what you don’t know about customer- perceived...

0732-2399/99/1801/0077$05.00Copyright q 1999, Institute for Operations Researchand the Management Sciences

Marketing Science/Vol. 18, No. 1, 1999

pp. 77–92

What You Don’t Know About Customer-Perceived Quality: The Role of Customer

Expectation Distributions

Roland T. Rust • J. Jeffrey Inman • Jianmin Jia • Anthony ZahorikOwen Graduate School of Management, Vanderbilt University, Nashville, Tennessee 37203,

[email protected] of Business, University of Wisconsin, Madison, Wisconsin 53706, [email protected]

Department of Marketing, Chinese University of Hong Kong, NT, Hong Kong, [email protected] Burke Institute, Nashville, Tennessee 37212, [email protected]

AbstractWe show that some of the most common beliefs aboutcustomer-perceived quality are wrong. For example, 1) it isnot necessary to exceed customer expectations to increasepreference, 2) receiving an expected level of bad service doesnot reduce preference, 3) rational customers may rationallychoose an option with lower expected quality, even if all non-quality attributes are equal, and 4) paying more attention toloyal, experienced customers can sometimes be counter-productive. These surprising findings make sense in retro-spect, once customer expectations are viewed as distribu-tions, rather than simple point expectations. That is, eachcustomer has a probability density function that describes therelative likelihood that a particular quality outcome will beexperienced. Customers form these expectation distributionsbased on their cumulative experience with the good or ser-vice. A customer’s cumulative expectation distribution maybe conceptualized as being a predictive density for the nexttransaction.

When combined with a diminishing returns (i.e., concave)utility function, this Bayesian theoretical framework resultsin predictions of: (a) how consumers will behave over time,and (b) how their perceptions and evaluations will change.In managerial terms, we conclude that customers considernot only expected quality, but also risk. This may help ex-plain why current measures of customer satisfaction (whichis highly related to expected quality) only partially predictfuture behavior. We find that most of the predictions of ourtheoretical model are borne out by empirical evidence fromtwo experiments. Thus, we conclude that our approach pro-vides a useful simplification of reality that successfully pre-dicts many aspects of the dynamics of consumer response toquality.

These findings are relevant to both academics and man-agers. Academics in the area of customer satisfaction and

service quality need to be aware that it may be insufficientto measure only the point expectation, as has always beenthe standard practice. Instead it may be necessary to measurethe uncertainty that the customer has with respect to the levelof service that will be received. Due to questionnaire lengthconstraints, it may not be practical for managers to includeuncertainty questions on customer satisfaction surveys. Nev-ertheless it is possible to build a proxy for uncertainty bymeasuring the extent of experience with the service/good,and this proxy can be used to partially control for uncertaintyeffects.

The findings of the study were obtained using 1) an ana-lytical model of customer expectation updating, based on aset of assumptions that are well-supported in the academicliterature, and 2) two behavioral experiments using humansubjects: a cross-sectional experiment, and a longitudinal ex-periment. Both the analytical model and the behavioral ex-periments were designed to investigate the effects that dis-tributions of expectations might have, and especially theeffects that might deviate from the predictions that wouldarise from a traditional point expectation model. The behav-ioral experiments largely confirmed the predictions of theanalytical model. As it turned out, the analytical model cor-rectly (in most cases) predicted behavioral effects that con-tradict some of the best-accepted “truisms” of customersatisfaction.

It is now clear that a more sophisticated view of customerexpectations is required—one that considers not only thepoint expectation but also the likelihood across the entiredistribution of possible outcomes. This distinction is not “justacademic,” because it results in predictable behavior that de-viates significantly from that which was traditionally ex-pected based on simpler models.(Quality; Customer Satisfaction Measurement; Customer Expec-tations; Customer Retention; Bayesian Updating; Customer Life-time Value)

WHAT YOU DON’T KNOW ABOUT CUSTOMER-PERCEIVED QUALITY:THE ROLE OF CUSTOMER EXPECTATION DISTRIBUTIONS

78 Marketing Science/Vol. 18, No. 1, 1999

1. IntroductionThe trade literature in quality and customer satisfac-tion abounds with rarely questioned platitudes. Someof the most often repeated and/or noncontroversialare:

• “It is necessary to exceed customer expectations.”• “If a customer expects a bad level of quality and receives

it, he/she will reduce his/her level of preference for thebrand.”

• “Given two equally-priced options, the customer willchoose the one with the higher expected quality.”

• “Management should always focus on its most loyalcustomers.”

We will show, using both an analytical model andbehavioral experiments, that all of these truisms areflawed. Management must instead adopt a more so-phisticated view of how customer expectations are up-dated, and how customer expectation updating relatesto preference and future choice behavior.

1.1. What’s Wrong with Customer SatisfactionMeasures?

Managers routinely conduct customer satisfaction sur-veys and use that information to produce explanatoryand predictive models of customer repurchase inten-tion, word-of-mouth intention, customer repurchasebehavior, and market share (Bolton and Drew 1991a,1991b; Boulding et al. 1993; Cronin and Taylor 1992,1994; Fornell 1992; Oliver and DeSarbo 1988;Parasuraman et al. 1985; Rust and Zahorik 1993; Teas1993).1 Such models may be used to evaluate the pro-jected profit impact of service improvement programs(Rust et al. 1995). Recently, some authors have ques-tioned the predictive ability of customer satisfactionmeasures (Gale 1997, Reichheld 1996). They have com-plained that many customers who report being “verysatisfied” or perceive quality to be “excellent” never-theless subsequently switch to a competitor. In general,while customer satisfaction measures and/or per-ceived quality measures are important predictors ofintentions and behavior, they often explain only 30–50% of the variance.

We argue that one of the reasons for the apparently

1Strictly speaking, many customer satisfaction surveys actually mea-sure perceived quality.

weak satisfaction-behavior link may be that customersrespond according to the perceived variance in service,in addition to expected quality. In other words, cus-tomers’ certainty about quality has an important im-pact. This perceived risk can be related to the varianceof a Bayesian predictive distribution of quality out-comes, resulting in a more multidimensional view ofcustomer expectations. This viewpoint results in someapparently counter-intuitive insights that have someimportant managerial implications.

1.2. Dynamics and the Decision TheoryPerspective

Because customer relationships unfold over time, it isimportant to clearly understand the dynamics of howquality perceptions are formed and updated as well ashow such perceptions influence customer retentionover time. With few exceptions (Anderson andSullivan 1993, Bolton and Drew 1991a, Boulding et al.1993), the research literature has not produced dy-namic models of how these changes occur. Researchershave begun to argue that a decision theory frameworkoffers a powerful way of conceptualizing the dynamicsof quality and customer retention (Anderson andSullivan 1993, Boulding et al. 1993). Bayesian decisiontheory (Berger 1985, Zellner 1971) provides a way ofunderstanding customer retention that has thus farbeen underutilized. This literature provides well-explored methods of describing how people incorpo-rate new information and form new expectations overtime. Thus, it can be used to generate testable predic-tions of behavior over time (Kahneman and Tversky1972).

Essentially there are two major issues in the field ofbehavioral decision research-risky choice behavior andprobability judgments. Expected utility can be deter-mined by mean and variance if the distribution is nor-mal or if the utility function is quadratic (Markowitz1959, 1987; Meyer 1987). Behavioral research on riskychoice has revealed discrepancies between actualchoice behavior and the prescriptions of the classic vonNeumann and Morgenstern (1944) utility axioms. Forexample, Payne (1973) argued that factors other thanmean and variance (or even higher moments) areneeded to capture risk perceptions and risky choicebehavior. Empirical studies have challenged the tra-ditional utility theory on several grounds (e.g.,

RUST, INMAN, JIA, AND ZAHORIKThe Role of Customer Expectation Distributions

Marketing Science/Vol. 18, No. 1, 1999 79

Kahneman and Tversky 1979; Bell 1982, 1985; Machina1987; Inman et al. 1997). This has resulted in a muchdeeper understanding of risky choice behavior, whilemaintaining much of the mathematical formalism ofthe normative approach (see Fishburn (1988) for a com-prehensive review of generalizations of expected util-ity theory).

Behavioral research on probability judgments hasproduced mixed results regarding the descriptivepower of Bayesian decision theory. In studies of prob-ability judgments, prior probability (or base-rate) in-formation is often ignored by subjects (Kahneman andTversky 1973), but sometimes it is utilized appropri-ately (Gigerenzer et al. 1988). Moreover, people havea tendency to seek confirmatory evidence in aggregat-ing various pieces of information, which may lead tothe failure to change one’s opinion in the face of non-supporting evidence (Wason 1960, Doherty et al. 1979).Another relevant issue is the underestimation of pos-terior probability, called “conservatism” (Edwards1968, Slovic and Lichtenstein 1971). That is, upon re-ceipt of new information, subjects revise their posteriorprobability estimates in the same direction as theBayesian model, but revisions are typically much lessextreme than those calculated from the Bayesian per-spective. In general, probabilistic judgments involvecontingent processing and heuristics, which some-times yield reasonable results and sometimes lead tosystematic biases (Tversky and Kahneman 1974,Fischhoff and Beyth-Marom 1983). People often utilizea variety of rules for making probabilistic inference indifferent environments, including both statistical andnonstatistical methods (Ginossar and Trope 1987).

Although previous research has generated severalempirical findings of risky choice behavior and prob-ability judgments, there is surprisingly little empiricalevidence regarding a dynamic decision model thatuses Bayesian updating for revising expectation distri-butions upon receipt of new information. In particular,previous research on probability judgments based onBayesian updating has used discrete probability infor-mation and/or nonbusiness contexts. In contrast, ourstudy focuses on the descriptive validity of a Bayesiandecision model using continuous probability infor-mation in the area of quality perceptions and customerretention.

1.3. Comparison with Previous ModelsBeginning with SERVQUAL (Parasuraman et al. 1988)service quality measurement instruments have univer-sally focused on a point estimate of quality. The im-plicit assumption has always been that a customer’s“perceived quality” or “expected quality” was a singlepoint. This viewpoint has also been shared by the re-cent dynamic models of customer satisfaction. Suchmodels involve the updating of expectations and qual-ity perceptions. For example, Boulding et al. (1993) em-ploy a linear updating scheme by which expectationsand cumulative quality perceptions are updated ac-cording to the most recent transaction quality percep-tion. Their model is analogous to an exponentialsmoothing model. Likewise, Bolton and Drew(1991a,b) employ linear updating functions. OnlyAnderson and Sullivan (1993) suggest (but do not de-velop) an updating approach that involves a distribu-tion rather than only a point estimate.

Our model is different in that it involves a fully de-veloped Bayesian updating scheme, in conjunctionwith a diminishing returns utility function. We will seethat the behavioral predictions arising from our modelare quite different in some cases from the predictionsarising from linear updating models. Our insights arisefrom the observation that experience with a brandmakes knowledge of the brand’s quality more com-plete, thereby reducing the customer’s risk. Becausethe customer’s utility function is concave (i.e., dimin-ishing returns), reduced risk is always good and in-creases the customer’s preference for the brand. Thisresults in some seemingly counter-intuitive results thatare nevertheless quite logical under closer inspection.Importantly, these results have key consequences formanagerial practice.

1.4. Plan of the PaperIn the next section we develop a theoretical frameworkthat can be used to describe the relationship betweencustomer retention and quality perceptions over time,based on the concepts of expected utility maximizationand Bayesian updating. This framework results in sev-eral propositions about how customers should behaveover time. We then present two behavioral experi-ments to test whether or not the theoretical proposi-tions hold up on actual behavior. Finally, we discuss



the implications of our results, as well as limitationsand directions for future research.

2. Model DevelopmentIn this section we present the relationship betweenquality and customer retention as a model in whichretention is based on the distribution of customer ex-pectations. Importantly, we allow the distribution tobe updated over time according to the quality per-ceived in a particular transaction. While previous re-search (e.g., Anderson and Sullivan 1993, Boulding etal. 1993) has pointed in this direction, this is the firstattempt to formulate these phenomena in a fullyBayesian framework, complete with operationalizedprior and posterior distributions of quality. We beginby presenting the assumptions on which our theoreti-cal model is based. We then present the theoretical for-mulation in formal detail. Finally, we provide a list ofnontrivial testable propositions that arise from themodel.

2.1. AssumptionsWe assume a scenario in which an individual with ex-isting expectations chooses a particular option, expe-riences an outcome, and then changes his/her expec-tations based on the outcome.

Assumption 1. Utility as a function of perceived qualityis continuous, twice differentiable, increasing and concave.This implies that customers suffer more from not havingtheir expectations met than they benefit from an equivalentpositive disconfirmation.

This assumption is borne out by prior empirical re-search in customer satisfaction and service quality(Anderson and Sullivan 1993, DeSarbo et al. 1994,Inman et al. 1997, Rust et al. 1995), and is also consis-tent with assumptions traditionally made in decisiontheory (Keeney and Raiffa 1976). This particular shapeof the utility function, well-documented in a variety ofacademic and proprietary studies, combines with theBayesian framework to produce interesting and man-agerially relevant, testable conclusions.

Assumption 2. The customer has a predictive distribu-tion of outcomes that reflects the relative likelihood that eachoutcome will occur.

This assumption says that the customer considersnot only the outcome s/he expects on average (i.e., thepoint expectation), but also the entire distribution ofpossible outcomes.

Assumption 3. In any purchase situation, the cus-tomer’s preference for an option increases with its expectedutility, and probability of choosing an option increases withpreference for that option.

The assumption is consistent with the economic the-ory of utility and the consumer viewpoint (Luce 1959,Manrai 1995, McFadden 1976) with randomness aris-ing from the variability of the predictive distribution.

Assumption 4. The customer has a prior distribution ofaverage quality for the product or service.

Through experience, a customer learns what qualitycan be expected on average. This prior distributionmay be formed on the basis of experience with otherbrands, on the basis of prior experience with the brandunder consideration, or both. Similar assumptions arecommonly made in a variety of Bayesian applications(e.g., Little 1966).

Assumption 5. The customer updates the prior distri-bution based on the perceived quality of the transaction us-ing a Bayesian updating process.

Such an updating mechanism has been suggested byprevious research in the service quality area (Andersonand Sullivan 1993, Boulding et al. 1993, Kopalle andLehmann 1995).2

Assumption 6. Perceived quality varies randomly acrosstransactions.

Even in a highly controlled environment such as amanufacturing assembly line, there are numerous un-controllable factors that can produce variability in thequality of the output (Burr 1976). It is notable that inservice, which is a growing majority of every devel-oped economy, variability is typically much greater

2We tested Assumption 5 in our first experiment by estimating therelative weight subjects placed on prior perceptions versus theweight placed on new information in updating the performance ex-pectation. The weight on new information was significantly positive.Thus, consistent with Assumption 5, subjects used the informationto update their perceptions of the expected performance.



than in manufacturing. While an entire field of study,statistical process control (Juran and Gryna 1980), hasarisen to identify and attempt to control the sources ofvariation in quality, some variability inevitably re-mains even in a process that is “under control”(Deming 1986). Another source of variation is percep-tion error on the part of the customer (Thurstone 1927).We recognize that prior research indicates that vari-ables such as prior expectations may influence thequality perception itself (e.g. Boulding et al. 1998,Oliver 1997), but to keep this initial model simpleenough to provide maximum insight, we suppress sec-ondary effects here.

2.2. Mathematical FormulationWe now build a formal mathematical structure thatincorporates the preceding assumptions. Let the cus-tomer have a known prior distribution p(Q) of the av-erage quality (Q) of the brand. Let p(Q) have a normaldistribution with mean l and variance s2 . 0, reflectingthe customer’s degree of uncertainty. Further, for aparticular transaction, we will denote the customer’sperceived quality as X. We assume that X is distributednormally, with mean Q and variance r2 . 0 repre-senting random variation arising from both variabilityof quality and errors in perception. It is clear from stan-dard Bayesian analysis (Berger 1985, p. 127) that thejoint density of Q and X is:

11 2 2h(x,Q) 4 (2prs) exp{11/2 [((Q 1 l) /s )

2 2` ((x 1 Q) /r )]} (1)

from which we can get the predictive (marginal distri-bution) of X:

`

p(x) 4 h(x,Q)dQ#1`

11/2 11 24 (2pq) (rs) exp{1(l 1 x)

2 2/[2(r ` s )]} (2)

where q 4 (r2 ` s2) / (r2s2). If we scale the units (with-out loss of generality) such that r2 ` s2 4 1, then thepredictive distribution becomes

11/2 2p(x) 4 (2p) exp{1(l 1 x) /2} (3)

which is immediately seen as being normal with meanl and variance one.

We now address what happens when a level of qual-ity, xt, is observed on the next transaction. We also de-fine the disconfirmation to be Dt 4 xt 1 l, following thetraditional definition (Oliver 1980). (It is worth notingthat the disconfirmation, although conceptualized as adifference, is not usually calculated mathematically,but rather observed, and measured, directly.) Againfrom standard Bayesian updating, the posterior distri-bution of Q, p(Q|xt), is calculated as:

p(Q|x ) 4 h(x ,Q)/p(x )t t t

1/24 (q/2p) exp{1(q/2)[Q 1 (1/q)

2 2 2((l/s ) ` (x /r ))] } (4)t

which is a normal distribution with mean l ` Ds2 andvariance q11 4 r2s2. Thus, the posterior mean in-creases (decreases) when the disconfirmation, D, ispositive (negative). Also the customer’s uncertaintydecreases, regardless of the outcome, since q11 , s2.The predictive density of the quality of the next trans-action, xt`1, given observed quality level xt, is

`

p(x |x ) 4 f(x |Q)p(Q|x )dQ (5)t`1 t t`1 t#1`

where f(xt`1|Q) is normal with mean l and variancer2. This predictive density is easily shown to be normalwith mean l ` Ds2 and variance r2 ` r2s2. The pre-dictive density’s new mean is bigger (smaller) if thedisconfirmation, D, is positive (negative), and its vari-ance is smaller regardless (since r2 ` r2s2 # 1), reflect-ing the greater certainty created by experience.

From (4) we can see that the mean of the posteriordistribution exceeds that of the prior distributionwhenever the disconfirmation, D, is positive and islower when D is negative. The same holds for the pre-dictive distributions. In other words, the expectations(and future predictions) move in the direction of theperceived level of quality. Note also from (4) that thevariance is reduced, as makes sense from having moreexperience. From Assumption 1, we assume a contin-uous, twice differentiable utility function U(x), with U8

. 0 and U9 , 0. From Assumption 3, we assume anexpected utility V, that is equal to *U(x)p(x)dx, wherep(x) is the predictive distribution of x. The expectedutility can be expressed in terms of mean and standarddeviation since the predictive distribution is normal.



For example, if we assume an exponential utility func-tion, such as is popularly used in modeling riskychoice problems, then the expected utility can be de-termined by a tradeoff between expected value andrisk (Jia and Dyer 1996):

2V 4 a ` a l 1 a r (6)i 1i 2i i 3i i

where Vi is the preference measure of brand i, li is theexpected performance of brand i, is the variance2ri

measuring perceived risk or uncertainty of brand i’sperformance, and a1i and a21, a3i . 0 are constants.This is a positive linear transformation of the certaintyequivalent form of the exponential expected utility.The ratio a2i/a3i measures customers’ risk tolerance,reflecting a tradeoff between the mean and variance indetermining preference.

Based on Assumption 3, the consumer chooses thebrand with the highest expected utility. Thus, a mul-tinomial logit model (Luce 1959, McFadden 1974) canbe used to estimate the choice likelihood for eachbrand:

p 4 exp(V ) exp(V ). (7)i i o j@j

2.3. Propositions Arising from the ModelFrom the model above, a number of empirically test-able propositions arise that warrant further scrutiny.(Proofs of all of the propositions are available in anAppendix that is available from the authors on re-quest.) Following are some of the propositions result-ing from the model that would seem to have a reason-able chance of being empirically falsified:

Proposition 1. If an option is chosen and a better thanexpected outcome is observed, then the probability of choos-ing that option will increase.

Intuitively, this proposition follows from the theo-retical formulation whereby the predictive distribu-tion’s mean increases (and the predictive distributionshrinks). Preference improves because quality nowseems better on average, and also less risky.

Proposition 2A. If the preferred option is chosen andan expected outcome is observed, then the probability ofchoosing that option will increase.

The logic here is that if the utility function is con-cave, reducing the variance of the predictive distribu-tion will increase the expected utility. Preference im-proves not because the option seems better, but ratherbecause it is less risky.

Proposition 2B. If a nonpreferred option is chosen (asis possible in our probabilistic choice model) and an expectedoutcome is observed, then the probability of choosing thatoption will increase.

Again, the shrinking predictive distribution in-creases the expected utility. We separate Propositions2A and 2B because 2B will not obviously hold even ifProposition 2A is supported. Shrinking variancemakes it more certain that the nonpreferred option willbe worse, which would seem to suggest a plausiblenon-Bayesian basis for which this proposition may nothold. As before, preference for the option improves be-cause the option seems less risky.

Proposition 3. A rational consumer may choose anequally-priced option for which the expected quality is worse.

The intuition behind this proposition is risk aver-sion. If one option has a larger variance for its predic-tive distribution, the uncertainty regarding this per-formance may induce many consumers to choose alower expected performance/less variance option. Dueto this, a less risky option can be preferable to a morerisky option with a higher mean.

Proposition 4. A worse than expected quality outcomemay still increase the probability of choice for that option.

The intuition for Proposition 4 is that there are twocountervailing effects at work. The worse than ex-pected outcome (all other things being equal) will tendto lower the predictive distribution and the expectedutility. However, at the same time the variance of thepredictive distribution is being reduced which will in-crease expected utility, given a concave utility func-tion. For a small enough negative disconfirmation, thepositive effect of the variance reduction will outweighthe negative effect of the lowered expectation, leadingto an increase in preference.

Proposition 5. A negative disconfirmation will evoke agreater relative change in preference relative to the case of



zero disconfirmation than will a positive disconfirmation ofequal magnitude.

As mentioned previously, there will be two effectson preference: a variance reduction effect and a dis-confirmation effect. Experience always reduces vari-ance, resulting (all other things being equal) in in-creased preference. Positive disconfirmation increasespreference, and negative disconfirmation (all otherthings being equal) decreases preference. Because ofthe concavity of the utility function, increasing abso-lute disconfirmation will create a disparity between theabsolute effect of a negative disconfirmation and thatof a positive disconfirmation. Proposition 5 refers tothe case of zero disconfirmation as the basis, whicheliminates the variance reduction effect (i.e., all caseshave the same reduced variance). Since there is only adisconfirmation effect at work, this results in a simplepattern of asymmetric changes in preference.

Proposition 6: Given diffuse priors, and an equalhistorically-observed mean and variance, a sufficiently largenegative disconfirmation will cause a greater preference shiftin a less experienced customer. (Also a positive disconfir-mation will cause a greater preference shift in a less expe-rienced customer.)

This effect occurs because more experience producesless uncertainty, but at a decreasing rate. In otherwords, experience has diminishing returns. Thereforethe value gained from experience is large at first, re-sulting in large changes in preference, but diminishes,eventually resulting in small changes in preference.This is consistent with other research that shows thatpreference shifts less for more experienced customers(Bolton 1998, Boulding et al. 1993, 1998).

While the theoretical decision model forms a math-ematically rigorous way of thinking about quality per-ceptions and customer retention, it is still only a math-ematical abstraction. To determine the usefulness ofthis approach, we must test whether customers’ be-havior over time corresponds with the model’s predic-tions. We now present results of two experiments inwhich the propositions were subjected to empiricaltest. We then discuss the implications of these resultsand directions for future research.

3. Longitudinal Experiment3.1. OverviewOne hundred and sixty undergraduate students at twolarge universities participated in a computerizeddecision-making exercise in return for extra credit. Theexperiment was designed to measure the effect of ex-pectation distributions on discrete choice and onchoice probability. The exercise consisted of the con-struction of a history of experiences among threebrands of camera batteries. Disconfirmation was ma-nipulated, while probability of choice, performance ex-pectation, and perceived variance in performance foreach brand of battery were measured at several pointsin the experiment.

The independent variables are 1) the disconfirma-tion (D 4 xt 1 l) between the actual and expectedperformance of the chosen battery, 2) the expected per-formance (l) of each battery, and 3) the perceived var-iance (r2) of each battery. Disconfirmation was manip-ulated, while expected performance and perceivedvariance were measured. Expected performance foreach battery was measured by the question “I’d expectthe next battery (A/B/C) to last hours,” while per-ceived variance for each battery was captured via sub-jects’ response to the question “About 95% of the bat-teries for Brand (A/B/C) last between and hours.”

Five different levels of disconfirmation were used:ten hours above expected, zero disconfirmation, andone, three, and ten hours below expected. The com-puter exercise interactively managed the amount ofdisconfirmation so that each subject was exposed totwo of the five different levels of disconfirmation inexperiences 4 and 7. For example, if the subject was tobe exposed to the three hour disconfirmation treat-ment, the computer provided outcome feedback forthe purchase by subtracting three hours from the sub-ject’s expectation for that brand.3

3.2. MethodWe used an unbalanced design with more subjects inthe zero and small disconfirmation (negative one andthree hours) conditions to increase statistical power forthose disconfirmation conditions. Sample sizes are

3The order of exposure to the amount of disconfirmation was coun-terbalanced.



Table 2 Longitudinal Experiment—Means Across Conditions

Shift in Choice Probability

Disconfirmation Level NOverallShift

LowExperience

HighExperience

`10 Hours 33 8.70** 14.31 3.410 Hours 112 5.21** 5.93 4.44Preferred option 40 2.88*Nonpreferred option 72 6.51**11 Hours 70 0.63 1.44 10.5213 Hours 72 10.42 0.76 11.41110 Hours 33 112.09** 113.81 110.47

**Shift statistically different than zero (p , 0.01).

*Shift statistically different that zero (p , 0.05).

Table 1 Longitudinal Experiment—Performance Feedback AcrossBrands

BrandsExp1

Exp2

Exp3

Exp4

Exp5

Exp6

Exp7

Exp8

Exp9

Exp10

1 72 59 67 E13–D1 74 57 E16–D2 72 56 692 63 69 60 E23–D1 59 68 E26–D2 67 64 653 66 60 64 E33–D1 65 67 E36–D2 59 67 63

Eir 4 Subject’s performance expectation for Brand i after round r.

Dj 4 Disconfirmation level for condition j.

shown in Table 2. Further, the brand labels were coun-terbalanced across subjects to control for order effects.Each subject was asked to imagine that s/he had re-cently purchased a brand of camera which used a spe-cial type of battery. A special type of battery was usedin order to mitigate effects of category familiarity (i.e.,so that each subject began the procedure with a diffuseexpectation). The subject was told that s/he had sam-pled each brand of battery three times and was ac-cordingly shown the first series of three experiences(see Table 1, columns 1–3). The subject was asked toprovide his/her choice probability for each battery andto give his/her performance expectation and perceivedvariance for each battery. At the beginning of thefourth experience, the subject was told that s/he hadpurchased one of the three brands4 (randomized acrosssubjects) and was exposed to the first disconfirmationcondition. Choice probability was then remeasured.The outcomes for the other two brands were thengiven and expected performance and perceived vari-ance in performance were measured.

Following the fourth experience, subjects completeda brief distractor task and were then given outcomefeedback on two more purchases of each brand (seeTable 1, columns 5 and 6). Performance expectationsand perceived variance were remeasured following thesixth experience. At the beginning of the seventh ex-perience, the subject was again told that s/he had pur-chased one of the three brands (different from the

4Subjects were told to imagine that they had purchased the brand,“perhaps because it was on sale or because you had a coupon forit.”

brand in the fourth experience) and was given out-come feedback for this brand. Choice probability wasremeasured. The subject was then given outcome feed-back on the other two brands to complete the seventhexperience. Following a second distractor task, subjectswere provided outcome feedback for three more pur-chases of each battery, yielding a final history of tenpurchases of each battery. Finally, expected perfor-mance and perceived variance were measured and thesubject was asked which battery she would choose.

3.3. ResultsTable 2 shows the means for each experimental con-dition. Most of our propositions regard shifts in choiceprobability at specific levels of disconfirmation. Prop-osition 1 predicts that the probability of choosing anoption will increase when the observed outcome isgreater than expected. This proposition is supported,as probability of choice increased by an average of 8.7points for the chosen option when the observed out-come exceeded the expected performance (t32 4 4.76,p , 0.01).5 As predicted, when consumers experiencedpositive disconfirmation, their probability of choosingthe option increased.

Proposition 2A predicts that if an expected outcomeis observed (i.e., zero disconfirmation) for the most

5Although most of our propositions are directional, to be conserva-tive we use a two-tailed test in our analysis.



preferred option, the probability of choosing the op-tion will increase. Proposition 2B makes a similar pre-diction, but for a nonpreferred option. Both proposi-tions are supported. The probability of choosing theoption increased when the option was the most pre-ferred option, shifting almost three points (t39 4 1.73,p , 0.10), while the probability of choosing the optionwhen it was not the most preferred option increasedover six points (t71 4 5.26, p , 0.01).

Proposition 3 predicts that subjects will not neces-sarily choose the brand with the highest expected per-formance. We tested Proposition 3 by asking subjectsto choose a brand after observing the ten experienceswith each brand. Using the multinomial logit model in(6) and (7), as one would anticipate, the expected per-formance exerts a significant effect on choice (a2 4

0.235, p , 0.01). Importantly, consistent with Proposi-tion 3, the perceived variance also exerts a significantimpact on subjects’ choice (a3 4 10.383, p , 0.01).Thus, a brand’s choice probability increases as the ex-pected performance increases and decreases as the per-ceived variance increases.

Proposition 4 predicts that an outcome that is worsethan expected may still increase the probability ofchoosing the option. We test this proposition by ex-amining the effect on choice probability of disconfir-mation levels of one, three, and ten hours less thanexpected. This proposition is not supported. The av-erage probability of choosing the option increasedslightly (i.e., in the expected direction) in the negativeone hour disconfirmation condition, but the increase isnot statistically significant (t69 4 0.50, NS). In the neg-ative three hour disconfirmation condition, the aver-age probability of choosing the option decreasedslightly, but again this decrease is not statistically sig-nificant (t71 4 10.34, NS).

Proposition 5 predicts that a negative disconfirma-tion will have a greater impact on preference than apositive disconfirmation of equal magnitude. We testthis by contrasting the 110 and `10 hour disconfir-mation groups. Let X1 be the difference between `10and zero disconfirmation effects and X2 be the differ-ence between zero and 110 disconfirmation effects.The variance of X1 and X2 is the sum of the varianceof the `10 and zero and 110 and zero disconfirmationeffects, respectively, and the variance of X1 1 X2 is the

sum of the variances of X1 and X2. The t-test of thedifference between X1 and X2 is statistically significant(t65 4 4.69, p , 0.01), supporting Proposition 5.

Since Proposition 6 states that an equivalent discon-firmation level results in a smaller shift in preferencefor more experienced customers, we conducted an AN-OVA with experience level (high/low), disconfirma-tion, and their interaction as independent variablesand the shift in preference as the dependent variable.6

As predicted by Proposition 6, the effect of experiencewas significant (F1,310 4 3.99, p , 0.05), with smallershifts at the higher experience level. Specifically, choiceprobability shifted by 2.7 points when experience wasrelatively low (four experiences) and shifted only 0.3points when experience was higher (seven experi-ences). Not surprisingly, the shift in choice probabilitywas an increasing function of the amount of disconfir-mation (F4,310 4 22.09, p , 0.01). Further, the interac-tion between experience level and disconfirmation wasmarginally significant (F4,310 4 2.02, p , 0.10).

In sum, we find support for Propositions 1, 2A, 2B,3, 5, and 6. Subjects’ probability of choosing an optionincreased if performance for that option met (Propo-sitions 2A and 2B) or exceeded (Proposition 1) the ex-pected outcome. Per Proposition 3, subjects did notnecessarily choose the brand with the greatest ex-pected performance. Rather, they balanced the brand’sexpected performance against its variability in perfor-mance. Further, while subjects’ probability of choosingan outcome was not adversely affected by an outcomethat was slightly below their prior expectations, theirprobability of choice did not increase (significantly) aspredicted by Proposition 4. The preference shifts cre-ated by negative vs. positive disconfirmations weregreater in the negative disconfirmation direction, aspredicted in Proposition 5. Proposition 6 was sup-ported as well—at a given disconfirmation level, moreexperienced subjects tended to update their expecta-tion to a lower degree than less experienced subjects.

While these results are encouraging, our first exper-iment is flawed in two respects. First, the longitudinal

6An analysis using prior preference as a covariate and preference asthe dependent variable produced almost identical results. For easeof exposition, we discuss the analysis with shift in preference as thedependent variable.



nature of the design may have increased the chancethat subjects saw that two factors were being manip-ulated and inferred the appropriate responses. How-ever, it is difficult to see how this can explain our asym-metry results. Second, we did not clearly demonstratethat, holding satisfaction/quality fixed, choice is influ-enced by the variance around this construct. Our sec-ond experiment addresses both of these concerns.

4. Cross-Sectional Experiment4.1. OverviewThe second experiment was a 2 (high/low amount ofexperience) 2 2 (zero/negative disconfirmation)between-subjects design. Two hundred and twentythree undergraduates participated in the experimentin return for course credit. Three subjects misunder-stood the directions and were eliminated from theanalysis, leaving a sample of 220 subjects.

4.2. Independent and Dependent MeasuresAs in the first experiment, our independent variablesare experience level (a proxy for s2) and disconfirma-tion (D). High experience subjects were exposed to 20experiences, while low experience subjects were ex-posed to only three experiences. In the zero disconfir-mation condition, subjects were shown an additionaloutcome that was equal to the mean of the previousexperiences. Subjects in the negative disconfirmationcondition were shown an outcome that was two stan-dard deviations below the mean of their previousexperiences.

Our primary dependent variable is perceived qual-ity. We adopted the measures of perceived qualityused by Boulding et al. (1993), a 100-point scale an-chored by “Very unfavorable” and “Very favorable.”This measure was taken immediately following the ex-perience manipulation (l) and again following the spe-cific visit outcome manipulation (l ` Ds2). We alsoused the measures developed by Boulding et al. (1993)to assess purchase intentions and likelihood to spreadword of mouth, 100-point scales anchored by “Veryunlikely” and “Very likely.” Like Boulding et al.(1993), we find that these two measures are highly cor-related (a 4 0.86), so we combine them in our analysis.

As manipulation checks, we took several measures

following the experience manipulation.7 First, we as-sessed perceived variability of the service and per-ceived consistency of the service. The first measureasked subjects to rate how variable they had found theservice to be, while the second asked them to rate howconsistent the service had been. We then asked subjectsto rate how sure they were of the average level of qual-ity of Cafe Au Lait and whether they were an occa-sional customer or a regular customer. Following thespecific visit manipulation, we asked subjects howclosely the experience on that occasion matched theirexpectations to assess subjective disconfirmation. Thiswas accomplished using a standard disconfirmationquestion, “Overall, how closely did your experiencewith Cafe Au Lait on this occasion match your expec-tations?,” measured on a 100-point “Much Worse ThanExpected” to “Much Better Than Expected” scale. Wealso measured subjects’ perceptions of how interestingand realistic they found the study, as well as their gen-der, age, and how many times they had stopped at acoffee shop in the last 30 days.

4.3. MethodSubjects were shown a scenario describing a coffeeshop, the Cafe Au Lait. They were told that the coffeeshop had recently opened near campus and that theyhad visited it (a) a few times in the low experiencecondition, or (b) every day for the past four weeks inthe high experience condition. They then read a verbaland graphical description of their series of experiences.In the low experience condition, subjects examined agraph of three experiences. In the high experience con-dition, subjects examined four graphs with one graphon each page. The first graph showed the first week,the second graph showed the first two weeks, and soon, so that by in the fourth graph the subjects couldsee the entire series of 20 past experiences. We werecareful to construct the graphs so that the mean (83)and the standard deviation (5.5) were identical be-tween the low and high experience conditions.

Subjects were run in groups so that we could controlthe amount of time (40 seconds) that each page was

7All of the manipulation checks suggest that the experience manip-ulation was successful and that service variability perceptions wereequivalent across groups. These are available from the authors.



Figure 1 Cross-sectional Experiment Shifts in Perceived Quality byCondition

Table 3 Cross-Sectional Experiment—Means Across Conditions(Standard Deviation in Parentheses)

ExperimentalCondition

InitialQualityPercep-

tions

UpdatedQualityPercep-

tions

Shift inQualityPercep-

tions

SubjectiveDiscon-firmation

BehavioralIntentions

High Experience/Zero Disconfirmation

81.2(6.9)

79.4(8.5)

11.8 57.9(10.5)

174.5(22.5)

High Experience/Negative Disconfirmation

80.2(7.2)

68.3(10.6)

111.9 36.3(12.2)

145.1(29.3)

Low Experience/Zero Disconfirmation

79.0(7.6)

78.2(7.3)

10.9 57.1(11.5)

154.6(24.0)

Low Experience/Negative Disconfirmation

78.7(7.7)

62.3(13.0)

116.4 40.7(11.5)

119.8(31.6)

viewed. For the experience manipulation to be suc-cessful, it is critical that the high experience subjectsinternalize the past experiences. Thus, we ensured thatthe high experience subjects examined each graph fora specified time to prevent them from speedingthrough the survey. After subjects examined thegraphs, they completed a series of measures describedin the next section. Following the first set of measures,subjects were exposed to the disconfirmation condi-tion. As before, they were timed so that each subjectexamined this stimulus for the same amount of time(25 seconds). They were then instructed to completethe remainder of the survey at their own pace, werethanked, and released.

4.4. Results

4.4.1. Updated Quality Perceptions. To analyzeour data we use analysis of covariance with updatedquality perceptions as the dependent variable and ini-tial quality perceptions and subjective disconfirmationas covariates (Cronbach and Furby 1970, Lord 1958).8

The fit for the model is relatively good, with an R2 of0.51 (see Table 3 for means). As predicted, the level ofexperience exerts a main effect on updated quality per-ceptions (t214 4 2.21, p , 0.05). Specifically, quality

8Importantly, the groups were equivalent in terms of both initialquality perceptions and subjective disconfirmation.

perceptions for high experience subjects updated lessthan those of low experience subjects (16.5 versus18.7 for high and low experience subjects, respec-tively. Importantly, effects of prior perceptions of qual-ity are controlled for in the analysis via the covariate(t214 4 8.64, p , 0.01).

Not surprisingly, the effect of disconfirmation onupdated quality perceptions is significant (t214 4 6.66,p , 0.01). Subjects who experienced no disconfirma-tion updated their quality perceptions less (11.8) thansubjects who experienced negative disconfirmation(114.2). Further, subjective disconfirmation exerted asignificant effect on updated quality perceptions (t214

4 2.81, p , 0.01), replicating recent findings in theservice quality literature (e.g., Anderson and Sullivan1993, Boulding et al. 1993, Inman et al. 1997).

The interaction between experience and disconfir-mation is significant (t214 4 2.57, p , 0.01). The shiftin quality perceptions across the four conditions isshown in Figure 1. Both high and low experience sub-jects updated their quality perceptions minimallywhen the outcome was equal to the mean of past ex-periences (11.8 for the high experience group and10.8 for the low experience group). However, the lowexperience group lowered their quality perceptionssignificantly more than the high experience group(116.4 versus 111.9 for the low and high experiencegroup, respectively) when the outcome was worsethan expected. Thus, consistent with prediction, ex-perience appears to “inoculate” consumers to some ex-tent against a single substandard outcome.



4.4.2 Behavioral Intentions. In examining the ef-fects of experience level and disconfirmation on be-havioral intentions, it is unclear whether or not theirimpact is mediated by quality perceptions. To test forpotential mediation, we use the method outlined byBaron and Kenny (1986). Specifically, we examine theeffect of experience (as a proxy for uncertainty) anddisconfirmation on the proposed mediator, updatedquality perceptions, and the dependent variable, be-havioral intentions, both with and without incorporat-ing the effect of the mediator. Perfect mediation isdemonstrated if the independent variable exerts a sig-nificant effect on the mediator as well as the dependentvariable but this effect becomes nonsignificant whenthe mediating variable is incorporated as a covariate.If the effect remains significant but the effect size sig-nificantly reduces, partial mediation is demonstrated.

We already presented evidence of the significant ef-fects of both experience and disconfirmation on up-dated quality perceptions. In the second ANOVA, us-ing behavioral intentions as the dependent variable,both experience (t216 4 6.22, p , 0.01) and disconfir-mation (t216 4 8.82, p , 0.01) demonstrate significantmain effects on behavioral intentions. Specifically, highexperience subjects state a much greater intention tovisit the service again and to tell others than do lowexperience subjects (160.9 vs. 137.2 for the high andlow experience groups, respectively). However, whenupdated quality perceptions are added to the model,the effect of experience is undiminished (t215 4 5.74, p, 0.01), while that of disconfirmation is greatly re-duced (t215 4 2.74, p , 0.01). Thus, these results sug-gest that cumulative experience exerts a direct influence onbehavioral intentions while the effect of disconfirmation islargely mediated by updated quality perceptions.

5. Discussion and Future Research5.1. Management ImplicationsBoth our analytical model and our empirical resultsshoot holes in some seemingly reasonable quality max-ims. Let us consider in particular the truisms from theIntroduction:

• “It is necessary to exceed customer expectations.”Both our analytical model and the longitudinal exper-iment contradict this (although the cross-sectional ex-periment, which is perhaps less sensitive to this effect,

does not). The longitudinal experiment showed signifi-cant positive preference shifts if customer expectationswere exactly met. The reason, based on the analyticalmodel, is that experience causes a shrinkage in the var-iance of the predictive distribution for the next trans-action. That is, experience with a brand leads to de-creased risk, and decreased risk leads to greaterpreference. So in fact, meeting expectations should un-ambiguously result in higher preference. The analyti-cal model also suggests that even not quite meeting ex-pectations might still increase preference, although thiseffect was not significant in the longitudinal experi-ment. Of course, exceeding customer expectations willstill be required if the company seeks to induce cus-tomer delight (Oliver et al. 1997), but lower levels ofperformance may also produce positive results.

• “If a customer expects a bad level of quality and receivesit, he/she will reduce his/her level of preference for thebrand.”The analytical model and both experiments contradictthis. In the cross-sectional experiment, subjects did notlower their quality perceptions of the service whentheir expectations were met. The longitudinal experi-ment showed significant positive preference shifts evenfor the nonpreferred option, indicating that even whenexpectations were low, meeting expectations raisedpreference. The analytical model explains why thispreference shift can occur. Again experience shrinksthe variance of the predictive distribution and reducesrisk, thereby increasing preference.

• “Given two equally-priced options, the customer willchoose the one with the higher expected quality.”Although this statement seems obviously true, it alsois contradicted by both the analytical model and thelongitudinal experiment. The reason, again, is risk. Ahigher expected quality can be outweighed by greaterperceived variability. Based on the logit analysis of thedata from the longitudinal experiment, perceived var-iance has a significant, negative impact on choice. Thisresulted in nearly half of our subjects choosing an op-tion with a lower expected quality.

• “Management should always focus on its most loyalcustomers.”Given that the most loyal customers are the most ex-perienced (which is consistent with the most typicalbehavioral definitions of loyalty), our research casts



doubt on this seemingly self-evident maxim. The cross-sectional experiment shows that disconfirmation hasthe biggest preference impact on less-experienced cus-tomers, and this phenomenon is supported theoreti-cally by the analytical model, for any nonnegative dis-confirmation, and for any sufficiently large negativedisconfirmation.9 This would seem to imply that man-agement should pay more attention to its newer (pre-sumably less loyal) customers, because those are thecustomers for which quality differences will have thegreatest impact. In other words, less experienced(loyal) customers are easier to lose, while more expe-rienced (loyal) customers are harder to lose, all otherthings being equal.

Several other managerial implications arise fromthis work. First, customer satisfaction and quality mea-surement surveys would benefit from including an ex-perience variable. This is because the degree to whichpreferences shift is dependent upon the degree of ex-perience, with more experienced customers beingmore difficult to shift. All other things being equal, formaximum shift of preference, less experienced custom-ers should be targeted. Further, in addition to measur-ing perceived quality, perceived variability and/orconsistency in quality is important to capture as well.Both factors influence overall quality perceptions andbehavioral intentions.

Our results suggest that it is insufficient for a goodor service to be perceived as better (i.e., a higher ex-pected quality) than its competitor. For example, newBrand A, even though it has a higher expected valuethan Brand B, may not be preferred to Brand B becauseof greater perceived variability resulting in its per-ceived value being worse. The implication to manage-ment launching a new product is that actual trial maybe more powerful than supplying information throughsuch methods as advertising. This suggests that cou-pons, promotions, and other ways to induce trial maybe necessary in the early stages of a product launch,even if a positive brand image has been already createdthrough advertising.

9Interestingly, there is an interval for which a negative disconfir-mation produces a smaller shift in preference for the less experiencedcustomer—whenever that customer’s positive preference shift fromvariance reduction is roughly equivalent to the negative preferenceshift from mean reduction.

Finally, worse-than-expected quality hurts morethan better-than-expected quality helps. This replicatesprevious findings by several researchers (e.g.Anderson and Sullivan 1993, DeSarbo et al. 1994, Rustet al. 1995). The managerial takeaway is that problemsshould be addressed first, and then positive opportu-nities. In other words, process improvement initiativesshould first focus on eliminating unsatisfactory serviceencounters (i.e., providing consistent quality) beforetrying to delight customers.

5.2. Contributions to the Quality LiteratureOur results contribute to the quality literature in fourrespects. First, they suggest that consumers are sensi-tive to not only the average performance of a product,but also to its variability around this mean. In otherwords, it is not sufficient for a brand to strive to in-crease its overall quality—it must also strive to reducethe risk of an outcome deviating from this performanceexpectation. Second, consumers are more likely to re-choose a brand (i.e., retention is increased) if the brandperforms as expected. Contrary to the traditional sat-isfaction model, meeting expectations can sometimesbe interpreted by consumers as a favorable outcome,as the experience provides more information about thebrand and reduces the perceived variance of thebrand’s future performance. Third, probability ofchoice is not necessarily adversely affected by an out-come that is slightly below expected. This is apparentlybecause the customer becomes less concerned withdownside risk if the disconfirmation is relatively mi-nor. Finally, if negative disconfirmation is relativelysubstantial, quality perceptions and behavioral inten-tions are negatively affected, but this effect is moder-ated by the amount of prior experience.

Figure 2 shows a graph of the mean shifts in choiceprobability across the five disconfirmation conditionsin the longitudinal experiment. One notes that con-sumers’ reaction to disconfirmation is nonlinear andthey appear more sensitive to negative disconfirmationthan to positive disconfirmation. While this curve issimilar to the effects one typically expects underKahneman and Tversky’s prospect theory (1979), ourtheoretical perspective and empirical results suggestthat the curve may be “choice-neutral”10 at a value less

10We call the curve “choice-neutral” at the point where observedquality results (on average) in unchanged probability of choice.



Figure 2 Shift in Choice Probability in Response to Disconfirmation

than zero disconfirmation. Specifically, the curve ischoice-neutral at a negative disconfirmation of abouttwo hours below expected, which is equivalent to ap-proximately one half of a standard deviation from themean. Importantly, the direction of this deviation isconsistent with our prediction (Proposition 4) that asmall negative disconfirmation can produce an in-crease in preference.

Our results suggest that the most important time fora brand to establish its quality perceptions in the mindsof consumers is at the time that consumers have littleprior experience with the category. In Bayesian terms,this is at the point when consumers hold a more diffuseprior expectation regarding the brand’s performance.Ironically, the time where many consumers are inex-perienced with the product is precisely the point in theproduct life cycle where the manufacturer is oftenstruggling with maintaining consistent product qual-ity. This may explain the Golder and Tellis (1993) find-ings regarding the failure of many pioneers. If the pio-neer struggles with quality, a subsequent (moreconsistent quality) entrant can gain a perceived qualityadvantage with consumers. Alternatively, once thepioneer has established a relatively high prior expec-tation of performance and a low perceived variability(through knowledge gained by consumption experi-ence), this imposes a barrier to later entrants into themarket since the pioneer finds customers are much eas-ier to retain. A subsequent entrant must have a signifi-cantly higher level of expected quality to counteract itshigher perceived variability and risk.

5.3. Future Research and LimitationsThere are many interesting ways in which this researchmight be extended. For example, the model might be

made richer (albeit more complicated) by the inclusionof effects not currently accounted for, such as allowingfor variables that affect perceived quality, or otherwisechange (bias) the perception and updating process. In-teresting work in this regard is already underway(Boulding et al. 1998). The existence of threshold effectsis another promising area for research. In other words,updating may only take place when absolute discon-firmation exceeds a certain threshold. Otherwise theobservation is simply viewed as an expected outcome.It is also possible that those thresholds may vary acrosscustomers. That might explain the fact that many ofour subjects did not update their preferences.

While our results suggest that, on average, consum-ers update their expectations in a largely Bayesianmanner, some consumers are probably more Bayesianthan others. In the first experiment, choice probabilityfor most subjects increased in the 10 hour disconfir-mation condition and decreased in the 110 hour dis-confirmation condition. However, in the zero, negativeone hour, and negative three hour disconfirmationconditions, the modal subjective probability shift waszero. Approximately half the subjects in both the zeroand negative one hour conditions reported that theirchoice probability would be unaffected, while a thirdof the subjects in the negative three hour conditiongave this response. Even allowing for measurement er-ror, some “updating heterogeneity” appears evidentamong consumers.

Additional research is warranted to explore the vari-ables that account for the heterogeneity in consumers’updating processes. For instance, the Boulding et al.(1993) notion of “should” expectations could be ex-tended to expectations of variability in quality. Further,individual differences in need for cognition (Cacioppoand Petty 1982) or tendency to engage in post-purchaseregret (e.g., Inman and Zeelenberg 1998) might be im-portant moderators of perception updating andchanges in probability of choice in response to discon-firmation. On the other hand, it is important to explorethe contexts in which consumers tend to form expec-tations of quality that are resistant to change.

Limitations of our study deserve mention. First, asin all research, one must be careful in overgeneralizingresults. However, we tested our updating model intwo quite different experimental designs and found



consistent support for our thesis. Second, we tested ourpropositions in lab contexts and our results are basedon self-reports. It is important to replicate this researchin other contexts to provide greater confidence in ourresults. For example, our subjects observed several it-erations of outcomes over a short period of time. In afield context, consumers’ priors might become morediffuse (due to forgetting) as the interpurchase timeincreases or become less diffuse as additional infor-mation is obtained from advertising or word of mouth.Finally we should note the assumption of normal dis-tributions. If customers’ priors and likelihood distri-butions are not normally distributed, then a more com-plicated Bayesian formulation would result. Thisresearch provides the basis for subsequent work in theimportant area of how consumers dynamically updatetheir quality perceptions and their preferences.11

ReferencesAnderson, Eugene W., Mary W. Sullivan. 1993. The antecedents and

consequences of customer satisfaction for firms. Marketing Sci.12 (Spring) 125–143.

Baron, Reuben M., David A. Kenny. 1986. The moderator-mediatorvariable distinction in social psychological research: Concep-tual, strategic, and statistical considerations. J. Personality andSoc. Psych. 51 (Dec.) 1173–1182.

Bell, David. 1982. Regret in decision making under uncertainty. Oper.Res. 30 961–981.

——. 1985. Disappointment in decision making under uncertainty.Oper. Res. 35 1–27.

Berger, James O. 1985. Statistical Decision Theory and Bayesian Anal-ysis. Second edition. Springer-Verlag, New York.

Bolton, Ruth, James Drew. 1991a. A longitudinal analysis of the im-pact of service changes on customer attitudes. J. Marketing 55(January) 1–9.

——, James Drew. 1991b. A multistage model of customers’ assess-ment of service quality and value. J. Consumer Res. 17 (March)375–384.

——. 1998. A dynamic model of the duration of the customer’s re-lationship with a continuous service provider: the role of sat-isfaction. Marketing Sci. 17 (1) 45–65.

Boulding, William, Ajay Kalra, Richard Staelin, Valarie A. Zeitham.1993. A dynamic process model of service quality. J. MarketingRes. 30 (February) 7–27.

——, Ajay Kalra, Richard Staelin. 1998. The quality double

11The authors thank Vanderbilt’s center for Service Marketing forproviding partial support for this project, and gratefully acknowl-edge the comments of Priya Raghubir, Steve Hoch, Rick Staelin, thearea editor and reviewers, and seminar participants at the Universityof Chicago.

whammy: The rich stay rich and the poor stay poor: Workingpaper. Duke University, Durham, NC.

Burr, Irving W. 1976. Statistical Quality Control Methods, Marcel Dek-ker, New York.

Cacioppo, John T., Richard E. Petty. 1982. The need for cognition. J.Personality and Soc. Psych. 42 (January) 116–131.

Cronbach, Lee J., Lita Furby. 1970. How we should measure‘change’—or should we? Psych. Bull. 74 (July) 68–80.

Cronin, J. Joseph, Steven A. Taylor. 1992. Measuring service quality:a reexamination and extension. J. Marketing 56 (July) 55–68.

——, Steven A. Taylor. 1994. SERVPERF versus SERVQUAL: rec-onciling performance-based and perceptions-minus-expectations measurement of service quality. J Marketing 58(January) 125–131.

Deming, W. Edwards. 1986. Out of the Crisis, MIT Center for Ad-vanced Engineering Study, Cambridge, MA.

DeSarbo, Wayne, Lenard Huff, Marcello Rolandelli, Jungwhan Choi.1994. On measurement of perceived service quality: A conjointmeasurement approach. Roland T. Rust and Richard L. OliverEds. Service Quality: New Directions in Theory and Practice. SagePublications, Thousand Oaks, CA, 199–220.

Doherty, Michael E., Clifford R. Mynatt, Ryan D. Tweney, MichaelD. Schiavo. 1979. Pseudodiagnosticity. Acta Psych. 43 111–121.

Edwards, Ward. 1968. Conservatism in human information process-ing. B. Kleinmuntz, ed. Formal Representation of Human Judg-ment. Wiley, New York.

Fischhoff, Baruch, Ruth Beyth-Marom. 1983. Hypothesis evaluationfrom a Bayesian perspective. Psych. Rev. 90 239–260.

Fishburn, Peter C. 1988. Nonlinear Preference and Utility Theory. JohnHopkins University Press, Baltimore, MD.

Fornell, Claes. 1992. A national customer satisfaction barometer: TheSwedish experience. J. Marketing 56 (January) 6–21.

Gale, Bradley. 1997. Satisfaction is not enough. Marketing News (Oc-tober 27) 18.

Gigerenzer, G., W. Hell, H. Blank. 1988. Presentation and content:the use of base-rates as continuous variable. J. Experiment.Psych. Human Perception and Performance 14 513–525.

Ginossar, Zvi, Yaacov Trope. 1987. Problem solving in judgment un-der uncertainty. J. Personality and Soc. Psych. 52 464–474.

Golder, Peter N., Gerald J. Tellis. 1993. Pioneering advantage: Mar-keting logic or marketing legend? J. Marketing Res. 30 (May)158–170.

Inman, J. Jeffrey, James S. Dyer, Jianmin Jia. 1997. A generalizedutility model of disappointment and regret effects on post-choice valuation. Marketing Sci. 16 (2) 97–111.

——, Marcel Zeelenberg. 1998. Wow, I could’ve had a V8!: The roleof regret in consumer choice. Working paper, University ofWisconsin, Madison, WI.

Jia, Jianmin, James S. Dyer. 1996. A standard measure of risk andrisk-value models. Management Sci. 42 1691–1705.

Juran, J. M., Frank M. Gryna, Jr. 1980. Quality Planning and Analysis:From Product Development through Use, McGraw-Hill, NewYork.

Kahneman, Daniel, Amos Tversky. 1972. Subjective probability: Ajudgment of representativeness. Cogn. Psych. 47 263–291.



——, ——. 1973. On the psychology of prediction. Psych. Rev. 80 237–251.

——, ——. 1979. Prospect theory: An analysis of decision under risk.Econometrica 47 263–290.

Keeney, Ralph L., Howard Raiffa. 1976. Decision with Multiple Objec-tives: Preferences and Value Tradeoffs, John Wiley & Sons, NewYork.

Kopalle, Praveen K., Donald R. Lehmann. 1995. The effects of ad-vertised and observed quality on expectations about new prod-uct quality. J. Marketing Res. 32 (August) 280–290.

Little, John D. C. 1966. A model of adaptive control of promotionalspending. Oper. Res. 14 (November–December) 1075–1098.

Lord, Frederic M. 1958. The utilization of unreliable differencescores. J. Ed. Psych. 49 (3) 150–152.

Luce, R. Duncan. 1959. Individual Choice Behavior: A Theoretical Anal-ysis, John Wiley & Sons, New York.

Machina, Mark J. 1987. Choice under uncertainty: problems solvedand unsolved. Econom. Perspectives 1(1) 121–154.

Manrai, Ajay K. 1995. Mathematical models of brand choice behav-ior. European J. Operational Res. 82 1–17.

Markowitz, H. M. 1959. Portfolio Selection, John Wiley & Sons, NewYork.

——. 1987. Mean-Variance Analysis in Portfolio Choice and Capital Mar-kets. Basil Blackwell, New York.

McFadden, Daniel. 1974. Conditional logit analysis and subjectiveprobability. Paul Zarembka, Ed. Frontiers in Econometrics. Aca-demic Press, New York, 105–12

——. 1976. Quantal choice analysis: A survey. Ann. Econom. Soc. Mea-surement 5 363–390.

Meyer, Jack. 1987. Two-moment decision models and expected util-ity maximization. Amer. Econom. Rev. 77 421–430.

Oliver, Richard L. 1980. A cognitive model of the antecedents andconsequences of satisfaction decisions. J. Marketing Res. 42 (No-vember) 460–469.

——, Wayne DeSarbo. 1988. Response determinants in satisfactionjudgments. J. Consumer Res. 14 (March) 495–507.

——. 1997. Satisfaction, McGraw-Hill, New York.——, Roland T. Rust, Sajeev Varki. 1997. Customer delight: foun-

dations, findings, and managerial insight. J. Retailing 73 (Fall)311–336.

Parasuraman, A., Valarie A. Zeithaml, Leonard Berry. 1985. A con-ceptual model of service quality and its implications for furtherresearch. J. Marketing 48 (Fall) 41–50.

Parasuraman, A., Valarie A. Zeithaml, Leonard Berry. 1988. SE-RVQUAL: A multiple-item scale for measuring consumer per-ceptions of service quality. J. Retailing 64 (1) 12–40.

Payne, John W. 1973. Alternative approaches to decision making un-der risk: moments versus risk dimensions. Psych. Bull. 80 (De-cember) 435–453.

Reichheld, Frederick F. 1996. Learning from customer defections.Harvard Bus. Rev. (March–April) 56–69.

Rust, Roland T., Anthony J. Zahorik 1993. Customer satisfaction, cus-tomer retention and market share. J. Retailing 69 (Summer) 193–215.

——, Anthony J. Zahorik, Timothy L. Keiningham. 1995. Return onquality (ROQ): making service quality financially accountable.J. Marketing 59 (April) 58–70.

Slovic, Paul, Sarah Lichtenstein. 1971. Comparison of Bayesian andregression approaches to the study of information processingin judgment. Organ. Behavior and Human Performance 6 649–744.

Teas, R. Kenneth. 1993. Expectations, performance evaluation, andconsumer perceptions of quality. J. Marketing 57 (October) 18–34.

Thurstone, L. L. 1927. A law of comparative judgment. Psych. Rev.34 273–286.

Tversky, Amos, Daniel Kahneman. 1974. Judgment under uncer-tainty: heuristics and biases. Science 185 1124–1131.

von Neumann, John, Oskar Morgenstern. 1944. Theory of Games andEconomic Behavior, Princeton University Press, Princeton, NJ.

Wason, P. C. 1960. On the failure to eliminate hypotheses in a con-ceptual task. Quart. J. Experiment. Psych. 12 129–140.

Zellner, Arnold. 1971. An Introduction to Bayesian Inference in Econo-metrics, John Wiley & Sons, New York.

This paper was received January 2, 1996, and has been with the authors 26 months for 2 revisions; processed by William Boulding.

what you don’t know about customer- perceived...

Documents