributeswill havea greaterimpacton repurchase
intentionsthanpositiveperformance.
Performance/35
Negativeand PositiveAttribute-Level

In the next sections, we presentresults from three largescale studies that test the preceding hypotheses. A unique
feature of all three studies is that they use data from commercial studies conducted by firms measuring satisfaction
with their productor service.

Study 1
Data for this study were providedby a large health maintenance organization (HMO) in the Northeastern United
States. The HMO collected these data as partof its ongoing
patient satisfaction measurementprogram. For this study,
the entire data set, which contained informationfrom 4517
telephone interviews conducted among patients enrolled
with the HMO, was used. These interviews were conducted
using Computer-Assisted-Telephone-Interview, which
enabled the interviewerto probe, clarify, and follow up on
responses to open-ended questions. Because responses to
the open-ended questions were a vital part of the measurement process, interviewerswere asked to exercise greatcaution in recordingand following up on them.
The sampling framefor the data set consisted of patients
who subscribedto 501 PCPoffices affiliatedwith the HMO.
A random sample was drawn from each PCP's patient list.
Thus, in effect, a proportionate-stratified
sample, with each
PCP as a stratum,was drawn. After contact with a household had been established, up to two members per household were interviewed as long as they evaluated different
PCPs. Thus, each household provided only one evaluation
per PCP.The qualified respondentwas an adult from a family whose members had visited a PCP within the past 12
months. All 4517 interviews were used in the analysis.
Although the exact responseratefor this survey could not be
ascertained,typical responseratesfor such surveys are more
than 70%. Input from the managersat the HMO and informal comparisons with data from other waves of the study
show that the sample is representativeof the population
served by the HMO.
Measures
The database contained measures for several constructs.
There were two dependentmeasuresfor this analysis: overall satisfaction and repurchaseintentions."Overallsatisfaction with the medical care at the PCP's office" was operationalized as a five-point scale (1 = very dissatisfied, 5 =
very satisfied). Repurchaseintentionor a patient'sintention
to switch to a different PCP was operationalizedas a yes or
no response to the following question: "Do you intend to
switch to a different primarycare physician when you next
have an opportunity?"Although both these measures are
single-item scales, there is considerableprecedentfor using
single-item measures in the context of large-scale satisfaction studies. LaBarberaand Mazursky (1983) discuss the
issue of using single- versus multi-itemscales for measuring
overall satisfaction and conclude that, in large-scale survey
research,the use of multi-itemscales actually may decrease
the quality of measurementratherthanenhanceit. Similarly,
in their large-scale study of driversof customer satisfaction
in the computer industry,Kekre, Krishnan,and Srinivasan
(1995) use a single-item measure for overall satisfaction.
36 / Journalof Marketing,January1998

Finally, Yi (1990), in his review of satisfaction research,

compares the test/retest reliabilities of several multi- and
single-item scales and finds that the test/retestreliabilityof
single-item scales is acceptable (.55 to .84). For these reasons, we believe the measuresare adequate.
Independentvariableswere createdon the basis of openended data providedby patients.In a free-thoughtelicitation
task duringthe telephone interview,respondentswere asked
about theirexperiences with variousfeaturesof theirvisit to
the PCP.The interviewerasked them to clarify and expand
their thoughts.
These thought listings provided the basis for the independentvariablesfor the first analysis.As shown in Table 1,
each thought was coded in a distinct category. When
queried,the projectmanagerfor this study indicatedthatthe
categories were determinedby the managersat the HMO in
conjunction with the project director. The process is to
review all the open-endeddatafrom a randomlychosen subset of interviews (usually 10%) and develop the categories
so that they are mutuallyexclusive and collectively exhaustive. Then the coders code a portionof the data using these
categories. After these data are coded, the coders and managers meet to see if the categories are "working,"that is, if
the categories are descriptive and capturethe "essence" of
what respondents said. Furthermore,when making categories, the managersensure that all attributesand thoughts
are coded. In any given study,fewer than2% of the thoughts
are forced into the "other/not specified" category. The
process of determining categories is done carefully, especially given the high cost of datacollection and the strategic
value of the informationcontained in these data.
For the purpose of this analysis, each category was
coded as 1 if a respondentgave a response in that category
and 0 otherwise. Three professionalcoders, each with more
than 2-3 years of coding experience, coded the categories.
However, each coder coded a separate set of respondents.
Because the coding already had been done before the data
set was made available for this study,no follow-up analyses
could be done to assess the reliability and agreementof the
coding. Therefore,to assess the agreementamong the three
coders, they were asked to code data from 337 respondents
from a subsequent wave of a similar study conducted for
another HMO in the same geographic area. For these 337
respondentsof the pilot study, the three coders were in perfect agreementfor 81% of the surveys.That is, coders coded
every category identically for 81% of the respondents.Furthermore,Coders I and 2 matchon 84.3% of the responses,
Coders 2 and 3 match on 85.5%, and Coders I and 3 match
on 85.2%. The associated Kendall's tau correlationwas as
follows: .84 for Coders 1 and 2, .85 for Coders I and 3, and
.85 for Coders 2 and 3 (all significant at p < .01).
With these limitations in mind, it can be said that such
an approachto understandingsatisfactionhas been used by
otherconsumersatisfactionresearchers(Bitner,Booms, and
Tetreault1990; Gardialet al. 1994). This approach,though
more unstructured than rating scales, is advantageous
because it capturesattributeperceptionsthat are not contaminatedby the preconceived notions that a researchercould
impose on the respondentsby cuing them regardingspecific
attributes.The thought categories then were classified fur-

ther as positively or negatively valenced. The positively

valenced items for each respondentwere summed to create
a new variable, POSITIVE. Similarly, a variable called
NEGATIVEwas createdby summing all the negative items
for each respondent.The positive/negative classification of
each thoughtcategory is indicatedin Table 1.
Analysis and Results
The general strategy employed for testing the proposed
asymmetry consists of computing separate estimates for
positive and negative performanceand statisticallycomparing the absolute magnitude of the positive and negative
estimates. The absolute magnitude is compared because
negative performance should be related negatively to the
dependentconstructs,and positive performancerelatedpositively to the dependent constructs. Therefore, to get a
proper comparison, the "signs" of the estimates must be
ignored.As an illustration,a test of the asymmetrybetween
the composite variables POSITIVE and NEGATIVE is

done, in which the coefficient for variable POSITIVE is

expected to have a positive sign, and the coefficient for the
variable NEGATIVEis expected to have a negative sign.
> [bposITIVEI sugTherefore, testing whether IbNEGATIVEI

gests the following hypothesis:

(1)

HA: IbNEGATIVEI> IbPOSITIVEI-

The correspondingnull hypothesis can be stated as follows:

(2)

Ho: IbNEGATIVEI < IbPOSITIVEI

To test HA, two rival models are compared to see

whether Ho can be rejectedstatisticallyin favor of HA.One
< IbPOSITIVEI
model formally constrainsIbNEGATIVEI
as sugin
whereas
in
the
the
coefficients
are
free
other,
gested Ho,
to vary as suggested in HA. Comparingthe performanceof
the constrained model to that of the unconstrainedmodel
provides a test of the asymmetry.If the model with the constraintis rejected in favor of the alterative model (i.e., Ho
is rejected in favor of HA), it can be concluded that the

TABLE 1
and
Their
Relative
Thought Categories
Impact on Overall Satisfaction with Medical Care
Positive Mentions

Negative Mentions
Coefficient

Thorough/attentive
Spends time withme/personable
Not afraidto referto anotherdoctor
Listens to his/her patients
Convenientoffice hours
Convenientoffice location
of doctor
Alwaysavailable/availability
No wait/sees patients rightaway

Coefficient

Matched Categories
.11***
Not thorough
.17****
Doesn't spend enough time/rushesyou in
and out
.07
Won'tgive referrals/trouble
getting referrals
.03
Not interested in patients/doesn'tlisten
.07
Office hours are not convenient
.02
Distance of office is inconvenient
.15***
Not available/isn'talways in office
.14**
Waittoo long/officetoo crowded

-.66****
-.40****
-.77****
-.94****
-.37****
.07
-.67****
-.38****

Unmatched Categories
Good, excellent doctorcare
Knowledge/educated/competent
Explainseverythingwell/answersquestions
Helpful
Good bedside manner
Professional
Friendly/nice/courteous
Concerned/caring/sympathetic
Good withkids
Makes me comfortable/easyto talkto
Verypatient
Sincere/honest/straightforward
Outgoing/goodpersonality/sense of humor
Is warm/kind
Is an old-fashionedtype of doctor
Calls back rightaway/keeps in
contact/followsthorough
Alwaysavailableduringemergency
Promptnessof service
Has been our doctorfor a long time
I like him/her
Staff is very nice/good
Cleanliness of office/building
*p < .10;<

**p.05** <

.26****
.02
.12***
.13*
.13
.13
.11***
.22****
.19****
.06
.15
.15
.25*
.19
.14

No followup/doesn'ttell you results

Doctor/staffoverbooksappointments
Poor staff/problemwithstaff
Problemswith HMO
Incorrectdiagnosis/unnecessarytesting

-.65****
-.60***
-.84****
-.35****
-.99****

.17***
.32
.02
.33****
.20***
.14**
.05

* **** < .
; p
001; Model R2 = .23, F = 30.9, error df = 4481, p < .0001.

Negativeand PositiveAttribute-Level
Performance/ 37

absolute magnitude of the coefficient for NEGATIVE is

> IbposITIVE
greater than POSITIVE, or that IbNEGATIVEI
Next, the results are described.
Negative versuspositive attributeperformance(H/a). A
dummy-variableregression was conducted using all of the
thought categories as independentvariablesand overall satisfaction as the dependent variable.2As shown in Table 1,
there were 30 positively valenced thoughtcategories and 13
negatively valenced thought categories. On the basis of
these, the following model was estimated:
(3)

Overall Satisfaction= Intercept+ Attpos ... + Attpos30

+ Attnegl ... + Attneg13

In this model, the impactof each attributeon overall satisfaction is calculated individually.Furthermore,positively
valenced attributementions are expected to have positive
signs, whereas negatively valenced attributementions are
expected to have negative signs. To test Hla, we must test
whether the absolute impact of a negative mention for an
attribute is greater than the absolute impact of a positive
mention for the same attribute.The model is significant (F =
30.9, p < .0001) and is reportedin Table 1.
Several findings of interestemerge from this table. First,
as shown in the first eight rows of Table 1, there were eight
attributesfor which there were both negative and positive
categories. Therefore,for these eight attributes,Hia could be
tested. For all eight pairs, the negatively valenced attributes
have a larger(in the absolute sense) coefficient than the corresponding positively valenced attributes.We used the statistical test explained previouslyto test whetherthe negative
coefficient for each pair of attributesis largerthan the corresponding positive coefficient for the same attribute;the
null hypothesis in Equation2 was rejectedfor seven of the
eight pairs at p < .001. For example, "always available/
availability of doctor"(.15, p < .01) has a smaller impacton
patient satisfaction than "not available/isn't always in his
office" (-.67, p < .0001). Similarresultscan be noted for (1)
"not afraid to refer to anotherdoctor"(.07, p < .26) versus
"won't give referrals/troublegetting referrals"(-.77, p <
.0001) and (2) "spends time with me/personable"(.17, p <
.0001) versus "doesn't spend enough time/rushesyou in and
out" (-.40, p < .0001). Only for "convenient office location," for which both the positive and negative coefficients
were nonsignificant, was the hypothesis not supported.
These results support Hia. However, because this study is
only able to provide a sense of the direction of a person's
evaluation on each attribute,a full test of H,, cannot be
done. In other words, because there are no data on thei,rgnitude of positive or negative performanceon each attribute,
additional testing for this hypothesis is required.Recall that
though Hla is stated in terms of varying levels of performance on each attribute,there are no data on varying levels
of performancein this study. Nevertheless, these results are
encouraging and point toward tthetype of investigationthat
is carriedout in Study 3.
2Acorrelationmatrixforall thoughtcategoriesshowedthehigh-

Second, differentattributesof the service encountercontribute differently to overall satisfaction. For example,
"cleanlinessof office/building" (.05, p = .78) is not consequential, whereas "no follow-up/doesn't tell you results"
(-.65, p < .0001) has a greaterimpacton overall satisfaction.
Several othercomparisonscan be made with similarresults.
These echo earlier findings (Bolton and Drew 1991) that
service attributesmay be core, facilitating, or supporting
and provideconfidence in our results.
Third, with an R2 of 23%, the overall variation
explained by the model appears to be low. However, this
could have resultedpartiallyfrom loss of informationresulting from the dichotomousnatureof the independentvariable
(Cohen 1983). Furthermore,because the main concern in
this study is to test the asymmetry hypothesis ratherthan
explain variationin the dependentvariable,the low R2 is not
particularlytroublesome.
Cumulative impact and diminishing sensitivity (H3a
and H3b).To test H3aand H3b,the following equation was
estimated:
+ b2(POSITIVE)2
(4) Overallsatisfaction= bi(POSITIVE)
+ b3(NEGATIVE)+ b4(NEGATIVE)2.

In Equation4, POSITIVEand NEGATIVEare the sums

of positively and negatively valenced mentionsof attributes,
respectively. Therefore, bl and b3 model the asymmetric
cumulative impact of positive and negative performanceon
overall satisfaction. To test for diminishing sensitivity, the
squaredtermsare introduced(cf. Hamilton 1992). Thus, for
positive performance,bl and b2 model diminishingsensitivity, whereas for negative performance, b3 and b4 model
diminishing sensitivity. These tests (for H3a and H3b) are
describedin detail next.
Cumulative impact of number of positive or negative
attributementions(H3a).As statedpreviously,bI and b3 can
be used to test the asymmetryfor the cumulative impact of
negative and positive mentionsof attributeson overall satisfaction statedin H3a.Because negative mentionsof attribute
performanceare hypothesized to have greater impacts on
overall satisfactionthanare positive mentions,lb3\shouldbe
greaterthan Ibli.Similar to the test of Hia, two rival models
are comparedto test whether the model with the constraint
Ib31< Ibil performsbetter or worse than the unconstrained
model. If the constrained model performs worse than the
unconstrainedmodel, the constraintis rejected, which supports|b31> Ibil.Results of the analysis are shown in Table 2.
The coefficients for POSITIVEand NEGATIVEare .29 and
-.72, respectively. Both are in the expected direction and
significantat p < .0001. Moreover,theirabsolute magnitude
is as hypothesized;the absolute value of the coefficient for
NEGATIVEis greaterthan the absolute value of the coefficient for POSITIVE.This was confirmed by comparingthe
constrainedand unconstrainedmodels as explained previously. The model with the constraintlb31< Ibll was rejected
in favor of the unconstrainedmodel, where lb3l> Ibjl(Fi,4477
= 50.37, p < .0001). Thus, H3ais supported;3cumulatively,

est con-elation to be .21. A majorityof the correlationswere non-

significant(p > .05) andbelow .10 in absolutemagnitude.

Thus,

multicollinearitydoes not appearto be an issue here. Furthermore,

thethoughtcategoriesarenot linearlydependenton one another.

38 / Journalof Marketing,January1998

3These models also were estimated using standardizedcoefficients. The results for the models with standardizedcoefficients
were practicallyidentical to the results reportedhere.

TABLE 2
Regression Results for Satisfaction
With Medical Care

Variable

was used alone, the function becomes a curve that "tapers

off' when the squaredterm (b2) is included in the equation.
This occurs because every increase in bl is countered by a
smaller (because of the squared term) decrease in b2. For
example, the estimate for bl is .28, whereasthe estimate for
b2 is -.04. Considerthe effect of increasingpositive performance for one and two attributes. In the absence of the
squared term, the beneficial impact of one attributeis .29
and of two attributesis .58; however, with b2, the beneficial
impact increases but at a decreasingrate. Using both bl and
b2, the beneficial impactof one attributeis .25, calculatedas
.29(1) - .04(1)2, whereas the impactof two attributesis .42,
calculatedas .29(2) - .04(2)2. Thus, inclusion of the squared
term capturesdiminishing sensitivity. Similar results apply
for b3 and b4 in the domainof negative performance.Results
based on this analysis are reportedin Table 2.
As seen in Table 2, all four coefficients are statistically
significant and in the predicted direction, which thereby
supports H3b. However, a key concern when using these
data to test H3bis that supportfor H3bmay be an artifactof
the way the variablesPOSITIVEand NEGATIVEwere constructed.Some of the attributesthat went into the construction of the variable may be inherently less importantor
salient than others. Thus, the test for H3bmay be simply an
artifactof the lower importanceor salience of some attributes used in the constructionof the cumulativevariablethan
others. Additional tests, described in the Appendix, were
used to test for this possibility. In these tests, the variables
POSITIVEand NEGATIVEwere constructedin eight differentways, and H3bwas tested eight differenttimes. Therefore, including the base case described in Table 3, a total of
nine tests were conducted for H3b.All nine tests supported
the diminishing sensitivity in the positive performance
domain, but only five supporteddiminishing sensitivity for
negative attributeperformance.Therefore,erringon the side

Standardized
Regression Standard
Coefficient
Error t-statistic p-value

Intercept
POSITIVE
NEGATIVE
POSITIVE2
NEGATIVE2

4.24
.29
-.72
-.04
.07

R2 = .23; F4,4477 = 330.65,

.02
.02
.05
.01
.02

199.48
13.87
-13.48
-7.78
3.01

.0001
.0001
.0001
.0001
.0030

p < .0001.

Test for constraint for Ho in Equation 2: F1,4477= 50.37, p <

.0001 (one-tailed test).

negative performance on attributes has a significantly

greater impact on overall satisfaction than positive performance on attributes.The regression coefficients show that
for this particularservice, the deleterious effect of an additional instance of negative performanceon attributesoutweighs the beneficial effect of positive performanceby a
marginof two to one (-.72 versus .29).
Diminishingsensitivity(H3b). This hypothesis suggests a
nonlinearrelationshipbetween overall satisfactionand performance on additionalattributes.This relationshipis modeled using the polynomial function describedin Equation4.
In Equation4, bl and b2 test for diminishing sensitivity in
the domain of positive performance. The linear positive
coefficient bl models the positive impact of performance,
whereas b2, the coefficient for the squared term, models
diminishing sensitivity by creatinga "dampening"effect on
the slope of bl. Instead of being a straight line as when bl

Logistic Regression

TABLE 3
for Intention to Switch to a Different PCP

Variable

Model 1
Attribute

Intercept

-1.50

Model 2
Overall
Satisfaction
4.17

(.07)
POSITIVE
NEGATIVE

-.66
(.05)
1.21
(.08)

Overall satisfaction
Chi-square (d.f.)
AIC

(.23)

3.14
(.24)
-.43
(.05)

708.7 (2)
(p < .0001)
2905.22

Model 3
Attributeand
Overall
Satisfaction

-1.45
(.06)
989.9 (1)
(p < .0001)
2606.10

.76
(.08)
-1.15
(.06)
1182.2 (3)
(p < .0001)
2417.71

% correctly classified

86.0%
88.1%
89.5%
Allcoefficientestimates are significantat p < .001; standarderrorsare in parentheses.
It could be argued that POSITIVEand NEGATIVE
are not 2 but 43 variablesbecause they are composed of multiplebinaryvariables.Therefore, the AICalso was recalculatedafterpenalizingModels 1 and 3 for43 ratherthan 2 variables.The results for
model fitwere virtuallyidentical to the one reportedin the table.

Negativeand PositiveAttribute-Level
Performance/ 39

of conservatism, the conclusion is that diminishing sensitiv-

ity holds for positive performancebut not for negative performanceon attributes.
These results suggest that overall satisfaction displays
diminishingsensitivity to each additionalinstanceof positive
performancebut not necessarilyto negative performance.In
other words, though the relative benefit of positive performance on additionalattributesmay decrease,each additional
instanceof negative performancemay be equallydeleterious.
Repurchase intentions (H4 and H5). Because the measure for switching intentions was a dichotomousvariable, a
logistic regressionanalysis was conductedto test H4and H5.
Three models predicting intention to switch to a different
PCP were estimated. Note that though the hypotheses are
stated in terms of repurchase intentions, results are discussed in terms of intention to switch to maintain consistency with the data. The models used for testing H4 and H5,
which are shown in Table 3, are as follows:
NEGATIVE};
(5) Model1:intentionto switch= f POSITIVE,
(6) Model2: intentionto switch= f{Overallsatisfaction);
NEGATIVE,
(7) Model3: intentionto switch= f{POSITIVE,
Overallsatisfaction}.
In Model 1, only POSITIVEand NEGATIVEwere used
as predictorvariables. In Model 2, only overall satisfaction
was used as a predictorvariable.The thirdmodel used POSITIVE, NEGATIVE,and overallsatisfactionas predictorvariables. Note that these models are concernedonly with asymmetry,not diminishingsensitivity;therefore,only the cumulative variablesPOSITIVEand NEGATIVEare used in these
models. Results from these models are reportedin Table 3.
Results for Model 1, the attributes-onlymodel, confirm
the hypothesis that negative attributesare more consequential for intention to switch than are positive attributesof the
service. Thus, H5 is supported. Next, all three models in
Table 3 were examined to test H4.The fit of the models was
examined using the Akaike InformationCriterion(AIC statistic) of model evaluation(Bollen 1989).The AIC evaluates
alternativemodels while accountingfor the numberof parameters in rival models; it penalizes models with many rather
than few free parameters.Models with smaller AIC values
provide a better fit. The AIC for each of the three models is
shown in Table 3.

The overall satisfactionmodel (Model 2) provides a better fit in predicting repurchase intentions than does the
attribute performance model (Model I); the AIC fit for
Model 2 is smaller thanthatof Model 1. However,Model 3,
which includes both attributeperformanceand overall satisfaction, outperformseither model. The value of the AIC for
Model 3 is smaller than either the overall satisfactionor the
attribute performance model. Thus, H4 is supported.4
4These results also were verified by dividing the data set in two
parts: one with 3517 and the other with 1000 people. The three
models were estimated with the first data set and then used to predict the value for intentionto switch in the holdoutsample of 1000
people. The coefficient estimates were virtually identical to the
ones reportedin Table 4, and the percentagescorrectly classified
for the three models were 65%, 69.9%, and 78.9%, respectively.
These results again supportthe hypotheses.

40 / Journalof Marketing,January1998

Attributeperformanceand overall satisfactionhave separate

and distinct effects on repurchaseintentions.
Summaryof resultsand limitations.The resultsfrom this
study support the hypothesized asymmetry and provide a
strong indicationof the existence of diminishingsensitivity,
at least on a cumulative basis. Negative performanceon
seven of eight pairs of attributeshas a greater impact on
overall satisfaction than does positive performance.This is
also true of the cumulative impact of negative versus positive attributes. Similarly, intentions to switch also are
affected asymmetrically by positive and negative performance. Overall satisfaction appearsto display diminishing
sensitivityto attributeperformancein the domainof positive
performancebut not negative performance;however, this
conclusion cannot be offered for different levels of performance within each attribute.Finally, both overall satisfaction and attributeperformancehave separate and distinct
impactson repurchaseintentions.
These results should be viewed in light of the study's
limitations.First, there were no data on expectationdisconfirmationat the attributelevel; this precludedtesting Hlb, a
limitation that is addressed in Study 2. Second, attributelevel performance was measured using 0,1 categories,
which precludedmeasurementof differentlevels of performance for an individualattribute.Thus, Hla and H2, which
are statedfor differentlevels of performancewithin an individual attribute,could not be tested. Study 3, in which performancelevels are measuredfor each attribute,fully examines Hia and H2. Third, Study 1 was conductedin a service
setting; therefore,it could be argued that the results of this
study would not necessarily generalize to a product-related
setting. Although there is no theoretical reason to expect
this, testing these resultsin a productsetting should increase
confidence in the generalizability.

Study 2
The purposeof this study is twofold: (1) to corroboratethe
findings of Study 1 in a different setting-automobiles-and (2) to test Hlb, which postulatesan asymmetric
impact of positive and negative disconfirmationon attributes on overall satisfaction. It is important to provide a separate test of Hlb, because both performance and disconfir-

mation have been found to have distinct impactson overall

satisfaction (Yi 1990).

Sample
Data for this study were obtained from a mail survey of
new car buyers throughout the United States. This survey is
part of an ongoing tracking study conducted by a domestic
automotive manufacturer. This study uses a mail survey,
and the response rate for a typical wave of this study ranges
from 36% to 40%, which is similar to the industry standard

for satisfaction surveys in general. For example, Mittal and

colleagues (1993) report results from a large-scale automotive study with a response rate of 33.5%, Hall (1995)
reports results from a patient satisfaction context with a
response rate of 41.1%, and LaBarbera and Mazursky
(1983) report results from grocery item shoppers with a
response rate of 25%.

The database included responses from 9359 new car

buyers in one wave of the larger study. These respondents
were selected randomlyfrom the largerdatabaseof 250,000
respondents that the organizationhas accumulated.A chisquare test on key demographicsand t-tests on the dependent variableshowed no differencebetween the largerdatabase and the random subsample. Furthermore,informal
comparisons with other industry databases and inspection
by managersin the automotiveresearchindustryconfirmed
that the sample was nationallyrepresentativeof automobile
buyers.
Measures
Overall satisfaction was measuredon a ten-point scale with
the following anchors: 1 = very dissatisfied and 10 = completely satisfied. Furthermore,disconfirmation measures
were available on ten attributesthathad been selected by the
firm on the basis of qualitativeresearch.Attribute-leveldisconfirmationwas measuredon a three-pointscale in which
respondentswere asked to indicate whetherperformanceon
each attributewas "betterthan what you expected," "about
what you expected," or "worse than what you expected."
Operationalizingdisconfirmationin this way facilitates
a comparison of the asymmetric impact of positive versus
negative disconfirmationin a mannerthat is consistent with
previous studies (cf. Oliver and DeSarbo 1988). This measure was coded as a dummy variable such that "betterthan
expected" was coded as (1,0), "worse than expected" was
coded as (0,1), and "aboutas expected" was coded as (0,0).
With this coding scheme, the asymmetric effects between
positive and negative disconfirmationon each attributecan
be compared; the neutral level serves as the "baseline"
against which the coefficients for "above" or "below"
expectations are computed (cf. Pedhazur 1982). In other
words, the estimates for the neutrallevel (aboutas expected)
for each attributeare confounded with the interceptterm.
Analysis and Results
A dummy-variableregression was conducted to test Hlb.
The specific equation that was estimated is
(8)Overall Satisfaction = Intercept + bll AttributeI (aboveexpect.)
+ b21 Attributel (below expect.)
+ ... b1 10 Attributelo (above
expect.)

+ b2 Io Attribute 10(belowexpect.)-

In Equation8, bli representsthe relative impact of positive disconfirmation on attributei, whereas b-i represents
the relative impact of negative disconfirmationon the same
attributeon overall satisfaction. Using this formulation,the
impact of each attribute on overall satisfaction was estimated. This data set had expectation/disconfirmationdata
on ten separateattributes,which resultedin 20 specific coefficients, one for positive and one for negative disconfirmation on each of the ten attributes.Results of the analysis are
shown in Table 4 and are discussed subsequently in two
parts. First, the substantive results related to HIb are discussed; second, the overall model, in terms of the variance
explained and the differentattributesincluded in the model,
is discussed.
Recall that Hib postulatesthat negative disconfirmation
on an attributewill have a greaterimpact than positive dis-

Regression

Attribute

TABLE 4
Results for Study 2
Dummy-Variable
Regression Coefficients
Positive
Negative
Disconfirmation Disconfirmation

Comfort
Ease of getting in
and out of vehicle
Driver'sseating
comfort
Ride quality
Maneuverability/
handling

-.86***

.24***

-.39***

.34***

-.29*
-.42**

.11*
.06n.s.

-.67***

.16**

Quietness of engine

-.39**

.07n.s.

-.62***
-1.43***
-.39**
-.32***

.08n.s
.42***
.01n.s
.08n.s.

Powerand pickup
Qualityof vehicle
Brakes operation
Transmission

R2 = .18; F20,9339 = 104.70,

p < .0001.

***p< .0001; **p < .01; *p < .05; n.s.p > .10.

confirmation on the same attribute.Table 4 supports this

hypothesis. For all ten attributes,the absolute value of the
coefficients is greaterfor negative disconfirmationthan for
positive disconfirmation.For example, the impact of negative disconfirmationon "comfort"is three times largerthan
the impact of positive disconfirmationon the same attribute
(-.86 versus .24). Negative disconfirmationon "quality of
vehicle" affects overall satisfaction three times more than
positive disconfirmation(-1.43 versus .42).
Equivalent results can be noted for other attributes,
though the magnitude of the asymmetry between positive
and negative disconfirmationvaries somewhat. For example, the magnitudeof the asymmetryis smaller for "ease of
getting in and out of vehicle" (-.39 versus .34) than for
"power and pickup"(-.62 versus .08). Overall, the results
supportHlb. More interesting,for many attributes,positive
disconfirmation has a nonsignificant impact (e.g., "ride
quality," "quietness of engine"). For these attributes, it
seems that there is no additionalbenefit in exceeding customer expectations, though the negative consequences of
not meeting expectations are relatively high. These results
also confirm the findings of Study 1 that differentattributes
have varying degrees of impact on overall satisfaction. For
example, the impactof negative disconfirmationon "quality
of vehicle" (-1.43, p < .0001) is much largerthan negative
disconfirmationon "driver'sseatingcomfort"(-.29, p < .05)
or "ride quality" (-.42, p < .01). Again, this suggests that
attributescould be classified as core or facilitatingattributes
on the basis of the magnitudeof the asymmetry.
However, the results of both Studies I and 2 must be
scrutinizedfurtherin light of the fact thatthe models used in
.bothstudies explain a small amount of variation(23% and
18%,respectively) in overall satisfaction.There are several
possible reasons for this. One explanation is that when
dummy variablesinstead of regularlyscaled items are used
to operationalize attribute-levelperformanceor disconfirmation, the overall varianceexplained by the model generally decreases because of !oss of informationfor the indePerformance/41
Negativeand PositiveAttribute-Level

pendentvariable(Cohen 1983; Cox 1980). In the context of

satisfactionliterature,only one study has examined the issue
of scale categories and explained variation. Wittink and
Bayer (1994) conducted a quasi-experimental study to
investigate different measurementsystems for satisfaction
research. They found that the system using dichotomized
variables for measuring attribute-level performance (e.g.,
high versus low) had a lower R2 than the one using a fivepoint scale for measuringattribute-levelperformance.Thus,
it can be arguedthatthe lower R2 is simply an artifactof the
measurementsystem adoptedhere; still, we cannot rule out
the alternative explanation that all possible attributesthat
explain overall satisfactionwere not included. This alternative explanation will be examined by comparingthe results
from Studies 2 and 3, because both pertainto overall satisfaction with an automobileand measuresimilar attributes.

Study 3
Study 3 included data from 13,759 respondentswho filled
out a satisfaction survey in the context of the automotive
industry. Before describing the study, it is important to
understandthe specific rationalebehind including Study 3.
Although the results of the previous studies indicate some
support of the central thesis-namely, the asymmetry and
diminishing returns-of this work, they do not test adequately for both asymmetry and diminishing returns for
each individual attribute(Hia and H2). For example, Study
1 shows diminishing sensitivity but only for all attributes
combined, whereas Study 2 is unableto test for diminishing
sensitivity. Therefore,to providedirectevidence for Hla and
H2, this thirdstudy is reported.
Sample
Analysis for Study 3 was based on a customer satisfaction
data set of a major automotive manufacturerin the United
States. This data set was assembled using a mail survey
methodology. Similarto the data in Study 2, these data were
collected as partof an ongoing trackingstudy to assess satisfaction with various aspects of the vehicle ownership
experience. A survey was mailed to people who owned their
vehicle for two or more years. The responserate for this survey varied from wave to wave but rangedbetween 38% and
42% for the last three waves.
In the survey,respondentsratedtheiroverall satisfaction
with the vehicle and evaluatedperformanceon six different
attributes:brakes, transmission,power and pickup, vehicle
quality, quietness, and interiorroominess. Data from 13,759
respondents who answered this section of the survey were
made available for this analysis. Similar to Study 2, the
demographic profile of these respondents was compared
with the demographic profile of the larger database from
which they were extracted. No statistically significant differences were found in terms of age, sex, income, race, or
geographic location (for all, p > .5).
Measures
There were two main constructs: overall satisfaction and
attribute-level performance.Overall satisfaction was measured on a ten-point scale (10 = completely satisfied, I =
42 / Journalof Marketing,January1998

completely dissatisfied). Performanceon each of the six

attributeswas measuredon an eight-pointscale (4/3 = very
satisfied; 2/1 = somewhat satisfied; -1/-2 = somewhat dissatisfied; and -3/-4 = very dissatisfied with performanceon
the attribute).Note that this scale is a satisfactionscale and
therefore does not measure performance directly at the
attributelevel. However, recent studies in the area of customer satisfaction show that though attribute-levelsatisfaction and performanceare conceptually distinct constructs,
their measures are highly correlated.For example, Spreng,
MacKenzie, and Olshavsky (1996) find high correlation
between attribute-level satisfaction and performance for
camcorders.In their study, the correlationbetween performance and satisfaction for "versatility"was .72 (p < .01),
and "picture"was .80 (p < .01). Furthermore,both performance and satisfaction ratingson each attributehad similar
correlations with overall satisfaction. Bitner and Hubbert
(1994, p. 85) find a high interfactor correlation (.94)
between satisfaction and quality/performanceand conclude
that "thisfinding providesevidence for the notion thatoverall service satisfaction and service quality may be tapping
elements of the same construct."Thus, for the sake of measurement,the use of the satisfactionscale as a surrogateof
attribute-level performance is reasonable for the current
study. However, this is in no way designed to suggest that
the conceptual distinction between performanceand satisfaction at the attributelevel should not be maintained.
Recall that the scale being used to measure attributelevel performanceis an eight-pointscale; on this scale, positive ratingson an attributereflect positive performanceon
that attribute, whereas negative ratings indicate negative
performanceon that attribute.For each attribute,there are
four levels or grades of positive performance(1, 2, 3, 4) and
four grades of negative performance(-1, -2, -3, -4). Thus,
the effect of differentlevels of performanceand diminishing
sensitivity can be modeled for each individualattribute.
Analysis and Results
The analyticstrategyfor this study was adaptedfromAnderson and Sullivan (1993). Accordingly, the asymmetricand
diminishing impact of each attributeon overall satisfaction
is modeled as follows:
(9)

OverallSatisfaction= Intercept
+ LN_PERFi
+ LP_PERFi ... LN_PERF6+ LP_PERF6.

In Equation 9, the measurement variable for each

attribute is decomposed into LP_PERF and LN_PERF,
where L denotes the natural logarithm of each measure,
P_PERF or N_PERF (for positive and negative performance, respectively). Thus, performanceon each attribute
is decomposed into positive and negative performanceas
follows: If performance on the first attribute is negative,
the measurement variable for negative attribute performance, LN_PERFI, is equal to -In (-PERFI),5 and
5Thenegativesign for PERFIfor negativeperformance
makes
the finaloutcomepositiveandallowsthe use of the naturallogarithmof the number.Forexample,if thereis a negativeperformanceratingof -4, thensettingthevariableto -(-4) yields+4, for
whichit is possibleto takethe naturallogarithm.Recallthatnaturallogarithms
do notexist for negativenumbers.

LP_PERFJis equal to zero. If performanceon an attribute

is positive, then LP_PERF is equal to In (PERFI), and
LN_PERFI is equal to zero. Thus, if the first attribute
received a rating of -4, LN_PERFi = -In {-(-4)}, and
LP_PERFI = 0. Conversely, if an attribute was rated 3,
LP_PERFI = ln{3}, and LN_PERFI = 0. Note that in this
analysis plan, two coefficients are estimated for each
attributefor a total of 12 coefficients.
This attribute-levelmeasurementplan accomplishestwo
goals. First,it ensures thatall of the coefficients will be positive; thus, for example, the coefficients for positive and
negative performanceon a given attributewill be positive.
This makes the interpretationof resultsmoreconvenientand
therefore managerially useful. Thus, if the coefficient for
negative performance is larger than the coefficient for
positive performance on an attribute, the hypothesized
asymmetry would be supported.Second, the naturallogarithm transformationcaptures diminishing returnor sensitivity (Anderson and Sullivan 1993). For example, if the
coefficient for positive performanceon an attributeis significant, it can be interpretedas supportingthe diminishing
sensitivity hypothesis for positive performance on that
attribute.
At this point, it is necessary to explain why this plan of
analysis ratherthan the one used in Equation4 was adopted
to model the asymmetry and diminishing sensitivity. First,
to use an analytic plan similar to Equation 4, four coefficients would have been requiredto model the asymmetry
and diminishing returnson each of the six attributes(positive, negative, positive2, and negative2 for each attribute);
thus, the use of a plan similar to Equation4 would necessitate the estimation of 24 coefficients. Equation 9, conversely, requiresthe estimationof only two coefficients per
attributefor a total of 12 coefficients. Second, the use of
squared terms in polynomial equations such as Equation4
introduces severe multicollinearity(e.g., positive and positive2 for an attribute are likely to be highly correlated).
Although such multicollinearitymay not be a serious issue
for a single attribute,when modeling several attributes,it
becomes difficult to detect and control for (e.g., Wittinkand
Bayer 1994). Thus, Equation9 not only providesa parsimonious representationof the model (12 instead of 24 terms),

but also circumventsissues relatedto multicollinearitythat

would arise because of the use of squaredterms along with
the base terms (e.g., positive and positive2) for several
attributes.Finally, the use of Equation9 not only complements the results of Study 1 by testing the asymmetry and
diminishingsensitivity for each individualattribute,but also
compares attribute-level results with those reported by
Andersonand Sullivan (1993).
Given this measurementand analysis plan, each coefficient in Equation 9 should be positive. The asymmetry is
tested by constrainingthe positive and negative coefficients
for each attributeto be equal (LN_PERFj= LP_PERFj),
determining whether the constraint can or cannot be
rejected,and reportingwhetherLN_PERFj> LP_PERFj.If
the constraintis rejectedand the absolute size of the coefficient for negative performanceis greaterthanthe coefficient
for positive performance,then the asymmetryis supported.
Furthermore,the naturallogarithmtransformationcaptures
diminishingsensitivity, and a statisticallysignificant coefficient indicates support for the diminishing hypothesis.
Results for these analyses are reportedin Table 5 and discussed next.
Asymmetrybetween positive and negative performance
(Hla). The results in Table5 show that all estimates are statistically significant. Forfive of the six attributes,Hla is supported because LN_PERF > LP_PERF and the constraint
LN_PERF = LP_PERF is rejected. A comparison of the
coefficients shows that the magnitudeof the asymmetry is
different for each attribute.Thus, the magnitude of the
asymmetry is much larger for "transmission"(.04 versus
.35) than for "quietness"(.11 versus .23). Similar to the
results of Study 2, "vehicle quality"has the largest impact
on overall satisfaction(even after two years of ownership).
Diminishing sensitivity. The logarithm transformation
facilitates testing the diminishingsensitivity hypothesis. As
the values get more extreme, a logarithmicfunction tapers
off and thus resembles the diminishing sensitivity curve.
Thus, significant coefficients that are based on the natural
logarithm indicate support for the diminishing sensitivity
hypothesis (cf. Anderson and Sullivan 1993). However,

TABLE 5
Regression Results for Study 3

Attribute
Brakes
Transmission
Power and pickup
Vehicle quality
Quietness

Interiorroominess

LP_PERF

LN PERF

(Regression
coefficients for
positive performance
on an attribute)
O.On.s.
.04*
.06**

(Regression
coefficients for
negative performance
on an attribute)
.24***
.35***
.21***

1.41***
.11***

(LN_PERF= LP_PERF)
Yes***
Yes***
Yes*

1.86***
.23***

Yes***
Yes*

.49***

.17*

Reject Constraint?

Yes***

R2 = .65; F12,13758= 2210.1, p < .0001.

***p< .0001;**p< .01;*p< .05;n.sp> .10.

Performance/ 43
Negativeand PositiveAttribute-Level

whetherthe logarithmictransformationaddedany additional

explanatorypower to that of a model without diminishing
returnsalso was checked. Anotherbaseline model in which
attributeperformancehad not undergonea logarithmictransformation also was estimated. Results from both models
were virtuallyidentical. For the baseline model-that is, the
model without the log-transformedvariables-R2 = .65 and
F12,13758= 2210.9 show equally good fit. A chow-test confirmedthatthe model with diminishingreturs did not fit significantlybetterthanthe model withoutdiminishingreturns.
Therefore,even thoughthe model with diminishingreturnsis
statistically significant, it can be concluded that it does not
add much to the explanatorypower of the model.

These results show that the relationship between

attribute-leveloutcomes and overall evaluationsis complex.
These results have several implications for research and
practice that are discussed next.
Implications for Research

Satisfaction research at an attribute level typically has

viewed attributes as unidimensional and classified them
accordingly.That is, most typologies classify attributeson
the basis of their weight or importancein determiningoverall satisfaction(cf. Bolton and Drew 1991; LaTourand Peat
1979). These results show that typologies that classify
attributeson the basis of their weight alone might not be
Summaryof results. The results supportthe asymmetry
enough. More specifically, there is a need to develop a
between different levels of negative and positive attribute
typology that enables researchers to understandwhy the
performancebut are inconclusive with regardto diminishing
magnitudeof the asymmetryis differentfor differentattribreturns.One importantpartof the results is thatfor "roomiutes. That is, is the observed asymmetryrelated to the utilness" a reversalof the hypothesizedasymmetryis observed.
ity-preserving or enhancing qualities of an attribute (cf.
The implications of this asymmetry for developing a theoKahn and Meyer 1991)? By their very nature, utility-preretical typology of attributesas they relate to overall satisserving attributes(e.g., correctdiagnosis by a doctor,safety
faction are discussed in a subsequentsection.
in air travel) seem to have a higher potential for negative
In addition, Study 3, which has a higher R2 than Study
disconfirmation,whereas utility-enhancingattributes(e.g.,
2, is based largely on the same attributesas Study 2. "Intedoctor is humorous, entertainmentaboard a flight) have a
rior roominess" is the only attributethat is not measured
higher potential for positive disconfirmation.Accordingly,
directly in Study 2. However, it can be arguedthat in Study
in Study 3, "interiorroominess"would be a utility-enhanc2, attributessuch as "comfort,""ease of getting in and out of
ing attribute,whereas the other attributesmight classify as
vehicle," "driver'sseating comfort,"and "ridequality"caputility-preservingattributes.Yet anotherreasoncould be the
ture most of the aspects of "interiorroominess"from Study
level of performanceand disconfirmationobserved in the
3. Similarly,though Study 1 has four times the attributesof
past. Thus, for airline safety, consumers typically observe
Study 3, the variance explained by Study 3 is much higher.
high positive performance/disconfirmation
(i.e., air travel is
Thus, it seems reasonableto infer that the lower R2 obtained
safe). If this expectation suddenly were disconrelatively
in Studies I and 2 is an artifactof informationloss due to
firmed negatively, then it would become salient and affect
coarse measurementscales ratherthanexclusion of relevant
overall satisfaction. If a customer frequents a restaurant
attributes.
where the service is consistently below expectations (i.e.,
negative performance)and receives bad service again on his
or her next visit, there would be little furtherimpact. However, if the customer suddenly received top-qualityservice,
We investigate the link between attribute-level perforit would have a great impact on overall satisfaction. This
mance/disconfirmation,satisfaction, and repurchaseintenline of logic also helps explain why there was a reversalof
tions. Using three large-scale data sets, several hypotheses
the hypothesized asymmetry for "interior roominess." It
about these relationshipswere tested. The results are sumcould be that in the past, automobilecustomersalways have
marized as follows:
expected low performance on interior roominess; thus,
*Both overall satisfactionand repurchaseintentionsare
encountering high performance on "interior roominess"
affectedasymmetrically
and
causes a higher level of disconfirmation(in the positive
by attribute-level
performance
disconfirmation.That is, negative performance/disconfir- direction). However, these ideas are
speculative and their
mationon an attributehas a greaterimpactthan positive
resolution is beyond the scope of this research.The develperformance/disconfirmation.
opment of a theoreticallydriven typology thatenables a res*Regardingdiminishing returns,the results are mixed. For all
olution
of these findings should be a welcome addition to
attributescombined, there is supportfor diminishing returns
satisfaction
research.
in the domain of positive performancebut not negative performance (Study 1). When tested for diminishingreturnsfor
Analytically, these results call into question previous
each individualattribute(Study 3), there is supportfor diminconceptualizations that treat the relationship between
ishing sensitivity for both positive and negative performance.
attribute-levelperformanceand overall satisfactionas symHowever, data also show that a linear conceptualizationfits
metric and linear(cf. LaTourand Peat 1979). More broadly,
the data as well as the diminishing returnconceptualization.
these results have implicationsfor multi-attributemodels of
*Finally, results show that, in additionto the impact mediated
consumer decision making and choice (cf. Kahnand Meyer
by satisfaction, attribute-level performance has a direct
1991; Wilkie and Pessemier 1973). Typical models of deciimpact on repurchaseintentions. This result calls into quession making do not distinguishbetween attributeson which
tion previous models that assume that the impact of performance on repurchaseintentionsis mediatedby overall evalua consumer has experienced positive or negative perforations (e.g., satisfaction).
mance. Depending on this past experience, the weights of

Discussion

44 / Journalof Marketing,January1998

the attributesin subsequent decisions could shift dramatically. For example, a consumer who has experienced positive performanceon comfort in the decision to purchasea
car may weight it less than another who has experienced
negative performanceon comfort. Previous experiences of
attribute-levelperformancethus serve as importantcontextual cues for subsequent decisions and choices. Applying
these findings to multi-attributemodels of consumerchoice
is a key directionfor furtherresearch.
Similarly,these results also point towarda careful reexaminationof satisfactionand its consequentbehaviors,such
as retentionand word-of-mouthactivity. Previousconceptualizations have assumed these relationshipsto be linear and
symmetric (Yi 1990). However, satisfaction's impact on
these behaviors could be asymmetric and nonlinear.High
levels of satisfaction might not increase retention,but high
levels of dissatisfaction might have a large and deleterious
impact on retention.An asymmetricconceptualizationalso
could help explain the curious finding that most firms are
experiencing, that is, high rates of customer defection
despite high rates of customer satisfaction (Reichheld
1996). High dissatisfaction with a product might prompt
high levels of information search for alternative brands,
whereas high levels of satisfaction might not decrease the
amount of informationsearch substantially;this asymmetry
has implicationsfor considerationset formationand choice.
Given the increased emphasis on linking satisfaction
research to the bottom line (Reichheld 1996), these issues
warrantcareful attention.
Finally, the finding that attribute-levelperformancealso
has a direct impact on repurchaseintentionsnot only is surprising but also calls into question previous models of satisfaction. These models (for a review, see lacobucci et al.
1996) presumethat attribute-levelevaluations affect behavioral intentions but only through an overall satisfaction
judgment. This finding raises several importantquestions
that need more research. First, the conditions under which
the direct impact of attribute-level outcomes is larger
(smaller) than the one mediatedthroughsatisfactionshould
be outlined. Second, it could be that the mediating role of
overall satisfaction is different for different attributesand
dependenton the magnitudeand directionof the asymmetry.
Developing a typology of attributes should help resolve
these issues.
Implications for Practice
The key implication for managers is that they no longer
should view attribute performance with a neutral eye.
Although positive and negative performanceon an attribute
are two sides of the same coin, each side of the coin buys a
different amountof overall satisfactionor repurchaseintentions. This asymmetry should be kept in mind in trying to
manage attribute-levelperformance.The strategic implication of this result is clear: To maximize overall satisfaction
and repurchase intentions, managers should optimize and
not maximize attribute-levelperformance.For example, for
any given attributeit is more importantto eliminatenegative
performancefirst and then focus on increasingperformance
in the positive direction. Similarly, in a given set of attrib-

utes, focus on attributesfor which customersare experiencing negative performance and then allocate resources to
maximizing performanceon attributesfor which consumers
are experiencing positive performance.
Concluding Comments
Despite the initial wave of skepticism with satisfaction
research among practitioners and academics, it now is
acknowledged that customer satisfaction and retention are
key strategic imperatives and not fads (e.g., Honomichl
1993; Reichheld 1996). Increasingly,researchersalso are
finding that the relationships between satisfaction and its
antecedents and consequences are complex (cf. Anderson
and Sullivan 1993; Coyne 1989; Oliva, Oliver, and Bearden
1995). Therefore, understandingthe complexity of these
interrelationshipshas become a key strategicimperativefor
most firms (e.g, Jones and Sasser 1995). We hope that this
work represents a small step in that direction and will
encourage further research aimed at understandingthese
complexities.

Appendix
Follow-up tests for H3bwere conducted to ensure that the
reported results are not due to the differing salience or
importance of some of the attributes used in the study.
Salience is operationalizedas the frequency with which an
attributeis mentioned (higher numberof mentions= higher
salience), whereas importance is operationalized as the
magnitudeof the regressionweight (largerregressioncoefficient = higher importance).To check for this possibility,
the cumulative variables were reconstructed, excluding
those attributesthat were not importantor salient.
To check for the salience bias, the first set of variables
was constructedfrom attributeson the basis of the number
of mentions. With number of mentions as a proxy for
salience, only those attributesthatreceived the top 3, 4, 5, or
6 mentions were included. To clarify, POSFRQ3 and
NEGFRQ3used only those attributesthat received the first,
second, and third highest number of mentions. POSFRQ4
and NEGFRQ4 also included the attributethat received the
fourthhighest numberof mentions, and so on. To check for
the importance bias, the cumulative variables were constructedon the basis of their imputedimportance.With the
regression weight in Table 2 as a proxy for an attribute's
importance, variables called POSREG5 and NEGREG5
were constructed using those attributes that had the five
highest regressionweights for the positive and negative categories. Similar variablesalso were constructedon the basis
of the top six and seven regressionweights.
These strategies of integrationare consistent with prior
theoreticaland empirical research(e.g., Dawes 1979). Furthermore,satisfaction researchers(e.g., Oliver 1993) have
used additive-linearcombinationsto integrateperformance
on several attributes.To summarize,eight additionalsets of
variables were constructed: POSFRQ or NEGFRQ3-6,
POSREG or NEGREG5-7, and POSWTD or NEGWTD5.
Next, the following model was estimated for each set of
variables:

Performance/ 45
Negativeand PositiveAttribute-Level

TABLE A1
Summary of Various Tests for Asymmetry and Diminishing Sensitivity
Restrictions
Test
Variable

R2

-.98***

.15*

.14

9.8,

-.78***

.07n.s.

.15

4.9, .0300

-.73***
-1.38***
-.78***

.02n.s.
.28***
.08n.s.

.17
.18
.18

-.13***
-.09***
-.08**

-1.22***
-1.21***
-.80***

.20***
.22***
.05n.s.

6.4,
50.9,
16.9,

.19
.21
.14

-.04***

-.72***

44.2, .0001
66.6, .0001
24.9, .0001

.07**

.23

50.4,

b2

POS3FRQ (salience)

.58***

-.16***

POS4FRQ (salience)

.53***

-.13***

POS5FRQ (salience)
POS5RG (importance)
POS6FRQ (salience)

.49***
.54***
.44***

-.11***
-.14***
-.09***

.52***
.46***
.35***
.29***

POS6RG (importance)
POS7RG (importance)
POSMTCH(importance)

POSALL (all attributes/initial case)

***p< .0001; **p< .001; *p < .05; n.s.p> .10.

(F-statistic,
significance)

b4

bl

Overall satisfaction= bi(POSITIVE)+ b2(POS[TIVE)2

+ b3(NEGATIVE)+ b4(NEGATIVE)2.
The results are described in Table Al. In all of the eight
additional cases plus the initial case, lb31> Ibll, which indicates that negative performance has a greater impact than
positive performance as shown in the columns labeled bi
and b2 in Table A 1. This result was significant at p < .03 in
all nine cases (see the column labeled "Restrictions Test" in
Table Al). These results show that the conclusion about H3
is not a methodological artifact.
Regarding diminishing sensitivity, both b2 and b4 have
the expected sign in all nine cases. Furthermore, b2 was sig-

b3

.0020
.0100
.0001
.0001

.0001

nificant for all of the nine models (p < .001), but b4 was significant only for five of the nine models. To ascertain the
"joint" statistical significance on the basis of these nine

models, a combined test such as Fisher's combined test

would be appropriate (Wolf 1986). However, because the
nine comparisons are not independent, such a test of significance cannot be done. Therefore, erring on the side of conservatism, we conclude that over the nine tests, b4 is not significant, whereas b2 is. Thus, diminishing sensitivity holds
for positive performance but not for negative performance
on attributes.

REFERENCES
Anderson, Eugene W. and Mary W. Sullivan (1993), "The
Antecedents and Consequences of Customer Satisfaction for
Firms,"MarketingScience, 12 (Spring), 125-43.
Bitner, Mary Jo, BernardH. Booms, and Mary Stanfield Tetreault
(1990), "The Service Encounter: Diagnosing Favorable and
Unfavorable Incidents,"Journal of Marketing, 54 (January),
71-84.
and Amy R. Hubbert(1994), "EncounterSatisfactionVersus Overall Satisfaction Versus Quality: The Customer's
Voice,"in Service Quality:New Directions in Theoryand Practice, RolandT. Rust and RichardL. Oliver, eds. London:Sage
Publications,72-94.
Bollen, KennethA. (1989), StructuralEquationswith Latent Variables. New York:John Wiley & Sons.
Bolton, Ruth N. and James H. Drew (1991), "A MultistageModel
of Customers' Assessments of Service Quality and Value,"
Journal of ConsumerResealrch,17 (March), 375-84.
Boulding, William, Ajay Kalra, Richard Staelin, and ValarieA.
Zeithaml (1993), "A Dynamic Process Model of Service Quality: From Expectations to Behavioral Intentions,"Journal of
MarketingResearrch,30 (February),7-27.
Cohen, Jacob (1983), "The Cost of Dichotomization,"Applied
7 (Summer), 249-53.
Psychological Measulremlent,
Cox, Eli P., III (1980), "The Optimal Number of Response Alternatives for a Scale: A Review,"Journal of MarketingResearch,
27 (November), 407-22.
Coyne, Kevin (1989), "Beyond Service Fads-Meaningful Strategies for the Real World,"Sloan ManagementReview, 39 (Summer), 69-76.

46 / Journalof Marketing,January1998

Dawes, Robyn M. (1979), "The RobustBeautyof ImproperLinear

Models in Decision Making," AmlericaniPsychologist, 34
(July), 571-82.
Einhorn,Hillel J. and Robin M. Hogarth(1981), "BehavioralDecision Theory: Processes of Judgment and Choice," Annual
Review of Psychology, 32, 53-88.
Feinberg,RichardA., RichardWiddows,MarlayaHirsch-Wyncott,
and Charles Trappey (1990), "Myth and Reality in Customer
Service: Good and Bad Service Sometimes Leads to Repurchase," Journal of CustomerSatisfaction, Dissatisfaction alnd
ComplainingBehavior, 3, 112-14.
Folkes, Valarie S. (1988), "Recent AttributionResearch in Consumer Behavior: A Review and New Directions,"Journal of
ColsunLmerResearch, 14 (March),548-65.
Gardial, Sarah Fisher, D. Scott Clemons, Robert B. Woodruff,
David W. Schumann, and Mary Jane Burns (1994), "Comparing Consumers'Recall of Prepurchaseand PostpurchaseProduct Evaluation Experiences,"Journal of ConsumerResearch,
20 (March),548-60.
Hall, Melvin F. (1995), "Patient Satisfaction or Acquiescence?
Comparing Mail and Telephone Survey Results," Jou1rnalof
Health Care Marketing, 15 (Spring), 54-61.
Hamilton,LawrenceC. (1992), Regression with Graphics:A Second Course in Applied Statistics. Belmont, CA: Duxbury
Press.
Hanson, Randy (1992), "Determining Attribute Importance,"
Quirk'sMarketingResearch Review, 6 (October), 16-18.
Honomichl, Jack (1993), "Spending on Customer Satisfaction
Continuesto Rise," MarketingNews, 27 (April 12), 17-18.

Iacobucci, Dawn, Amy L. Ostrom, BridgetteM. Braig, and Alexa

Bezjian-Avery(1996), "A CanonicalModel of ConsumerEvaluations and Theoretical Bases of Expectations,"Advances in
Services Marketingand Management,5, 1-44.
Jones, Thomas 0. and W. Earl Sasser, Jr. (1995), "Why Satisfied
Customers Defect," Harvard Business Review, 73 (November/December),88-99.
Kahn, BarbaraE. and RobertJ. Meyer (1991), "ConsumerMultiattributeJudgmentsunderAttributeWeight Uncertainty,"Journal of ConsumerResearch, 17 (March),508-22.
Kahneman,Daniel and Amos Tversky (1979), "ProspectTheory:
An Analysis of Decisions Under Risk," Econometrica, 47
(March), 263-91.
Kekre, Sunder, Mayuram S. Krishnan, and Kannan Srinivasan
(1995), "Drivers of Customer Satisfaction for Software Products: Implications for Design and Service Support,"Management Science, 41 (September), 1456-70.
LaBarbera,Priscilla A. and David Mazursky(1983), "A Longitudinal Assessment of Consumer Satisfaction/Dissatisfaction:
The Dynamic Aspect of the Cognitive Process," Journal of
MarketingResearch, 20 (November), 393-404.
LaTour,Stephen A. and Nancy C. Peat (1979), "Conceptualand
Methodological Issues in Consumer SatisfactionResearch,"in
Advances in Consumer Research, Vol. 6, William D. Perrault
Jr., ed. Ann Arbor, MI: Association for Consumer Research,
431-37.
Mittal, Vikas, Jerome M. Katrichis, Frank Forkin, and Mark
Konkel (1993), "Does Satisfaction With Multi-AttributeProducts Vary Over Time? A Performance Based Approach," in
Advances in ConsumerResearch, Vol. 21, Chris T. Allen and
Deborah Roedder John, eds. Provo, UT: Association for Consumer Research, 412-17.
Oliva, Terence A., Richard L. Oliver, and William O. Bearden
(1995), "The Relationship Among Consumer Satisfaction,
Involvement, and Product Performance,"Behavioral Science,
40 (April), 104-32.
, and lan C. MacMillan (1992), "A Catastrophe
Model for Developing Service SatisfactionStrategies,"Journal
of Marketing, 56 (July), 83-95.

Oliver, Richard L. (1993), "Cognitive, Affective, and Attribute

Bases of the Satisfaction Response," Journal of Consumer
Research, 20 (December), 418-30.
and WayneS. DeSarbo(1988), "ResponseDeterminantsin
Satisfaction Judgments,"Journal of Consumer Research, 14
(March),495-507.
Ostrom,Amy and Dawn Iacobucci (1995), "ConsumerTrade-Offs
and the Evaluationof Services,"Journal of Marketing,59 (January), 17-28.
Parasuraman,A., ValarieZeithaml, and LeonardL. Berry (1988),
"SERVQUAL:A Multiple-ItemScale for MeasuringConsumer
Perceptions of Service Quality," Journal of Retailing, 64
(Spring), 12-40.
Pedhazur, Elazar J. (1982), Multiple Regression in Behavioral
Research, 2d ed. New York:Holt, Rinehartand Winston.
Peeters, G. and J. Czapinski(1990), "Positive-Negative Asymmetry in Evaluations: The Distinction Between Affective and
InformationalNegativity Effect," European Review of Social
Psychology, 1, 33-60.

Reichheld, Fredrick(1996), The LoyaltyEffect: The HiddenForce

Behind Growth, Profits, and Lasting Value. Boston: Harvard
Business School Press.
Spreng, Richard A., Scott B. MacKenzie, and Richard W.
Olshavsky (1996), "A Reexaminationof the Determinantsof
ConsumerSatisfaction,"Journalof Marketing,60 (July), 15-32.
Taylor,Shelly (1982), "TheAvailabilityBias in Social Perception,"
in JudgmentUnder Uncertainty:Heuristics and Biases, Daniel
Kahnemanet al., eds. Cambridge:CambridgeUniversityPress,
190-200.
Wilkie, William L. and EdgarA. Pessemier(1973), "Issuesin Marketing's Use of Multi-AttributeAttitude Models," Journal of
MarketingResearch, 10 (November), 428-41.
Wittink,Dick R. and LeonardR. Bayer (1994), "The Measurement
Imperative,"MarketingResearch,6 (Winter), 14-23.
Wolf, Fredric M. (1986), Meta-Analysis: Quantitative Methods for

Research Synthesis.London:Sage Publications.

Yi, Youjae(1990), "A CriticalReview of ConsumerSatisfaction,"
in Review of Marketing 1990, Valarie A. Zeithaml, ed. Chicago:

AmericanMarketingAssociation, 68-123.

Performance/47
Negativeand PositiveAttribute-Level