Jsr303012.qxd

Service Recovery Paradox: A Meta-Analysis
Celso Augusto de Matos, Jorge Luiz Henrique and Carlos Alberto Vargas Rossi The online version of this article can be found at: http://jsr.sagepub.com/cgi/content/abstract/10/1/60 can be found at:
Journal of Service Research
Additional services and information for
Citations
Service Recovery Paradox:
A Meta-Analysis
Celso Augusto de Matos
Jorge Luiz Henrique
Carlos Alberto Vargas Rossi
School of Management, Federal University of Rio Grande do Sul (PPGA-EA-UFRGS)
The Service Recovery Paradox (SRP) has emerged as an The Service Recovery Paradox (SRP) is a peculiar important effect in the marketing literature. However, effect in the services marketing literature and has been empirical research testing the SRP has produced mixed conceptually defined as a situation in which a customer’s results, with only some studies supporting this paradox. postfailure satisfaction exceeds prefailure satisfaction Because of these inconsistencies, a meta-analysis was (McCollough and Bharadwaj 1992). When reviewing this conducted to integrate the studies dealing with the SRP literature, the theoretical paper by Hart, Heskett, and and to test whether studies’ characteristics influence the Sasser (1990) is one of those frequently cited, especially results. The analyses show that the cumulative mean the statement that “a good recovery can turn angry, frus- effect of the SRP is significant and positive on satisfac- trated customers into loyal ones. It can, in fact, create more tion, supporting the SRP, but nonsignificant on repur- goodwill than if things had gone smoothly in the first chase intentions, word-of-mouth, and corporate image, place” (p. 148). In this way, recovery encounters would suggesting that there is no effect of the SRP on these vari- mean an opportunity for service providers to increase ables. Additional analyses of moderator variables find customer retention (Hart, Heskett, and Sasser 1990).
that design (cross-sectional versus longitudinal), subject The topic of SRP has been of great importance for (student versus nonstudent), and service category (hotel, managers and researchers. Given that failure is one of the restaurant, and others) influence the effect of SRP on sat- main reasons that drive customer-switching behavior, isfaction. Finally, implications for managers and direc- understanding recovery is relevant because a successful tions for future research are presented. recovery may lead to customer retention, which willaffect company profitability (McCollough, Berry, and Keywords:
service recovery; Service Recovery Paradox; Yadav 2000). On the other hand, there has always been a meta-analysis; moderation analysis question in the service literature as to whether highrecovery efforts can really create greater satisfaction The authors are thankful for the support provided by the Brazilian Funding Council for Research (CNPq and CAPES) and the Graduate School of Management. The authors also would like to thank various authors who sent their recent articles on the topic ofservice recovery, including Chihyung Ok, David A. Cranage, Stefan Michel, Steven H. Seggie, and Vincent P. Magnini. The authorsare also grateful to Professor Frank L. Schmidt and Professor David B. Wilson for their support in answering questions about meth-ods of meta-analysis and to the editor and three anonymous reviewers of JSR for their insightful comments.
Journal of Service Research, Volume 10, No. 1, August 2007 60-77DOI: 10.1177/1094670507303012 2007 Sage Publications when compared to the situation of no failure (Etzel and using subgroup meta-analysis and hierarchical moderator analysis (Hunter and Schmidt 2004) to provide insights However, empirical studies investigating the SRP have regarding studies’ characteristics that might moderate the produced results that vary considerably in terms of statis- effects of the SRP. Finally, we present a discussion with tical significance, direction, and magnitude. Although theoretical and managerial implications, limitations, and some studies provide support for the SRP (Hocutt, Bowers, and Donavan 2006; Hocutt and Stone 1998;Magnini et al. 2007; Maxham and Netemeyer 2002;McCollough 2000; Michel 2001; Michel and Meuter THE SERVICE RECOVERY PARADOX
2006; Smith and Bolton 1998), others have found no sup-port (Andreassen 2001; Halstead and Page 1992; Hocutt, The SRP is defined as the situation in which postre- Chakraborty, and Mowen 1997; Maxham 2001; covery satisfaction is greater than that prior to the service McCollough, Berry, and Yadav 2000; Ok, Back, and failure when customers receive high recovery perfor- Shanklin 2006; Zeithaml, Berry, and Parasuraman 1996).
mance (Maxham 2001; McCollough 1995; McCollough These conflicting results might be a consequence of a and Bharadwaj 1992; Smith and Bolton 1998). In this number of factors, from different methodological aspects context, effective service recovery may lead to higher sat- in the studies to certain conditions moderating the para- isfaction compared to the service that was correctly per- dox. In this respect, some variables have been proposed formed the first time, and recovery encounters would in the service recovery literature as potential moderators mean an opportunity for service providers to increase of the paradox, including severity of the failure, prior customer retention (Hart, Heskett, and Sasser 1990).
failure with the firm, stability of the cause of the failure, Based on the disconfirmation framework (McCollough, and perceived control (Magnini et al. 2007).
Berry, and Yadav 2000; Oliver 1997), the paradox is related Along with these conflicting results, “there is a con- to a secondary satisfaction following a service failure in siderable body of conjecture and intuition pertaining to which customers compare their expectations for recovery to the existence of the service recovery paradox” their perceptions of the service recovery performance. If (Andreassen 2001, p. 40). Thus, this inconsistency with there is a positive disconfirmation, that is, if perceptions of regard to the effect of the SRP suggests the need for a service recovery performance are greater than expectations, meta-analysis to provide both a systematic review and a a paradox might emerge (secondary satisfaction becomes quantitative integration of all the available SRP research.
greater than prefailure satisfaction). Otherwise, in the case A meta-analysis can provide insights into these inconsis- of a negative disconfirmation, there is a double negative tencies by accumulating effects across studies after effect, as service failure is followed by a flawed recovery adjusting for the studies’ main artifacts (i.e., measure- (Bitner, Booms, and Tetreault 1990; McCollough, Berry, ment and sampling error), identifying measurement and and Yadav 2000; Smith and Bolton 1998).
sample characteristics that affect the support/nonsupport The paradox can also be justified by the script theory of the SRP, and also testing the generalizability of the and the commitment–trust theory for relationship mar- results (Farley, Lehmann, and Sawyer 1995).
keting (Magnini et al. 2007). Script theory proposes that Through meta-analysis we aim to (a) reflect on the there is a common sequence of acts in a service delivery, different methodological approaches used to test the SRP, in such a way that employees and customers share simi- (b) map the dependent variables that have been consid- lar beliefs regarding the expected order of events and ered when the SRP is tested in the literature, (c) reveal their respective roles in the process (Bitner, Booms, and which of these dependent variables support the SRP, (d) Mohr 1994). If a service failure occurs, it works as a investigate which methodological differences across the deviation from the predicted script and produces an studies moderate the results for the effects of the SRP, increased sensitivity in the customer regarding the failure and (e) identify research questions worthy of future and the redress process. Because of this, satisfaction with empirical investigations regarding the SRP.
the recovery process becomes more relevant than satis- First, we present a theoretical background about the faction with the initial attributes in influencing the final SRP to guide the meta-analysis. Second, we discuss the cumulative satisfaction (Bitner, Booms, and Tetreault procedures for building the database, computing, and integrating the effect sizes. Third, we present a quantita- Because an excellent service recovery has a direct tive summary that includes the adjusted cumulative mean impact on how much consumers trust the firm (Kau and values of the effect of service recovery on dependent Loh 2006; Tax, Brown, and Chandrashekaran 1998), variables and test whether the paradox is supported or there is also a foundation for the Service Recovery not. Fourth, we conduct a more detailed examination, Paradox in Morgan and Hunt’s (1994) commitment-trust JOURNAL OF SERVICE RESEARCH / August 2007 Hypothesis 1c: There is a significant positive SRP Meta-Analytic Framework of the SRP
Hypothesis 1d: There is a significant positive SRP - Method (survey × experiment)- Design (cross-sectional × Conflicting results have been found in the literature, with some studies supporting the SRP (Hocutt, Bowers, and Donavan 2006; Hocutt and Stone 1998; Magnini et al. 2007; Maxham and Netemeyer 2002; McCollough 2000; Michel 2001; Michel and Meuter 2006; Smith and Bolton 1998) and others not supporting this effect (Andreassen 2001; Halstead and Page 1992; Hocutt, Chakraborty, and Mowen 1997; Mattila 1999; Maxham 2001; McCollough, Berry, and Yadav 2000; Ok, Back, - Prior failure experience- Stability attributions and Shanklin 2006; Zeithaml, Berry, and Parasuraman 1996). Also, there is support for the notion that the SRPis more likely when service failure causes low harm, indi- NOTE: dashed line (- - -) indicates path not tested in the meta-analysis.
cating that recovery strategies may be more effectivewhen the failure is perceived by the customers as less theory for relationship marketing (Magnini et al. 2007).
severe (Magnini et al. 2007; Mattila 1999; Smith and In this view, both service recovery efforts and relation- Bolton 1998). These contingencies are discussed later in ship marketing focus on customer satisfaction, trust, and commitment. Trust exists when one party has confidence Another possible explanation for the mixed findings in another’s reliability and integrity (Moorman, Zaltman, might be related to the nature of the paradox (Michel and and Deshpande 1992; Morgan and Hunt 1994). As fail- Meuter 2006). In this view, it is considered that the SRP is ures contribute to create insecurity in the customers and a very rare event (Boshoff 1997), that only a minority of affect trust in the firm, an effective service recovery can dissatisfied customers complains (Singh 1990), and that be an opportunity to make customers feel that the firm is only few recoveries lead to customer satisfaction (Kelley, able and willing to correct the problem. As a result, a fair Hoffman, and Davis 1993). As a result, it becomes very conflict resolution may have a positive impact on con- difficult to achieve a large sample of customers who have received a very satisfactory recovery,2 and this requirement Although most studies test the Service Recovery may have an influence on nonsignificant results presented Paradox for satisfaction and repurchase intentions,1 there in the literature (Michel and Meuter 2006).
are also studies considering the paradox for word-of-mouth(Hocutt, Bowers, and Donavan 2006; Kau and Loh 2006;Maxham 2001; Maxham and Netemeyer 2002; Ok, Back, Moderators
and Shanklin 2006), corporate image (Andreassen 2001;Kwortnik 2006), trust (Kau and Loh 2006), quality The mixed findings can also be caused by certain con- (McCollough 1995), complaint intentions (Hocutt, ditions moderating the paradox. For example, although the Chakraborty, and Mowen 1997), switching intentions, pay- theoretical definition of the SRP seems to be convergent in more intentions, and external response (Zeithaml, Berry, the literature, the same is not true for the operationaliza- and Parasuraman 1996). Figure 1 provides an overarching tions of the concept. Although some authors use a conceptual framework for our meta-analysis and also syn- between-subjects approach, comparing a recovery group thesizes key insights from previous studies and discussions with a no-failure control group (Hocutt, Bowers, and about the SRP in the extant literature. If the SRP exists, a Donavan 2006; Kau and Loh 2006; McCollough 1995; service failure that is followed by a high recovery effort McCollough, Berry, and Yadav 2000; Michel and Meuter should produce outcome variables that are higher when 2006; Ok, Back, and Shanklin 2006), others use a within- compared to a situation in which no failure occurred. Based subjects approach, comparing different measures from the on this framework, we expect the following: same subject before and after a failure and/or recovery(Magnini et al. 2007; Maxham 2001; Maxham and Hypothesis 1a: There is a significant positive SRP Netemeyer 2002; Smith and Bolton 1998). These differ- ences also extend to the type of research design used in Hypothesis 1b: There is a significant positive SRP terms of experiment or survey approach, cross-sectional or longitudinal measures, student or nonstudent subjects, single or multiple items measuring dependent variables, Hoffman, and Davis 1993). For instance, a recovery scenario- or non-scenario-based research, and the different action (e.g., apology or compensation) might increase manipulated factors (in case of experiments).
customer satisfaction after a delay in waiting in a line.
Assmus, Farley, and Lehmann (1984) suggested that But what if this delay has caused a serious consequence four categories of characteristics might help identify sys- for the customer (e.g., he missed his flight after a delay at tematic patterns in a meta-analysis: research context, the hotel desk)? It is unlikely that a recovery action model specification, measurement methods, and estima- would be able to either bring the customer to the original tion procedure. However, because our unit of analysis level of satisfaction or, even more improbable, increase was bivariate correlations, we seek systematic differ- ences in the study characteristics. This procedure iscommon in meta-analysis using correlations (e.g., Pan Hypothesis 3a: The SRP is more (less) likely to and Zinkhan 2006). In our investigation, we examine four occur when the customer perceives the failure as potential moderators: method (survey versus experi- ment), design (cross-sectional versus longitudinal),subject (student versus nonstudent), and service category Given that customers usually have a history of interac- (hotel versus restaurant versus others).
tions with the firm, their cumulative satisfaction, as Based on the methodological differences across the opposed to a transaction-specific satisfaction, is based on studies, we propose the following (see Figure 1): their evaluations of multiple experiences with the firm overtime (Bolton and Drew 1991). In this way, satisfactory Hypothesis 2a: The SRP effects differ in studies recoveries may yield paradoxical gains only in the short using survey methods versus those using exper- run, and customers will likely infer that multiple failures are because of problems inherent to the firm (Maxham and Hypothesis 2b: The SRP effects differ in studies Netemeyer 2002). Hence, when a customer experiences a using cross-sectional designs versus those using second failure, he or she is more likely to attribute the cause of that problem to the firm than when the customer Hypothesis 2c: The SRP effects differ in studies experienced failure for the first time (Magnini et al. 2007; using student subjects versus those using non- Maxham and Netemeyer 2002). Thus, we propose: Hypothesis 2d: The SRP effects differ across stud- Hypothesis 3b: The SRP is more likely to occur ies using different service categories.
when the customer experiences the failure forthe first time when compared to the situation in Regarding the boundary conditions for the SRP effects, which there has already been a previous service some theoretical variables have also been proposed as potential moderators in the service recovery literature,including severity of the failure, prior failure with the firm, Another influencing factor is the stability of the cause of stability of the cause of the failure, and perceived control the failure. Stability attributions refer to customers’ infer- (Magnini et al. 2007). Most of these contingencies, how- ences about whether similar failures are likely to occur in the ever, have been tested only in the more recent literature future, given the customers’ dissatisfaction with a product or (with severity of the failure3 being an exception), which service (Blodgett, Granbois, and Walters 1993; Folkes precluded their empirical assessment as moderators in our 1984). When customers experience a service failure, they meta-analysis. Nevertheless, they are included in Figure 1 ask themselves whether the failure has temporary (i.e., as propositions to be investigated further in future unstable) or permanent (i.e., stable) causes, and if they think research. Their rationale for proposing the various theoret- that the problem has stable causes (i.e., it is likely to occur again), then they will try to avoid this service provider in the Studies support the notion that it is harder to recover future (Folkes 1984). Studies have indicated that consumers from high-magnitude failures (Magnini et al. 2007; who perceive a service failure as more stable present lower Mattila 1999; McCollough 1995; Smith and Bolton repatronage intentions (Folkes 1984, 1988). Smith and 1998) or that the perceived harm caused by the failure Bolton (1998) found similar results. In their findings, if a interacts with the recovery effort to influence customer customer believed that the unavailability of the requested satisfaction (McCollough, Berry, and Yadav 2000). It has food item was because of a consistent omission of the been found that the higher the magnitude or severity of restaurant, he or she would be less satisfied and less the failure, the lower the overall customer satisfaction likely to repatronize this restaurant. Hence, customers are (Mattila 1999; Weun, Beatty, and Jones 2004), just as less more likely to forgive failures with unstable (temporary) favorable recoveries tend to be more memorable (Kelley, causes (Kelley, Hoffman, and Davis 1993; Magnini et al.
JOURNAL OF SERVICE RESEARCH / August 2007 2007) and to express a situation of recovery paradox in All identified studies were then examined in terms of this context (Magnini et al. 2007). Thus, we expect: the following relevant variables: authors, year, journal,service category (hotel, restaurant, and others), method Hypothesis 3c: The SRP is more (less) likely when (survey versus experiment), subjects (students versus customers perceive that the failure is less (more) nonstudents), number of compared groups, number of factors manipulated or measured, dependent variables(satisfaction, repurchase intentions, word-of-mouth, Finally, attributions related to whether the firm had trust, image, quality, intentions to complain, switching much or little control over the occurrence of the failure intentions, pay-more intentions, and external response), also influence the recovery paradox. When customers per- reliabilities for the dependent variables, and effect sizes.
ceive that the firm had little control over the service failure,they are more likely to comprehend and forgive the prob- Effect Size Computation
lem (Maxham and Netemeyer 2002). This is in agreementwith studies showing that the perceived reason for a prod- Our meta-analytic procedure followed common guide- uct’s failure is an important predictor of how consumers lines for meta-analysis of experimental studies (Lipsey and react (Folkes 1984). For instance, complainants who Wilson 2001), in which standardized mean differences believed that firms were responsible for the failure were (Cohen’s d) are computed first and then converted to cor- more likely to expect redress (e.g., apologies, refunds).
relation coefficients (r). This procedure is the same as that Also, customers who attribute failures to controllable fac- employed by other meta-analyses in the marketing litera- tors will probably be more dissatisfied with the failure and ture (e.g., Brown and Stayman 1992; Eisend 2006). We less forgiving in their evaluations. Indeed, it has been selected the correlation coefficient, r, as the effect-size found that an SRP is more (less) likely to occur when the metric because it is easier to interpret and a scale-free mea- customer perceives that the firm had little (much) control sure. Also, the correlation coefficient is the mostly used over the cause of the failure (Magnini et al. 2007). Based effect size in meta-analyses in the marketing literature (e.g., DelVecchio, Henard, and Freling 2006; Eisend 2004,2006; Franke and Park 2006; Janiszewski, Noel, and Hypothesis 3d: The SRP is more (less) likely when Sawyer 2003; Palmatier et al. 2006; Pan and Zinkhan customers perceive that the firm had little 2006). As we included in our data set both surveys and (much) control over the cause of the failure.
experiments, we could integrate them by using r as thecommon effect size. Positive (negative) values of the cor-relation coefficient indicate the presence (absence) of the SRP. This procedure followed recommendations by Lipseyand Wilson (2001, pp. 14, 173) for conducting meta-analysis Search Process and Sampling Frame
with group contrasts (in experiments or surveys) and isbased on the following rationale: Studies were identified by a computerized biblio- graphic search. Databases included Blackwell Synergy, 1. The SRP refers to the effect that an outcome Elsevier Science Direct, Ebsco, Emerald Insight, Infotrac variable (e.g., satisfaction) is greater for a cus- College, Proquest, Scopus, Thompson Gale, Wilson Web, tomer that has experienced a failure and a high and Google Scholar. First, we searched for the terms recovery effort when compared to a customer “service failure” and “service recovery” in keywords and abstracts. Then we narrowed our search by “service 2. The recovery group is considered as the treat- recovery paradox” in abstracts, keywords, or full text.
ment group and the no failure group as the con- Using this procedure, we found a total of 319 articles ranging from 1987 to 2006. By searching Proquest and 3. “The contrast between the experimental and Google, we could also access 14 dissertations on the control group on the values of an outcome vari- research topic, leading to a total of 333 studies. Of this able is interpreted as the effect of treatment”(Lipsey and Wilson 2001, p. 14). Treatment in total, 42 (12.6%) were theoretical papers and the remain- our case is the high recovery effort.
ing 291 investigated service failure and/or recovery empirically. Among these studies, 21 were identified as testing the SRP empirically and were chosen for the analy- 5. Then, if satisfaction has a significantly higher sis, producing a total of 24 observations (independent mean value in the experimental group (high samples) in our data set (see Table 1).
recovery) when compared to the control group Studies Included in the Meta-Analysis
Effect Sizes for the Relationship of SRP and … NOTE: Total number of effect sizes: 45. Sat = satisfaction; Rep = repurchase intentions; Wom = word-of-mouth; Tru = trust; Ima = image; Qua =quality; IntC = intentions to complain; Swi = switching intentions; Pay = pay more intentions; Ext = external response.
a. Classified as outlier.
b. Experiment 3 was used.
(no failure), the SRP is present and ES is positive.
All together, using this approach, 45 effect sizes were Otherwise, if satisfaction is higher in the condi- available for the purpose of our meta-analysis. As pre- tion of no failure, then ES is negative and reflects sented in Table 1, most studies reported multiple effect a situation of an inverse SRP. Finally, if there are sizes, particularly if we consider the first two outcome no significant differences between satisfaction in variables (satisfaction and repurchase intentions), for the conditions of high recovery and no failure, which there are more frequencies of effects (31 of the then ES is close to zero and the SRP is null.
total 45). However, there were no significant mean dif- 6. In conclusion, a positive effect size reflects a positive effect of the treatment (high recovery ferences for satisfaction when comparing studies report- ing single or multiple effect sizes (t = .383, p < .706).
When considering repurchase intentions, only one study However, direct calculation of effect size (as present presented a single effect size (ES = –.52), and for all oth- above in Item 4) is uncommon because rarely is there ers, there were multiple effect sizes (n = 11; mean ES = enough information. Under those circumstances, effects .025). Mean differences in this case were not significant sizes are obtained through a range of statistical data (e.g., at the .05 significance level (t = 2.042, p < .068).
Student’s t, F ratios with one df in the numerator, χ2) by For most cases (35 of the 45 effect sizes), articles pre- means of the formulas given by Lipsey and Wilson sented ANOVA results, including mean values of the out- come variable (e.g., satisfaction) for the treatment group JOURNAL OF SERVICE RESEARCH / August 2007 Descriptive Statistics for Effect Size Integration
NOTE: LCI = lower confidence interval; UCI = upper confidence interval; Sig. = significance; Min. = minimum; Max. = maximum; Sat = satisfaction;Rep = repurchase intentions; Wom = word-of-mouth; Image = corporate image. Fail-safe number attenuated at .05. NC means the file drawer N wasnot calculated because the 95% confidence interval contained zero (effect size with significance p > .05).
a. Number of studies.
b. Number of observations.
c. Combined N over all independent samples.
(high recovery effort) and for the control group (no fail- Homogeneity of the effect size distribution was tested ure), sample size for each one, and the t or F statistic.
by the Q statistic, which is distributed as a chi-square This was enough for computing the standardized differ- (Hedges and Olkin 1985).5 If the null hypothesis of ence and transforming it into r. In some cases (3 of 45), homogeneity is rejected, it indicates that the variability in nonparametric tests were presented (e.g., chi-square) and effect sizes is larger than it would be expected from sam- the appropriate formula was used for conversion. In the pling error, or in other words, differences in effect sizes remaining cases (7 of 45), authors provided either com- may be attributed to factors other than sampling error plete information for the direct calculation (mean, stan- alone, maybe moderating variables related to studies dard deviation, and sample size of each group) or F value characteristics (Lipsey and Wilson 2001, p. 115).
and group sizes (without group means).
For each relationship in which homogeneity of effect size was rejected (i.e., heterogeneity evidence), an analy- Effect Size Integration
sis of moderating effects was performed, considering thestudy characteristics that were coded based on informa- Effect size integration (mean, significance, and confi- tion provided in the articles. These variables included dence intervals) was performed following common guide- service category (hotels, restaurants, and others), method lines (Lipsey and Wilson 2001). Because the true (survey versus experiment), design (cross-sectional ver- relationship between variables is mainly influenced by sus longitudinal), subjects (students versus nonstudents), sampling and measurement error, correlations were first scenario (used or not), sample size (total, control group, weighted by the inverse variance and then by the inverse and treatment group),6 number of items used to measure variance corrected for measurement error (cf. Lipsey and the dependent variable,7 and reliability of the dependent Wilson 2001, p. 110).4 Thus, our effect size integration is variable. These data were available for most studies.
presented in three stages, based first on observed correla- When information about the number of subjects in each tions, then on correlations corrected for sampling error, experimental group was not available, we used the mean and finally on correlations corrected for both sampling and group size by dividing the total sample size by the number measurement error. A confidence interval is presented for of groups in the experiment (in 6 of the 24 observations).
each effect size, and it is significant when it does not In some studies, authors mentioned in tables’ notes the include zero. Significance for the mean effect size can also sizes of minimum and maximum groups. In these cases, be tested by z statistic (p < .05 if z > 1.96). When the mean we used the minimum as the sample size for both the treat- effect size is significant, a fail-safe N (also known as file ment and the control group (in 7 of the 24 observations).
drawer N) is calculated, estimating the number of non- When studies measured dependent variables using sin- significant and unavailable studies that would be necessary gle items8 or when reliability values were unavailable,9 to bring the cumulated effect size to a nonsignificant value these reliabilities were estimated using the Spearman- (known as the “file drawer problem”; Rosenthal 1979).
Brown procedure suggested by Hunter and Schmidt (2004, This statistic is an indication of how robust results are.
p. 311), a common approach in meta-analyses in market- Table 2 presents a summary of this information.
ing (e.g., Grewal et al. 1997). Among the 19 studies providing effect size for satisfaction, 5 were based on sin- studies remain in the “file drawer,” especially because gle items and had an average reliability estimated as .701.
this is a recent research topic in the services marketing In the same way, among the 12 studies providing effect size for repurchase intentions, 5 studies did not provide Conversely, the cumulated effect of the SRP on repur- information on reliability (2 were based on single items chase intentions is negative and not significant at the .05 and had reliability estimated as .66, 2 were based on two significant level (r = –.07, p < .068), not supporting items and had reliability estimated as .795, and 1 used four Hypothesis 1b and suggesting that there is no SRP effect items and had reliability estimated as .886).
on repurchase intentions. Confidence intervals rangedfrom –.143 to –.002, indicating a small effect. These find-ings were based on 12 independent observations and ANALYSIS AND RESULTS
7,788 subjects. Because of this small number of observa-tions and the low magnitude of the average effect size, a Table 2 presents the results for the integration of effect relatively small number was found for the fail-safe N, sizes of the SRP on satisfaction, repurchase intentions, indicating that 5 unpublished studies with null results word-of-mouth, and image. Note that only those depen- would be needed to reduce the average effect size to the dent variables with 2 or more observations are presented level of .05. Compared with satisfaction, it is more likely (trust, quality, intentions to complain, switching inten- that this number of unpublished studies exists. Therefore, tions, pay-more intentions, and external response are not these results do not support Hypothesis 1b and indicate submitted to further analyses because they were based on that repurchase intentions are not increased by a high In this stage of the analysis, 1 of the 24 studies pre- The integrated effect sizes of word-of-mouth and image sented in Table 1 (Oh 2003) had to be excluded because were also negative but not significant, and they were based it was classified as an outlier. Its sample size was 30,905 on a smaller number of observations (six and two, respec- for the experimental group and 17,278 for the control tively). These results do not support Hypotheses 1c and 1d.
group (the mean sample size for the remaining studies Because of these null results, the file drawer number is not was 90 for the experimental group and 418 for the con- calculated for word-of-mouth and image.
trol group). If included, this study would produce an A heterogeneous subset of effect sizes was obtained inverse variance weight of 11,053, whereas the mean both for satisfaction and for repurchase intentions (see Q value for the other studies was only 23. In summary, the test of homogeneity in Table 2), indicating that moderating study was excluded because it would bias the cumulated variables might help explain the variance in the effect effect size.10 All of the subsequent analyses were per- sizes. Hence, we tested whether there were differences in formed after the exclusion of this study from the data set.
effect sizes across study characteristics or, in other words, The integrated effect size (corrected for sampling and if the coded study characteristics were significant modera- measurement error) shows a positive and significant tors. These analyses are presented in the next section.
value for satisfaction (r = .125, p < .017), supporting theexistence of the SRP for satisfaction (Hypothesis 1a), Moderating Effects
with a medium effect size (Lipsey and Wilson [2001]consider r ≤ .10 as small effect; .10 < r < .40 as medium A common procedure for testing whether studies’ effect; and r ≥ .40 as large effect). This indicates that sat- characteristics can explain variability in the effect sizes is isfaction increases after a high service-recovery effort.
regression analysis, in which effect sizes are entered as The confidence interval ranged from .032 to .217, sug- dependent variables and moderators as independent vari- gesting a small to medium effect. These results were ables (e.g., Eisend 2006; Szymanski and Henard 2001).
based on 18 independent observations and 7,502 However, this procedure may be limited if there are only subjects. Fail-safe N suggests that 27 studies with non- few observations for each level of the moderators and/or significant effect size would be needed to reduce the there is small number of effect sizes. In this case, confi- cumulated effect size to a level of just significant (a level dence in the results is threatened by low statistical power of .05 was used as “just significant,” similar to Grewal and capitalization on sampling error (Hunter and et al. 1997).11 In other words, to bring the significant Schmidt 2004, p. 70). This was the case in our data set, effect of the SRP on satisfaction down to the level of just as there were only 18 effect sizes for satisfaction and 12 significant at α = .05, it would be necessary to find 27 studies with null results to be included in our analysis.
Because of this, we conducted a subgroup meta-analysis, This is rather a small number, but it is somewhat unlikely comparing the mean effect size between the levels of that with only 18 studies identified for satisfaction, 27 each moderator, a common procedure in meta-analyses in JOURNAL OF SERVICE RESEARCH / August 2007 Effects of Moderator Variables
marketing (see, e.g., Geyskens, Steenkamp, and Kumar or in the other categories (–.057), with significant 1998; Grewal et al. 1997; Palmatier et al. 2006; Pan and differences (p < .000 in both cases), which support Zinkhan 2006). Also, an additional moderator analysis Hypothesis 2d for satisfaction. In repurchase intentions, was conducted following the hierarchical method pro- the difference between hotels versus others was repli- posed by Hunter and Schmidt (2004). In this method, cated (p < .002), and the studies conducted in restaurants moderator variables are considered in combination to presented higher effect sizes when compared to studies in avoid confounding of correlated moderators.
other categories (p < .056), giving partial support for In Table 3, we present the results of the moderator Hypothesis 2d for repurchase intentions. These findings analyses, based on the subgroup meta-analyses. These indicated that the context in which the SRP is investi- results show the mean corrected effect size (both for mea- gated might also influence the results.14 surement and sampling error) in each level of the moder- Following a recommendation by Hunter and Schmidt ators, together with the number of studies and the test of (2004, p. 424) that subgroup meta-analysis may yield con- mean differences. Our moderator variables included founded results if moderators are correlated, we conducted service category (hotels, restaurants, and others), method an additional test for moderators, using hierarchical analy- (Survey × Experiment), design (Cross-Sectional × sis. In this analysis, moderators are considered together.
Longitudinal), subjects (Students × Nonstudents), and We provide results from this analysis in Table 4, consider- ing the three moderators dealing with the methodological As indicated in Table 3, the use of surveys or exper- differences. One difficulty we encountered in this analysis iments did not change the direction of effect sizes, was the small number of studies in each cell when consid- either in satisfaction or in repurchase intentions, not ering the eight groups (a combination of 2 × 2 × 2).
supporting Hypothesis 2a. On the other hand, longitudi- Despite this limitation, we could make five comparisons nal studies tended to present relatively higher means for for the satisfaction effect sizes and one for the repurchase satisfaction when compared to cross-sectional studies intentions. In these analyses, two factors are fixed and the (.192 versus .042, p < .056). This factor did not have levels of the third factor are compared.
influence on the effect sizes of repurchase intentions, In Table 4, we first compare cross-sectional versus not supporting Hypothesis 2b for this variable. Also, longitudinal studies between experiments conducted with studies using student samples presented higher mean- students (.086 versus .239). Results did not suggest a sig- effect sizes for satisfaction when compared to those nificant difference (p < .129). In the sequence, the same using nonstudent samples (.184 versus .060, p < .095).
comparison (cross-sectional versus longitudinal) was Thus, there was support for Hypothesis 2c for satisfac- made between experiments conducted with nonstudents tion only at the .10 significance level. Findings sug- and a relatively higher effect size of satisfaction was gested no difference between Students × Nonstudents in found for longitudinal studies (–.149 versus .156, p < the repurchase intentions, not supporting Hypothesis 2c .085). When we compared cross-sectional versus longitu- dinal between surveys conducted with nonstudents (.147 Considering service categories, studies conducted in versus .103), no significant difference was found (p < hotels presented relatively higher effect sizes for satisfac- .399). This is an indication that the difference between tion (.495) when compared to those in restaurants (.044) cross-sectional versus longitudinal (suggested in Table 3) Hierarchical Moderator Analysis: Cross-Tabulation of Effect Sizes Among Moderators
NOTE: Values in each cell represent number of effect sizes, mean effect size, and significance; em-dashes indicate no data. Values of repurchase inten-tions are in parentheses and values of satisfaction are outside parentheses. For example, there were two effect sizes of satisfaction in the categorySurvey/Nonstudent/Cross-sectional, with a mean effect size of .147 and significance of .453. NC = significance not calculated when only one effectsize was available.
a. Contrast 1: .086 versus .239, p < .129.
b. Contrast 2: –.149 versus .156, p < .085.
c. Contrast 3: .147 versus .103, p < .399.
d. Contrast 4: .086 versus –.149, p < .088.
e. Contrast 5: .239 versus .156, p < .334.
f. Contrast 6: –.092 versus .10, p < .137.
is likely to be higher when experiments are conducted mixed results in the SRP literature, a meta-analysis can with nonstudents rather than with students.
help in understanding these inconsistencies by accumu- Another comparison was made between students ver- lating results after adjusting for measurement and sam- sus nonstudents, considering experiments using cross- pling error and by identifying study characteristics that sectional design. In this case, a relatively higher effect may account for the variability in effect sizes. Although size of satisfaction was found for the studies using the SRP is a relatively recent topic in the services mar- students (.086 versus –.149, p < .088). This difference keting literature (first empirical studies begin in the was not statistically significant when we made this same 1990s), a total of 21 studies (24 independent samples) comparison in experiments using longitudinal approach.
could be included in the meta-analysis and used for the Therefore, the difference in satisfaction effect sizes between students versus nonstudents (suggested in Table Our primary results reveal support for the SRP only in 3) seems to be stronger for experiments using a cross- the case of satisfaction, with a mean adjusted effect size sectional (rather than longitudinal) approach.
of .125, which was significant at the 5% level. This is A similar analysis was conducted for the effect sizes interpreted as a medium effect size (Lipsey and Wilson of repurchase intentions. However, the limitation of a 2001), with a confidence interval ranging from .032 to small sample size in each cell was more severe in this .217. Thus, from the cumulative studies reviewed by our case, as the total number of effect sizes was smaller for this variable (n = 12). As a consequence, only one com- increases after a high service-recovery effort, suggesting parison was possible: cross-sectional versus longitudinal the existence of the SRP for this variable. This finding for surveys using nonstudents (–.092 versus .100, p < supports the notion that a customer’s postfailure satisfac- .137). However, no significant difference was found.
tion exceeds prefailure satisfaction (McCollough andBharadwaj 1992).
Based on this result, are recovery encounters really DISCUSSION
good opportunities for service providers to increase cus-tomer retention, as recommended by Hart, Heskett, and The meta-analysis presented in this study provides a Sasser (1990)? The empirical integration provided by our systematic review and a quantitative integration of the meta-analysis suggests a negative answer, in that the SRP effects of high recovery efforts on the dependent vari- effect was not evident for the repurchase intentions vari- ables (satisfaction, repurchase intentions, word-of- able at the 5% significance level. Even if we consider a mouth, and corporate image), revealing the cumulative 10% significance level, the small mean effect size with a effect of the SRP on these variables. Because there are negative sign (–.072) suggests that the SRP effect on JOURNAL OF SERVICE RESEARCH / August 2007 repurchase intentions runs counter to what the SRP pre- Results suggested that effect sizes were not significantly dicts. In other words, customers’ postfailure repurchase different across studies using experiments or surveys, intentions are likely to be lower than or equal to their pre- either for satisfaction or for repurchase intentions. This is an indication that this methodological decision did not These results are interesting because they show that the affect the support/nonsupport of the SRP in the related SRP works for satisfaction but not for repurchase inten- studies, in agreement with what is predicted by Michel tions. Customers are willing to make a positive evaluation of a firm providing a high recovery effort, but they are not When comparing longitudinal versus cross-sectional likely to repatronize this firm. Why would this happen? studies, it was found that longitudinal studies provided A possible explanation is that satisfied customers are stronger evidence for the SRP in satisfaction. This differ- not necessarily loyal (Reichheld 1994). A meta-analysis ence was further investigated in the hierarchical moderator of customer satisfaction found, for instance, that it analysis. It was found not only that longitudinal studies explains less than 25% of the variance in repurchase provided stronger support for the SRP effect for satisfac- intentions (Szymanski and Henard 2001). In agreement tion (contrary to the prediction of Michel and Meuter with this rationale, a recent study in e-retailing by Forbes, 2006), but also that the difference between cross-sectional Kelley, and Hoffman (2006) found that customers are not versus longitudinal was likely to be higher when experi- likely to repurchase once a failure has been experienced, ments were conducted with nonstudents, rather than with even if they are satisfied with the recovery effort. These find- students.15 This finding is interesting because it suggests a ings might be influenced by the low switching costs that possible interaction between moderators.
customers experience in online shopping. Nevertheless, Another influencing factor in the effect of SRP on sat- they are further evidence of the importance of consider- isfaction was the use or nonuse of students as research ing satisfaction levels and switching levels in combina- subjects (significance was found only at the 10% level).
It was found that students were more likely to support the Another possible explanation is the following line of reasoning: When evaluating postrecovery satisfaction, A possible explanation may be that nonstudents are customers can be more influenced by the recovery usually more experienced customers, and because of this, process itself and the positive rewards that it may provide they are less likely to experience a positive disconfirma- (e.g., a free service, a compensation for the failure), but tion if they have higher expectations. Furthermore, when evaluating their likelihood of repurchasing from students are usually researched out of the purchasing the same firm, customers might think that their original environment and, therefore, their satisfaction-evaluation desired result was not accomplished in the purchasing process may not be influenced by their past experiences, process (the company did not provide a correct service in as in the case of “real” customers. Hence, satisfaction the first time), and therefore, it is not worth repurchasing evaluations of nonstudents will likely be lower than those from this firm. In agreement with this, a recent study by of students, and in the case of a service failure, this pat- Magnini et al. (2007) supports the notion that customers tern may be maintained because students out of their pur- who have experienced previous failure do not experience chasing context might not be as severe in their the SRP effect (i.e., the SRP is more likely to occur when evaluations as the nonstudents, who will also be closer to it is the firm’s first failure with the customer).
their previous experiences and encounters with the firm.
The accumulated effect size was not significant for These differences between students versus nonstu- word-of-mouth and corporate image. A limitation in this dents were more pronounced in cross-sectional experi- case was the reduced number of available studies for test- ments when compared to longitudinal experiments, ing the SRP effect for these variables. As a consequence suggesting again a possible interaction between modera- of the nonsignificant result and the homogeneous effect tors. In summary, these results should make us reflect on sizes, they were not included in the subsequent analysis the following: If researchers conduct cross-sectional experiments, then a difference in support for the SRP Further analyses of homogeneity of effect sizes for might exist when using students or nonstudents; but satisfaction and repurchase intentions suggested possible when conducting longitudinal experiments, no signifi- moderators, as heterogeneous effect sizes were found for cant difference exists between students or nonstudents.
both variables. By using subgroup meta-analysis and As longitudinal studies provided stronger effect sizes for hierarchical moderating analysis, we first identified three the SRP, this is an indication that longitudinal experi- factors that moderated the SRP effect for satisfaction ments are more likely to provide support for the SRP, (design, subject, and service category) and one factor whether having students as respondents or not. This find- influencing repurchase intentions (service category).
ing contributes to the SRP literature by shedding light on the question “Are methodological aspects of the studies by β). The power of an experiment is thus defined as influencing their support/nonsupport for the SRP?” Our (1 – β) and interpreted as the probability that a statistical meta-analysis seems to offer an affirmative answer and test will correctly reject a false null hypothesis (Cohen suggests that future studies should focus on these bound- 1988). Although Cohen (1988) recommends using .80 as the threshold for statistical power, there is also a sugges- Finally, service category was also a significant moder- tion for using .50 for the social sciences, in which errors ator, with influence both on satisfaction and on repur- are less likely to have life-threatening consequences chase intentions. It was found that studies in hotels provided higher support for the SRP effect on satisfaction It has been argued that including the statistical power when compared to all other categories. This should also discussion in the context of a meta-analysis can con- be a point of future debate: “Are there differences across tribute to enhance the reliability of the meta-analysis service categories that facilitate the support for the (Muncer, Craigie, and Holmes 2003). Following the SRP?” For example, when a service failure occurs in a guidelines presented by these authors, we (a) used the context of the hospitality industry, most customers will mean effect size computed for the included studies to seek a solution to their problem, and therefore, there may estimate the average statistical power of the combined be a greater tendency toward redress-seeking behavior in studies and (b) estimated the statistical power of each this context (McCollough 2000), and they may also have study, indicating the ability to detect an effect size of the a higher likelihood of receiving recovery service. In this magnitude of the mean effect size obtained in the meta- case, customers cannot easily look for another service analysis (estimated as population effect size), given the provider or interrupt their travel plans. As mentioned sample size and the significance level of .05. We used above, switching costs may be an influencing factor in G*Power 3.0® (Faul et al. in press) for these analyses and this context. The same factor may explain why hotels and the results are presented and discussed below.
restaurants scored higher in effect sizes for repurchase The mean effect size of satisfaction (r = .125; d = intentions. Future studies could investigate this proposi- .251), in conjunction with the mean sample size of the tion further and test, for instance, the differences in the SRP effects for customers with high versus low switch- level of .05, produced a statistical power of .42. This value is below the recommended .80 threshold level but An additional analysis of possible moderators indicated relatively close to the level of .50 (Muncer, Craigie, and that studies using more reliable scales did not provide sup- Holmes 2003). For the studies of repurchase intentions port (or provided weaker support) for the SRP (a signifi- (r = –.072; d = –.145; n cant negative correlation was found) both for satisfaction tical power was estimated as .32. This analysis indicates and repurchase intentions. As expected, the same influence that the average statistical power was smaller in the 12 was found for the number of items used to measure the studies of repurchase intentions when compared to the 18 dependent variables. On the other hand, sample size was not significantly related (at the 5% level) to the effect sizes A deeper investigation of the statistical power of the in these two dependent variables, although a correlation of individual studies showed that power varied between .11 .43 (significant at the 10% level) was found between treat- and .98 (mean = .30; SD = .23; n = 18) for the satisfaction ment group and satisfaction, indicating that studies using variable and between .07 and .80 (mean = .22; SD = .20; larger samples for the service recovery group were more n =12) for repurchase intentions. There was no significant likely to provide support for the SRP. This is in agreement difference between these two means (t = .91, sig. = .37), with Michel and Meuter’s (2006) argument that, once the indicating that statistical power was not statistically differ- SRP is a very rare event (Boshoff 1997), it becomes very ent between the observations of satisfaction and repur- difficult to achieve a large sample of customers who have chase intentions. Note that these means are not weighted received a very satisfactory recovery, and this limitation by any factor, justifying why they are different from the may be responsible for the nonsignificant results presented previous mean statistical power presented for satisfaction in the literature. Indeed, as we found, studies with a larger (.42) and repurchase intentions (.32), when average statis- number of respondents in the recovery group tend to pre- tical power was computed directly from the final average sent greater support for the SRP and have higher statistical effect size weighted by sample size and reliability.
power, as discussed in the next section.
We also recomputed statistical power after excluding from the database those studies with power lower than .5 Statistical Power16
and checked whether there were major changes in theestimates of the mean effect size (weighted by sample Statistical power is related to the probability of not size and reliability), significance and confidence inter- rejecting a false null hypothesis (Type II error, defined vals. For satisfaction, four studies produced a mean effect JOURNAL OF SERVICE RESEARCH / August 2007 size of r = .334, with sig. = .018, and confidence intervals the different stages of loyalty (e.g., the SRP may exist between .20 and .47. We can note that these studies pro- during the cognitive stage but not the action stage).
vide stronger support for the SRP in the case of satisfac- Future studies are needed to provide further investiga- tion (r changed from .125 to .334). Considering only tion of the reasons behind this difference of support of these four studies with power greater than .50, we should the SRP in satisfaction but not in repurchase intentions.
conclude that there is a stronger SRP effect for satisfac- In this sense, switching costs may be one of the relevant tion, although it is still a medium effect.
factors accounting for this difference, as suggested by the For repurchase intentions, only one study presented findings of Forbes, Kelley, and Hoffman (2006). The dif- power higher than .50 (power = .80). This study pre- ference between “satisfaction level” and “satisfaction sented an effect size of r = –.10, lower bound = –.20, strength” (Chandrashekaran et al. 2007) can also con- upper bound = .00. These values are similar to the ones obtained from the 12 integrated studies (see Table 2: r = Another possibility is that studies investigating repur- –.072, lower bound = –.143, upper bound = .002), indi- chase intentions may have used approaches that tend to cating that including the studies with lower power did not give no support for the SRP. For example, 6 out of 12 of change the interpretation of the results in the cumulated the effect sizes for repurchase intentions came from cross-sectional surveys with nonstudents. It was sug- Overall, these power analyses indicate that the studies gested in the moderator analysis of satisfaction that (a) conducted to test the SRP and included in our meta-analy- studies using cross-sectional design produce relatively sis have relatively low statistical power, which can be one smaller effect sizes when compared to longitudinal additional reason why “conflicting results” are found approach, and (b) studies with nonstudents produced regarding the existence or nonexistence of the SRP. Our lower effect sizes when compared to students use. Given analyses showed, for example, that the mean effect size for that 50% of the effect sizes for repurchase intentions used satisfaction was greater (r = .33) when only studies with this design, it may be possible that a SRP was not sup- acceptable power (higher than .50) were retained in the ported for repurchase intentions because of these analysis. This suggests a stronger effect (r = .33) than methodological differences across the studies.
when studies with lower power are also included in the Our moderation analyses were intended to be only an meta-analysis (r = .125). Because of the limited number of exploratory investigation of the possible methodological studies, however, we could not include the power estimates aspects that might have an influence on the effects of as one of the variables in our moderation analysis.
SRP, especially because of the limited number of studiesin each cell in our hierarchical moderation procedure Implications and Further Research
(Hunter and Schmidt 2004). Because of this, futureresearch is necessary to provide further investigation of The main implication of this meta-analysis is to show that the SRP effect is more likely to occur for satisfaction Also, although theoretical moderators have been sug- than for repurchase intentions. This result challenges us gested in the SRP literature (see Magnini et al. 2007), to understand why satisfied customers are not neces- because only few studies have tested them empirically, sarily loyal. In a recent investigation of this topic, they could not be included as variables in our data set,19 Chandrashekaran et al. (2007), by decomposing satisfac- as illustrated in Figure 1. Examples of these moderators tion in two factors (satisfaction level and satisfaction include severity of the failure, prior failure with the firm, strength), have theorized that weakly held satisfaction stability of the cause of the failure, and perceived does not translate into loyalty and that only strongly held company control. Hence, we suggest that these modera- satisfaction is able to translate into loyalty. Based on this tors be further investigated in future SRP studies.
rationale, future SRP studies could check the influence Our revision of the mean effect sizes when taking sta- that satisfaction strength might exert as a possible mod- tistical power into account, as suggested by Muncer, erator (i.e., the SRP effect may occur for repurchase Craigie, and Holmes (2003), indicates the relevance of intentions when it occurs for satisfaction and this satis- considering statistical power in the context of the studies testing the SRP, especially for the satisfaction variable, Also, the lack of support for the SRP effect on repur- whose findings suggested a stronger effect of the SRP.
chase intentions may suggest differences of the SRP on Thus, future studies of the SRP should consider statisti- the diverse stages of the loyalty pyramid (cognitive → cal power a priori so as to be able to achieve greater con- affective → conative → action; Oliver 1997). Because fidence in the results of the significance tests.
the studies included in our meta-analysis did not take this Our statistical power analysis can also be used to sug- factor into account empirically, future studies should be gest sample sizes for future studies. Based on the inte- conducted to investigate whether the SRP can exist for grated effect size as an estimate of the effect size of the considered population, we can calculate the required indicated that customer satisfaction after a high recovery sample size that future studies should use to be able to effort is greater when compared to that satisfaction prior detect the desired effect. For example, for satisfaction, to the service failure. However, the same is not true for considering sig. = .05, power = .8, it would be necessary the customer repurchase intentions. Because of this, to have a sample size of 251 in each group (experimental service managers should make every effort to provide and control) to be able to identify an effect size of r = services correctly on the first time, rather than permitting .125 (d = .251) as significant. This value would be failures and then trying to respond with superior recov- reduced to 123 if we considered the statistical power in ery. This view has already been advocated by single the suggested level of .50 (Muncer, Craigie, and Holmes studies (e.g., Andreassen 2001; McCollough, Berry, and 2003). On the other hand, for the repurchase intentions, it Yadav 2000), but our meta-analysis makes this argument would be necessary to have a sample size of 748 in each stronger, given that the meta-analysis provides an inte- of the two groups to identify an effect size of r = –.072 grative review and a quantitative integration of the con- (d = –.145) with a statistical power of .8. If we lowered the power to .5, the sample size would be reduced to 367.
Second, trust is considered a key variable when man- As expected, for a given power, it is necessary to have a aging customer relationships (see Morgan and Hunt larger sample size to identify smaller effect sizes.
1994). In the context of service failure and recovery, it is Another relevant area for further research includes the expected that satisfaction with service recovery would cognitive/affective mechanisms behind the SRP. Because lead to the building of trust. However, because results there is support for the SRP effect on satisfaction, as sug- have demonstrated that customers who were initially sat- gested by the integrated results of the meta-analysis, and isfied with the service expressed greater trust when com- satisfaction involves cognitive and affective dimensions pared to the satisfied complainants, not supporting the in prepurchase, purchase, and postpurchase phases of SRP in trust (Kau and Loh 2006), a service failure seems consumption (e.g., Westbrook 1980), future studies to be a serious threat to trust. From the manager’s point should investigate further the cognitive and/or affective of view, it is critical to manage customer trust in the mechanisms driving the SRP effect. Although previous service provider. Trust can be built and/or enhanced with studies have investigated the influence of the cognitive a company providing a reliable service over time. Hence, and/or affective factors on satisfaction with service service failure should be avoided also because of its neg- recovery (e.g., Andreassen 2000), there is a lack of stud- ative impact on customer trust. Indeed, research investi- ies examining the cognitive and affective dimensions in gating why customers stay, given a switching dilemma, has suggested that the most important reason (out of all As discussed by Parasuraman (2007), future research the 28 revealed reasons) was “lack of a critical incident,” should also investigate whether there is an optimal mix of or in other words, customers stayed because a negative reliability versus recovery investments or, in other words, critical event had not occurred (Colgate et al. 2007).
how much should be invested in delivering reliable Thus, the service provider should perform as promised if service rather than in superior recovery when problems customers’ perceived confidence is expected to be occur. What are the main variables in this context and strengthened. This confidence can be increased by invest- what are their influences? These questions are also rele- ing in customers’ feeling of comfort, trust in the service vant for future research on the service-recovery context.
provider, satisfaction with the service provider, familiar- Finally, we would suggest that more studies should be ity with the service provider, history with the current conducted to investigate not only satisfaction and repur- service provider, and lack of negative critical incidents chase intentions as main recovery outcomes but also impor- (Colgate et al. 2007). Service managers should invest in tant constructs like word-of-mouth and corporate image, for these factors to earn the customer’s confidence.
which only a limited number of studies were found in the lit- Nevertheless, achieving 100% service reliability can be erature. Other relevant variables include trust, quality, inten- impossible or cost prohibitive in most settings. Thus, in tions to complain, and switching behavior. Testing the SRP case of a failure, companies should strive to provide a effect on the customers’ actual behavior rather than on their service recovery of high performance anyway because a behavioral intentions alone would also contribute to the cur- delight with the recovery can contribute in moving the cus- rent state of the knowledge about the SRP.
tomer up in the loyalty hierarchy (cognitive → affective →conative → action), as suggested by Andreassen (2001). In Managerial Implications
other words, the effect of the SRP on repurchase intentionsmay be influenced by the loyalty stage in which the cus- Our meta-analysis has a number of implications for tomer is found. This prediction needs to be confirmed service managers. First, the reviewed studies of the SRP by future studies, and the findings from such studies will JOURNAL OF SERVICE RESEARCH / August 2007 provide service managers a deeper understanding of this Limitations
process. In the interim, we suggest that if a service failureoccurs, the first step should be to provide a recovery of Meta-analyses offer several benefits, but they also high performance in such a way to restore customer satis- have intrinsic limitations, which are common in most faction and achieve customer delight with the recovery meta-analytic studies in the marketing literature. We dis- process. The second step should be to consider whether cuss the main limitations of our study below.
this customer’s manifested satisfaction would be translated First, our analyses are based on secondary data, and into future loyalty with the company. In this stage, man- therefore, we cannot use information other than those agers could estimate through surveys the customer’s satis- presented in the articles. For example, we could test the faction strength (i.e., the strength with which the methodological moderators presented in Figure 1 but not satisfaction judgment is held; Chandrashekaran et al.
the theoretical moderators because very few studies 2007) in such a way as to target those customers with tested the SRP considering these categories. Also, we had weakly held satisfaction, as they are less likely to turn their to consider a mean group size when studies did not satisfaction into loyalty (Chandrashekaran et al. 2007). For inform the exact size of the groups being compared.
instance, these customers could receive long-term benefits Moreover, only satisfaction and repurchase intentions (e.g., discounts based on history with the company), which presented relatively high frequency of studies and could would increase their switching costs, and the company be entered in further analysis of moderators. For would use the future service encounters as opportunities to example, even though there was reference for other increase the customer satisfaction and make it more recovery outcomes (e.g., trust, quality, intentions to com- plain, and switching intentions), they could not be Third, service managers should be cognizant of the dif- included in our effect-size integration.
ferences of the SRP across service settings. Findings from Second, although there were a large number of studies the meta-analysis indicated a stronger effect size of satis- investigating service failure and recovery (about 300 were faction in hotels, compared to all other categories, sug- identified), a relatively small number of them tested the gesting that SRP is more likely in this setting. This may be SRP empirically and could be included in the analysis.
because of the high-contact characteristic of the hospitality Even with the recommendation that 10 or more studies industry. Furthermore, this is a service of relatively longer should be an acceptable minimum number (see note 12), duration because even customers who stay for a short time we should be careful in interpreting the results based on in a hotel (e.g., 1 day) may engage in a series of service a small number of studies, especially regarding the mod- encounters. Also, given a service failure, the customer of a erator analysis. We recognize that the presented modera- hotel may be more prone to engage in redress seeking, as tor analysis has a more exploratory perspective.
he or she would not like (or be able) to change his or her Nevertheless, results from the moderator analysis can schedule (e.g., cancel a business meeting or a vacation help researchers to design new studies that address the plan) because of this failure. Interestingly, there was also a boundary conditions for the SRP effect.
higher likelihood of the SRP effect in repurchase inten- Finally, studies included in the meta-analysis pre- tions in the hotel setting. Thus, managers dealing with fail- sented relatively limited statistical power in general.
ures in a hotel context should also be able to provide Because of this, the number of studies limited the con- recoveries of high performance so as to boost customer servative procedure of calculating more robust effect satisfaction and repatronage intentions.
sizes only from studies with acceptable statistical power.
In addition, service managers should also monitor cus- When we implemented this procedure for satisfaction, tomers’ word-of-mouth, which can be very negative if a only 4 studies remained, with the remaining 14 studies failure occurs and the company is not able to provide a sat- having low power. This is a clear indication that future isfactory recovery (i.e., a “double deviation,” as termed by studies in this context should consider statistical power a Bitner, Booms, and Tetreault 1990). Indeed, the findings priori and determine the minimum sample size required from the meta-analysis showed that there was a negative average effect size for word-of-mouth resultant from thesix independent observations reviewed. As recommendedby Andreassen (2001), positive word-of-mouth from exist- CONCLUSION
ing customers can make the company more attractive inthe eyes of the new customers. Negative word-of-mouth Notwithstanding the presented limitations, the find- derived from unsatisfactory recoveries, on the other hand, ings from this meta-analysis contribute to a greater could push to competitors not only potential new understanding of the SRP by (a) estimating its cumulated customers but also existing customers.
mean effect for the key dependent variables (satisfaction and repurchase intentions), (b) testing how studies char- most journals will not publish a meta-analysis unless it contains at least acteristics might influence these results, and (c) suggest- 5 to 7 studies. So with 10 or more, you are OK” (personal communica-tion, October 16, 2006). Professor David B. Wilson (author of the book Practical Meta-Analysis; see Lipsey and Wilson 2001) also said, Meta-analyses should not be viewed as conclusive or “Minimum number of studies: 2. Of course, this limits the analyses that as a substitute of new primary research but only as a you can do” (personal communication, October 14, 2006).
13. Method and scenario are analyzed together because all studies methodological tool that makes a temporary “balance using experiments are based on scenarios, and none of the surveys used sheet” of the current state of the knowledge in a given area. Its main contribution to science is to help 14. To achieve a better understanding of the variability of the effect sizes, we also checked if effect sizes were influenced by scale reliabil- researchers to direct their next wave of research toward ity, number of items measuring dependent variables, and sample size.
still-unexplored questions and boundary conditions.
We found a significant correlation between effect size and scale relia- In this spirit, we hope that the results reported by our bility (–.62 in satisfaction and –.57 in repurchase intentions), with neg-ative values indicating that studies using more reliable scales did not meta-analysis provide managers and researchers with provide support (or provided weaker support) for the SRP. As expected, inspirations for designing new in-depth and extensive significant positive correlations were found between number of items investigations that will keep advancing the services mar- measuring the dependent variable and scale reliability (.59 in satisfac-tion and .81 in repurchase intentions). As a result, the number of items measuring satisfaction was also negatively related to the satisfactioneffect size (r = –.47) and the same was true for repurchase intentions(r = –.57). Sample size (total, treatment, and control) was not signifi- cantly correlated with effect sizes either in satisfaction or in repurchaseintentions (the highest correlation was found in the pair treatmentgroup–satisfaction, r = .43, p < .075). A regression analysis of these 1. Among our 24 studies, 10 used the term “repurchase intentions” variables on the effect sizes produced no significant results for either as dependent variables and only 2 considered “loyalty” (Kau and Loh satisfaction or repurchase intentions. However, these regression results 2006; Zeithaml, Berry, and Parasuraman 1996). Because these two might not be reliable because they may be influenced by low statistical studies worked with loyalty intentions, we considered these variables power and capitalization on chance because of the small number of together, named as “repurchase intentions.” observations (see Hunter and Schmidt 2004, p. 70).
2. Indeed, we noticed in our data analysis process that the treatment 15. We could not test if surveys with nonstudents (rather than with group was relatively smaller (mean = 90) when compared to the control students) produce greater difference between cross-sectional versus longitudinal because there were no observation for the cells “survey 3. Even severity of the failure could not be empirically evaluated in our moderation analysis because only 4 studies tested SRP considering 16. We are very thankful to Reviewer C for suggesting the inclusion Low × High severity. With missing values in the other 20 studies, this variable could not be integrated in our moderation analyses.
17. These are mean values based on 18 observations with total sam- 4. The adjustments for unreliability can be applied directly to the ple size of 7,218 in the control group (7,218 / 18 = 401) and 1,015 in the inverse variance weights using w’ = w × (r ). In this approach, the effect experimental group (1,015 / 18 = 57). For repurchase intentions, 12 obser- size is corrected both for sampling and measurement error (cf. Lipsey vations had a total sample size of 6,686 in the control group (6,686 / 12 = 558) and 1,601 in the experimental group (1,601 / 12 = 134).
18. We also checked if there was a correlation between power and effect sizes of the individual studies. This correlation was not signifi- cant for either satisfaction or repurchase intentions.
19. Most of the studies did not test these moderators or provide information that could allow the authors to classify the study in one cat- 6. For most of the studies the total sample was greater than the sum egory or another. Because of this, we chose to tabulate only information of the treatment and the control group. This was because each study tested relationships other than the paradox (e.g., a given study used a 2 × 2 × 2 experiment with a control group, and the paradox was testedcomparing one of the eight experimental groups with the controlgroup). In this case, we coded three variables: (a) total sample size from REFERENCES
the nine groups, (b) size of treatment group, and (c) size of the control *Asterisks denote studies used in meta-analysis.
7. It is expected that measures with more items produce greater reli- ability (Nunnally and Bernstein 1994) and then stronger effect sizes.
Achrol, Ravi S. (1991), “Evolution of the Marketing Organization: New 8. Seven out of the 24 studies used single items for any of the mea- Forms for Turbulent Environments,” Journal of Marketing, 55 (4), sured variables; a total of nine observations of the 45 values in our data set.
9. There were only four missing values of reliability for studies using multiple items in dependent variables.
Andreassen, Tor Wallin (2000), “Antecedents to Satisfaction with Service 10. The cumulated effect size for satisfaction would become –.135 Recovery,” European Journal of Marketing, 34 (1/2), 156-75.
if Oh (2003) were included in the data set. The observed effect size for *——— (2001), “From Disgust to Delight—Do Customers Hold a Grudge?” Journal of Service Research, 4 (1), 39-49.
11. The formula for the fail-safe number is k × (r – r ) / r , where k Assmus, Gert, John U. Farley, and Donald R. Lehmann (1984), “How is the accumulated number of studies, r is the mean effect size, and r is Advertising Affects Sales: Meta-Analysis of Econometric Results,” the critical effect size, or the “just significant” level (Hunter and Journal of Marketing Research, 21 (1), 65-74.
Schmidt 2004, p. 501; Lipsey and Wilson 2001, p.166).
Bitner, Mary Jo, Bernard H. Booms, and Mary S. Tetreault (1990), “The 12. We asked experts in meta-analysis what would be the minimum Service Encounter: Diagnosing Favorable and Unfavorable number of studies needed to conduct a meta-analysis. Professor Frank Incidents,” Journal of Marketing, 54 (1), 71-84.
L. Schmidt (author of the book Methods of Meta-Analysis; see Hunter ———, ———, and L. A. Mohr (1994), “Critical Service Encounters: and Schmidt 2004) told us, “The really bare minimum is 2 studies, but The Employee’s Viewpoint,” Journal of Marketing, 58 (3), 95-106.
JOURNAL OF SERVICE RESEARCH / August 2007 Blodgett, Jeffrey G., Donald H. Granbois, and Rockney G. Walters Hedges, Larry V. and Ingram Olkin (1985), Statistical Methods for (1993), “The Effects of Perceived Justice on Complainants’ Meta-Analysis. Orlando, FL: Academic Press.
Negative Word-of-Mouth Behavior and Repatronage Intentions,” *Hocutt, Mary Ann, Goutam Chakraborty, and John Mowen (1997), Journal of Retailing, 69 (4), 399-428.
“The Impact of Perceived Justice on Customer Satisfaction and Bolton, Ruth N. and James H. Drew (1991), “A Multistage Model of Intention to Complain in a Service Recovery,” Advances in Customers’ Assessments of Service Quality and Value,” Journal of Consumer Research, 24, 457-63.
Consumer Research, 17 (March), 375-84.
*——— and Thomas H. Stone (1998), “The Impact of Employee *Boshoff, Christo (1997), “An Experimental Study of Service Recovery Empowerment on the Quality of a Service Recovery Effort,” Options,” International Journal of Service Industry Management, 8 Journal of Quality Management, 3 (1), 117-32.
*———, Michael R. Bowers, and D. Todd Donavan (2006), “The Art Brown, Steven and Douglas M. Stayman (1992), “Antecedents and of Service Recovery: Fact or Fiction?” Journal of Services Consequences of Attitude toward the Ad: A Meta-Analysis,” Journal of Consumer Research, 19 (1), 34-51.
Hunter, John E. and Frank L. Schmidt (2004), Methods of Meta- Chandrashekaran, Murali, Kristin Rotte, Stephen S. Tax, and Rajdeep Analysis: Correcting Error and Bias in Research Findings, 2nd ed.
Grewal (2007), “Satisfaction Strength and Customer Loyalty,” Journal of Marketing Research, 44 (1), 153-63.
Janiszewski, Chris, Hayden Noel, and Alan G. Sawyer (2003), “A Cohen, Jacob (1988), Statistical Power Analysis for the Behavioral Meta-Analysis of the Spacing Effect in Verbal Learning: Sciences, 2nd ed. Hillsdale, NJ: Lawrence Erlbaum.
Implications for Research on Advertising Repetition and Consumer Colgate, Mark, Vicky T. U. Tong, Christina K. C. Lee, and John U.
Memory,” Journal of Consumer Researcher, 30 (1), 138-49.
Farley (2007), “Back from the Brink: Why Customers Stay,” Jones, Thomas O. and W. Earl Sasser, Jr. (1995), “Why Satisfied Journal of Service Research, 9 (3), 211-28.
Customers Defect,” Harvard Business Review, 73 (November- DelVecchio, Devon, David H. Henard, and Traci H. Freling (2006), “The Effect of Sales Promotion on Post-Promotion Brand *Kau, Ah-Keng and Elizabeth Wan-Yium Loh (2006), “The Effects of Preference: A Meta-Analysis,” Journal of Retailing, 82 (3), 203-13.
Service Recovery on Consumer Satisfaction: A Comparison Eisend, Martin (2004), “Is It Still Worth to be Credible? A Meta- between Complainants and Non-Complainants,” Journal of Analysis of Temporal Patterns of Source Credibility Effects in Services Marketing, 20 (2), 101-11.
Marketing,” in Advances in Consumer Research, Vol. 31, Barbara Kelley, Scott W., K. Douglas Hoffman, and Mark A. Davis (1993), “A Kahn and Mary Frances Luce, eds. Urbana, IL: Association of Typology of Retail Failures and Recoveries,” Journal of Retailing, “Two-Sided Advertising: A Meta-Analysis,” *Kwortnik, Robert J. (2006), “Shining Examples of Service when the International Journal of Research in Marketing, 23 (2), 187-98.
Lights Went Out: Hotel Employees and Service Recovery during Etzel, Michael J. and Bernard I. Silverman (1981), “A Managerial the Blackout of 2003,” Journal of Hospitality and Leisure Perspective on Directions for Retail Customer Dissatisfaction Research,” Journal of Retailing, 57 (3), 124-36.
Lipsey, Mark W. and David B. Wilson (2001), Practical Meta-Analysis.
Farley, John U., Donald R. Lehmann, and Alan Sawyer (1995), “Empirical Marketing Generalization Using Meta-Analysis,” *Magnini, Vincent P., John B. Ford, Edward P. Markowski, and Earl D.
Marketing Science, 14 (3), 36-46.
Honeycutt (2007), “The Service Recovery Paradox: Justifiable Faul, Franz, Edgar Erdfelder, Albert-Georg Lang, and Axel Buchner (in Theory or Smoldering Myth?” Journal of Services Marketing, press), “G*Power 3: A Flexible Statistical Power Analysis Program for the Social, Behavioral, and Biomedical Sciences,” Behavior *Mattila, Anna S. (1999), “An Examination of Factors Affecting Research Methods, http://www.psycho.uni-duesseldorf.de/abteil Service Recovery in a Restaurant Setting,” Journal of Hospitality & Tourism Research, 23 (3), 284-98.
Folkes, Valerie (1984), “Consumer Reactions to Product Failure: An *Maxham, James G., III (2001), “Service Recovery’s Influence on Attributional Approach,” Journal of Consumer Research, 10 Consumer Satisfaction, Positive Word-of-Mouth, and Purchase Intentions,” Journal of Business Research, 54 (1), 11-24.
——— (1988), “Recent Attribution Research in Consumer Behavior: A *——— and Richard G. Netemeyer (2002), “A Longitudinal Study Review and New Directions,” Journal of Consumer Research, 14 of Complaining Customer’s Evaluations of Multiple Service Failures and Recovery Efforts,” Journal of Marketing, 66 (4),57-71.
Forbes, Lukas P., Scott W. Kelley, and K. Douglas Hoffman (2006), “Typologies of E-Commerce Retail Failures and Recovery *McCollough, Michael A. (1995), “The Recovery Paradox: A Strategies,” Journal of Services Marketing, 19 (5), 280-92.
Conceptual Model and Empirical Investigation of CustomerSatisfaction and Service Quality Attitudes after Service Failure and Franke, George R. and Jeong-Eun Park (2006), “Salesperson Adaptive Recovery,” doctoral dissertation, Texas A&M University, College Selling Behavior and Customer Orientation: A Meta-Analysis,” Journal of Marketing Research, 43 (4), 693-702.
*——— (2000), “The Effect of Perceived Justice and Attributions Geyskens, Inge, Jan-Benedict E. M. Steenkamp, and Nirmalya Kumar regarding Service Failure and Recovery on Post-Recovery (1998), “Generalizations about Trust in Marketing Channel Customer Satisfaction and Service Quality Attitudes,” Journal of Relationships Using Meta-Analysis,” International Journal of Hospitality & Tourism Research, 24 (4), 423-47.
Research in Marketing, 15 (3), 223-49.
——— and Sundar G. Bharadwaj (1992), “The Recovery Paradox: An Grewal, Dhruv, Sukumar Kavanoor, Edward F. Fern, Carolyn Costley, Examination of Consumer Satisfaction in Relation to and James Barnes (1997), “Comparative versus Noncomparative Disconfirmation, Service Quality, and Attribution Based Theories,” Advertising: A Meta-Analysis,” Journal of Marketing, 61 (4), 1-15.
in Marketing Theory and Applications, Chris T. Allen et al., eds.
*Halstead, Diane and Thomas J. Page (1992), “The Effects of Chicago: American Marketing Association, 119.
Satisfaction and Complaining Behavior on Consumer Repurchase *———, Leonard L. Berry, and Manjit S. Yadav (2000), “An Empirical Intentions,” Journal of Consumer Satisfaction, Dissatisfaction and Investigation of Customer Satisfaction after Service Failure and Complaining Behavior, 5, 1-11.
Recovery,” Journal of Service Research, 3 (2), 121-37.
Hart, Christopher W., James L. Heskett, and W. Earl Sasser, Jr. (1990), *Michel, Stefan (2001), “Analyzing Service Failures and Recoveries: A “The Profitable Art of Service Recovery,” Harvard Business Process Approach,” International Journal of Service Industry *——— and Matthew L. Meuter (2006), “The Service Recovery Tax, Stephen S., Stephen W. Brown, and Murali Chandrashekaran Paradox: True but Overrated?” working paper, www.dienstleis (1998), “Customer Evaluations of Experiences: Implications for tungsmarketing.ch/documents/IJSIMParadoxMichelMeuter.pdf.
Relationship Marketing,” Journal of Marketing, 62 (2), 60-76.
Moorman, Christine, Gerald Zaltman, and Rohit Deshpande (1992), Weun, Seungoog, Sharon E. Beatty, and Michael A. Jones (2004), “The “Relationships between Providers and Users of Market Research: Impact of Service Failure Severity on Service Recovery Evaluations The Dynamics of Trust within and between Organizations,” Journal and Post-Recovery Relationships,” Journal of Services Marketing, of Marketing Research, 29 (3), 314-39.
Morgan, Robert and Shelby Hunt (1994), “The Commitment-Trust Westbrook, Robert A. (1980), “Intrapersonal Affective Influences upon Theory of Marketing Relationships,” Journal of Marketing, Consumer Satisfaction with Products,” Journal of Consumer Muncer, Steven, Mark Craigie, and Joni Holmes (2003), “Meta- *Zeithaml, Valerie A., Leonard L. Berry, and A. Parasuraman (1996), Analysis and Power: Some Suggestions for the Use of Power in “The Behavioral Consequences of Service Quality,” Journal of Research Synthesis,” Understanding Statistics, 21 (1), 1-12.
Nunnally, Jum and Ira Bernstein (1994), Psychometric Theory. New Celso Augusto de Matos is a marketing doctoral candidate at the
*Oh, Haemoon (2003), “Reexamining Recovery Paradox Effects and School of Management, Federal University of Rio Grande do Sul Impact Ranges of Service Failure and Recovery,” Journal of (PPGA-EA-UFRGS), Brazil. His main research interests lie in the Hospitality & Tourism Research, 27 (4), 402-18.
areas of consumer behavior in services, attitude formation and *Ok, Chihyung, Ki-Joon Back, and Carol W. Shanklin (2006), “Service Recovery Paradox: Implications from an Experimental Study in a change, and marketing research. His research has been published Restaurant Setting,” Journal of Hospitality & Leisure Marketing, 14 in the Journal of Consumer Marketing, International Journal of Consumer Studies, ACR Conference, and in a number of Brazilian Oliver, Richard L. (1997), Satisfaction: A Behavioral Perspective on the journals and proceedings. He can be contacted at: Consumer. New York: McGraw-Hill.
Palmatier, Robert W., Rajiv P. Dant, Dhruv Grewal, and Kenneth R.
Evans (2006), “Factors Influencing the Effectiveness ofRelationship Marketing: A Meta-Analysis,” Journal of Marketing, Jorge Luiz Henrique is a marketing doctoral candidate at the
School of Management, Federal University of Rio Grande do Sul Pan, Yue and George M. Zinkhan (2006), “Determinants of Retail (PPGA-EA-UFRGS), Brazil. His research focuses on consumer Patronage: A Meta-Analytical Perspective,” Journal of Retailing, 82 behavior in services and relationship marketing. He is a marketing manager at Banco do Brasil (Bank of Brazil). His research has been Parasuraman, A. (2007), “Modeling Opportunities in Service Recovery published in the Journal of Internet Banking and Commerce, and Customer-Managed Interactions,” Marketing Science, 25 (6),590-93.
Global Information Technology Management (GITM), The Reichheld, Frederick F. (1994), “Loyalty and the Renaissance of Business Association of Latin American Studies (BALAS), Marketing,” Marketing Management, 2 (4), 10-21.
International Association for Management of Technology (IAMOT) Rosenthal, Robert (1979), “The ‘File Drawer Problem’ and Tolerance conferences, and in a number of Brazilian journals and proceedings.
for Null Results,” Psychological Bulletin, 86, 638-41.
Singh, J. (1990), “Voice, Exit and Negative Word-of-Mouth Behaviors: An Investigation across Three Service Categories,” Journal of the Carlos Alberto Vargas Rossi is a professor of marketing at the
Academy of Marketing Science, 18 (1), 1-15.
School of Management, Federal University of Rio Grande do Sul *Smith, Amy K. and Ruth N. Bolton (1998), “An Experimental (PPGA-EA-UFRGS), Brazil. His research interests are consumer Investigation of Service Failure and Recovery: Paradox or Peril?” behavior and marketing theory. His research has been published Journal of Service Research, 1 (1), 65-81.
in the Journal of Consumer Marketing, International Journal of Szymanski, David M. and David H. Henard (2001), “Customer Satisfaction: A Meta-Analysis of the Empirical Evidence,” Journal Consumer Studies, ACR, AMA and EMAC conferences, and in a of the Academy of Marketing Science, 29 (1), 16-35.
number of Brazilian journals and proceedings.

Source: http://cavrossi.com.br/arquivo/artigo/arquivos/8.pdf