The Asymmetric Impact of Negative and Positive Attribute-Level Performance on Overall Satisfaction and Repurchase Intentions Author(S): Vikas Mittal, William T

Home , Vol (command)

The Asymmetric Impact of Negative and Positive Attribute-Level Performance on Overall Satisfaction and Repurchase Intentions Author(s): Vikas Mittal, William T. Ross, Jr. and Patrick M. Baldasare Source: 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: 18/12/2014 21:27

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

. 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 [email protected].

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

http://www.jstor.org

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions VikasMittal, William T. Ross, Jr., & PatrickM. Baldasare

The AsymmetricImpact of Negative and Positive Attribute-Level Performanceon Overall Satisfaction and RepurchaseIntentions 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: metric and linear relationship between attribute-levelper- formanceand constructs ?In an effortto increasecustomer satisfaction, an automotive dependent such as overall satisfac- manufacturerconsistently has increasedperformance on tion and purchase intentions. However, what if, at the engine power. However,results show that, after a while, attribute level, performance and satisfaction were linked increasesin the ratingsfor performanceon enginepower do asymmetrically?The answer to this question is important not yield correspondingchanges in satisfactionratings. for academics and managersalike. Although academics (cf. Designerswonder if they shouldkeep investing resources to Anderson and Sullivan 1993; Oliva, Oliver, and Bearden enhanceengine power. 1995) and practitioners(cf. Coyne 1989) recognize that the *A physicianhistorically has receivedhigh rat- performance overall satisfactionfunction might not be linear and/orsym- ings on attributessuch as politeness,timeliness, reliability of metric, no such conclusion can be offered at the attribute service,and correct diagnosis. In the current survey, she finds thatperformance ratings on timelinessare down,but are up level. Yet, it is at the attributelevel that managersmust make on all the otherattributes. However, she also finds a sharp decisions. declinein heroverall satisfaction ratings. She is puzzledas to Therefore,our objective is to investigate-theoretically why negativeperformance on a single attributeis not offset and empirically-the existence and natureof the asymmet- on a hostof otherattributes. by positiveperformance ric and nonlinear response of satisfaction to attribute-level These scenarios highlight typical dilemmas that man- performance.In the first section, we discuss the need for an agers face at customer-drivenorganizations where the goals attribute-levelconceptualization of the performance/satis- are to design productsand services with attributesthat max- faction relationship.That is followed by a discussion of the imize customer satisfaction.These goals generally are oper- asymmetricand nonlinearnature of the relationshipand the ationalized by multiple-regressionmodels that identify key development of hypotheses. Then we present results from attributes into which managers should invest resources to three studies that test the hypotheses. enhance overall satisfaction (e.g., Bolton and Drew 1991; Hanson 1992; Wittink and Bayer 1994). The assumption such driver"models is Multi-AttributeProducts underlying "key that there is a sym- and Satisfaction There are several reasons-theoretical and VikasMittal is a assistant GraduateSchool of managerial-to visiting professor,Kellogg use multi-attributemodels in the context of customer satis- Management,Northwestern University. William T.Ross, Jr. is anassociate professor,Temple University. Patrick M. Baldasare is Presidentand Chief faction (cf. LaTour and Peat 1979; Wilkie and Pessemier ExecutiveOfficer of The ResponseCenter, a divisionof TeleSpectrum 1973). First, consumers are more likely to render evalua- WorldWide,Inc. The authors acknowledge the helpful comments and sug- tions of their postpurchaseexperiences of satisfaction at an gestionsof EugeneAnderson, Terry Oliva, D. Raghavarao, Soumya Roy, attributelevel ratherthan at the productlevel (Gardialet al. andfour reviewers.The authors thank Frank anonymous Forkin,Mark G. 1994). Second, an attribute-based approach enables Konkel,and Alan Rogers of ConsumerAttitude Research, Inc./Research researchersto conceptualize commonly observed DataAnalysis, Inc. for theautomotive data, and James Cuthrell, phenom- providing ena, such as consumers mixed feelings toward KristineGupta, Michael Gerhart, Jerry Katrichis, and E. Monsoon for their experiencing support. a productor service. A consumer can be both satisfied and dissatisfied with different aspects of the same product.

Journal of Marketing Vol. 62 (January 1998), 33-47 Negativeand Positive Attribute-LevelPerformance / 33

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions Although such phenomena are not easy to model in an next sections, the theoretical logic is developed along two overall satisfaction approach, the attribute-level approach lines of reasoning.One is based on prospecttheory (Kahne- provides a simple and elegant solution: Mixed feelings man and Tversky 1979), and the other is rooted in cognitive toward a productexist because a consumer may be satisfied researchthat examines the memorabilityof positive versus with one attributebut dissatisfied with another.For exam- negative events. ple, in a restaurant,a customer may be highly satisfied with the food but highly dissatisfied with the service at the same Prospect Theory time. Third, an attribute-level approach to satisfaction Prospecttheory (Kahnemanand Tversky 1979) is a descrip- affords researchers a higher level of specificity and diag- tive theoryin which all of the alternativesthat a personfaces nostic usefulness compared with the productlevel or "over- are reduced to a series of prospects that are evaluated inde- all" approach(LaTour and Peat 1979). For example, Para- pendently on the basis of an S-shaped value function (see suraman,Zeithaml, and Berry (1988) show that measuring Figure 1). various attributesor dimensions of a service provides a bet- As depicted in Figure 1, prospect theory postulates that ter understandingof global constructs,such as service qual- peoples'judgments display (1) referencedependence (carri- ity. Similarly, Boulding and colleagues (1993) suggest that ers of value are gains and losses from a referencepoint) and quality is multidimensional and different dimensions of (2) loss aversion (the function is steeper in the negative than quality are averaged together in some fashion to producean in the positive domain). In addition, evaluations display overall assessment of quality. However, both works treat diminishing sensitivity (marginal values of both gains and the relationship between attributes and overall quality as losses decrease with their size). The two key propertiesof linear and symmetric. Fourth, the higher specificity and the value function for our discussion are loss aversion and diagnostic usefulness of multi-attributemodels also ensures diminishing sensitivity. that academic research will find more welcome applica- The loss aversion built into prospecttheory suggests that tions among managers who generally work at the attribute losses loom largerthan gains (Einhom and Hogarth 1981). level ratherthan at an overall level (e.g., Hanson 1992; Wit- Psychologically, a one-unit loss is weighted more than an tink and Bayer 1994). Finally, there is some evidence that equal amountof gain. In a satisfactioncontext, negative out- attribute-level performance/disconfirmation and overall comes on attributeperformance should carrymore weight in satisfaction are qualitatively different constructs (Oliva, the overall satisfactionjudgment than equal amountsof pos- Oliver, and Bearden 1995), and, if treated interchangeably, itive outcomes on attributeperformance. For example, if a "global customer satisfaction responses may mask specific car's mileage were to decrease by 10 miles per gallon, it product issues" (Oliva, Oliver, and Bearden 1995, p. 26). would have a greaterimpact on the overall satisfactionjudg- Thus, studying satisfaction at the attribute level can help ment than if the car's mileage were to increase by 10 miles extend both conceptual and empirical understandingof the per gallon. Thus, negative performanceon an attributewill phenomenon. loom largerthan positive performanceon the same attribute. On the basis of these theories, we propose The Asymmetric Effect of Product Performance and Satisfaction: FIGURE 1 Asymmetric Impact of Attribute-Level Theory and Hypotheses Performance When examining the theoretical and analytical importance of the link between attribute-levelperformance and overall satisfaction, it is importantto recognize that the 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 correspond- ing 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 finds that Negative Positive (1993) (1) attribute-levelsatis- Performance II Performance faction and dissatisfaction significantly affect overall satis- on Attribute on Attribute faction with a product(automobile) and a service (an under- graduate course offering) and (2) attribute dissatisfaction has a larger weight than attributesatisfaction for the prod- uctI but not for the service. However, no theoreticalmotiva- tion for the observed disparity between the impact of i attributesatisfaction and dissatisfaction is provided. In the

tStandardizedloadings were as follows: attributesatisfaction OverallSatisfaction/ and dissatisfactionfor the automobile(.135 and -.175, respec- RepurchaseIntentions tively)and the course (.130 and-.059, respectively).

34 Journalof Marketing,January 1998

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions H ,,:Negative performance on an attributewill havea greater H3b:Overall satisfaction will displaydiminishing sensitivity to impacton overallsatisfaction than positive performance additionalinstances of negative or positive performance. on the sameattribute. In otherwords, each additionalinstance of positiveor Hlb:Negative disconfirmationon an attributewill have a negativeperformance should have a smallerimpact on greaterimpact on overallsatisfaction than positive dis- overallsatisfaction. confirmationon the sameattribute. Repurchase Intentions In addition, on the basis of prospect theory, overall sat- Attribute-level should affect satisfaction and isfaction also should display diminishing sensitivity toward performance intentions Ostrom and attributeperformance. That is, at high (low) levels of per- repurchase differently (e.g., Iacobucci 1995). One reason could be that satisfaction and formance, positive (negative) performanceon an attribute intentions are different constructs. should not affect satisfaction as dramaticallyas it does at repurchase qualitatively Satisfaction be a with and lower levels of performance.This hypothesis is similar to may merely judgment cognitive affective dimensions, whereas intentions also the diminishing returns hypothesis in classical economics repurchase have a behavioral Based on the consumer's and also is depicted in Figure 1. Therefore, component. goals (e.g., Mittal et al. 1993), performance on a certain Overallsatisfaction will H2: displaydiminishing sensitivity to attributemay become crucial for intentions but in the of for a repurchase changes magnitude performance given not satisfaction. For consider the case of a attribute.In otherwords, at levelsof or example, patient high positive neg- who is satisfied with all of the service ativeperformance on an attribute,overall satisfaction will aspects provided by be affectedless thanat intermediatelevels of performance his or her primary care physician (PCP), except that the on thatattribute. patient has now relocated to another area far from the PCP's office. When it comes time to renew and choose a Memory for Negative Versus Positive Instances PCP, the patient might indicate high overall satisfaction with the PCP but still Researchers distinguish between transaction-based and might choose another PCP, because cumulativesatisfaction (Bitner and Hubbert1994). Cumula- performance on a critical attribute has changed. In other the on "distance of tive satisfaction typically is conceptualized as an overall words, though performance PCP's office from home" judgment based on several transactions with a product or had a negative impact on repurchase it has a small or service. To the extent that all aspects of the transactionsare intentions, virtually no impact on overall not equally accessible (Gardialet al. 1994), the overall sat- satisfaction. One reason for such a discrepancy can be found in the isfactionjudgment will vary on the basis of the accessibility attributionliterature (Folkes 1988). Here, the of a given aspect. Memory accessibility is a function of patient may attribute"poor" performanceon "distance of PCP's office from stimulus salience, among other things (Taylor 1982). Evi- home" to him or herself or to causes dence shows that negative informationis more perceptually other than the PCP and thus not change his or her overall satisfaction salient than positively valenced information,is given more rating; nevertheless, the repurchase intentions of the weight than positive information, and elicits a stronger patient change. Thus, physiological response than positive information (Peeters H4:Overall satisfaction and attribute-levelperformance will and Czapinski 1990). haveseparate and distinct effects on repurchaseintentions. Similar psychological operations should occur for cus- However, the relative of overall satisfaction tomer satisfaction because satisfaction is linked to memory- magnitude and attribute should be driven based processing (Yi 1990). To the extent that attributes performance contextually, and no are made in that sim- with negative performancewill be more perceptuallysalient specific predictions regard.H4 states that attribute-level and overall satis- than attributes with positive performance, attributes with ply performance faction will have and distinct on negative performanceshould have a greater impact on the separate impacts repurchase intentions. cumulative satisfactionjudgment. Thus, within a given set Next, we an effect of attribute-level of attributes,the relative impact of each attributewill be propose asymmetric on intentions. Previous asymmetric.Consequently, when combined, attributeswith performance repurchase research Oliva, Oliver, and MacMillan shows that over- negative performanceshould have a greaterimpact on over- (e.g., 1992) all satisfaction and are related to all satisfaction than their correspondingattributes with pos- performance nonlinearly intentions or and itive performancecombined. On the basis of the previous repurchase loyalty. Feinberg colleagues found that across several discussion, we hypothesize that (1990, p. 113) productcategories "the probability of repurchase was not isomorphic with attributeswith H3a: Cumulatively, negative performancewill either positive or negative service Research have on overallsatisfaction than attrib- experiences." greaterimpacts also shows that satisfaction and dissatisfactionhave differ- uteswith positive performance. ent affective consequences (Oliver 1993), which may be Along similar lines, a diminishing sensitivity hypothesis related differentiallyto repurchaseintentions. Furthermore, between overall satisfaction and performance on various the same psychological principlesthat operate in the context attributescan be proposed. In a given set of attributes,each of satisfaction also should operate in the context of repur- additional instance of positive performanceon an attribute chase intentions.Therefore, will have a smaller impact than the other attributes.Con- intentionswill be affected each additional instance of H5:Repurchase asymmetricallyby versely, negative performance attribute-levelperformance. Negative on should have a smaller performance correspondingly negative impact on attributeswill havea greaterimpact on repurchaseinten- overall satisfaction.Thus, tionsthan positive performance.

Negativeand Positive Attribute-LevelPerformance /35

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions In the next sections, we present results from three large- Finally, Yi (1990), in his review of satisfaction research, scale studies that test the preceding hypotheses. A unique compares the test/retest reliabilities of several multi- and feature of all three studies is that they use data from com- single-item scales and finds that the test/retestreliability of mercial studies conducted by firms measuring satisfaction single-item scales is acceptable (.55 to .84). For these rea- with their productor service. sons, we believe the measures are adequate. Independentvariables were createdon the basis of open- ended data providedby patients.In a free-thoughtelicitation Study 1 task duringthe telephone interview,respondents were asked Data for this study were provided by a large health mainte- about theirexperiences with various featuresof their visit to nance organization (HMO) in the Northeastern United the PCP.The interviewerasked them to clarify and expand States. The HMO collected these data as partof its ongoing their thoughts. patient satisfaction measurementprogram. For this study, These thought listings provided the basis for the inde- the entire data set, which contained informationfrom 4517 pendentvariables for the first analysis. As shown in Table 1, telephone interviews conducted among patients enrolled each thought was coded in a distinct category. When with the HMO, was used. These interviews were conducted queried,the projectmanager for this study indicatedthat the using Computer-Assisted-Telephone-Interview, which categories were determinedby the managersat the HMO in enabled the interviewer to probe, clarify, and follow up on conjunction with the project director. The process is to responses to open-ended questions. Because responses to review all the open-endeddata from a randomlychosen sub- the open-ended questions were a vital part of the measure- set of interviews (usually 10%) and develop the categories ment process, interviewerswere asked to exercise greatcau- so that they are mutuallyexclusive and collectively exhaus- tion in recordingand following up on them. tive. Then the coders code a portion of the data using these The sampling frame for the data set consisted of patients categories. After these data are coded, the coders and man- who subscribedto 501 PCP offices affiliatedwith the HMO. agers meet to see if the categories are "working,"that is, if A random sample was drawn from each PCP's patient list. the categories are descriptive and capture the "essence" of Thus, in effect, a proportionate-stratifiedsample, with each what respondents said. Furthermore,when making cate- PCP as a stratum, was drawn. After contact with a house- gories, the managersensure that all attributesand thoughts hold had been established, up to two members per house- are coded. In any given study,fewer than 2% of the thoughts hold were interviewed as long as they evaluated different are forced into the "other/not specified" category. The PCPs. Thus, each household provided only one evaluation process of determining categories is done carefully, espe- per PCP.The qualified respondentwas an adult from a fam- cially given the high cost of data collection and the strategic ily whose members had visited a PCP within the past 12 value of the informationcontained in these data. months. All 4517 interviews were used in the analysis. For the purpose of this analysis, each category was the exact rate for this could not be Although response survey coded as 1 if a respondentgave a response in that category rates such ascertained,typical response for surveys are more and 0 otherwise. Three professionalcoders, each with more than 70%. from the at the HMO and infor- Input managers than 2-3 years of coding experience, coded the categories. mal with data from other waves of the comparisons study However, each coder coded a separate set of respondents. show that the is sample representativeof the population Because the coding already had been done before the data served the HMO. by set was made available for this study, no follow-up analyses could be done to assess the reliability and agreementof the Measures coding. Therefore,to assess the agreementamong the three The database contained measures for several constructs. coders, they were asked to code data from 337 respondents There were two dependentmeasures for this analysis: over- from a subsequent wave of a similar study conducted for all satisfaction and repurchaseintentions. "Overallsatisfac- another HMO in the same geographic area. For these 337 tion with the medical care at the PCP's office" was opera- respondentsof the pilot study, the three coders were in per- tionalized as a five-point scale (1 = very dissatisfied, 5 = fect agreementfor 81 % of the surveys. That is, coders coded very satisfied). Repurchaseintention or a patient's intention every category identically for 81% of the respondents.Fur- to switch to a different PCP was operationalizedas a yes or thermore,Coders I and 2 match on 84.3% of the responses, no response to the following question: "Do you intend to Coders 2 and 3 match on 85.5%, and Coders I and 3 match switch to a different primarycare physician when you next on 85.2%. The associated Kendall's tau correlationwas as have an opportunity?"Although both these measures are follows: .84 for Coders 1 and 2, .85 for Coders I and 3, and single-item scales, there is considerableprecedent for using .85 for Coders 2 and 3 (all significant at p < .01). single-item measures in the context of large-scale satisfac- With these limitations in mind, it can be said that such tion studies. LaBarberaand Mazursky (1983) discuss the an approachto understandingsatisfaction has been used by issue of using single- versus multi-itemscales for measuring other consumersatisfaction researchers(Bitner, Booms, and overall satisfaction and conclude that, in large-scale survey Tetreault 1990; Gardial et al. 1994). This approach,though research, the use of multi-itemscales actually may decrease more unstructured than rating scales, is advantageous the quality of measurementrather than enhance it. Similarly, because it captures attributeperceptions that are not conta- in their large-scale study of drivers of customer satisfaction minatedby the preconceived notions that a researchercould in the computer industry, Kekre, Krishnan,and Srinivasan impose on the respondentsby cuing them regardingspecific (1995) use a single-item measure for overall satisfaction. attributes.The thought categories then were classified fur-

36 / Journalof Marketing,January 1998

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions ther as positively or negatively valenced. The positively done, in which the coefficient for variable POSITIVE is valenced items for each respondentwere summed to create expected to have a positive sign, and the coefficient for the a new variable, POSITIVE. Similarly, a variable called variable NEGATIVE is expected to have a negative sign. NEGATIVEwas createdby summing all the negative items Therefore, testing whether IbNEGATIVEI> [bposITIVEI sug- for each respondent.The positive/negative classification of gests the following hypothesis: each thoughtcategory is indicated in Table 1. (1) HA: IbNEGATIVEI> IbPOSITIVEI-

Analysis and Results The correspondingnull hypothesis can be stated as follows: The general strategy employed for testing the proposed (2) Ho: IbNEGATIVEI < IbPOSITIVEI asymmetry consists of computing separate estimates for positive and negative performanceand statisticallycompar- To test HA, two rival models are compared to see ing the absolute magnitude of the positive and negative whether Ho can be rejected statisticallyin favor of HA. One estimates. The absolute magnitude is compared because model formally constrainsIbNEGATIVEI < IbPOSITIVEI as sug- negative performance should be related negatively to the gested in Ho, whereas in the other, the coefficients are free dependentconstructs, and positive performancerelated pos- to vary as suggested in HA. Comparingthe performanceof itively to the dependent constructs. Therefore, to get a the constrained model to that of the unconstrainedmodel proper comparison, the "signs" of the estimates must be provides a test of the asymmetry.If the model with the con- ignored. As an illustration,a test of the asymmetrybetween straint is rejected in favor of the alterative model (i.e., Ho the composite variables POSITIVE and NEGATIVE is is rejected in favor of HA), it can be concluded that the

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

Positive Mentions Negative Mentions Coefficient Coefficient

Matched Categories Thorough/attentive .11*** Not thorough -.66**** Spends time with me/personable .17**** Doesn't spend enough time/rushesyou in and out -.40**** Not afraidto referto anotherdoctor .07 Won'tgive referrals/troublegetting referrals -.77**** Listens to his/her patients .03 Not interested in patients/doesn'tlisten -.94**** Convenientoffice hours .07 Office hours are not convenient -.37**** Convenientoffice location .02 Distance of office is inconvenient .07 Alwaysavailable/availability of doctor .15*** Not available/isn'talways in office -.67**** No wait/sees patients rightaway .14** Waittoo long/officetoo crowded -.38****

Unmatched Categories Good, excellent doctor care .26**** No followup/doesn't tell you results -.65**** Knowledge/educated/competent .02 Doctor/staffoverbooks appointments -.60*** Explainseverything well/answers questions .12*** Poor staff/problemwith staff -.84**** Helpful .13* Problemswith HMO -.35**** Good bedside manner .13 Incorrectdiagnosis/unnecessary testing -.99**** Professional .13 Friendly/nice/courteous .11*** Concerned/caring/sympathetic .22**** Good with kids .19**** Makes me comfortable/easyto talk to .06 Verypatient .15 Sincere/honest/straightforward .15 Outgoing/goodpersonality/sense of humor .25* Is warm/kind .19 Is an old-fashionedtype of doctor .14 Calls back rightaway/keeps in contact/followsthorough .17*** Alwaysavailable during emergency .32 Promptness of service .02 Has been our doctorfor a long time .33**** I like him/her .20*** Staff is very nice/good .14** Cleanliness of office/building .05 < ** * **** *p .10;< **p.05 < ; p < . 001; Model R2 = .23, F = 30.9, error df = 4481, p < .0001.

Negativeand Positive Attribute-LevelPerformance / 37

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions absolute magnitude of the coefficient for NEGATIVE is Second, differentattributes of the service encountercon- greater than POSITIVE, or that IbNEGATIVEI> IbposITIVE tribute differently to overall satisfaction. For example, Next, the results are described. "cleanliness of office/building" (.05, p = .78) is not conse- whereas Negative versus positive attributeperformance (H/a). A quential, "no follow-up/doesn't tell you results" dummy-variableregression was conducted using all of the (-.65, p < .0001) has a greaterimpact on overall satisfaction. thought categories as independentvariables and overall sat- Several other comparisonscan be made with similarresults. isfaction as the dependent variable.2As shown in Table 1, These echo earlier findings (Bolton and Drew 1991) that there were 30 positively valenced thoughtcategories and 13 service attributes may be core, facilitating, or supporting negatively valenced thought categories. On the basis of and provide confidence in our results. these, the following model was estimated: Third, with an R2 of 23%, the overall variation explained by the model appears to be low. However, this (3) Overall Satisfaction = Intercept+ Attpos ... + Attpos30 could have resulted from loss of informationresult- + Attnegl ... + Attneg13 partially ing from the dichotomousnature of the independentvariable In this the of each attributeon overall sat- model, impact (Cohen 1983). Furthermore,because the main concern in isfaction is calculated Furthermore, individually. positively this study is to test the asymmetry hypothesis rather than valenced attribute mentions are to have expected positive explain variationin the dependentvariable, the low R2 is not whereas valenced attributementions are signs, negatively particularlytroublesome. expected to have negative signs. To test Hla, we must test Cumulative and whether the absolute impact of a negative mention for an impact diminishing sensitivity (H3a and To test and the attribute is greater than the absolute impact of a positive H3b). H3a H3b, following equation was estimated: mention for the same attribute.The model is significant (F = 30.9, p < .0001) and is reportedin Table 1. (4) Overallsatisfaction = bi(POSITIVE)+ b2(POSITIVE)2 Several findings of interestemerge from this table. First, + b3(NEGATIVE)+ b4(NEGATIVE)2. as shown in the first eight rows of Table 1, there were eight In Equation4, POSITIVEand NEGATIVEare the sums attributesfor which there were both negative and positive of positively and negatively valenced mentionsof attributes, categories. Therefore,for these eight attributes,Hia could be respectively. Therefore, bl and b3 model the asymmetric tested. For all eight pairs, the negatively valenced attributes cumulative impact of positive and negative performanceon have a larger(in the absolute sense) coefficient than the cor- overall satisfaction. To test for diminishing sensitivity, the responding positively valenced attributes.We used the sta- squaredterms are introduced(cf. Hamilton 1992). Thus, for tistical test explained previouslyto test whetherthe negative positive performance,bl and b2 model diminishingsensitiv- coefficient for each pair of attributesis larger than the cor- ity, whereas for negative performance, b3 and b4 model responding positive coefficient for the same attribute;the diminishing sensitivity. These tests (for H3a and H3b) are null hypothesis in Equation2 was rejected for seven of the described in detail next. eight pairs at p < .001. For example, "always available/ Cumulative impact of number of positive or negative availability of doctor"(.15, p < .01) has a smaller impact on attributementions (H3a). As stated previously,bI and can patient satisfaction than "not available/isn't always in his b3 be used to test the asymmetry for the cumulative of office" (-.67, p < .0001). Similar resultscan be noted for (1) impact negative and positive mentions of attributeson overall satis- "not afraid to refer to anotherdoctor" (.07, p < .26) versus faction stated in H3a.Because negative mentionsof attribute "won't give referrals/troublegetting referrals"(-.77, p < performanceare hypothesized to have greater impacts on .0001) and (2) "spends time with me/personable"(.17, p < overall satisfactionthan are positive mentions, lb3\should be .0001) versus "doesn't spend enough time/rushesyou in and greaterthan Ibli. Similar to the test of two rival models out" (-.40, p < .0001). Only for "convenient office loca- Hia, are to test whether the model with the constraint tion," for which both the positive and negative coefficients compared Ib31< Ibil performs better or worse than the unconstrained were nonsignificant, was the hypothesis not supported. model. If the constrained model performs worse than the These results support Hia. However, because this study is unconstrainedmodel, the constraintis rejected, which only able to provide a sense of the direction of a sup- person's ports > Results of the are shown in Table 2. evaluation on each attribute, a full test of cannot be |b31 Ibil. analysis H,, The coefficients for POSITIVEand NEGATIVEare .29 and done. In other words, because there are no data on thei,rg- -.72, respectively. Both are in the expected direction and nitude of positive or negative performanceon each attribute, significant at p < .0001. Moreover,their absolute magnitude additional testing for this hypothesis is required.Recall that is as hypothesized;the absolute value of the coefficient for though Hla is stated in terms of varying levels of perfor- NEGATIVEis than the absolute value of the coeffi- mance on each attribute,there are no data on varying levels greater cient for POSITIVE.This was confirmed the of performancein this study. Nevertheless, these results are by comparing constrained and unconstrainedmodels as encouraging and point toward tthe type of investigation that explained previously. The model with the constraint < was is carried out in Study 3. lb31 Ibll rejected in favor of the unconstrainedmodel, where lb3l> Ibjl(Fi,4477 = 50.37, p < .0001). Thus, H3ais supported;3cumulatively, 2Acorrelation matrix for all thoughtcategories showed the highest con-elation to be .21. A majorityof the correlationswere nonsignificant(p > .05) and below .10 in absolutemagnitude. Thus, 3These models also were estimated using standardizedcoeffi- multicollinearitydoes not appearto be an issue here. Furthermore, cients. The results for the models with standardizedcoefficients the thoughtcategories are not linearlydependent on one another. were practicallyidentical to the results reportedhere.

38 / Journalof Marketing,January 1998

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions TABLE 2 was used alone, the function becomes a curve that "tapers Regression Results for Satisfaction off' when the squaredterm (b2) is included in the equation. With Medical Care This occurs because every increase in bl is countered by a smaller (because of the squared term) decrease in b2. For Standardized example, the estimate for bl is .28, whereas the estimate for Regression Standard is -.04. Consider the effect of Variable Coefficient Error t-statistic b2 increasingpositive perfor- p-value mance for one and two attributes. In the absence of the Intercept 4.24 .02 199.48 .0001 squared term, the beneficial impact of one attributeis .29 POSITIVE .29 .02 13.87 .0001 and of two attributesis .58; however, with the beneficial NEGATIVE -.72 .05 -13.48 .0001 b2, increases but at a rate. POSITIVE2 -.04 .01 -7.78 .0001 impact decreasing Using both bl and NEGATIVE2 .07 .02 3.01 .0030 b2, the beneficial impactof one attributeis .25, calculatedas .29(1) - .04(1)2, whereas the impact of two attributesis .42, R2 = .23; = 330.65, p < .0001. F4,4477 calculatedas .29(2) - .04(2)2. Thus, inclusion of the squared Test for constraint for Ho in Equation 2: F1,4477= 50.37, p < term captures diminishing sensitivity. Similar results apply .0001 (one-tailed test). for b3 and b4 in the domain of negative performance.Results based on this analysis are reportedin Table 2. As seen in Table 2, all four coefficients are statistically negative performance on attributes has a significantly significant and in the predicted direction, which thereby greater impact on overall satisfaction than positive perfor- supports H3b. However, a key concern when using these mance on attributes.The regression coefficients show that data to test H3bis that supportfor H3bmay be an artifactof for this particularservice, the deleterious effect of an addi- the way the variablesPOSITIVE and NEGATIVEwere con- tional instance of negative performanceon attributesout- structed.Some of the attributesthat went into the construc- weighs the beneficial effect of positive performanceby a tion of the variable may be inherently less important or marginof two to one (-.72 versus .29). salient than others. Thus, the test for H3bmay be simply an artifactof the lower importanceor salience of some attrib- Diminishing (H3b). This a sensitivity hypothesis suggests utes used in the constructionof the cumulativevariable than nonlinear between overall satisfaction and relationship per- others. Additional tests, described in the were formance on additionalattributes. This is mod- Appendix, relationship used to test for this possibility. In these tests, the variables eled using the polynomial function described in 4. Equation POSITIVE and NEGATIVEwere constructed in eight dif- In 4, bl and b2 test for in Equation diminishing sensitivity ferent ways, and H3bwas tested eight differenttimes. There- the domain of The positive performance. linear positive fore, including the base case described in Table 3, a total of coefficient bl models the of positive impact performance, nine tests were conducted for H3b.All nine tests supported whereas the coefficient for the b2, squared term, models the diminishing sensitivity in the positive performance diminishing sensitivity by creating a "dampening"effect on domain, but only five supporteddiminishing sensitivity for the of slope bl. Instead of being a straight line as when bl negative attributeperformance. Therefore, erring on the side

TABLE 3 Logistic Regression for Intention to Switch to a Different PCP

Model 3 Model 2 Attribute and Model 1 Overall Overall Variable Attribute Satisfaction Satisfaction Intercept -1.50 4.17 3.14 (.07) (.23) (.24) POSITIVE -.66 -.43 (.05) (.05) NEGATIVE 1.21 .76 (.08) (.08) Overall satisfaction - -1.45 -1.15 (.06) (.06) 708.7 Chi-square (d.f.) (2) 989.9 (1) 1182.2 (3) (p < .0001) (p < .0001) (p < .0001) AIC 2905.22 2606.10 2417.71 % correctly classified 86.0% 88.1% 89.5% All coefficientestimates are significantat p < .001; standarderrors are in parentheses. It could be argued that POSITIVEand NEGATIVEare not 2 but 43 variablesbecause they are composed of variables.There- the AICalso was multiplebinary fore, recalculatedafter penalizing Models 1 and 3 for43 ratherthan 2 variables.The results for model fit were virtuallyidentical to the one reportedin the table.

Negativeand Positive Attribute-LevelPerformance / 39

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions of conservatism, the conclusion is that diminishing sensitiv- Attributeperformance and overall satisfactionhave separate ity holds for positive performancebut not for negative per- and distinct effects on repurchaseintentions. formance on attributes. Summaryof resultsand limitations.The resultsfrom this These results suggest that overall satisfaction displays study support the hypothesized asymmetry and provide a sensitivity to each additionalinstance of diminishing positive strong indicationof the existence of diminishingsensitivity, but not necessarilyto negative In performance performance. at least on a cumulative basis. Negative performanceon other words, though the relative benefit of positive perfor- seven of eight pairs of attributeshas a greater impact on mance on additionalattributes may decrease,each additional overall satisfaction than does positive performance.This is instanceof negative performancemay be equally deleterious. also true of the cumulative impact of negative versus posi- Repurchase intentions (H4 and H5). Because the mea- tive attributes. Similarly, intentions to switch also are sure for switching intentions was a dichotomous variable, a affected asymmetrically by positive and negative perfor- logistic regression analysis was conductedto test H4 and H5. mance. Overall satisfaction appears to display diminishing Three models predicting intention to switch to a different sensitivity to attributeperformance in the domainof positive PCP were estimated. Note that though the hypotheses are performancebut not negative performance;however, this stated in terms of repurchase intentions, results are dis- conclusion cannot be offered for different levels of perfor- cussed in terms of intention to switch to maintain consis- mance within each attribute.Finally, both overall satisfac- tency with the data. The models used for testing H4 and H5, tion and attribute performance have separate and distinct which are shown in Table 3, are as follows: impacts on repurchaseintentions. (5) Model 1: intentionto switch= f POSITIVE,NEGATIVE}; These results should be viewed in light of the study's limitations. First, there were no data on expectationdiscon- (6) Model2: intentionto switch= f{Overallsatisfaction); firmationat the attributelevel; this precludedtesting Hlb, a (7) Model3: intentionto switch= f{POSITIVE,NEGATIVE, limitation that is addressed in Study 2. Second, attribute- Overallsatisfaction}. level performance was measured using 0,1 categories, which measurementof differentlevels of In Model 1, only POSITIVEand NEGATIVEwere used precluded performance for an individual attribute. and which as predictorvariables. In Model 2, only overall satisfaction Thus, Hla H2, are stated for differentlevels of within an indi- was used as a predictorvariable. The thirdmodel used POSI- performance vidual could not be tested. in which TIVE, NEGATIVE,and overall satisfactionas predictorvari- attribute, Study 3, per- formancelevels are measuredfor each exam- ables. Note that these models are concernedonly with asym- attribute,fully ines and 1 was conducted in a metry,not diminishing sensitivity;therefore, only the cumu- Hia H2. Third, Study service lative variablesPOSITIVE and NEGATIVEare used in these setting; therefore,it could be argued that the results of this would not models. Results from these models are reportedin Table 3. study necessarily generalize to a product-related Results for Model 1, the attributes-onlymodel, confirm setting. Although there is no theoretical reason to expect these results in a should the hypothesis that negative attributesare more consequen- this, testing productsetting increase tial for intention to switch than are positive attributesof the confidence in the generalizability. service. Thus, H5 is supported. Next, all three models in Table 3 were examined to test H4.The fit of the models was 2 examined using the Akaike InformationCriterion (AIC sta- Study The of this is twofold: to corroboratethe tistic) of model evaluation (Bollen 1989). The AIC evaluates purpose study (1) of 1 in a different alternativemodels while accountingfor the numberof para- findings Study setting-automo- biles-and to test which an meters in rival models; it penalizes models with many rather (2) Hlb, postulates asymmetric of and disconfirmationon attrib- than few free parameters.Models with smaller AIC values impact positive negative utes on overall satisfaction. It is to a provide a better fit. The AIC for each of the three models is important provide sep- shown in Table 3. arate test of Hlb, because both performance and disconfirmation have been found to have distinct on overall The overall satisfaction model (Model 2) provides a bet- impacts satisfaction ter fit in predicting repurchase intentions than does the (Yi 1990). attribute performance model (Model I); the AIC fit for Sample Model 2 is smaller than thatof Model 1. However, Model 3, which includes both attributeperformance and overall satis- Data for this study were obtained from a mail survey of faction, outperformseither model. The value of the AIC for new car buyers throughout the United States. This survey is Model 3 is smaller than either the overall satisfactionor the part of an ongoing tracking study conducted by a domestic attribute performance model. Thus, H4 is supported.4 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 4These results also were verified by dividing the data set in two for satisfaction surveys in general. For example, Mittal and one with 3517 and the other with 1000 The three parts: people. colleagues (1993) report results from a large-scale automo- models were estimated with the first data set and then used to pre- tive study with a response rate of 33.5%, Hall (1995) dict the value for intentionto switch in the holdout sample of 1000 reports results from a patient satisfaction context with a people. The coefficient estimates were virtually identical to the ones reported in Table 4, and the percentagescorrectly classified response rate of 41.1%, and LaBarbera and Mazursky for the three models were 65%, 69.9%, and 78.9%, respectively. (1983) report results from grocery item shoppers with a These results again supportthe hypotheses. response rate of 25%.

40 / Journalof Marketing,January 1998

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions The database included responses from 9359 new car TABLE 4 buyers in one wave of the larger study. These respondents Regression Results for Study 2 were selected randomlyfrom the largerdatabase of 250,000 respondents that the organizationhas accumulated. A chi- Dummy-Variable Coefficients square test on key demographicsand t-tests on the depen- Regression dent variable showed no difference between the largerdata- Negative Positive base and the random subsample. Furthermore,informal Attribute Disconfirmation Disconfirmation comparisons with other industry databases and inspection Comfort -.86*** .24*** by managersin the automotiveresearch industry confirmed Ease of getting in that the sample was nationallyrepresentative of automobile and out of vehicle -.39*** .34*** buyers. Driver'sseating comfort -.29* .11* Measures Ride quality -.42** .06n.s. Maneuverability/ Overall satisfaction was measuredon a ten-point scale with handling -.67*** .16** the following anchors: 1 = very dissatisfied and 10 = com- Quietness of engine -.39** .07n.s. pletely satisfied. Furthermore,disconfirmation measures Power and pickup -.62*** .08n.s were available on ten attributesthat had been selected by the Qualityof vehicle -1.43*** .42*** firm on the basis of research.Attribute-level dis- Brakes operation -.39** .01n.s qualitative Transmission -.32*** .08n.s. confirmationwas measuredon a three-pointscale in which respondentswere asked to indicate whetherperformance on R2 = .18; F20,9339 = 104.70, p < .0001. each attributewas "betterthan what you expected," "about ***p< .0001; **p < .01; *p < .05; n.s.p > .10. what you expected," or "worse than what you expected." Operationalizingdisconfirmation in this way facilitates confirmation on the same attribute.Table 4 this a comparison of the asymmetric impact of positive versus supports For all ten attributes,the absolute value of the negative disconfirmationin a mannerthat is consistent with hypothesis. coefficients is for disconfirmationthan for previous studies (cf. Oliver and DeSarbo 1988). This mea- greater negative disconfirmation.For the of sure was coded as a dummy variable such that "betterthan positive example, impact negative disconfirmationon "comfort"is three times than expected" was coded as (1,0), "worse than expected" was larger the of disconfirmationon the same attribute coded as (0,1), and "aboutas expected" was coded as (0,0). impact positive (-.86 versus .24). Negative disconfirmationon of With this coding scheme, the asymmetric effects between "quality vehicle" affects overall satisfaction three times more than positive and negative disconfirmationon each attributecan disconfirmation versus be compared; the neutral level serves as the "baseline" positive (-1.43 .42). results can be noted for other against which the coefficients for "above" or "below" Equivalent attributes, though the of the between expectations are computed (cf. Pedhazur 1982). In other magnitude asymmetry positive and disconfirmationvaries somewhat. For exam- words, the estimates for the neutrallevel (about as expected) negative the of the is smaller for "ease of for each attributeare confounded with the interceptterm. ple, magnitude asymmetry getting in and out of vehicle" (-.39 versus .34) than for Analysis and Results "power and pickup" (-.62 versus .08). Overall, the results supportHlb. More interesting,for attributes, A dummy-variable regression was conducted to test many positive Hlb. disconfirmation has a "ride The specific equation that was estimated is nonsignificant impact (e.g., quality," "quietness of engine"). For these attributes, it (8)Overall Satisfaction = Intercept + bll AttributeI (aboveexpect.) seems that there is no additional benefit in exceeding cus- + b21 Attributel (below expect.) tomer expectations, though the negative consequences of + ... b1 10 Attributelo (above expect.) not meeting expectations are relatively high. These results + b2 Io Attribute 10(below expect.)- also confirm the findings of Study 1 that differentattributes In Equation8, bli representsthe relative impact of posi- have varying degrees of impact on overall satisfaction. For tive disconfirmation on attributei, whereas b-i represents example, the impactof negative disconfirmationon "quality the relative impact of negative disconfirmationon the same of vehicle" (-1.43, p < .0001) is much largerthan negative attributeon overall satisfaction. Using this formulation,the disconfirmationon "driver'sseating comfort"(-.29, p < .05) impact of each attribute on overall satisfaction was esti- or "ride quality" (-.42, p < .01). Again, this suggests that mated. This data set had expectation/disconfirmationdata attributescould be classified as core or facilitatingattributes on ten separateattributes, which resultedin 20 specific coef- on the basis of the magnitudeof the asymmetry. ficients, one for positive and one for negative disconfirma- However, the results of both Studies I and 2 must be tion on each of the ten attributes.Results of the analysis are scrutinizedfurther in light of the fact thatthe models used in shown in Table 4 and are discussed subsequently in two .both studies explain a small amount of variation(23% and parts. First, the substantive results related to HIb are dis- 18%, respectively) in overall satisfaction.There are several cussed; second, the overall model, in terms of the variance possible reasons for this. One explanation is that when explained and the differentattributes included in the model, dummy variables instead of regularlyscaled items are used is discussed. to operationalize attribute-level performanceor disconfir- Recall that Hib postulates that negative disconfirmation mation, the overall variance explained by the model gener- on an attributewill have a greaterimpact than positive dis- ally decreases because of !oss of informationfor the inde-

Negativeand Positive Attribute-LevelPerformance /41

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions pendent variable(Cohen 1983; Cox 1980). In the context of completely dissatisfied). Performanceon each of the six satisfaction literature,only one study has examined the issue attributeswas measuredon an eight-point scale (4/3 = very of scale categories and explained variation. Wittink and satisfied; 2/1 = somewhat satisfied; -1/-2 = somewhat dis- Bayer (1994) conducted a quasi-experimental study to satisfied; and -3/-4 = very dissatisfied with performanceon investigate different measurementsystems for satisfaction the attribute).Note that this scale is a satisfactionscale and research. They found that the system using dichotomized therefore does not measure performance directly at the variables for measuring attribute-level performance (e.g., attributelevel. However, recent studies in the area of cus- high versus low) had a lower R2 than the one using a five- tomer satisfaction show that though attribute-levelsatisfac- point scale for measuringattribute-level performance. Thus, tion and performanceare conceptually distinct constructs, it can be arguedthat the lower R2 is simply an artifactof the their measures are highly correlated.For example, Spreng, measurementsystem adopted here; still, we cannot rule out MacKenzie, and Olshavsky (1996) find high correlation the alternative explanation that all possible attributesthat between attribute-level satisfaction and performance for explain overall satisfaction were not included. This alterna- camcorders. In their study, the correlationbetween perfor- tive explanation will be examined by comparing the results mance and satisfaction for "versatility"was .72 (p < .01), from Studies 2 and 3, because both pertainto overall satis- and "picture"was .80 (p < .01). Furthermore,both perfor- faction with an automobile and measure similar attributes. mance and satisfaction ratings on each attributehad similar correlations with overall satisfaction. Bitner and Hubbert 3 (1994, p. 85) find a high interfactor correlation (.94) Study between satisfaction and quality/performanceand conclude 3 included data from who Study 13,759 respondents filled that "this finding provides evidence for the notion thatover- out a satisfaction in the context of the automotive survey all service satisfaction and service quality may be tapping Before industry. describing the study, it is important to elements of the same construct."Thus, for the sake of mea- understandthe rationalebehind 3. specific including Study surement,the use of the satisfaction scale as a surrogateof the results of the studies indicate some Although previous attribute-level performance is reasonable for the current of the central the and support thesis-namely, asymmetry study. However, this is in no way designed to suggest that returns-of this do not test ade- diminishing work, they the conceptual distinction between performanceand satis- quately for both asymmetry and diminishing returns for faction at the attributelevel should not be maintained. each individual attribute and For (Hia H2). example, Study Recall that the scale being used to measure attribute- 1 shows but for all attributes diminishing sensitivity only level performanceis an eight-point scale; on this scale, pos- combined, whereas 2 is unable to test for Study diminishing itive ratings on an attributereflect positive performanceon Therefore,to direct evidence for and sensitivity. provide Hla that attribute, whereas negative ratings indicate this third is negative H2, study reported. performanceon that attribute.For each attribute,there are four levels or of Sample grades positive performance(1, 2, 3, 4) and four grades of negative performance(-1, -2, -3, -4). Thus, for 3 was based on a customer satisfaction Analysis Study the effect of different levels of performanceand data set of a diminishing major automotive manufacturerin the United sensitivity can be modeled for each individualattribute. States. This data set was assembled using a mail survey methodology. Similar to the data in Study 2, these data were Analysis and Results collected as of an to assess sat- part ongoing trackingstudy The analytic strategyfor this study was adaptedfrom Ander- isfaction with various of the vehicle aspects ownership son and Sullivan (1993). Accordingly, the and A was mailed to asymmetric experience. survey people who owned their diminishing impact of each attributeon overall satisfaction vehicle for two or more years. The response rate for this sur- is modeled as follows: vey varied from wave to wave but rangedbetween 38% and 42% for the last three waves. (9) OverallSatisfaction = Intercept+ LN_PERFi + LP_PERFi ... + In the survey, respondentsrated their overall satisfaction LN_PERF6 LP_PERF6. with the vehicle and evaluated performanceon six different In Equation 9, the measurement variable for each attributes:brakes, transmission,power and pickup, vehicle attribute is decomposed into LP_PERF and LN_PERF, quality, quietness, and interiorroominess. Data from 13,759 where L denotes the natural logarithm of each measure, who respondents answered this section of the survey were P_PERF or N_PERF (for positive and negative perfor- made available for this analysis. Similar to Study 2, the mance, respectively). Thus, performanceon each attribute demographic profile of these respondents was compared is decomposed into positive and negative performanceas with the demographic profile of the larger database from follows: If performance on the first attribute is negative, which were they extracted. No statistically significant dif- the measurement variable for negative attribute perfor- ferences were found in terms of age, sex, income, race, or mance, LN_PERFI, is equal to -In (-PERFI),5 and geographic location (for all, p > .5). 5Thenegative sign for PERFIfor negativeperformance makes Measures the finaloutcome positive and allows the use of the naturalloga- rithmof the number.For if thereis a There were two main constructs: overall satisfaction and example, negativeperfor- mancerating of -4, thensetting the variableto -(-4) +4, for attribute-level Overall satisfaction was yields performance. mea- whichit is possibleto takethe naturallogarithm. Recall that nat- sured on a ten-point scale (10 = completely satisfied, I = urallogarithms do not exist for negativenumbers.

42 / Journalof Marketing,January 1998

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions LP_PERFJis equal to zero. If performanceon an attribute but also circumvents issues related to multicollinearitythat is positive, then LP_PERF is equal to In (PERFI), and would arise because of the use of squaredterms along with LN_PERFI is equal to zero. Thus, if the first attribute the base terms (e.g., positive and positive2) for several received a rating of -4, LN_PERFi = -In {-(-4)}, and attributes.Finally, the use of Equation 9 not only comple- LP_PERFI = 0. Conversely, if an attribute was rated 3, ments the results of Study 1 by testing the asymmetry and LP_PERFI = ln{3}, and LN_PERFI = 0. Note that in this diminishing sensitivity for each individualattribute, but also analysis plan, two coefficients are estimated for each compares attribute-level results with those reported by attributefor a total of 12 coefficients. Andersonand Sullivan (1993). This attribute-levelmeasurement two plan accomplishes Given this measurementand analysis plan, each coeffi- First, it ensures that all of the coefficients will be goals. pos- cient in Equation 9 should be positive. The asymmetry is itive; thus, for the coefficients for and example, positive tested by constrainingthe positive and negative coefficients on a attributewill be negative performance given positive. for each attribute to be equal = This makes the of results more convenient and (LN_PERFj LP_PERFj), interpretation determining whether the constraint can or cannot be therefore managerially useful. Thus, if the coefficient for rejected, and reportingwhether LN_PERFj> LP_PERFj.If negative performance is than the coefficient for larger the constraintis rejectedand the absolute size of the coeffi- positive performance on an attribute, the hypothesized cient for negative performanceis greaterthan the coefficient asymmetry would be supported. Second, the natural loga- for positive performance,then the asymmetry is supported. rithm transformationcaptures diminishing returnor sensi- Furthermore,the naturallogarithm transformation captures tivity (Anderson and Sullivan 1993). For example, if the diminishing sensitivity, and a statisticallysignificant coeffi- coefficient for positive performanceon an attributeis sig- cient indicates support for the diminishing nificant, it can be as the hypothesis. interpreted supporting diminishing Results for these are in Table 5 and dis- for on that analyses reported sensitivity hypothesis positive performance cussed next. attribute. At this point, it is necessary to explain why this plan of Asymmetrybetween positive and negative performance analysis ratherthan the one used in Equation4 was adopted (Hla). The results in Table 5 show that all estimates are sta- to model the asymmetry and diminishing sensitivity. First, tistically significant. For five of the six attributes,Hla is sup- to use an analytic plan similar to Equation 4, four coeffi- ported because LN_PERF > LP_PERF and the constraint cients would have been to model the required asymmetry LN_PERF = LP_PERF is rejected. A comparison of the and returnson each of the six attributes diminishing (posi- coefficients shows that the magnitudeof the asymmetry is and negative2 for each tive, negative, positive2, attribute); different for each attribute. Thus, the magnitude of the the use of a similar to 4 would necessi- thus, plan Equation asymmetry is much larger for "transmission"(.04 versus tate the estimation of 24 coefficients. con- Equation 9, .35) than for "quietness" (.11 versus .23). Similar to the the estimation of two coefficients versely, requires only per results of Study 2, "vehicle quality" has the largest impact attribute for a total of 12 coefficients. Second, the use of on overall satisfaction (even after two years of ownership). squared terms in polynomial equations such as Equation4 introduces severe multicollinearity(e.g., positive and posi- Diminishing sensitivity. The logarithm transformation tive2 for an attribute are likely to be highly correlated). facilitates testing the diminishingsensitivity hypothesis. As Although such multicollinearitymay not be a serious issue the values get more extreme, a logarithmicfunction tapers for a single attribute,when modeling several attributes,it off and thus resembles the diminishing sensitivity curve. becomes difficult to detect and control for (e.g., Wittinkand Thus, significant coefficients that are based on the natural Bayer 1994). Thus, Equation9 not only provides a parsimo- logarithm indicate support for the diminishing sensitivity nious representationof the model (12 instead of 24 terms), hypothesis (cf. Anderson and Sullivan 1993). However,

TABLE 5 Regression Results for Study 3 LP_PERF LN PERF (Regression (Regression coefficients for coefficients for Reject Constraint? positive performance negative performance Attribute on an attribute) on an attribute) (LN_PERF= LP_PERF) Brakes O.On.s. .24*** Yes*** Transmission .04* .35*** Yes*** Power and pickup .06** .21*** Yes* Vehicle quality 1.41*** 1.86*** Yes*** Quietness .11*** .23*** Yes* Interiorroominess .49*** .17* Yes*** R2 = .65; F12,13758= 2210.1, p < .0001. ***p< .0001;**p < .01;*p < .05;n.sp > .10.

Negativeand PositiveAttribute-Level Performance / 43

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions whetherthe logarithmictransformation added any additional These results show that the relationship between explanatorypower to that of a model without diminishing attribute-leveloutcomes and overall evaluationsis complex. returnsalso was checked. Another baseline model in which These results have several implications for research and attributeperformance had not undergonea logarithmictrans- practice that are discussed next. formation also was estimated. Results from both models were virtuallyidentical. For the baseline model-that is, the Implications for Research model without the log-transformedvariables-R2 = .65 and Satisfaction research at an attribute level has F12,13758= 2210.9 show equally good fit. A chow-test con- typically viewed attributes as unidimensional and classified them firmedthat the model with diminishingreturs did not fit sig- That is, most attributeson nificantly better than the model without diminishingreturns. accordingly. typologies classify the basis of their or in over- Therefore,even thoughthe model with diminishingreturns is weight importance determining all satisfaction Bolton and Drew 1991; LaTourand Peat statistically significant, it can be concluded that it does not (cf. 1979). These results show that that add much to the explanatorypower of the model. typologies classify attributeson the basis of their weight alone might not be results. The results the Summaryof support asymmetry enough. More specifically, there is a need to develop a between different levels of and attribute negative positive typology that enables researchers to understandwhy the but are inconclusive with to performance regard diminishing magnitudeof the asymmetryis differentfor differentattrib- returns.One of the results is that for "roomi- importantpart utes. That is, is the observed asymmetryrelated to the util- ness" a reversal of the is observed. hypothesizedasymmetry ity-preserving or enhancing qualities of an attribute (cf. The of this for a theo- implications asymmetry developing Kahn and Meyer 1991)? By their very nature, utility-pre- retical of attributesas relate to overall satis- typology they serving attributes(e.g., correct diagnosis by a doctor, safety faction are discussed in a section. subsequent in air travel) seem to have a higher potential for negative In addition, 3, which has a R2 than Study higher Study disconfirmation, whereas utility-enhancingattributes (e.g., 2, is based on the same attributesas 2. "Inte- largely Study doctor is humorous, entertainmentaboard a flight) have a rior roominess" is the attributethat is not measured only higher potential for positive disconfirmation.Accordingly, in 2. However, it can be that in directly Study argued Study in Study 3, "interiorroominess" would be a utility-enhanc- 2, attributessuch as "comfort,""ease of in and out of getting ing attribute,whereas the other attributesmight classify as vehicle," "driver's and "ride seating comfort," quality"cap- utility-preservingattributes. Yet anotherreason could be the ture most of the of "interiorroominess" from aspects Study level of performanceand disconfirmationobserved in the 3. 1 has four times the attributesof Similarly, though Study past. Thus, for airline safety, consumers typically observe 3, the variance 3 is much Study explained by Study higher. high positive performance/disconfirmation(i.e., air travel is Thus, it seems reasonableto infer that the lower R2 obtained relatively safe). If this expectation suddenly were discon- in Studies I and 2 is an artifact of informationloss due to firmed negatively, then it would become salient and affect coarse measurementscales rather than exclusion of relevant overall satisfaction. If a customer a restaurant attributes. frequents where the service is consistently below expectations (i.e., negative performance)and receives bad service again on his Discussion or her next visit, there would be little furtherimpact. How- ever, if the customer suddenly received top-qualityservice, We investigate the link between attribute-level perfor- it would have a great impact on overall satisfaction. This mance/disconfirmation,satisfaction, and repurchaseinten- line of logic also helps explain there was a reversal of tions. three data sets, several why Using large-scale hypotheses the for "interior roominess." It about these were tested. The results are sum- hypothesized asymmetry relationships could be that in the automobile marized as follows: past, customersalways have expected low performance on interior roominess; thus, *Both overall satisfactionand repurchaseintentions are encountering high performance on "interior roominess" affected asymmetricallyby attribute-levelperformance and causes a higher level of disconfirmation (in the positive disconfirmation.That is, negative performance/disconfir- direction). However, these ideas are and their mationon an attributehas a than speculative greaterimpact positive resolution is beyond the of this research.The devel- performance/disconfirmation. scope opment of a theoreticallydriven that enables a res- *Regarding returns,the results are mixed. For all typology diminishing olution of these should be a welcome addition to attributescombined, there is support for returns findings diminishing satisfaction research. in the domain of positive performancebut not negative performance (Study 1). When tested for diminishing returnsfor Analytically, these results call into question previous each individualattribute (Study 3), there is supportfor dimin- conceptualizations that treat the relationship between for both and negative ishing sensitivity positive performance. attribute-levelperformance and overall satisfactionas sym- However, data also show that a linear fits conceptualization 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 addition to the mediated impact consumer decision making and choice Kahn and by satisfaction, attribute-level performance has a direct (cf. Meyer 1991; Wilkie and Pessemier impact on repurchaseintentions. This result calls into ques- 1973). Typical models of deci- tion previous models that assume that the impact of perfor- sion making do not distinguish between attributeson which mance on repurchaseintentions is mediatedby overall evalu- a consumer has experienced positive or negative perfor- ations (e.g., satisfaction). mance. Depending on this past experience, the weights of

44 / Journalof Marketing,January 1998

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions the attributesin subsequent decisions could shift dramati- utes, focus on attributesfor which customersare experienc- cally. For example, a consumer who has experienced posi- ing negative performance and then allocate resources to tive performanceon comfort in the decision to purchasea maximizing performanceon attributesfor which consumers car may weight it less than another who has experienced are experiencing positive performance. negative performanceon comfort. Previous experiences of attribute-levelperformance thus serve as importantcontex- Concluding Comments tual cues for decisions and choices. subsequent Applying Despite the initial wave of skepticism with satisfaction these to multi-attributemodels of consumerchoice findings research among practitioners and academics, it now is is a direction for furtherresearch. key acknowledged that customer satisfaction and retention are these results also toward a careful reex- Similarly, point key strategic imperatives and not fads (e.g., Honomichl aminationof satisfaction and its such consequent behaviors, 1993; Reichheld 1996). Increasingly, researchers also are as retentionand word-of-mouth Previous activity. conceptu- finding that the relationships between satisfaction and its alizations have assumed these to be linear and relationships antecedents and consequences are complex (cf. Anderson satisfaction's on symmetric (Yi 1990). However, impact and Sullivan 1993; Coyne 1989; Oliva, Oliver, and Bearden these behaviors could be and nonlinear. asymmetric High 1995). Therefore, understandingthe complexity of these levels of satisfaction not increase retention,but might high interrelationshipshas become a key strategicimperative for levels of dissatisfaction have a and deleterious might large most firms (e.g, Jones and Sasser 1995). We hope that this on retention.An also impact asymmetric conceptualization work represents a small step in that direction and will could the curious that most firms are help explain finding encourage further research aimed at understandingthese that rates of customer defection experiencing, is, high complexities. despite high rates of customer satisfaction (Reichheld 1996). High dissatisfaction with a product might prompt high levels of information search for alternative brands, Appendix whereas levels of satisfaction not decrease the high might Follow-up tests for H3b were conducted to ensure that the amount of informationsearch this substantially; asymmetry reported results are not due to the differing salience or has for considerationset formationand choice. implications importance of some of the attributes used in the study. Given the increased on satisfaction emphasis linking Salience is operationalizedas the frequency with which an research to the bottom line these issues (Reichheld 1996), attributeis mentioned (higher numberof mentions = higher warrantcareful attention. salience), whereas importance is operationalized as the the that attribute-level Finally, finding performancealso magnitudeof the regression weight (larger regression coef- has a direct on intentionsnot is sur- impact repurchase only ficient = higher importance).To check for this possibility, but also calls into models of satis- prising question previous the cumulative variables were reconstructed, excluding faction. These models a see (for review, lacobucci et al. those attributesthat were not importantor salient. 1996) presume that attribute-levelevaluations affect behav- To check for the salience bias, the first set of variables ioral intentions but only through an overall satisfaction was constructedfrom attributeson the basis of the number This raises several judgment. finding importantquestions of mentions. With number of mentions as a proxy for that need more research. the conditions under which First, salience, only those attributesthat received the top 3, 4, 5, or the direct of attribute-level outcomes is impact larger 6 mentions were included. To clarify, POSFRQ3 and than the one (smaller) mediated throughsatisfaction should NEGFRQ3used only those attributesthat received the first, be outlined. it could be that the role of Second, mediating second, and third highest number of mentions. POSFRQ4 overall satisfaction is different for different attributesand and NEGFRQ4 also included the attributethat received the on the dependent magnitudeand directionof the asymmetry. fourth highest numberof mentions, and so on. To check for a of Developing typology attributes should help resolve the importance bias, the cumulative variables were con- these issues. structedon the basis of their imputed importance.With the regression weight in Table 2 as a proxy for an attribute's for Practice Implications importance, variables called POSREG5 and NEGREG5 The key implication for managers is that they no longer were constructed using those attributes that had the five should view attribute performance with a neutral eye. highest regression weights for the positive and negative cat- Although positive and negative performanceon an attribute egories. Similar variablesalso were constructedon the basis are two sides of the same coin, each side of the coin buys a of the top six and seven regression weights. different amount of overall satisfactionor repurchaseinten- These strategies of integrationare consistent with prior tions. This asymmetry should be kept in mind in trying to theoretical and empirical research(e.g., Dawes 1979). Fur- manage attribute-levelperformance. The strategic implica- thermore, satisfaction researchers(e.g., Oliver 1993) have tion of this result is clear: To maximize overall satisfaction used additive-linearcombinations to integrateperformance and repurchase intentions, managers should optimize and on several attributes.To summarize,eight additionalsets of not maximize attribute-levelperformance. For example, for variables were constructed: POSFRQ or NEGFRQ3-6, any given attributeit is more importantto eliminate negative POSREG or NEGREG5-7, and POSWTD or NEGWTD5. performancefirst and then focus on increasingperformance Next, the following model was estimated for each set of in the positive direction. Similarly, in a given set of attrib- variables:

Negativeand PositiveAttribute-Level Performance / 45

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions TABLE A1 Summary of Various Tests for Asymmetry and Diminishing Sensitivity

Restrictions Test

(F-statistic, Variable bl b2 b3 b4 R2 significance) POS3FRQ (salience) .58*** -.16*** -.98*** .15* .14 9.8, .0020 POS4FRQ (salience) .53*** -.13*** -.78*** .07n.s. .15 4.9, .0300 POS5FRQ (salience) .49*** -.11*** -.73*** .02n.s. .17 6.4, .0100 POS5RG (importance) .54*** -.14*** -1.38*** .28*** .18 50.9, .0001 POS6FRQ (salience) .44*** -.09*** -.78*** .08n.s. .18 16.9, .0001 POS6RG (importance) .52*** -.13*** -1.22*** .20*** .19 44.2, .0001 POS7RG (importance) .46*** -.09*** -1.21*** .22*** .21 66.6, .0001 POSMTCH(importance) .35*** -.08** -.80*** .05n.s. .14 24.9, .0001 POSALL (all attributes/initial case) .29*** -.04*** -.72*** .07** .23 50.4, .0001 ***p< .0001; **p< .001; *p < .05; n.s.p> .10.

Overall satisfaction= bi(POSITIVE) + b2(POS[TIVE)2 nificant for all of the nine models (p < .001), but b4 was sig- + b3(NEGATIVE)+ b4(NEGATIVE)2. nificant only for five of the nine models. To ascertain the statistical The results are described in Table Al. In all of the eight "joint" significance on the basis of these nine a combined test additional cases plus the initial case, lb31> Ibll, which indi- models, such as Fisher's combined test cates that negative performance has a greater impact than would be appropriate (Wolf 1986). However, because the nine are not positive performance as shown in the columns labeled bi comparisons independent, such a test of signif- and b2 in Table A 1. This result was significant at p < .03 in icance cannot be done. Therefore, erring on the side of con- all nine cases (see the column labeled "Restrictions Test" in servatism, we conclude that over the nine tests, b4 is not sig- Table Al). These results show that the conclusion about H3 nificant, whereas b2 is. Thus, diminishing sensitivity holds is not a methodological artifact. for positive performance but not for negative performance Regarding diminishing sensitivity, both b2 and b4 have on attributes. the expected sign in all nine cases. Furthermore, b2 was sig-

REFERENCES Anderson, Eugene W. and Mary W. Sullivan (1993), "The Dawes, Robyn M. (1979), "The Robust Beauty of ImproperLinear Antecedents and Consequences of Customer Satisfaction for Models in Decision Making," AmlericaniPsychologist, 34 Firms,"Marketing Science, 12 (Spring), 125-43. (July), 571-82. Bitner, Mary Jo, BernardH. Booms, and Mary Stanfield Tetreault Einhorn,Hillel J. and Robin M. Hogarth(1981), "BehavioralDeci- (1990), "The Service Encounter: Diagnosing Favorable and sion Theory: Processes of Judgment and Choice," Annual Unfavorable Incidents,"Journal of Marketing, 54 (January), Review of Psychology, 32, 53-88. 71-84. Feinberg,Richard A., RichardWiddows, MarlayaHirsch-Wyncott, and Amy R. Hubbert(1994), "EncounterSatisfaction Ver- and Charles Trappey (1990), "Myth and Reality in Customer sus Overall Satisfaction Versus Quality: The Customer's Service: Good and Bad Service Sometimes Leads to Repur- Voice," in Service Quality:New Directions in Theoryand Prac- chase," Journal of CustomerSatisfaction, Dissatisfaction alnd tice, Roland T. Rust and Richard L. Oliver, eds. London: Sage ComplainingBehavior, 3, 112-14. Publications,72-94. Folkes, Valarie S. (1988), "Recent AttributionResearch in Con- KennethA. Bollen, (1989), StructuralEquations with Latent Vari- sumer Behavior: A Review and New Directions,"Journal of ables. New York:John Wiley & Sons. ColsunLmerResearch, 14 (March),548-65. Bolton, Ruth N. and James H. Drew (1991), "A Multistage Model Gardial, Sarah Fisher, D. Scott Clemons, Robert B. Woodruff, of Customers' Assessments of Service Quality and Value," David W. Schumann, and Mary Jane Burns (1994), "Compar- Journal of ConsumerResealrch, 17 (March), 375-84. ing Consumers'Recall of Prepurchaseand PostpurchaseProd- Boulding, William, Ajay Kalra, Richard Staelin, and Valarie A. uct Evaluation Experiences,"Journal of ConsumerResearch, Zeithaml (1993), "A Dynamic Process Model of Service Qual- 20 (March),548-60. From to ity: Expectations Behavioral Intentions,"Journal of Hall, Melvin F. (1995), "Patient Satisfaction or Acquiescence? MarketingResearrch, 30 (February),7-27. Comparing Mail and Telephone Survey Results," Jou1rnal Jacob of Cohen, (1983), "The Cost of Dichotomization," Applied Health Care Marketing, 15 (Spring), 54-61. Psychological Measulremlent,7 (Summer), 249-53. Hamilton, Lawrence C. (1992), Regression with Graphics: A Sec- Cox, Eli P., III "The Number of (1980), Optimal Response Alter- ond Course in Applied Statistics. Belmont, CA: Duxbury natives for a Scale: A Review,"Journal of MarketingResearch, Press. 27 407-22. (November), Hanson, Randy (1992), "Determining Attribute Importance," Coyne, Kevin (1989), "Beyond Service Fads-Meaningful Strate- Quirk'sMarketing Research Review, 6 (October), 16-18. for the Real gies World,"Sloan ManagementReview, 39 (Sum- Honomichl, Jack (1993), "Spending on Customer Satisfaction 69-76. mer), Continuesto Rise," MarketingNews, 27 (April 12), 17-18.

46 / Journalof Marketing,January 1998

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions Iacobucci, Dawn, Amy L. Ostrom, Bridgette M. Braig, and Alexa Oliver, Richard L. (1993), "Cognitive, Affective, and Attribute Bezjian-Avery(1996), "A Canonical Model of ConsumerEval- Bases of the Satisfaction Response," Journal of Consumer uations and Theoretical Bases of Expectations,"Advances in Research, 20 (December), 418-30. Services Marketingand Management,5, 1-44. and Wayne S. DeSarbo(1988), "Response Determinantsin Jones, Thomas 0. and W. Earl Sasser, Jr. (1995), "Why Satisfied Satisfaction Judgments,"Journal of Consumer Research, 14 Customers Defect," Harvard Business Review, 73 (Novem- (March),495-507. ber/December),88-99. Ostrom,Amy and Dawn Iacobucci (1995), "ConsumerTrade-Offs Kahn, BarbaraE. and Robert J. Meyer (1991), "ConsumerMulti- and the Evaluationof Services,"Journal of Marketing,59 (Jan- attributeJudgments under Attribute Weight Uncertainty,"Jour- uary), 17-28. nal of ConsumerResearch, 17 (March),508-22. Parasuraman,A., ValarieZeithaml, and Leonard L. Berry (1988), Kahneman,Daniel and Amos Tversky (1979), "ProspectTheory: "SERVQUAL:A Multiple-ItemScale for MeasuringConsumer An Analysis of Decisions Under Risk," Econometrica, 47 Perceptions of Service Quality," Journal of Retailing, 64 (March), 263-91. (Spring), 12-40. Kekre, Sunder, Mayuram S. Krishnan, and Kannan Srinivasan Pedhazur, Elazar J. (1982), Multiple Regression in Behavioral (1995), "Drivers of Customer Satisfaction for Software Prod- Research, 2d ed. New York:Holt, Rinehartand Winston. ucts: Implications for Design and Service Support,"Manage- Peeters, G. and J. Czapinski (1990), "Positive-Negative Asymme- ment Science, 41 (September), 1456-70. try in Evaluations: The Distinction Between Affective and LaBarbera,Priscilla A. and David Mazursky(1983), "A Longitu- InformationalNegativity Effect," European Review of Social dinal Assessment of Consumer Satisfaction/Dissatisfaction: Psychology, 1, 33-60. The Dynamic Aspect of the Cognitive Process," Journal of Reichheld, Fredrick(1996), The Loyalty Effect: The Hidden Force MarketingResearch, 20 (November), 393-404. Behind Growth, Profits, and Lasting Value. Boston: Harvard LaTour,Stephen A. and Nancy C. Peat (1979), "Conceptualand Business School Press. Methodological Issues in Consumer Satisfaction Research,"in Spreng, Richard A., Scott B. MacKenzie, and Richard W. Advances in Consumer Research, Vol. 6, William D. Perrault Olshavsky (1996), "A Reexaminationof the Determinantsof Jr., ed. Ann Arbor, MI: Association for Consumer Research, ConsumerSatisfaction," Journal of Marketing,60 (July), 15-32. 431-37. Taylor,Shelly (1982), "TheAvailability Bias in Social Perception," Mittal, Vikas, Jerome M. Katrichis, Frank Forkin, and Mark in JudgmentUnder Uncertainty:Heuristics and Biases, Daniel Konkel (1993), "Does Satisfaction With Multi-AttributeProd- Kahnemanet al., eds. Cambridge:Cambridge University Press, ucts Vary Over Time? A Performance Based Approach," in 190-200. Advances in Consumer Research, Vol. 21, Chris T. Allen and Wilkie, William L. and EdgarA. Pessemier (1973), "Issues in Mar- Deborah Roedder John, eds. Provo, UT: Association for Con- keting's Use of Multi-AttributeAttitude Models," Journal of sumer Research, 412-17. MarketingResearch, 10 (November), 428-41. Oliva, Terence A., Richard L. Oliver, and William O. Bearden Wittink,Dick R. and LeonardR. Bayer (1994), "The Measurement (1995), "The Relationship Among Consumer Satisfaction, Imperative,"Marketing Research, 6 (Winter), 14-23. Involvement, and Product Performance,"Behavioral Science, Wolf, Fredric M. (1986), Meta-Analysis: Quantitative Methods for 40 (April), 104-32. Research Synthesis. London: Sage Publications. , and lan C. MacMillan (1992), "A Catastrophe Yi, Youjae (1990), "A Critical Review of Consumer Satisfaction," Model for Developing Service SatisfactionStrategies," Journal in Review of Marketing 1990, Valarie A. Zeithaml, ed. Chicago: of Marketing, 56 (July), 83-95. American MarketingAssociation, 68-123.

Negativeand Positive Attribute-LevelPerformance /47

This content downloaded from 140.120.52.117 on Thu, 18 Dec 2014 21:27:50 PM All use subject to JSTOR Terms and Conditions