COMPARISON OF RELATIVE IMPORTANCE OF SERVICE RECOVERY
ATTRIBUTES IN AN ONLINE SELF-SERVICE TECHNOLOGY CONTEXT
AND OFFLINE SERVICE RECOVERY
A Conjoint Approach
By
JEROEN HILLEMANS
University of Groningen
Faculty of Economics and Business
Msc Marketing Management & Marketing Research
1
ABSTRACT
Service recovery is considered an important part of building and improving a company’s relationship with their customers. Based on previous literature, this research presents an empirical analysis of the role several of the most prevalent service recovery attributes play in an online self-service technology context and compares these results to a traditional setting of service recovery. Furthermore, several moderating variables are included in the analysis to determine if segments can be identified. This study is split up in three projects and three separate databases with data collected in two different countries (the Netherlands & the United States) were used for the analysis. Data for the first two studies was collected through individually completed questionnaires that contained two experimental conditions (severity of the service failure and relationship quality) as well as a conjoint task. The last study was designed to determine whether or not evaluations of service recovery changed based on the context (online/offline). Using this data, a traditional conjoint analysis (both aggregate and segmented analyses) was performed in order to find out which attributes are the most important. Overall results show that there are no significant differences regarding the
importance of service recovery attributes between online- and offline service recovery (study 3). However, the relative importance of these attributes does change when there is a chance customers are harmed by a service failure (e.g. the service recovery does not return them to their starting point). In the case customers are at least returned to their starting point,
timeliness is the most important, followed by attentiveness and credibility (study 1). However, if this is not the case, redress becomes the most important attribute by far, followed by
timeliness (study 2 & 3). Managerial implications of this research are directed at e-tailers who can use the segments that were found to determine an optimal (online) service recovery. Furthermore, this research provides a basis for other researchers to further study (online) service recoveries and the importance of recovery attributes while also stipulating the importance of considering the power of redress when customers are harmed during a service failure.
Keywords: Online service recovery, service recovery, self-service technology, recovery
2
1.Introduction
In today’s service-focused society it is becoming increasingly important for service organizations to handle service failures in a way that is satisfactory for their customers. These service recoveries are actions a company takes in response to a service failure (Gronroos, 1988). Service failures and failed recoveries are among the most important reasons for customers to switch service providers (Keaveney, 1995; Smith, Bolton & Wagner, 1999). Furthermore, research shows that 48% of consumers tell ten or more people when they experience bad service as opposed to only 23% when the service was satisfactory (Dixon, Freeman & Toman, 2010). This means that quality service recoveries are a crucial element in building relationships with customers and preventing customer defection and can even be seen as a substitute for offensive marketing techniques such as advertising or price reduction (Fornell & Wernerfelt, 1987).
According to Berry and Parasuraman (1991) only 50% of customers that experience a service failure are satisfied with the recovery provided by the company. This indicates that even though research shows that service recovery can be a very profitable endeavor (Hart, Heskett & Sasser, 1990), (e-tail) companies do not always pay enough attention to it yet. Furthermore, the number of customers that are satisfied with the recovery provided by the company is even lower in the case of self-service technologies (Forbes, 2008). A self-service technology (SST) can be described as a technological interface that allows consumers to produce a good or service themselves, without any direct contact with the service provider (Meuter, Ostrom, Roundtree & Bitner, 2000). Examples of SST’s include internet purchasing, ATM machines, ticket machines and telephone banking (Forbes, 2008).
3 Taking these previous arguments into account, it is also surprising there are so few large-scale studies that have focused on service recovery in a SST context. Despite the staggering growth in online spending and online SST’s (Forbes, Kelley & Hoffman, 2005), research on service recovery issues has mostly been limited to traditional service encounters, which are service situations in which the consumer and the service provider interact with each other face-to-face. For an overview of research on traditional service recovery, see Orsingher, Valentini & De Angelis (2010) or Gelbrich & Roschk (2011). Failure and recovery issues regarding SST’s that were already considered in previous research include: Typologies of failure and recovery strategies in a SST context (Forbes et al., 2005; Forbes,2008), investigation of satisfying and dissatisfying incidents in SST’s (Meuter et al., 2000), outcomes associated with recovery strategies in online SST’s (Fan, Wu & Wu, 2010; Sousa & Voss, 2009; Wang, Lin & Phongkusolchit, 2010) and factors influencing the choice for a particular SST (Gelderman, Gheisen & van Diemen, 2011). However, even though offline service recovery has been studied extensively, the relative importance of different attributes has not been covered yet in offline- or online service recovery.
In order to gain more insight in the importance of service recovery attributes, this study will focus on both offline and online service recovery, which leads to the following main research question that will be covered in this paper:
‘Which attributes of service recovery are the most important when considering satisfaction with the service recovery in an online self-service technology context when compared to offline service recovery?’
This research question has been split up into several sub-questions focused on both offline and online service recovery which will help answer the main question. These are:
1. Which are the most prevalent service recovery attributes according to literature? 2. What are the differences between online and offline service recovery and how
important is each attribute in the online context when considered relative to the offline context?
3. Which factors have a moderating effect on the relationship between the recovery attributes and satisfaction with the service recovery?
4 Considering the arguments presented in the previous section, this study’s main contribution to the existing literature is to provide insight in the importance of service recovery attributes. In order for marketers to take full advantage of the opportunities online SST’s provide, it is important to gain more knowledge about the relative importance of service recovery attributes in this specific context, which will also allow practitioners to take steps helping them to provide customers with the ‘ideal’ service recovery. This study will also look at several relevant moderators, to see whether or not these have an impact on the relative importance of the recovery attributes.
Currently, there is a lack of understanding of how customers evaluate and react to recovery efforts by online firms and what is considered a successful recovery in this context (Holloway & Beatty, 2003; McCollough, Berry & Yadav, 2000). Also, Sousa & Voss (2009) argue that it is important to understand what attributes differentiate recovery that leads to high satisfaction from recovery that leads to low satisfaction. As was mentioned previously, only a handful of papers have investigated the impact of different aspects of service recovery on postcomplaint customer behavior (e.g. Smith et al., 1999; Tax, Brown & Chandrashekaran, 1998) and none of these focus on the importance of these attributes in an online context, or offline context for that matter. Being able to quantify the costs and benefits of service recovery actions will have a large impact on customer loyalty and customer retention theory (Davidow, 2003). Therefore, the purpose of this research is to fill this gap in current literature by finding out which of the service recovery attributes are the most important to consumers by using a conjoint approach. This means that, on the one hand, this paper will provide a theoretical basis for service recovery in an online SST context (but also considering the offline side), while on the other hand, it also provides useful practical knowledge for e-tail marketing managers.
5 the form of theoretical- and practical implications, the limitations of the study and suggestions for further research will be given.
2. Theory
This section will deal with relevant previous literature in the field of service failure and recovery, which includes previous research on traditional (offline) service recovery and research in a SST context. Differences in service recovery between off- and online services will be discussed and the relevant concepts will be defined for the purpose of this research. Finally, the service recovery attributes found in literature will be discussed as well as the moderating variables (see figure 1 below for the conceptual framework).
2.1 Service failure and recovery
6 loss’. This provides a clear explanation of the general dynamic of a service recovery encounter and is applicable even in the context of this paper.
As was described before, this paper focuses on the differences in service recovery between offline- and online service recovery. Meuter et al. (2000) define self-service technologies as a technological interface that allows consumers to produce a good or service themselves, without any direct contact with the service provider. Therefore, examples of online SST’s include online banking and web shops. Due to the impersonal nature of online SST’s there are several differences between offline and online service recovery. These differences will be discussed later on, but first some theoretical background of offline service recovery will be outlined in the next section.
2.2 Offline service recovery
7 Another interesting theoretical notion in traditional service recovery research is the service recovery paradox. This means that customers who experience a failure followed by a high quality recovery action are prone to have a higher (or at least the same) satisfaction level than if no failure would have occurred (McCollough et al., 2000). Although some studies have found results confirming this recovery paradox (e.g. Bitner et al., 1990; Kelley, Hoffman & Davis, 1993), others have found opposite results (e.g. Berry, Zeithaml & Parasuraman, 1990; Fornell, 1992). However, regardless of the amount of research either confirming or disconfirming the recovery paradox, a theoretical explanation for this phenomenon has not been provided yet. McCollough et al. (2000) were one of the few that directly compared post recovery satisfaction of consumers that experienced a failure or error-free service, while also controlling for possible confounds affecting satisfaction ratings. However, they also found a lack of support for the recovery paradox effect. They argue that the effect can only be present when full or near-full service recovery is a possibility. Although the service recovery paradox will not be dealt with directly in this paper, it is something to keep in mind when interpreting the results.
2.3 Differences between online and offline service recovery
Recent developments in the field of services has lead to a stronger emphasis on online services and by extension service recovery (Sousa & Voss, 2009; Wang et al., 2010). Previous literature found that the Internet does not fundamentally change traditional marketing principles (Barwise, Elberse & Hammond, 2002) which means that insights gathered from offline service recovery research are also relevant in an online context (Holloway & Beatty, 2003). However, there are still several prominent differences that are important for this paper and will therefore be discussed in this section.
8 Second, service failures in an online environment are relatively objective and obvious (Meuter et al., 2000) when it concerns a technology failure, since you can immediately see it on your computer screen when something goes wrong which leads to a greater transparency concerning these kinds of failures online (Forbes et al., 2005). Other types of failures might take longer to identify (e.g. something is ordered online, but the wrong product is sent). However, the transparency associated with online technology failures means customers have an easier time objectively identifying this type of service failure and because the consumer has the ability to exit the relationship with a click of the mouse (Holloway & Beatty, 2003), minimizing and effectively recovering from service failures is crucial to online retailers (Wang et al.,2010).
A third important difference is the difference in perceived severity of service failures online compared to traditional ones. Since many customers are relatively new to online shopping, problems are more likely to occur on the Internet. Furthermore, Forbes et al. (2005) argue that consumers perceive online service failures as more severe than their offline counterparts. This can also be related to the absence of interpersonal contact that was mentioned before, since the presence of an employee cannot relieve the severity of the service failure in this setting (Forbes et al., 2005). Another reason for the perceived severity of these failures is that online failures tend to be either process or technology failures (Holloway & Beatty, 2003; Meuter et al., 2000) which are both considered to be severe failures from the customer’s point of view. A process failure causes large problems for consumers, because they are under the impression that the transaction was completed, while in reality, a process failure occurred that may not be discovered until later (Meuter et al., 2000). Technology failures are similarly critical because they prevent customers from using the technology (Meuter et al., 2000). Keeney (1999) states that these kind of failures are especially bad because consumers expect convenience and control in an online environment.
9 Finally, the circumstances surrounding an online service failure can be quite different from traditional settings. Holloway & Beatty (2003) mentioned credit card security, privacy, on-time delivery, ease of navigation and a general perceived insecurity of the Internet as circumstances that are significantly different from offline settings. This also means different types of failures will occur online. As was mentioned before, the most common failures include technology and process failures online and research by Holloway & Beatty (2003) shows that the most common types of failures include (1) delivery problems, (2) web site design problems, (3) payment problems, (4) security problems, (5) problems with product quality, and (6) problems with customer service. Of these categories, they found that delivery problems were the most common, which means products either arrived later than promised, a wrong item was delivered or the product was delivered to the wrong address. These findings can partly be explained by the fact that consumers cannot inspect the ordered products until they arrive at their home, which makes these kind of failures inherent to online SST’s (Forbes et al., 2005). The next section will deal with the different service recovery attributes that are relevant for online service recovery.
2.4 Service recovery attributes
10 These are redress, timeliness, credibility, apology, attentiveness and facilitation. This empirically tested model successfully differentiated between these attributes and satisfaction and is also the first paper that discussed more than three attributes at once, which makes it very suitable to use as the basis for a framework. Because there is a lack of consensus in this research field regarding the service recovery attributes, several of these are labeled differently across studies. However, the dimensions discussed by Davidow (2000) and reviewed by Davidow (2003) cover all attributes as discussed by Boshoff (1999), Smith et al. (1999), Wang et al. (2010) & Wirtz & Matilla (2004) and will therefore be used as the basis for this paper’s framework (figure 1).
The dependent variable in the model is satisfaction with the service recovery since this is the first outcome of the different service recovery attributes (Wirtz & Matilla, 2004). Note that often postcomplaint behavior such as word of mouth and repurchase intentions are used as outcomes of service recovery studies (Davidow, 2003; Orsingher et al., 2010). However, research shows that recovery satisfaction fully mediates the relationship between recovery attributes and behavioral intentions (Wirtz & Matilla, 2004). Before we proceed to discuss the service recovery attributes in detail in the next section, two important remarks have to be made about the model.
First, it is important to note that facilitation, which can be described as: ’the policies, procedures and structures a company has in place to support customers engaging in
complaints and communications (Davidow, 2003)’, will be excluded from this research since it is not an attribute of a service recovery, but relates more to the policy of the focal company. Since the goal of this paper is to find out what the ideal composition of an online service recovery is, policies and regulations cannot be taken into account even though previous research does indicate it has a positive effect on satisfaction (Boshoff & Leong, 1998).
11 recovery preferences and that it is important for service providers to match their recovery effort to the customer’s preference. This issue makes the framework revolving around service recovery attributes a better fit. Also, Davidow (2003) argues that the studies that used these justice categories might have mislabeled these dimensions since a consumer’s perception of justice is a subjective feeling and does not take into account the actual action initiated by the organization. Since the purpose of this article is finding out which of the service recovery attributes are the most important to consumers in an online context when compared to a traditional context, some of the justice concepts from previous studies have been re-categorized. This means, for example, that studies operationalizing procedural justice according to response speed are discussed in the timeliness section here. The impact of each service recovery attribute and the expected moderators according to previous literature are discussed next.
12
Redress
Redress is the most popular attribute of service recovery in research. The basic idea of redress is that the consumer must at least be returned to their starting point (before the service failure) or else they will still be dissatisfied (Davidow, 2003). Redress can be defined as: ‘the benefits or response outcome that a customer receives from the organization in response to the complaint’ (Davidow, 2003:232). Previous research showed that redress can be used to restore equity to a service relationship when a service failure has occurred (Walster, Berscheid & Walster, 1973). One of the few studies that covered service recovery attributes in an online setting are Wang et al. (2010). They found a positive effect of redress (magnitude of service recovery) on post-recovery satisfaction, loyalty and positive WOM. Redress is often cited as one of the most important service recovery attributes related to recovery satisfaction (Estelami, 2000; Davidow, 2003; Hocutt, Chakraborty & Mowen, 1997; McCollough, 2000) and previous research shows that redress is expected by customers when a failure occurs (Johnston & Fern, 1999). However, a study in the restaurant industry found that overspending on redress is not necessarily effective (Mack, Mueller, Crotts & Broderick, 2000). It is important to note that all these studies were focused on traditional services and their results might not translate completely to the online channel.
Timeliness
13
relevant (Davidow, 2003). Another angle was taken by Boshoff (1997) whom suggests timeliness is not a significant factor, unless the delay is especially long. This means that timeliness may not be the most important service recovery attribute, unless the delay is beyond expectations. However, these results might not apply to online SST’s as Wang et al. (2010) argue that especially for electronic service companies timeliness can be a crucial attribute due to the low switching costs for consumers online. Furthermore, as was mentioned before, failures in the online environment tend to be either process or technology failures (Holloway & Beatty, 2003; Meuter et al.,2000) which are both service failures of a nonmonetary nature. This suggests that timeliness might have a more important role online in comparison to traditional service settings because, as was mentioned above, timeliness is only a significant factor if there was no money involved (Gilly & Gelb, 1982). However, Smith et al. (1999) found that timeliness had a greater effect on satisfaction with the service recovery when an outcome failure occurs, which is basically the opposite of the previous argument. This indicates literature is conflicted on the role of timeliness and this empirical study should provide some indication of the importance of timeliness in online settings.
Credibility
14 found that redress without an explanation could be considered as an admission of guilt and would therefore reduce satisfaction.
Apology
According to Davidow (2000) an apology can be thought of as a psychological compensation for the consumer. It is formally defined as: ‘an acknowledgement by the organization of the complainant’s distress’ (Davidow, 2003:232) and usually is an attempt to inform the consumer that the company accepts responsibility for the failure and to express regret for the problem that has occurred (Conlon & Murray, 1996). An apology communicates politeness, courtesy, concern, effort and empathy to consumers that are affected by a service failure (Hart et al., 1990; Kelley, Hoffman & Davis, 1995). It is important to note that only studies that considered apology a separate attribute instead of being part of the redress dimension will be considered here.
Smith et al. (1999) found that apology had an indirect positive effect on satisfaction with the service recovery (through interactional justice). However, results on the apology dimension are mixed, especially when looking at the effect of apology on other outcomes, such as repurchase intentions and customer satisfaction. Therefore, the apology dimension might be the most useful when used in conjunction with the other attributes. An example of this is the reported interaction effect between redress and apology. For example, Boshoff (1997) found that an apology without redress was not significant, while Goodwin & Ross (1992) argued that an apology affected satisfaction more when redress was involved.
Attentiveness
15 the service recovery. McCollough (2000) complemented those findings by showing that effort and courtesy both significantly influence complaint handling satisfaction.
2.5 Moderating variables
The strength of the relationships that are depicted in the conceptual model (figure 1) may differ due to moderating influences of other variables. In this study several of these will be taken into account and this section will deal with previous literature on these variables. First of all, the cumulative online purchasing experience of consumers will be taken into account, which can be described as ‘the total purchase frequency and volume for an individual consumer across all previous online exchange transactions’ (Holloway, Wang & Parish, 2005:55). Previous research shows that as consumers gain more experience with the online channel, their attitudes and behavior (Holloway et al., 2005) as well as their expectations of service quality will change (Cadotte, Woodruff & Jenkins, 1987). Parasuraman, Zeithaml & Berry (1985) found that when consumers do not have a lot of experience, their expectations are usually focused on what they feel a service provider should do for them, instead of having realistic expectations based on experience. Holloway et al. (2005) studied this phenomenon in the context of online purchasing experience. They found that more experienced online users have more realistic expectations of the potential for service failures and were less dissatisfied when a service failure occurred, while the same service failure had a more severe impact for the less experienced purchasers. Therefore, previous online purchasing experience acts as a buffer for service failures (Holloway et al., 2005), which indicates that the strength of the relationship between recovery attributes and satisfaction with the service recovery might vary when a user is more or less experienced.
16 Third, several demographics are taken into account in the model, because they might have an influence on the evaluation of service recoveries. The most important of these might be age, since there are significant differences in online buying behavior when comparing between age categories (Leppel & McClosky, 2011; Sorce, Perotti & Wildrick, 2005). However, concerning buying behavior, there have been conflicting results in previous literature. Donthu & Garcia (1999) found that older internet users were more likely to buy products online when compared to younger users, even though younger users had a more positive attitude towards online shopping, while Joines, Scherer & Scheufele (2003) found that younger users are more likely to buy online. Sorce et al. (2005) added that younger consumers did not buy more online, but do search for more products than older consumers. Furthermore, older consumers had a higher likelihood of being frustrated when searching for specific product information on the internet when compared to younger users (Leppel & McClosky, 2011). Since elderly people are more easily frustrated when buying online, it seems likely that they value attributes like attentiveness more than younger people. The idea behind this is that they are already frustrated easily when searching for products online, but if a service failure occurs, they will need additional reassurance when compared to younger people. These differences might moderate the relationship between the recovery attributes and satisfaction with the service recovery. Furthermore, in an offline context older people are more loyal when dealing with service providers, compared to younger people (Cambra-Fierro, Berbel-Pineda, Ruiz-Benitez & Carrasco, 2011). Also, Varela-Neira, Vazquez-Cassieles & Iglesias (2010) argue that customer age is negatively related to the intensity of negative emotions experienced due to a service failure. This indicates that it is probable that older people will be more flexible in dealing with the service failure. Both of these arguments indicate that demographics is also a relevant moderator for the offline service recovery setting.
17 Finally, relationship quality will be taken into account. There are two streams concerning relationship quality and its effect regarding service recovery. On the one hand, several researchers have found that a high relationship quality acts as some sort of buffer when a failure occurs, which means customers are more likely to forgive the company for a service failure (Singh & Sirdeshmukh, 2000; Tax et al., 1998). On the other hand, it was also found that a high relationship quality may enhance a customer’s reaction following a service failure because they feel betrayed (Holloway, Wang & Beatty, 2009; Kelley & Davis, 1994). Furthermore, high relationship quality customers will respond less strongly to an organization’s recovery efforts (Holloway et al., 2009). These findings suggest that the strength of the relationship between the attributes and satisfaction with the recovery might vary due to relationship quality, indicating a moderating influence. Finally, an overview of which moderating variables will be taken into account in what study can be found in table 1. As can be seen, all variables will be taken into account in study 1. However, online cumulative purchasing experience and technology readiness are not relevant in the offline setting.
TABLE 1.
Moderating variables per study
Moderating variable: Online (study 1) Offline (study 2)
Online cumulative purchasing experience
x
Technology readiness x
Demographics x x
Severity of the failure x x
Relationship quality x x
3.Methodology
18 differences between the studies will be discussed in a separate section. After these two studies were performed, it became clear a third study was necessary to distinguish between the different effects that were found. This study will be elaborated on in the results section to avoid confusion between the three studies.
3.1 Traditional conjoint analysis
This research employs a traditional conjoint analysis (TCA) in order to determine which of the service recovery attributes are the most important in an online SST setting. TCA is appropriate in this context for several reasons. First, TCA matches the goals of this study because it assesses how consumers balance benefits that are associated with product or service attributes (e.g. Wittink & Cattin, 1989; Wittink, Vriens & Burhenne, 1994). Second, the downsides associated with the technique are not relevant given the research design used in this study since people do not ‘buy’ an online service recovery. Therefore, it is fine that TCA only covers the preference formation stage of the decision process and is unable to distinguish between buyers and non-buyers (Backhaus, Hillig & Wilken, 2007). Finally, the dependent variable in this study is ‘satisfaction with the service recovery’ and TCA allows us to use a metric rating scale (Hair, Black, Babin & Anderson, 2006). Furthermore, the limited number of attributes used in this study means TCA is appropriate because it allows for a maximum of nine attributes (Hair et al., 2006).
3.2 Data collection procedure
Empirical data for study was gathered by using a sample of Dutch consumers and specifically consumers that have bought products online. In order to attain a sufficiently heterogeneous sample the snowball sampling procedure was used through social media (Facebook and LinkedIn). Data was collected through individually completed questionnaires filled out online.
3.3 Measurement scales
19 covers both frequency and quantity of online shopping. Furthermore, technology readiness was measured by a measurement scale called the ‘technology readiness index’ developed by Parasuraman (2000). Pre-testing of the scenarios lead to the following levels of each condition. Low severity means the product cannot be delivered in the scheduled delivery period while high severity means the web shop delivered the wrong product to the consumer. The scenarios for relationship quality are based on a measurement scale by Holloway et al. (2009). Low relationship quality indicates the respondent is unfamiliar with the web shop and does not know yet if it can be trusted or is honest while high relationship quality means the consumer feels like they are a loyal customer and the organization can be trusted.
3.4 Experimental design
A 2x2 between subjects factorial design was used to manipulate the effects of severity of the failure (low versus high) and relationship quality (low versus high). Each subject was randomly exposed to one of these four failure scenarios and within these the respondents had to complete a conjoint task of fifteen profiles within which they had to rate their satisfaction with the depicted service recovery on a ten point scale. Regardless of the scenario respondents were exposed to, each of them completed the same conjoint task. The scenarios can be found in appendix 2.
Scenarios were chosen for several reasons. First, it eliminates problems with observations or manipulation of actual service failure/recovery, which includes expenses, ethical considerations and unwillingness to intentionally expose customers to service failures (Smith et al., 1999). Second, they also argue that the use of scenarios minimizes memory bias as well as rationalization tendencies which is common in self-reports of service failure (Smith et al., 1999). Third, it reduces problems with individual differences in response styles and personal circumstances since respondents all face the same situation more or less (Wirtz & Bateson, 1999).
20 to ensure sufficient discrimination between them, but not by so much that it becomes unbelievable. The levels of the attributes were either based on literature or common sense. For example, the levels for credibility, apology and attentiveness indicated the attribute was either present or not present. The descriptions of these attributes were based on the definitions provided by Davidow (2003) discussed previously. Redress meant the complaint would be resolved for all customers but the ‘high’ level of the attribute would add some extra benefit as compensation. Pre-testing showed a discount of 50% was appropriate. Timeliness was also determined by pre-testing which showed that recovery within a week was considered desirable for online services while a month was outside the norm, but not unbelievable.
Also, each attribute has a linear preference function since they each only have two levels of which the higher level should reasonably have a higher utility. Because of this, the loss of accuracy is minimal while maintaining statistical efficiency using a linear preference function (Hair et al., 2006). Table 2 describes the levels of each attribute in more detail.
TABLE 2.
Recovery attributes and their levels Attribute: Levels:
Redress High: The organization provides you with compensation that covers the expenses
caused by the service failure and a discount coupon for 50% off your next purchase
Low: The organization provides you with compensation that covers the expenses
caused by the service failure
Timeliness High: Your complaint is resolved within the week Low: The resolution of your complaint takes a month
Credibility High: The organization takes responsibility for the failure and provides an explanation Low: The organization does not explicitly take responsibility for the failure and does
not provide an explanation
Apology Yes: The company apologizes for the failure No: The company does not apologize for the failure
Attentiveness High: The employee handling your complaint shows respect, empathy and a
willingness to listen and makes a real effort to make sure you are satisfied
Low: The employee handling your complaint does the minimum necessary to resolve
your complaint, but does not show empathy and a willingness to listen.
21 Furthermore, a fractional factorial design called a ‘Resolution V design’ was used to determine which sets of attributes would be shown to the respondent. This is a design in which no direct effect or factor interaction is aliased with other direct effects or two-factor interactions (Montgomery, 1997). What this means is that half of the possible profiles are eliminated while maintaining orthogonality ( because we have five attributes with two levels each). Since we have a limited amount of attributes with only two levels each the highest possible resolution can be used in this research design which allows for less restrictive assumptions regarding interactions that are negligible (Montgomery, 1997). In the end this led to sixteen profiles being included in the research design. Finally, one profile was excluded1 since all attributes had the lowest level which could lead to response bias (Hair et al., 2006), therefore the final research design contained fifteen recovery profiles.
3.5 Data analysis
As was argued before, TCA was used to analyze the importance of the service recovery attributes. Utilities of each attribute were found using a multiple regression approach using satisfaction with the service recovery as the dependent variable. Furthermore, latent class analysis was performed to see if heterogeneity was present in the estimated parameters. Segmenting the respondents based on the moderating variables proposed in our model means that companies can modify their behavior to the type of customer they are dealing with (Green & Krieger, 1991) if heterogeneity is indeed present. Finally, predictive validity was ensured by determining three hold-out sets at random. This means that twelve sets were used for the estimation of the model. The hold-out sets were used to check how accurate the estimated model predicts these hold-out sets (Hair et al., 2006).
3.6 Differences between Study 1 and Study 2
Although the studies are very similar methodologically, there are some important differences that are useful to discuss. This section will provide a short overview of these differences. First, data collection was done in a different country (the United States) through a survey website, namely ‘Mechanical Turk’. This website allows researchers to distribute surveys for a small fee.
1 Please note that the profile with the highest level for each attribute was already excluded due to the nature
22 Second, measurement instruments were similar across studies. The two experimental conditions (severity of the failure and relationship quality) were worded in the same way between the two studies. The only difference being the context of the scenario (online vs offline). For the scenarios in study 2, please refer to appendix 3. Next, technology readiness and online cumulative shopping experience were not included in the offline study, because they were not considered relevant in that context. Demographics were measured in the same way.
Third, an additional level for redress was added to the design used in study 2 after reviewing the preliminary results of study 1. In this study redress had two levels (see table 2 above) of which the lowest level returned the customer to their starting point before the starting failure. Davidow (2002) argued that complainers should at least be returned to their starting point before the dissatisfaction triggered by the service failure can be negated. Otherwise, they will still be dissatisfied with the response. However, due to the way the previous design was set up, it is possible there was not enough variance in these levels leading to a suppressed relative importance for an attribute that is often cited in literature as one of the most important attributes of service recovery. Therefore an additional level was added to the redress attribute, covering the state in which the consumer is not returned to their starting level.
23
4 Results Study 1 ‘online’
This chapter will deal with the results from the aggregate conjoint analysis (using regression) as well as the results from the latent class analysis. This was done in order to determine if there is any heterogeneity in the estimated parameters based on the moderating variables. However, this section will start with a description of the sample that was used, followed by a discussion on scale purification for technology readiness using multi-item scales. Finally, the results of the aggregate and segmented conjoint analyses as well as the validation of the model will be discussed.
4.1 Sample ‘online’
The online survey resulted in a total of 238 respondents of which 39 respondents had to be omitted due to a very high number of missing values (<10% of the questionnaire completed), which resulted in a final sample of 199 respondents. In order to obtain a reliable result for a conjoint analysis with a moderate amount of attributes (<9) a sample size of 200 is usually recommended (Hair et al., 2006). As can be seen in table 3, the sample consists of a range of age groups (although the age group of 20 through 30 dominates the sample with a share of 49.3%). Furthermore, 50.8% of the online shoppers were male while 49.2% were female. Moreover, the sample seems to be relatively active in the online shopping environment. The majority of the respondents report to purchase products online either weekly (9.5%), monthly (40.7%) or quarterly (33.7%) which shows that the sample is relatively active in the online shopping environment.
TABLE 3.
Characteristics of the sample (N=199)
Number of respondents Percentage Gender Male 101 50.8% Female 98 49.2% Age <20 years 58 29.1% 20 – 30 years 98 49.3% 31 – 40 years 10 5.0% 41-50 years 10 5.0% >50 years 23 11.6%
Frequency Online Purchase
Weekly 19 9.5%
Monthly 81 40.7%
24
Yearly 32 16.1%
Never 0 0.0%
Quantity Online Purchase
1-3 products 69 34.7%
4-7 products 59 29.6%
7-10 products 26 13.1%
10 or more products 45 22.6%
Average Spending per Purchase
Less than €20 33 16.6%
20-€60 119 59.8%
€61-€100 31 15.6%
More than €100 16 8.0%
Average Spending per 6 Months Less than €50 29 14.6% €50-€225 92 46.2% €226-€400 36 18.1% €401-€575 24 12.1% More than €575 18 9.0%
4.2 Measurement purification ‘online’
Before the results of the conjoint analysis can be discussed, it is important to purify the measurement scales that consist of multiple items (Hair et al., 2006). In the case of this study, this applies to the moderating variable technology readiness2. Therefore, reliability and validity of the measurement instrument that was used for the construct technology readiness will be assessed by using convergent validity, discriminant validity and face validity (Hair et al., 2006). As was mentioned before, the items used for technology readiness were previously validated and originate from the ‘technology readiness index’ developed by Parasaruman (2000). Furthermore, all questions in the scale used a 5-point Likert scale (1 = Completely disagree; 5 = Completely agree). The different types of validity will be discussed in this section in order to enable summated scales for technology readiness in which a number of items are transformed into a composite measure.
Convergent validity means that separate items within a construct should converge or, in other words, share a high percentage of variance (Hair et al., 2006). The measure most often used for this type of validity is Cronbach’s Alpha (Peterson, 1994), which determines the internal reliability of separate items within a construct. Although there is some debate on
2 The moderating variable ‘cumulative online shopping experience’ is also a multi-item scale, but since the
25 this topic in academic literature, a Cronbach’s Alpha of .7 is usually regarded as the minimum for a reliable scale (Peterson, 1994).
Before the Cronbach’s Alpha was calculated, some of the items were recoded in such a way that all questions were measured in the same direction. After determining the Cronbach’s Alpha for all items in the measurement scale of technology readiness, it appears the internal reliability of the entire scale is somewhat low (α = 0.653). This is especially relevant since it concerns a previously validated scale, which usually implies a higher value for Cronbach’s Alpha is required (Peterson, 1994).
Since the internal reliability of the entire scale is insufficient, a factor analysis with varimax rotation was performed in order to determine the underlying constructs. Varimax rotation was used since it is an orthogonal rotation method that maximizes the sum of variances of required loadings for the factor matrix and seems to give a clearer separation of factors when compared to other orthogonal rotation methods (Hair et al., 2006; Malhotra, 2010).
After examining the initial results of the factor analysis, it appeared some variables had communalities that were insufficient. These items were omitted from the analysis in a stepwise manner after which the analysis was performed again. The item ‘we should be careful with replacing important human tasks with technology, since technology can break down or cause mistakes’ was omitted from the analysis (communality = 0.375) after which all communalities were above .5.
The results show that Bartlett’s test of sphericity was significant (2
/df=4653.526/55, p = .000) which indicates there are at least some correlations between the variables included in the analysis (Hair et al., 2006). The Kaiser-Meyer-Olkin measure of sampling adequacy (MSA) is .666 which indicates the items are at least partly explained by the other variables. Furthermore, none of the individual MSA’s fall below .5 which means the sampling adequacy is sufficient (Hair et al.,2006). Table 4 shows the eigenvalues and extracted variance per factor.
TABLE 4.
Eigenvalues and extracted variance Factor Eigenvalue Cumulative variance extracted
1 2.511 22.8%
2 1.699 38.3%
3 1.446 51.4%
26
5 .901 70.0%
6 .739 76.7%
Table 4 indicates that a four factor solution is the most suitable since adding more factors means the added eigenvalue falls below one. The chosen solution also extracts the minimum of 60% variance (Hair et al.,2006) and the four factor solution also matches the factors that were found by Parasuraman (2000) which provides additional theoretical proof to choose a four factor solution. The initial solution showed some cross loadings which lead us to interpret the Varimax rotated solution instead. Table 5 shows the final factor solution as well as the Cronbach’s Alpha of the four factors and the value when an item would be deleted.
TABLE 5.
Factor Solution and Reliability
Items Factor Loadings Alpha If Item
Deleted
Cronbach Alpha For Scale
Factor Analysis Technology Readiness
Innovativeness (Factor 1): .790
‘Other people come to you for advice about new technology’ .880 .612 ‘In general, you are among the first in your circle of friends to
acquire new technology when it appears’
.840 .727 ‘You can usually figure out new high-tech products and services
without any help from others’
.761 .789
Optimism (Factor 2): .593
‘Technology gives you more freedom of mobility’ .774 .427 ‘Technology gives people more control over their daily lives’ .766 .455 ‘Products and services that use the newest technology are much
more convenient to use’
.585 .583
Security (Factor 3): .531
‘You do not consider it safe to give out credit card information over the internet’
.719 .424 ‘Any business transaction that you do electronically should be
confirmed later with something in writing’
.703 .486 ‘If you provide information to a machine or over the internet, you
can never be sure it really gets to the right place’
.696 .384
Comfort (Factor 4): .335
‘It is embarrassing when you have trouble with a high tech gadget while people are watching’
.773 -
‘Technology always seems to fail at the worst possible time’ .699 -
27 .79), which lead us to apply weighted factor scores instead of summated scales. This enhances internal reliability by using factor loadings as a weight factor in the calculation of the summated variables. Thus making sure the items with the highest correlation to the factor are also assigned the highest weight.
The four factors that were found are identical to the ones Parasuraman (2000) found in his study. They are innovativeness, optimism, security and comfort regarding technology. Please note that the original constructs were called insecurity and discomfort, but since we recoded these variables to make sure all items are measured in the same direction, they were renamed.
4.3 Aggregate conjoint analysis ‘online’
This section will discuss the aggregate conjoint analysis which will give an initial indication of the importance of each attribute for the entire sample. The multiple regression with ‘satisfaction with the service recovery’ as the dependent variable and the five recovery attributes that were discussed previously as independent variables showed an aggregate R² of .2106 which indicates 21% of the variance in the dependent variable is explained by the chosen attributes. Table 6 shows the results of the aggregate analysis, including the relative importance of each attribute and their significance. All five attributes significantly explain the consumer’s satisfaction with the service recovery (p<0.001). The results show that timeliness is the most important attribute (33.4%) followed by attentiveness (21.9%), while redress is the least important attribute (10.1%) in the case of an online service recovery. The model was re-run three times in which the maximum number of iterations was increased from the standard number of 250 up to 500 which was done to ensure the model was stable. The results showed minimal variation which indicates a stable model. Next, a segmented conjoint analysis will be performed in order to see if any differences in the parameters per segment can be detected.
TABLE 6.
Output aggregate conjoint analysis
28
4.4 Segmented conjoint analysis ‘online’
In order to check if there are differences in the evaluation of service recoveries based on several moderating variables (see the conceptual model in figure 1) a segmented conjoint analysis was run in LatentGold. This means the moderating variables were used as covariates in the analysis to find potential segments of consumers that have similar utilities (Hair et al., 2006; Malhotra, 2010). The questions regarding cumulative online shopping experience were recoded into dummy variables, which resulted in thirteen dummy variables (three variables per item and four for total expenditures, since there were four categories except for total expenditures which consisted of five categories). This was done because only continuous or nominal variables could be used in the analysis.
In this case 21 covariates were added to the model (age, gender, thirteen dummies for online shopping experience, four variables for technology readiness and a variable for relationship quality and severity of the failure) of which five were numeric and sixteen were nominal.
First, models ranging from one class to five classes were run in order to determine the most appropriate number of classes when considering the available data. The model fit likelihood ratio chi-squared statistic (L2) is significant for all five class models at p < .001, which indicates that the estimated frequencies are similar to the observed frequencies (Hair et al., 2006). In order to determine what model has the best fit, several information criteria of different strictness will be used (AIC, BIC, AIC3) and the results can be found in table 7.
The model with the lowest information criteria should be chosen since lower criteria indicate a better fit (Hair et al., 2006). However, even though all three criteria weight fit and parsimony of the model by adjusting the Log Likelihood (LL) to adjust for the amount of
TABLE 7 Model selection
Model LL BIC(LL) AIC(LL) AIC3(LL df P-value R²
1 class -4568.5 9211.0 9164.9 9178.9 185 .000 0.2106
2 class -4230.0 8724.7 8560.0 8610.0 149 .000 0.4594
3 class -4064.1 8583.4 8300.1 8386.1 113 .000 0.5096
4 class -3951.8 8549.5 8147.7 8269.7 77 .000 0.5260
5 class -3872.0 8580.4 8060.0 8218.0 41 .000 0.5613
29 parameters in the model (Hair et al., 2006) the BIC is the most restrictive in this regard, followed by the AIC and the AIC3. Since the BIC is often used to determine the correct number of classes (Hair et al., 2006), it will be used here to make a final choice which leads us to choose a four class model (BIC: 8549.48). Another reason for this choice is the distribution of importance regarding the recovery attributes. The four class solution provides the clearest differentiation regarding relative importance between classes when compared to other solutions (see table 8). Furthermore, the size of the classes is also distributed the most evenly in this solution (a solution including more classes also has a higher number of classes that are significantly smaller than the others).
TABLE 8
Relative importance recovery attributes per segment Relative importance
Attribute Class 1 Class 2 Class 3 Class 4
Attentiveness 23.36% 18.75% 24.26% 17.93%
Credibility 18.69% 17.14% 18.48% 21.85%
Apology 13.85% 19.75% 19.37% 15.05%
Timeliness 35.32% 33.36% 27.12% 26.43%
Redress 8.78% 11.00% 10.78% 18.75%
Note: the highest relative importance per attribute is underlined
The differences in relative importance of the recovery attributes will be discussed in detail below within the description of the segments. Also, the final segmented analysis shows a relatively equal distribution of size among the different classes, except for class four which is the smallest (respectively 34%, 33.5%, 26.3% and 6.1%). The results show that nine of the 21 covariates significantly differentiated between the segments. These significant covariates will be used to describe the four segments next and can be found in table 9. The table in appendix 1 depicts all parameters regarding the covariates.
30
Optimism 0.01
Security 0.03
Comfort 0.019
Dummy Quantity online purchases 3 0.0082 Dummy Frequency online purchases 2 0.035 Dummy Frequency online purchases 3 0.011
Class 1- Technologically insecure youngsters
When looking at the relative difference between the segments, table 8 shows that segment 1 has the strongest focus on timeliness and attentiveness (35.3% and 23.4%), while apology and redress are less important (13.9% and 8.8%). This segment is likely to contain younger people (β = -0.02) when compared to the other segments. The parameters of the covariates relationship quality and severity of the failure for this segment were found to be close to zero (β = 0,05 and β = 0.09 respectively) indicating these do not contribute a lot to the prediction of class membership. Also, this segment is the least likely to be optimistic about technology (β = -0.62) which means this segment feels technology does not add freedom or mobility to their daily lives and they prefer products and services with traditional technology. Furthermore, this segment is the most likely to be secure regarding technology (β = 0.46), which indicates they do feel safe using technology. However, their score for comfort around technology is average when compared to the other segments (β = 0.08). This segment is the second most likely to purchase ten or more products online each half year (β = 1.20) and the most likely to purchase products online quarterly (β = 2.49). In summary, this is a younger segment that is relatively pessimistic about technology, but does feel secure around it. Also, they purchase an above average quantity of products online. When it comes to service recovery in an online context, they strongly focus on timeliness and attentiveness.
Class 2- Young average Joe’s
31 likely to be comfortable around technology (β= -0.32). Values regarding online shopping frequency all focus around zero, indicating they do not contribute a lot to predicting class membership. In summary, this is a young segment with average values regarding online shopping experience, a high probability to have a low technology readiness and value an apology and timeliness of the service recovery above all.
Class 3- Loyal and comfortable seniors
Attentiveness is the most important attribute in this segment (24.3%), when compared to the other segments, while apology also scores relatively high (19.4%). People with a higher age are most likely to belong to this segment (β= 0.04). Furthermore, this segment has the highest likelihood to have a high relationship quality with the organization (β= 0.61) and when a service failure with a high severity occurs, a customer is least likely to belong in this segment when compared to the other segments (β= -0.58). Customers here are relatively secure (β= 0.42) and the most likely to be comfortable with technology (β= 0.55). Also, customers that buy ten or more products each half-year are most likely to belong in this segment (β= 2.14) and buy products online the most frequently (β= 2.55). This means people that are older, with a relatively high security and comfort when it comes to technology and a high relationship quality are the most likely to belong to this segment. When trying to perform service recovery for this segment, the focus should be on attentiveness and apology; while timeliness is also still one of the most important attributes.
Class 4- Low budget optimists
32
4.5 Validation ‘online’
In order to assess the predictive validity of the segmented conjoint analysis three holdout sets were used. The predicted value of the grade for the service recovery these respondents would give according to our model is compared to the actual value the respondent provided. This study will use the mean absolute percentage error (MAPE) to show the predictive validity of our model, because it is widely accepted as a predictive validity measure and it is easily related to since it is depicted as a percentage (Leeflang, Wittink, Wedel & Naert, 2000; Swanson, Tayman & Bryan, 2007). The MAPE for our final model is 27.4%, which is fairly high. A possible explanation for this might be that a small portion of the observations had a very large difference between the observed and predicted values (27 out of 597 observations), while the rest only had a small error percentage. It is possible these outliers strongly influences the MAPE score, since a few extreme values can mean the MAPE understates predictive validity (Hoaglin, Mosteller and Tukey, 1983). In order to circumvent these problems with the MAPE, another measure was calculated for the model, namely the symmetric mean absolute prediction error (SMAPE), which adjusts the formula to be less sensitive to outliers (Armstrong, 1985). The SMAPE showed a prediction error of 22.9%.
5. Results Study 2 ‘offline’
The results for the second study (about offline service recovery) will be discussed here. First, the results from the aggregate conjoint analysis will be shown, followed by the segmented conjoint analysis (although it appears there is hardly any heterogeneity in the parameters. However, we will start with some descriptive of the sample that was used. In the end, the results of both models as well as the validation of these will be elaborated on.
5.1 Sample ‘offline’
33 (e.g. only fives for all sets). This means the final sample included 219 respondents, which is a reasonable sample size to obtain reliable estimates using conjoint analysis with a moderate number of attributes (<9) according to Hair et al. (2006). The sample consists of 132 men (60.3%) and 87 women (39.7%) while the age distribution starts at 18 and stops at 61. The average age of the sample is 30.
5.2 Aggregate conjoint analysis ‘offline’
The results of the aggregate conjoint analysis for the offline study will be discussed here. Based on this, an initial indication of the importance of each attribute for the entire sample can be determined. The model was estimated using multiple regression in which ‘satisfaction with the service recovery’ was the dependent variable and the five recovery attributes as independent variables showed an aggregate R² of .329 which indicates 33% of the variance in the dependent variable is explained by the service recovery attributes. Table 10 shows the results of the aggregate analysis, which includes the relative importance of each attribute and their significance. All five attributes significantly explain the consumer’s satisfaction with the service recovery (p<0.001). As can be seen, redress is the most important attribute (37.9%) followed by timeliness (22.1%). The least important attribute is apology in the offline setting (9.8%). Stability of the model was checked by re-running it three times in which the maximum number of iterations was increased from the standard number of 250 up to 500. Results of these models show minimal variation which indicates a stable model. In order to see if there is heterogeneity in the parameters due to the proposed mediators, a segmented conjoint analysis will be performed next.
TABLE 10.
Output aggregate conjoint analysis offline
34
5.3 Segmented conjoint analysis ‘offline’
After the aggregate analysis was estimated, a segmented conjoint analysis was run in LatentGold in order to check if there are differences in the evaluation of service recovery attributes based on moderating variables. The procedure for the segmented analysis was similar to the one presented in study 1. However, significantly less covariates were added to this segmented model. In this case four covariates were added to the final model (age, gender, relationship quality and severity of the failure) of which only age was considered numeric.
Before interpreting the model, we first estimated models for one through five classes to determine the most appropriate number of classes. For all five models the model fit likelihood ratio chi-squared statistic (L²) is significant at p<.001. The results can be found in table 11 which shows the information criteria that were also used in study 1 as well as R² of the models.
Since lower criteria indicate a better fit, the model with the lowest score on these criteria will be chosen (Hair et al., 2006). As was explained before, not all criteria are equally strict in ‘punishing’ additional parameters. The BIC is the most restrictive in this sense which makes it an adequate tool for model selection here, considering a higher number of classes makes interpretation exponentially more difficult. Based on the BIC a four class model is deemed most appropriate (BIC: 8271.2). Also, the two- , three- and five class solutions were examined to determine which model had the clearest differences in relative importance of attributes. In the end, the four class model showed the most variation in relative.
Next, table 12 shows the relative importance for each attribute per class in an offline setting. These will be discussed in more detail when segments are described below. The final segmented analysis showed a somewhat unequal distribution of size in the different classes (44.8%, 27.8%, 18.6% and 8.8%). Unfortunately, the results show that none of our covariates
TABLE 11 Model selection (offline)
Model LL BIC(LL) AIC(LL) AIC3(LL df P-value R²
1 class -4323.8 8728.4 8677.5 8692.5 204 .000 0.33
2 class -4097.6 8383.9 8265.3 8300.3 184 .000 0.53
3 class -3997.1 8290.6 8104.2 8159.2 164 .000 0.57
4 class -3933.5 8271.2 8017.0 8092.0 144 .000 0.58
5 class -3898.6 8309.2 7987.3 8082.3 124 .000 0.60
35 significantly differentiated between the segments (p>.1), which means these variables cannot be used to describe our segments. Because of this, the classes can only be described based on their differences regarding recovery attribute evaluation.
TABLE 12
Relative importance recovery attributes per segment (offline) Relative importance
Attribute Class 1 Class 2 Class 3 Class 4
Attentiveness 11.7% 17.8% 13.4% 2.1%
Credibility 18.4% 15.6% 12.8% 18.1%
Apology 10.2% 8.7% 11.1% 5.5%
Timeliness 20.2% 25.7% 26.4% 19.6%
Redress 39.6% 32.3% 36.3% 54.7%
Note: highest importance is underlined
As can be seen from table 12, redress is the most important attribute in this offline setting across all classes. There are some differences in percentages for the other attributes, but overall it seems only class 4 is very different from the rest. The largest class with 44.8% is class 1 and has an importance for redress of almost 40%. Timeliness is the next most important followed by credibility of the organization. Compared to this, class 2 (27.8%) values redress slightly lower (32.3%), while timeliness is slightly more important (25.7%). Furthermore, attentiveness is the most important for class 2 when compared to the other classes (17.8%). Class 3 (18.6%) shows a comparable picture, but they value attentiveness (13.4%) and credibility (12.8%) somewhat lower which results in a higher importance for redress (36.3%). Finally, class 4 (8.8i%) shows the most extreme distribution of relative importance. Redress accounts for more than half of this segment’s evaluation (54.7%) while apology (5.5%) and attentiveness (2.1%) are almost irrelevant.
5.4 Validation ‘offline’
36 438 observations). After removing individual observations with a individual MAPE larger than 2, the overall MAPE was reduced to 19.9%. Again, the SMAPE was also calculated, which makes the formula less sensitive to outliers. Indeed, it appears the SMAPE is comparable to the MAPE without the most extreme outliers (19.5%).
6. Results Study 3 ‘comparison’
After examining the results of both study 1 and 2, it became clear another experimental study was necessary to distinguish between the effects that were found. As was expected, redress was a more important attribute in the offline context (study 2). However, there are some additional explanations for these findings. First, an extra level for the redress attribute was added for study 2. Considering that attributes with more levels usually have a higher relative importance than others (Krieger et al., 1998), this might explain the difference in findings. Also, study 2 was conducted in a different country (the US), which could also explain the difference in importance for redress. In order to disentangle these effects, an additional experiment is conducted.
The methodology for this experiment was very similar to study 2, which is why this was not discussed separately in the methodology section for clarity’s sake. However, there were some differences. First, only two scenarios were used in this study (online vs. offline), instead of four. This also meant the scenario text was adapted to not include references to relationship quality (our previous moderator). However, severity of the failure was still described, but it was held constant at the lower level for severity (the product arrives too late). Also, a question was added as a manipulation check concerning the type of store the problem had occurred at. All respondents in the final sample correctly identified the type of store their scenario mentioned. The rest of the experimental design was the same as study 2. This means respondents were asked to evaluate twelve types of service recovery based on the five
recovery attributes.
